Merge branch 'develop' into ck_tile/fa_bwd_v3

CMakeLists.txt

@@ -97,10 +97,9 @@ if(DL_KERNELS)
     add_definitions(-DDL_KERNELS)
     set(CK_ENABLE_DL_KERNELS "ON")
 endif()
-
-if(INSTANCES_ONLY)
-    add_definitions(-DINSTANCES_ONLY)
-    set(CK_ENABLE_INSTANCES_ONLY "ON")
+option(CK_USE_CODEGEN "Enable codegen library" OFF)
+if(CK_USE_CODEGEN)
+    add_definitions(-DCK_USE_CODEGEN)
 endif()
 
 include(getopt)
@@ -127,7 +126,17 @@ rocm_setup_version(VERSION ${version})
 list(APPEND CMAKE_PREFIX_PATH ${CMAKE_INSTALL_PREFIX} ${CMAKE_INSTALL_PREFIX}/llvm ${CMAKE_INSTALL_PREFIX}/hip /opt/rocm /opt/rocm/llvm /opt/rocm/hip "$ENV{ROCM_PATH}" "$ENV{HIP_PATH}")
 
 message("GPU_TARGETS= ${GPU_TARGETS}")
-
+message("GPU_ARCHS= ${GPU_ARCHS}")
+if(GPU_ARCHS)
+    #disable GPU_TARGETS to avoid conflicts, this needs to happen before we call hip package
+    unset(GPU_TARGETS CACHE)
+    unset(AMDGPU_TARGETS CACHE)
+endif()
+if(GPU_TARGETS)
+    set(USER_GPU_TARGETS 1)
+else()
+    set(USER_GPU_TARGETS 0)
+endif()
 find_package(hip)
 # No assumption that HIP kernels are launched with uniform block size for backward compatibility
 # SWDEV-413293 and https://reviews.llvm.org/D155213
@@ -135,56 +144,39 @@ math(EXPR hip_VERSION_FLAT "(${hip_VERSION_MAJOR} * 1000 + ${hip_VERSION_MINOR})
 message("hip_version_flat=${hip_VERSION_FLAT}")
 
 message("checking which targets are supported")
-#This is the list of targets to be used in case GPU_TARGETS is not set on command line
-#These targets will be filtered and only supported ones will be used
-#Setting GPU_TARGETS on command line will override this list
-if(NOT PROFILER_ONLY)
-    if(NOT ENABLE_ASAN_PACKAGING)
-        #build CK for all supported targets
-        if(NOT WIN32 AND ${hip_VERSION_FLAT} LESS 600300000)
-            # WORKAROUND: compiler does not yet fully support gfx12 targets, need to fix version above
-            rocm_check_target_ids(DEFAULT_GPU_TARGETS
-                TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102")
-        else()
-            rocm_check_target_ids(DEFAULT_GPU_TARGETS
-                TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
-        endif()
+#In order to build just the CK library (without tests and examples) for all supported GPU targets
+#use -D GPU_ARCHS="gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201" 
+#the GPU_TARGETS flag will be reset in this case in order to avoid conflicts.
+#
+#In order to build CK along with all tests and examples it should be OK to set GPU_TARGETS to just 1 or 2 similar architectures.
+if(NOT ENABLE_ASAN_PACKAGING)
+    if(NOT WIN32 AND ${hip_VERSION_FLAT} LESS 600300000)
+        # WORKAROUND: compiler does not yet fully support gfx12 targets, need to fix version above
+        set(CK_GPU_TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102")
     else()
-        #build CK only for xnack-supported targets
-        rocm_check_target_ids(DEFAULT_GPU_TARGETS
-            TARGETS "gfx908:xnack+;gfx90a:xnack+;gfx940:xnack+;gfx941:xnack+;gfx942:xnack+")
-        set(GPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
+        set(CK_GPU_TARGETS "gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201")
     endif()
 else()
-    add_definitions(-DPROFILER_ONLY)
-    set(GPU_TARGETS "" CACHE STRING "" FORCE)
-    if(GPU_TARGETS)
-        message(FATAL_ERROR "For PROFILE_ONLY build, please do not set GPU_TARGETS, use GPU_ARCH = gfx90, gfx94, gfx10, gfx11 or gfx12")
-    endif()
-    if(GPU_ARCH MATCHES "gfx90")
-        rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx908;gfx90a")
-    elseif(GPU_ARCH MATCHES "gfx94")
-        rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx940;gfx941;gfx942")
-    elseif(GPU_ARCH MATCHES "gfx10")
-        rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1030")
-    elseif(GPU_ARCH MATCHES "gfx11")
-        rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1100;gfx1101;gfx1102")
-    elseif(GPU_ARCH MATCHES "gfx12")
-        rocm_check_target_ids(DEFAULT_GPU_TARGETS TARGETS "gfx1200;gfx1201")
-    else()
-        message(FATAL_ERROR "For PROFILE_ONLY build, please specify GPU_ARCH as gfx90, gfx94, gfx10, gfx11 or gfx12")
-    endif()
-    set(GPU_TARGETS "${DEFAULT_GPU_TARGETS}" CACHE STRING " " FORCE)
+    #build CK only for xnack-supported targets when using ASAN
+    set(CK_GPU_TARGETS "gfx908:xnack+;gfx90a:xnack+;gfx940:xnack+;gfx941:xnack+;gfx942:xnack+")
 endif()
 
-message("Supported GPU_TARGETS= ${DEFAULT_GPU_TARGETS}")
-
-if(GPU_TARGETS)
-    message("Building CK for the following targets: ${GPU_TARGETS}")
+#if user set GPU_ARCHS on the cmake command line, overwrite default target list with user's list
+#otherwise, if user set GPU_TARGETS, use that set of targets
+if(GPU_ARCHS)
+    set(CK_GPU_TARGETS ${GPU_ARCHS})
 else()
-    message("Building CK for the default targets: ${DEFAULT_GPU_TARGETS}")
+    if(USER_GPU_TARGETS)
+        set(CK_GPU_TARGETS ${GPU_TARGETS})
+    endif()
 endif()
 
+#make sure all the targets on the list are actually supported by the current compiler
+rocm_check_target_ids(SUPPORTED_GPU_TARGETS
+        TARGETS ${CK_GPU_TARGETS})
+
+message("Building CK for the following targets: ${SUPPORTED_GPU_TARGETS}")
+
 if (GPU_TARGETS)
     if (GPU_TARGETS MATCHES "gfx9")
         add_definitions(-DCK_USE_XDL)
@@ -557,8 +549,7 @@ ENDFOREACH()
 add_custom_target(instances DEPENDS utility;${CK_DEVICE_INSTANCES}  SOURCES ${INSTANCE_FILES})
 add_subdirectory(library)
 
-if(NOT DEFINED INSTANCES_ONLY)
- if(NOT DEFINED PROFILER_ONLY)
+if(NOT GPU_ARCHS AND USER_GPU_TARGETS)
    rocm_package_setup_component(tests
         LIBRARY_NAME composablekernel
         PACKAGE_NAME tests # Prevent -static suffix on package name
@@ -569,24 +560,18 @@ if(NOT DEFINED INSTANCES_ONLY)
         PACKAGE_NAME examples
    )
    add_subdirectory(example)
-   add_subdirectory(test)
-
-   rocm_package_setup_component(profiler
-        LIBRARY_NAME composablekernel
-        PACKAGE_NAME ckprofiler
-   )
-   add_subdirectory(profiler)
-  else()
-    #When building PROFILER_ONLY, label the package with GPU_ARCH
-    rocm_package_setup_component(profiler
-       LIBRARY_NAME composablekernel
-       PACKAGE_NAME ckprofiler_${GPU_ARCH}
-    )
-    add_subdirectory(profiler)
-  endif()
+   if(BUILD_TESTING)
+	   add_subdirectory(test)
+   endif()
 endif()
 
-if(NOT DEFINED PROFILER_ONLY AND (GPU_TARGETS MATCHES "gfx9" OR DEFINED INSTANCES_ONLY))
+rocm_package_setup_component(profiler
+    LIBRARY_NAME composablekernel
+    PACKAGE_NAME ckprofiler
+)
+add_subdirectory(profiler)
+
+if(CK_USE_CODEGEN AND (GPU_TARGETS MATCHES "gfx9" OR GPU_ARCHS))
   add_subdirectory(codegen)
 endif()
 

Jenkinsfile

@@ -100,7 +100,15 @@ def getDockerImage(Map conf=[:]){
         dockerArgs = dockerArgs + " --no-cache "
     }
     echo "Docker Args: ${dockerArgs}"
-    def image = getDockerImageName()
+    def image
+    if ( params.BUILD_LEGACY_OS && conf.get("docker_name", "") != "" ){
+        image = conf.get("docker_name", "")
+        echo "Using legacy docker: ${image}"
+    }
+    else{
+        image = getDockerImageName()
+        echo "Using default docker: ${image}"
+    }
     //Check if image exists 
     def retimage
     try 
@@ -125,7 +133,9 @@ def buildDocker(install_prefix){
     def image_name = getDockerImageName()
     echo "Building Docker for ${image_name}"
     def dockerArgs = "--build-arg BUILDKIT_INLINE_CACHE=1 --build-arg PREFIX=${install_prefix} --build-arg CK_SCCACHE='${env.CK_SCCACHE}' --build-arg compiler_version='${params.COMPILER_VERSION}' --build-arg compiler_commit='${params.COMPILER_COMMIT}' --build-arg ROCMVERSION='${params.ROCMVERSION}' --build-arg DISABLE_CACHE='git rev-parse ${params.COMPILER_VERSION}' "
-
+    if(params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
+        dockerArgs = dockerArgs + " --no-cache "
+    }
     echo "Build Args: ${dockerArgs}"
     try{
         if(params.BUILD_DOCKER){
@@ -259,6 +269,7 @@ def cmake_build(Map conf=[:]){
             """)
         sh cmd3
     }
+
     // reduce parallelism when compiling, clang uses too much memory
     def nt = nthreads()
     def cmd
@@ -273,7 +284,7 @@ def cmake_build(Map conf=[:]){
         }
         else{
             setup_cmd = conf.get("setup_cmd", "${cmake_envs} cmake ${setup_args}   .. ")
-            build_cmd = conf.get("build_cmd", "${build_envs} dumb-init make -j${nt} ${config_targets}")
+            build_cmd = conf.get("build_cmd", "${build_envs} make -j${nt} ${config_targets}")
         }
         cmd = conf.get("cmd", """
             ${setup_cmd}
@@ -292,8 +303,8 @@ def cmake_build(Map conf=[:]){
     dir("build"){
         //build CK
         sh cmd
-        //run tests
-        if(!setup_args.contains("NO_CK_BUILD")){
+        //run tests except when NO_CK_BUILD or BUILD_LEGACY_OS are set
+        if(!setup_args.contains("NO_CK_BUILD") && !params.BUILD_LEGACY_OS){
             if (setup_args.contains("gfx90a") && params.NINJA_BUILD_TRACE){
                 sh "/ninjatracing/ninjatracing .ninja_log > ck_build_trace.json"
                 archiveArtifacts "ck_build_trace.json"
@@ -309,7 +320,7 @@ def cmake_build(Map conf=[:]){
     if (package_build == true && (env.BRANCH_NAME == "develop" || env.BRANCH_NAME == "amd-master")) {
         archiveArtifacts artifacts: "build/*.deb", allowEmptyArchive: true, fingerprint: true
     }
-    if (params.RUN_CK_TILE_TESTS){
+    if (params.RUN_CK_TILE_FMHA_TESTS){
         try{
             archiveArtifacts "perf_fmha_fwd_*.log"
             archiveArtifacts "perf_fmha_bwd_*.log"
@@ -330,11 +341,19 @@ def buildHipClangJob(Map conf=[:]){
         env.HSA_ENABLE_SDMA=0
         checkout scm
 
-        def image = getDockerImageName() 
+        def image
+        if ( params.BUILD_LEGACY_OS  && conf.get("docker_name", "") != "" ){
+            image = conf.get("docker_name", "")
+            echo "Using legacy docker: ${image}"
+        }
+        else{
+            image = getDockerImageName()
+            echo "Using default docker: ${image}"
+        }
         def prefixpath = conf.get("prefixpath", "/opt/rocm")
 
         // Jenkins is complaining about the render group 
-        def dockerOpts="--rm --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
         if (conf.get("enforce_xnack_on", false)) {
             dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
         }
@@ -352,7 +371,7 @@ def buildHipClangJob(Map conf=[:]){
         def retimage
         (retimage, image) = getDockerImage(conf)
 
-        gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
+        gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
             withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
                 timeout(time: 48, unit: 'HOURS')
                 {
@@ -393,7 +412,7 @@ def runCKProfiler(Map conf=[:]){
         def prefixpath = conf.get("prefixpath", "/opt/rocm")
 
         // Jenkins is complaining about the render group 
-        def dockerOpts="--rm --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
         if (conf.get("enforce_xnack_on", false)) {
             dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
         }
@@ -407,7 +426,7 @@ def runCKProfiler(Map conf=[:]){
         def variant = env.STAGE_NAME
         def retimage
 
-        gitStatusWrapper(credentialsId: "${status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
+        gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
             try {
                 (retimage, image) = getDockerImage(conf)
                 withDockerContainer(image: image, args: dockerOpts) {
@@ -512,11 +531,20 @@ def Build_CK(Map conf=[:]){
         env.DOCKER_BUILDKIT=1
         checkout scm
 
-        def image = getDockerImageName() 
+        def image
+        if ( params.BUILD_LEGACY_OS  && conf.get("docker_name", "") != "" ){
+            image = conf.get("docker_name", "")
+            echo "Using legacy docker: ${image}"
+        }
+        else{
+            image = getDockerImageName()
+            echo "Using default docker: ${image}"
+        }
+
         def prefixpath = conf.get("prefixpath", "/opt/rocm")
 
         // Jenkins is complaining about the render group 
-        def dockerOpts="--rm --device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+        def dockerOpts="--device=/dev/kfd --device=/dev/dri --group-add video --group-add render --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
         if (conf.get("enforce_xnack_on", false)) {
             dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
         }
@@ -524,6 +552,9 @@ def Build_CK(Map conf=[:]){
         if (params.COMPILER_VERSION == "amd-staging" || params.COMPILER_VERSION == "amd-mainline-open" || params.COMPILER_COMMIT != ""){
             dockerOpts = dockerOpts + " --env HIP_CLANG_PATH='/llvm-project/build/bin' "
         }
+        if(params.BUILD_LEGACY_OS){
+            dockerOpts = dockerOpts + " --env LD_LIBRARY_PATH='/opt/Python-3.8.13/lib' "
+        }
         def video_id = sh(returnStdout: true, script: 'getent group video | cut -d: -f3')
         def render_id = sh(returnStdout: true, script: 'getent group render | cut -d: -f3')
         dockerOpts = dockerOpts + " --group-add=${video_id} --group-add=${render_id} "
@@ -532,7 +563,7 @@ def Build_CK(Map conf=[:]){
         def variant = env.STAGE_NAME
         def retimage
 
-        gitStatusWrapper(credentialsId: "${env.status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
+        gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
             try {
                 (retimage, image) = getDockerImage(conf)
                 withDockerContainer(image: image, args: dockerOpts) {
@@ -629,7 +660,7 @@ def process_results(Map conf=[:]){
     def prefixpath = "/opt/rocm"
 
     // Jenkins is complaining about the render group 
-    def dockerOpts="--rm --cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
+    def dockerOpts="--cap-add=SYS_PTRACE --security-opt seccomp=unconfined"
     if (conf.get("enforce_xnack_on", false)) {
         dockerOpts = dockerOpts + " --env HSA_XNACK=1 "
     }
@@ -637,7 +668,7 @@ def process_results(Map conf=[:]){
     def variant = env.STAGE_NAME
     def retimage
 
-    gitStatusWrapper(credentialsId: "${env.status_wrapper_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
+    gitStatusWrapper(credentialsId: "${env.ck_git_creds}", gitHubContext: "Jenkins - ${variant}", account: 'ROCm', repo: 'composable_kernel') {
         try {
             (retimage, image) = getDockerImage(conf)
         }
@@ -651,7 +682,7 @@ def process_results(Map conf=[:]){
         timeout(time: 1, unit: 'HOURS'){
             try{
                 dir("script"){
-                    if (params.RUN_CK_TILE_TESTS){
+                    if (params.RUN_CK_TILE_FMHA_TESTS){
                         try{
                             unstash "perf_fmha_fwd_gfx942.log"
                             unstash "perf_fmha_bwd_gfx942.log"
@@ -704,10 +735,11 @@ def process_results(Map conf=[:]){
 
 //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
 CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true
-                                              0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true
+                                              0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true;RUN_CODEGEN_TESTS=true
                                               0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
                                               0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
-                                              0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false''' : ""
+                                              0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false
+                                              0 13 * * * % BUILD_LEGACY_OS=true''' : ""
 
 pipeline {
     agent none
@@ -774,6 +806,10 @@ pipeline {
             name: "RUN_GROUPED_CONV_LARGE_CASES_TESTS",
             defaultValue: false,
             description: "Run the grouped conv large cases tests (default: OFF)")
+        booleanParam(
+            name: "RUN_CODEGEN_TESTS",
+            defaultValue: false,
+            description: "Run codegen tests (default: OFF)")
         booleanParam(
             name: "RUN_CK_TILE_FMHA_TESTS",
             defaultValue: false,
@@ -794,6 +830,10 @@ pipeline {
             name: "NINJA_BUILD_TRACE",
             defaultValue: false,
             description: "Generate a ninja build trace (default: OFF)")
+        booleanParam(
+            name: "BUILD_LEGACY_OS",
+            defaultValue: false,
+            description: "Try building CK with legacy OS dockers: RHEL8 and SLES15 (default: OFF)")
     }
     environment{
         dbuser = "${dbuser}"
@@ -802,7 +842,7 @@ pipeline {
         dbsshport = "${dbsshport}"
         dbsshuser = "${dbsshuser}"
         dbsshpassword = "${dbsshpassword}"
-        status_wrapper_creds = "${status_wrapper_creds}"
+        ck_git_creds = "${ck_git_creds}"
         gerrit_cred="${gerrit_cred}"
         DOCKER_BUILDKIT = "1"
     }
@@ -890,7 +930,30 @@ pipeline {
                         execute_args = """ ../script/cmake-ck-dev.sh  ../ gfx90a && \
                                            make -j64 test_grouped_convnd_fwd_large_cases_xdl && \
                                            ./bin/test_grouped_convnd_fwd_large_cases_xdl"""
-                   }
+                    }
+                    steps{
+                        buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
+                        cleanWs()
+                    }
+                }
+            }
+        }
+        stage("Run Codegen Tests")
+        {
+            parallel
+            {
+                stage("Run Codegen Tests on gfx90a")
+                {
+                    when {
+                        beforeAgent true
+                        expression { params.RUN_CODEGEN_TESTS.toBoolean() }
+                    }
+                    agent{ label rocmnode("gfx90a")}
+                    environment{
+                        setup_args = "NO_CK_BUILD"
+                        execute_args = """ CXX=/opt/rocm/llvm/bin/clang++ cmake ../codegen && \
+                                           make -j64 check"""
+                    }
                     steps{
                         buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
                         cleanWs()
@@ -915,7 +978,7 @@ pipeline {
                                            make -j64 tile_example_fmha_fwd tile_example_fmha_bwd && \
                                            cd ../ &&
                                            example/ck_tile/01_fmha/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx90a """
-                   }
+                    }
                     steps{
                         buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
                         cleanWs()
@@ -934,7 +997,7 @@ pipeline {
                                            make -j64 tile_example_fmha_fwd tile_example_fmha_bwd && \
                                            cd ../ &&
                                            example/ck_tile/01_fmha/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx942 """
-                   }
+                    }
                     steps{
                         buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
                         cleanWs()
@@ -946,7 +1009,6 @@ pipeline {
         {
             parallel
             {
-
                 stage("Run CK_TILE_GEMM Tests on gfx90a")
                 {
                     when {
@@ -960,12 +1022,11 @@ pipeline {
                                            make -j64 tile_example_gemm_basic && \
                                            cd ../ &&
                                            example/ck_tile/03_gemm/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx90a """
-                   }
+                    }
                     steps{
                         buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
                         cleanWs()
                     }
-
                 }
                 stage("Run CK_TILE_GEMM Tests on gfx942")
                 {
@@ -980,7 +1041,7 @@ pipeline {
                                            make -j64 tile_example_gemm_basic && \
                                            cd ../ &&
                                            example/ck_tile/03_gemm/script/run_full_test.sh "CI_${params.COMPILER_VERSION}" "${env.BRANCH_NAME}" "${NODE_NAME}" gfx942 """
-                   }
+                    }
                     steps{
                         buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
                         cleanWs()
@@ -988,15 +1049,54 @@ pipeline {
                 }
             }
         }
+
 		stage("Build CK and run Tests")
         {
             parallel
             {
+                stage("Build CK with RHEL8")
+                {
+                    when {
+                        beforeAgent true
+                        expression { params.BUILD_LEGACY_OS.toBoolean() }
+                    }
+                    agent{ label rocmnode("gfx90a") }
+                    environment{
+                        def docker_name = "${env.CK_DOCKERHUB_PRIVATE}:ck_rhel8_rocm6.3"
+                        setup_args = """ -DGPU_TARGETS="gfx942" \
+                                         -DCMAKE_CXX_FLAGS=" -O3 " \
+                                         -DCK_USE_ALTERNATIVE_PYTHON=/opt/Python-3.8.13/bin/python3.8 """
+                        execute_args = " "
+                    }
+                    steps{
+                        Build_CK_and_Reboot(setup_args: setup_args, config_targets: " ", no_reboot:true, build_type: 'Release', docker_name: docker_name)
+                        cleanWs()
+                    }
+                }
+                stage("Build CK with SLES15")
+                {
+                    when {
+                        beforeAgent true
+                        expression { params.BUILD_LEGACY_OS.toBoolean() }
+                    }
+                    agent{ label rocmnode("gfx90a") }
+                    environment{
+                        def docker_name = "${env.CK_DOCKERHUB_PRIVATE}:ck_sles15_rocm6.3"
+                        setup_args = """ -DGPU_TARGETS="gfx942" \
+                                         -DCMAKE_CXX_FLAGS=" -O3 " \
+                                         -DCK_USE_ALTERNATIVE_PYTHON=/opt/Python-3.8.13/bin/python3.8 """
+                        execute_args = " "
+                    }
+                    steps{
+                        Build_CK_and_Reboot(setup_args: setup_args, config_targets: " ", no_reboot:true, build_type: 'Release', docker_name: docker_name)
+                        cleanWs()
+                    }
+                }
                 stage("Build CK for all gfx9 targets")
                 {
                     when {
                         beforeAgent true
-                        expression { params.RUN_FULL_QA.toBoolean() }
+                        expression { params.RUN_FULL_QA.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
                     }
                     agent{ label rocmnode("gfx90a") }
                     environment{
@@ -1018,7 +1118,7 @@ pipeline {
                 {
                     when {
                         beforeAgent true
-                        expression { params.RUN_FULL_QA.toBoolean() }
+                        expression { params.RUN_FULL_QA.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
                     }
                     agent{ label rocmnode("gfx942") }
                     environment{
@@ -1038,7 +1138,7 @@ pipeline {
                 {
                     when {
                         beforeAgent true
-                        expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
+                        expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
                     }
                     agent{ label rocmnode("gfx90a") }
                     environment{
@@ -1058,16 +1158,16 @@ pipeline {
                 {
                     when {
                         beforeAgent true
-                        expression { params.BUILD_INSTANCES_ONLY.toBoolean() && !params.RUN_FULL_QA.toBoolean() }
+                        expression { params.BUILD_INSTANCES_ONLY.toBoolean() && !params.RUN_FULL_QA.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
                     }
                     agent{ label rocmnode("gfx90a") }
                     environment{
                         execute_args = """ cmake -D CMAKE_PREFIX_PATH=/opt/rocm \
                                            -D CMAKE_CXX_COMPILER="${build_compiler()}" \
                                            -D CMAKE_BUILD_TYPE=Release \
-                                           -D INSTANCES_ONLY=ON \
-                                           -DCMAKE_CXX_FLAGS=" -O3 " .. && make -j64 """
-                   }
+                                           -D GPU_ARCHS="gfx908;gfx90a;gfx940;gfx941;gfx942;gfx1030;gfx1100;gfx1101;gfx1102"  \
+                                           -D CMAKE_CXX_FLAGS=" -O3 " .. && make -j64 """
+                    }
                     steps{
                         buildHipClangJobAndReboot(setup_cmd: "",  build_cmd: "", no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
                         cleanWs()
@@ -1077,7 +1177,7 @@ pipeline {
                 {
                     when {
                         beforeAgent true
-                        expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
+                        expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
                     }
                     agent{ label rocmnode("gfx1030") }
                     environment{
@@ -1097,7 +1197,7 @@ pipeline {
                 {
                     when {
                         beforeAgent true
-                        expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
+                        expression { !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
                     }
                     agent{ label rocmnode("gfx1101") }
                     environment{
@@ -1117,7 +1217,7 @@ pipeline {
                 {
                     when {
                         beforeAgent true
-                        expression { params.BUILD_GFX12.toBoolean() && !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() }
+                        expression { params.BUILD_GFX12.toBoolean() && !params.RUN_FULL_QA.toBoolean() && !params.BUILD_INSTANCES_ONLY.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
                     }
                     agent{ label rocmnode("gfx1201") }
                     environment{
@@ -1144,7 +1244,7 @@ pipeline {
                 {
                     when {
                         beforeAgent true
-                        expression { params.RUN_PERFORMANCE_TESTS.toBoolean() }
+                        expression { params.RUN_PERFORMANCE_TESTS.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
                     }
                     options { retry(1) }
                     agent{ label rocmnode("gfx90a")}
@@ -1165,7 +1265,7 @@ pipeline {
                 stage("Process results"){
                     when {
                         beforeAgent true
-                        expression { params.RUN_PERFORMANCE_TESTS.toBoolean() }
+                        expression { params.RUN_PERFORMANCE_TESTS.toBoolean() && !params.BUILD_LEGACY_OS.toBoolean() }
                     }
                     agent { label 'mici' }
                     steps{

README.md

@@ -90,7 +90,13 @@ Docker images are available on [DockerHub](https://hub.docker.com/r/rocm/composa
     ```
 
     If you don't set `GPU_TARGETS` on the cmake command line, CK is built for all GPU targets
-    supported by the current compiler (this may take a long time).
+    supported by the current compiler (this may take a long time). 
+    Tests and examples will only get built if the GPU_TARGETS is set by the user on the cmake command line.
+
+    NOTE: If you try setting `GPU_TARGETS` to a list of architectures, the build will only work if the 
+    architectures are similar, e.g., `gfx908;gfx90a`, or `gfx1100;gfx1101;gfx11012`. Otherwise, if you 
+    want to build the library for a list of different architectures,
+    you should use the `GPU_ARCHS` build argument, for example `GPU_ARCHS=gfx908;gfx1030;gfx1100;gfx942`.
 
 4. Build the entire CK library:
 
@@ -137,10 +143,6 @@ crash. In such cases, you can reduce the number of threads to 32 by using `-j32`
 
 Additional cmake flags can be used to significantly speed-up the build:
 
-* `INSTANCES_ONLY` (default is OFF) must be set to ON in order to build only the instances and library
-  while skipping all tests, examples, and profiler. This is useful in cases when you plan to use CK as a
-  dependency and don't plan to run any examples or tests.
-
 * `DTYPES` (default is not set) can be set to any subset of "fp64;fp32;fp16;fp8;bf16;int8" to build
   instances of select data types only. The main default data types are fp32 and fp16; you can safely skip
   other data types.

cmake/Embed.cmake

@@ -233,6 +233,8 @@ function(add_embed_library EMBED_NAME)
     else()
         target_sources(${EMBED_NAME} INTERFACE $<TARGET_OBJECTS:${INTERNAL_EMBED_LIB}>)
     endif()
-    target_include_directories(${EMBED_NAME} INTERFACE "${EMBED_DIR}/include")
+    target_include_directories(${EMBED_NAME} INTERFACE 
+	    $<BUILD_INTERFACE:${EMBED_DIR}/include>
+	        $<INSTALL_INTERFACE:include/ck>)
 endfunction()
 

codegen/CMakeLists.txt

@@ -1,3 +1,6 @@
+cmake_minimum_required(VERSION 3.16)
+project(composable_kernel_host)
+
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
@@ -5,49 +8,51 @@ set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 set(CK_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/..)
 
-add_compile_options(-std=c++17)
-find_package(hip)
-add_custom_target(codegen)
+find_package(ROCM)
+include(ROCMInstallTargets)
+include(ROCMTest)
 
-# add include directories
-include_directories(BEFORE
-            ${PROJECT_BINARY_DIR}/include
-            ${PROJECT_SOURCE_DIR}/include
-            ${PROJECT_SOURCE_DIR}/library/include
-            ${HIP_INCLUDE_DIRS}
-            )
+rocm_setup_version(VERSION 1.0)
 
 list(APPEND CMAKE_MODULE_PATH ${CK_ROOT}/cmake)
 include(Embed)
 file(GLOB_RECURSE KERNEL_FILES CONFIGURE_DEPENDS
-	${CK_ROOT}/include/ck/*.hpp)
-#printouts fot debug purposes
-#message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
-#message(STATUS "RELATIVE: ${CK_ROOT}/include")
+    ${CK_ROOT}/include/ck/*.hpp)
+# printouts fot debug purposes
+# message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
+# message(STATUS "RELATIVE: ${CK_ROOT}/include")
 add_embed_library(ck_headers ${KERNEL_FILES} RELATIVE ${CK_ROOT}/include)
 
-file(GLOB SOURCES CONFIGURE_DEPENDS src/*.cpp)
+add_compile_options(-std=c++17)
 
-##message(STATUS "SOURCE_FILES: ${SOURCES}")
+file(GLOB SOURCES CONFIGURE_DEPENDS src/*.cpp)
 # TODO: Use object library
 add_library(ck_host STATIC ${SOURCES})
 target_link_libraries(ck_host PRIVATE ck_headers)
 
-set_target_properties(ck_host PROPERTIES
-	LINKER_LANGUAGE CXX
-        POSITION_INDEPENDENT_CODE ON)
+set_target_properties(ck_host PROPERTIES 
+    LINKER_LANGUAGE CXX
+    POSITION_INDEPENDENT_CODE ON)
 
-target_include_directories(ck_host PUBLIC
-    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
-)
+# target_include_directories(ck_host PUBLIC
+#     $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+# )
 
 add_executable(ck-template-driver driver/main.cpp)
 target_link_libraries(ck-template-driver ck_host)
 
-rocm_install(
+rocm_install_targets(
     TARGETS ck_host ck_headers
-    EXPORT ck_hostTargets
+    EXPORT ck_host_targets
+    INCLUDE include
+    PRIVATE
+)
+rocm_export_targets(
+    EXPORT ck_host_targets
+    NAMESPACE composable_kernel::
 )
-rocm_install(DIRECTORY include/ck DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
 
-add_subdirectory(test)
+if(BUILD_TESTING)
+    add_subdirectory(test)
+endif()
+

codegen/test/CMakeLists.txt

@@ -1,22 +1,25 @@
 list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
 add_subdirectory(rtc)
 file(GLOB TEST_SRCS CONFIGURE_DEPENDS *.cpp)
-if(NOT INSTANCES_ONLY)
-  foreach(TEST_SRC ${TEST_SRCS})
-    set_source_files_properties(${TEST_SRC} PROPERTIES LANGUAGE HIP)
-    get_filename_component(BASE_NAME ${TEST_SRC} NAME_WE)
-    add_executable(codegen_test_${BASE_NAME} ${TEST_SRC})
-    if(CK_USE_ALTERNATIVE_PYTHON)
-      target_link_options(codegen_test_${BASE_NAME} PRIVATE -lstdc++fs)
-    endif()
-    add_dependencies(codegen codegen_test_${BASE_NAME})
-    add_dependencies(tests codegen_test_${BASE_NAME})
-    add_dependencies(check codegen_test_${BASE_NAME})
-    add_test(NAME codegen_test_${BASE_NAME} COMMAND codegen_test_${BASE_NAME})
-    message("adding test codegen_test_${BASE_NAME}")
-    target_link_libraries(codegen_test_${BASE_NAME} ck_rtc ck_host)
-    target_include_directories(codegen_test_${BASE_NAME} PUBLIC ${CK_ROOT}/codegen/test/include)
+
+# TODO: These tests need to be refactored to remove dependency on main ck
+# headers and device compilation.
+set(TESTS_REQUIRE_DEVICE_COMPILE 
+  grouped_conv_fwd_multiple_d_v1
+  grouped_conv_fwd_multiple_d_v2
+  grouped_conv_fwd_multiple_d_v3
+  grouped_conv_fwd_multiple_d_v4
+)
+find_package(hip)
+
+foreach(TEST_SRC ${TEST_SRCS})
+  get_filename_component(BASE_NAME ${TEST_SRC} NAME_WE)
+  rocm_add_test_executable(codegen_test_${BASE_NAME} ${TEST_SRC})
+  target_link_libraries(codegen_test_${BASE_NAME} ck_rtc ck_host)
+  target_include_directories(codegen_test_${BASE_NAME} PUBLIC include)
+  if(BASE_NAME IN_LIST TESTS_REQUIRE_DEVICE_COMPILE)
+    target_link_libraries(codegen_test_${BASE_NAME} hip::device)
     target_include_directories(codegen_test_${BASE_NAME} PUBLIC ${CK_ROOT}/include)
     target_include_directories(codegen_test_${BASE_NAME} PUBLIC ${CK_ROOT}/library/include)
-  endforeach()
-endif()
+  endif()
+endforeach()

codegen/test/rtc/CMakeLists.txt

@@ -1,4 +1,6 @@
+find_package(hip)
 file(GLOB RTC_SOURCES CONFIGURE_DEPENDS src/*.cpp)
 add_library(ck_rtc ${RTC_SOURCES})
 target_include_directories(ck_rtc PUBLIC include)
 target_link_libraries(ck_rtc PUBLIC hip::host)
+target_link_libraries(ck_rtc PUBLIC -lstdc++fs)

codegen/test/rtc/include/rtc/compile_kernel.hpp

@@ -2,14 +2,14 @@
 #define GUARD_HOST_TEST_RTC_INCLUDE_RTC_COMPILE_KERNEL
 
 #include <rtc/kernel.hpp>
-#include <ck/filesystem.hpp>
+#include <rtc/filesystem.hpp>
 #include <string>
 
 namespace rtc {
 
 struct src_file
 {
-    CK::fs::path path;
+    fs::path path;
     std::string_view content;
 };
 

codegen/test/rtc/include/rtc/filesystem.hpp

@@ -0,0 +1,60 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#ifndef GUARD_TEST_HOST_RTC_FILESYSTEM_HPP
+#define GUARD_TEST_HOST_RTC_FILESYSTEM_HPP
+
+#include <string>
+#include <string_view>
+
+// clang-format off
+#if defined(CPPCHECK)
+  #define RTC_HAS_FILESYSTEM 1
+  #define RTC_HAS_FILESYSTEM_TS 1
+#elif defined(_WIN32)
+  #if _MSC_VER >= 1920
+    #define RTC_HAS_FILESYSTEM 1
+    #define RTC_HAS_FILESYSTEM_TS 0
+  #elif _MSC_VER >= 1900
+    #define RTC_HAS_FILESYSTEM 0
+    #define RTC_HAS_FILESYSTEM_TS 1
+  #else
+    #define RTC_HAS_FILESYSTEM 0
+    #define RTC_HAS_FILESYSTEM_TS 0
+  #endif
+#elif defined(__has_include)
+  #if __has_include(<filesystem>) && __cplusplus >= 201703L
+    #define RTC_HAS_FILESYSTEM 1
+  #else
+    #define RTC_HAS_FILESYSTEM 0
+  #endif
+  #if __has_include(<experimental/filesystem>) && __cplusplus >= 201103L
+    #define RTC_HAS_FILESYSTEM_TS 1
+  #else
+    #define RTC_HAS_FILESYSTEM_TS 0
+  #endif
+#else
+  #define RTC_HAS_FILESYSTEM 0
+  #define RTC_HAS_FILESYSTEM_TS 0
+#endif
+// clang-format on
+
+#if RTC_HAS_FILESYSTEM
+#include <filesystem>
+#elif RTC_HAS_FILESYSTEM_TS
+#include <experimental/filesystem>
+#else
+#error "No filesystem include available"
+#endif
+
+namespace rtc {
+
+#if RTC_HAS_FILESYSTEM
+namespace fs = ::std::filesystem;
+#elif RTC_HAS_FILESYSTEM_TS
+namespace fs = ::std::experimental::filesystem;
+#endif
+
+} // namespace rtc
+
+#endif // GUARD_RTC_FILESYSTEM_HPP_

codegen/test/rtc/include/rtc/tmp_dir.hpp

@@ -2,13 +2,13 @@
 #define GUARD_HOST_TEST_RTC_INCLUDE_RTC_TMP_DIR
 
 #include <string>
-#include <ck/filesystem.hpp>
+#include <rtc/filesystem.hpp>
 
 namespace rtc {
 
 struct tmp_dir
 {
-    CK::fs::path path;
+    fs::path path;
     tmp_dir(const std::string& prefix = "");
 
     void execute(const std::string& cmd) const;

codegen/test/rtc/src/compile_kernel.cpp

@@ -1,4 +1,4 @@
-#include "rtc/hip.hpp"
+#include <rtc/hip.hpp>
 #include <rtc/compile_kernel.hpp>
 #include <rtc/tmp_dir.hpp>
 #include <stdexcept>
@@ -70,9 +70,9 @@ kernel compile_kernel(const std::vector<src_file>& srcs, compile_options options
 
     for(const auto& src : srcs)
     {
-        CK::fs::path full_path   = td.path / src.path;
-        CK::fs::path parent_path = full_path.parent_path();
-        CK::fs::create_directories(parent_path);
+        fs::path full_path   = td.path / src.path;
+        fs::path parent_path = full_path.parent_path();
+        fs::create_directories(parent_path);
         write_string(full_path.string(), src.content);
         if(src.path.extension().string() == ".cpp")
         {
@@ -86,7 +86,7 @@ kernel compile_kernel(const std::vector<src_file>& srcs, compile_options options
     td.execute(compiler() + options.flags);
 
     auto out_path = td.path / out;
-    if(not CK::fs::exists(out_path))
+    if(not fs::exists(out_path))
         throw std::runtime_error("Output file missing: " + out);
 
     auto obj = read_buffer(out_path.string());

codegen/test/rtc/src/tmp_dir.cpp

@@ -31,10 +31,10 @@ std::string unique_string(const std::string& prefix)
 }
 
 tmp_dir::tmp_dir(const std::string& prefix)
-    : path(CK::fs::temp_directory_path() /
+    : path(fs::temp_directory_path() /
            unique_string(prefix.empty() ? "ck-rtc" : "ck-rtc-" + prefix))
 {
-    CK::fs::create_directories(this->path);
+    fs::create_directories(this->path);
 }
 
 void tmp_dir::execute(const std::string& cmd) const
@@ -43,6 +43,6 @@ void tmp_dir::execute(const std::string& cmd) const
     std::system(s.c_str());
 }
 
-tmp_dir::~tmp_dir() { CK::fs::remove_all(this->path); }
+tmp_dir::~tmp_dir() { fs::remove_all(this->path); }
 
 } // namespace rtc

docs/reference/API_Reference_Guide.rst

@@ -12,12 +12,6 @@ API reference guide
 This document contains details of the APIs for the Composable Kernel (CK) library and introduces
 some of the key design principles that are used to write new classes that extend CK functionality.
 
-=================
-Using CK API
-=================
-
-This section describes how to use the CK library API.
-
 =================
 CK Datatypes
 =================

docs/sphinx/requirements.in

@@ -1,2 +1,2 @@
-rocm-docs-core==1.8.1
+rocm-docs-core==1.8.2
 sphinxcontrib-bibtex==2.6.3

docs/sphinx/requirements.txt

@@ -103,7 +103,7 @@ requests==2.32.3
     # via
     #   pygithub
     #   sphinx
-rocm-docs-core==1.8.1
+rocm-docs-core==1.8.2
     # via -r requirements.in
 six==1.16.0
     # via pybtex

example/01_gemm/common.hpp

@@ -21,6 +21,7 @@
 #include "ck/library/utility/host_tensor_generator.hpp"
 #include "ck/library/utility/literals.hpp"
 #include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+#include "ck/library/reference_tensor_operation/gpu/reference_gemm.hpp"
 
 struct ProblemSize final
 {
@@ -28,9 +29,9 @@ struct ProblemSize final
     ck::index_t N = 4096;
     ck::index_t K = 4096;
 
-    ck::index_t StrideA = 4096;
-    ck::index_t StrideB = 4096;
-    ck::index_t StrideC = 4096;
+    ck::index_t StrideA = 0;
+    ck::index_t StrideB = 0;
+    ck::index_t StrideC = 0;
 };
 
 struct ProblemSizeStreamK final
@@ -39,9 +40,9 @@ struct ProblemSizeStreamK final
     ck::index_t N = 4096;
     ck::index_t K = 4096;
 
-    ck::index_t StrideA = 4096;
-    ck::index_t StrideB = 4096;
-    ck::index_t StrideC = 4096;
+    ck::index_t StrideA = 0;
+    ck::index_t StrideB = 0;
+    ck::index_t StrideC = 0;
 
     ck::index_t NumSKBlocks = -1;
 };
@@ -51,9 +52,9 @@ struct ProblemSizeStreamK_universal final
     ck::index_t N = 4096;
     ck::index_t K = 4096;
 
-    ck::index_t StrideA = 4096;
-    ck::index_t StrideB = 4096;
-    ck::index_t StrideC = 4096;
+    ck::index_t StrideA = 0;
+    ck::index_t StrideB = 0;
+    ck::index_t StrideC = 0;
 
     ck::index_t Grid_size   = -1; // defaults to max occupancy
     ck::index_t Streamk_sel = 1;  // defaults to 1-tile SK
@@ -65,9 +66,9 @@ struct ProblemSizeSplitK final
     ck::index_t N = 4096;
     ck::index_t K = 4096;
 
-    ck::index_t StrideA = 4096;
-    ck::index_t StrideB = 4096;
-    ck::index_t StrideC = 4096;
+    ck::index_t StrideA = 0;
+    ck::index_t StrideB = 0;
+    ck::index_t StrideC = 0;
 
     ck::index_t KBatch = 1;
 };
@@ -125,7 +126,7 @@ bool parse_cmd_args<ProblemSize>(int argc,
     }
     else
     {
-        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+        std::cerr << "arg1: verification (0=no, 1=CPU and GPU)" << std::endl
                   << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
                   << std::endl
                   << "arg3: time kernel (0=no, 1=yes)" << std::endl
@@ -175,7 +176,7 @@ bool parse_cmd_args<ProblemSizeStreamK_universal>(int argc,
     else
     {
         std::cerr
-            << "arg1: verification (0=no, 1=yes)" << std::endl
+            << "arg1: verification (0=no, 1=CPU and GPU)" << std::endl
             << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)" << std::endl
             << "arg3: time kernel (0=no, 1=yes)" << std::endl
             << "arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC" << std::endl
@@ -224,7 +225,7 @@ bool parse_cmd_args<ProblemSizeStreamK>(int argc,
     }
     else
     {
-        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+        std::cerr << "arg1: verification (0=no, 1=CPU and GPU)" << std::endl
                   << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
                   << std::endl
                   << "arg3: time kernel (0=no, 1=yes)" << std::endl
@@ -274,7 +275,7 @@ bool parse_cmd_args<ProblemSizeSplitK>(int argc,
     }
     else
     {
-        std::cerr << "arg1: verification (0=no, 1=yes)" << std::endl
+        std::cerr << "arg1: verification (0=no, 1=CPU and GPU)" << std::endl
                   << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)"
                   << std::endl
                   << "arg3: time kernel (0=no, 1=yes)" << std::endl

example/01_gemm/gemm_dl_fp16.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -32,6 +32,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_dl_fp32.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -32,6 +32,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_dl_int8.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -32,6 +32,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDl
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_dpp_fp16.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -34,6 +34,9 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmDpp
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::
+    ReferenceGemm<ALayout, BLayout, CLayout, ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_wmma_fp16.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -68,6 +68,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmWmma_CShuffle
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_bf16.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -33,6 +33,20 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceComputeType     = float;
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp,
+                                                                             ReferenceComputeType,
+                                                                             ReferenceComputeType>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_bf16_rtn.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -34,6 +34,20 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceComputeType     = float;
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp,
+                                                                             ReferenceComputeType,
+                                                                             ReferenceComputeType>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_fp16.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -47,6 +47,17 @@ using DeviceGemmInstance = DeviceGemmInstance1;
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_fp16_fp8.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -42,6 +42,17 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
                                                                         CElementOp,
                                                                         ComputeType>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_fp16_v2.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -46,6 +46,17 @@ using DeviceGemmInstance =
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_fp64.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -41,6 +41,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl
                                                                             BElementOp,
                                                                             CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_fp8.cpp

@@ -37,6 +37,20 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceComputeType     = float;
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp,
+                                                                             ReferenceComputeType,
+                                                                             ReferenceComputeType>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_fp8_bf8.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -44,6 +44,17 @@ using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataTyp
                                                                         ComputeTypeA,
                                                                         ComputeTypeB>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_int8.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -33,6 +33,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_lds_direct_load_fp16.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
 
@@ -53,6 +53,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_lds_direct_load_fp32.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2023-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include <iostream>
 
@@ -52,6 +52,17 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/gemm_xdl_streamk.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -44,6 +44,17 @@ using DeviceGemmInstance = DeviceGemmStreamK;
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_streamk_example(argc, argv); }

example/01_gemm/gemm_xdl_wavelet_fp16.cpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "common.hpp"
 
@@ -37,6 +37,17 @@ using DeviceGemmInstance = DeviceGemmInstance;
 using ReferenceGemmInstance = ck::tensor_operation::host::
     ReferenceGemm<ADataType, BDataType, CDataType, AccDataType, AElementOp, BElementOp, CElementOp>;
 
+using ReferenceGemmInstanceGPU = ck::tensor_operation::device::ReferenceGemm<ALayout,
+                                                                             BLayout,
+                                                                             CLayout,
+                                                                             ADataType,
+                                                                             BDataType,
+                                                                             CDataType,
+                                                                             AccDataType,
+                                                                             AElementOp,
+                                                                             BElementOp,
+                                                                             CElementOp>;
+
 #include "run_gemm_example.inc"
 
 int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/01_gemm/run_gemm_example.inc

@@ -173,6 +173,7 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 
     Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
     Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
+    Tensor<CDataType> c_m_n_device_ref_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
 
     std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
     std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
@@ -193,6 +194,8 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
     DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
     DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
     DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize());
+    DeviceMem c_m_n_device_ref_buf(sizeof(CDataType) *
+                                   c_m_n_device_ref_result.mDesc.GetElementSpaceSize());
 
     a_m_k_device_buf.ToDevice(a_m_k.mData.data());
     b_k_n_device_buf.ToDevice(b_k_n.mData.data());
@@ -325,14 +328,18 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
     std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
               << gemm.GetTypeString() << std::endl;
 
+    bool pass = true;
+
     if(config.do_verification)
     {
+        // CPU verification
         auto ref_gemm    = ReferenceGemmInstance{};
         auto ref_invoker = ref_gemm.MakeInvoker();
 
         auto ref_argument = ref_gemm.MakeArgument(
             a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op);
 
+        std::cout << "Running verification on CPU." << std::endl;
         ref_invoker.Run(ref_argument);
 
 #ifdef BUILD_INT4_EXAMPLE
@@ -346,15 +353,42 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 #else
         c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
 
-        return ck::utils::check_err(c_m_n_device_result,
-                                    c_m_n_host_result,
-                                    "Error: Incorrect results!",
-                                    get_rtol<CDataType>(),
-                                    get_atol<CDataType>());
+        pass &= !ck::utils::check_err(c_m_n_device_result,
+                                      c_m_n_host_result,
+                                      "Error: Incorrect results!",
+                                      get_rtol<CDataType>(),
+                                      get_atol<CDataType>());
 #endif
+
+        // GPU verification
+        auto ref_gemm_gpu    = ReferenceGemmInstanceGPU{};
+        auto ref_invoker_gpu = ref_gemm_gpu.MakeInvoker();
+
+        auto ref_argument_gpu = ref_gemm_gpu.MakeArgument(
+            static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
+            static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
+            static_cast<CDataType*>(c_m_n_device_ref_buf.GetDeviceBuffer()),
+            M,
+            N,
+            K,
+            a_element_op,
+            b_element_op,
+            c_element_op);
+
+        std::cout << "Running verification on GPU." << std::endl;
+        ref_invoker_gpu.Run(ref_argument_gpu, StreamConfig{});
+
+        c_m_n_device_ref_buf.FromDevice(c_m_n_device_ref_result.mData.data());
+        c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data());
+
+        pass &= !ck::utils::check_err(c_m_n_device_result,
+                                      c_m_n_device_ref_result,
+                                      "Error: Incorrect results!",
+                                      get_rtol<CDataType>(),
+                                      get_atol<CDataType>());
     }
 
-    return true;
+    return !pass;
 }
 
 bool run_gemm_example(int argc, char* argv[])

example/01_gemm/run_gemm_example_streamk_v2.inc

@@ -117,9 +117,9 @@ bool run_gemm(const ProblemType& problem_size, const ExecutionConfig& config)
 
     auto f_get_default_stride =
         [](std::size_t row, std::size_t col, ck::index_t stride, auto layout) {
-            if(stride == -1)
+            if(stride == 0)
             {
-                // give a chance if stride is -1, return a default packed stride
+                // give a chance if stride is 0, return a default packed stride
                 if constexpr(std::is_same_v<decltype(layout), ck::tensor_layout::gemm::RowMajor>)
                 {
                     return static_cast<std::size_t>(col);

example/44_elementwise_permute/CMakeLists.txt

@@ -5,3 +5,4 @@ add_example_executable(example_elementwise_permute_4D_fp32_col elementwise_permu
 add_example_executable(example_elementwise_permute_4D_fp16_col elementwise_permute_4D_fp16_col.cpp)
 add_example_executable(example_elementwise_binary_4D_fp16 elementwise_binary_4D_fp16.cpp)
 add_example_executable(example_elementwise_trinary_4D_fp16 elementwise_trinary_4D_fp16.cpp)
+add_example_executable(elementwise_scale_permute_amax_2D_fp16_fp8 elementwise_scale_permute_amax_2D_fp16_fp8.cpp)

example/44_elementwise_permute/elementwise_scale_permute_amax_2D_fp16_fp8.cpp

@@ -0,0 +1,247 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_elementwise_dynamic_vector_dims_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+
+#include "ck/library/reference_tensor_operation/cpu/reference_elementwise.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/utility/reduction_enums.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+using F8  = ck::f8_t;
+
+using InputDataType  = F16;
+using ScaleDataType  = F32;
+using OutputDataType = F8;
+
+static constexpr ck::index_t NumDim = 2;
+
+constexpr ck::ReduceTensorOp ReduceOpId = ck::ReduceTensorOp::MAX;
+constexpr bool PropagateNan             = true;
+constexpr bool OutputIndex              = false;
+
+using ReduceOperation = typename ck::reduce_binary_operator<ReduceOpId>::opType;
+
+struct ScalePassThrough
+{
+    ScalePassThrough(const float alpha = 1.f) : alpha_(alpha) {}
+
+    __host__ __device__ constexpr void
+    operator()(OutputDataType& y0, OutputDataType& y1, const InputDataType& x0) const
+    {
+        y0 = ck::type_convert<OutputDataType>(ck::type_convert<ScaleDataType>(x0) * alpha_);
+        y1 = y0;
+    }
+
+    const ScaleDataType alpha_;
+};
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using UnaryAbs    = ck::tensor_operation::element_wise::UnaryAbs;
+
+using DeviceElementwisePermuteInstance = ck::tensor_operation::device::DeviceElementwiseImpl<
+    ck::Tuple<InputDataType>,                  // InDataTypeTuple
+    ck::Tuple<OutputDataType, OutputDataType>, // OutDataTypeTuple
+    ScalePassThrough,                          // Elementwise
+    NumDim,                                    // NumDim
+    256,                                       // BlockSize
+    128,                                       // M0PerBlock
+    128,                                       // M1PerBlock
+    8,                                         // M0PerThread
+    8,                                         // M1PerThread
+    ck::Sequence<1, 0>,                        // ThreadClusterArrangeOrder
+    ck::Sequence<8>,                           // InScalarPerVectorSeq
+    ck::Sequence<8, 1>>;                       // OutScalarPerVectorSeq
+
+using DeviceReduceInstance =
+    ck::tensor_operation::device::DeviceReduceMultiBlock<OutputDataType,
+                                                         OutputDataType,
+                                                         OutputDataType,
+                                                         NumDim,
+                                                         NumDim,
+                                                         ReduceOperation,
+                                                         UnaryAbs,
+                                                         PassThrough,
+                                                         ck::InMemoryDataOperationEnum::Set,
+                                                         PropagateNan,
+                                                         OutputIndex,
+                                                         false, // HaveIndexInputIfOutputIndex
+                                                         1024,  // BlockSize
+                                                         1,     // MThreadClusterSize
+                                                         1024,  // KThreadClusterSize
+                                                         1,     // MThreadSliceSize
+                                                         16,    // KThreadSliceSize
+                                                         1,     // InSrcVectorDim
+                                                         16,    // InSrceVectorSize
+                                                         1>;    // OutDstVectorSize
+
+void reference_scale_permute_amax(Tensor<InputDataType>& input,
+                                  Tensor<OutputDataType>& host_output_scaled_casted_transposed,
+                                  Tensor<OutputDataType>& host_output_scaled_casted,
+                                  Tensor<OutputDataType>& host_output_amax,
+                                  const float scale)
+{
+    ScalePassThrough out_element_op(scale);
+    const ck::index_t M = input.GetLengths()[0];
+    const ck::index_t K = input.GetLengths()[1];
+    for(ck::index_t m = 0; m < M; m++)
+    {
+        for(ck::index_t k = 0; k < K; k++)
+        {
+            OutputDataType y0, y1;
+            out_element_op(y0, y1, input(m, k));
+
+            host_output_scaled_casted(m, k)            = y0;
+            host_output_scaled_casted_transposed(m, k) = y1;
+            const OutputDataType y_fabs =
+                ck::type_convert<OutputDataType>(ck::math::abs(ck::type_convert<float>(y0)));
+            host_output_amax(0) = ck::math::max(y_fabs, host_output_amax(0));
+        }
+    }
+}
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    bool time_kernel     = true;
+
+    const float scale = 2.f;
+
+    ck::index_t M = 1024;
+    ck::index_t K = 1024;
+
+    if(argc == 3)
+    {
+        M = std::stoi(argv[1]);
+        K = std::stoi(argv[2]);
+    }
+
+    std::array<ck::index_t, 2> dims        = {M, K};
+    std::array<ck::index_t, 2> in_strides  = {K, 1};
+    std::array<ck::index_t, 2> out_strides = {1, M};
+
+    Tensor<InputDataType> input(dims, in_strides);
+    Tensor<OutputDataType> output_scaled_casted_transposed(dims, out_strides);
+    Tensor<OutputDataType> output_scaled_casted(dims, in_strides);
+    Tensor<OutputDataType> output_amax({1});
+
+    input.GenerateTensorValue(GeneratorTensor_3<InputDataType>{0.0, 1.0});
+
+    DeviceMem input_dev_buf(sizeof(InputDataType) * input.mDesc.GetElementSpaceSize());
+    DeviceMem output_scaled_casted_transposed_dev_buf(
+        sizeof(OutputDataType) * output_scaled_casted_transposed.mDesc.GetElementSpaceSize());
+    DeviceMem output_scaled_casted_dev_buf(sizeof(OutputDataType) *
+                                           output_scaled_casted.mDesc.GetElementSpaceSize());
+    DeviceMem output_amax_dev_buf(sizeof(OutputDataType) * output_amax.mDesc.GetElementSpaceSize());
+
+    input_dev_buf.ToDevice(input.mData.data());
+
+    std::array<const void*, 1> inputs = {input_dev_buf.GetDeviceBuffer()};
+    std::array<void*, 2> outputs      = {output_scaled_casted_transposed_dev_buf.GetDeviceBuffer(),
+                                    output_scaled_casted_dev_buf.GetDeviceBuffer()};
+
+    std::cout << "Input: " << input.mDesc << std::endl;
+    std::cout << "Scale: " << scale << std::endl;
+    std::cout << "Output scaled casted transposed: " << output_scaled_casted_transposed.mDesc
+              << std::endl;
+    std::cout << "Output scaled casted: " << output_scaled_casted.mDesc << std::endl;
+    std::cout << "Output amax: " << output_amax.mDesc << std::endl;
+
+    auto launch_transpose_scale = [&]() {
+        auto transposeScale = DeviceElementwisePermuteInstance{};
+        auto argument       = transposeScale.MakeArgumentPointer(dims,
+                                                           {in_strides},
+                                                           {out_strides, in_strides},
+                                                           inputs,
+                                                           outputs,
+                                                           ScalePassThrough{scale});
+
+        if(!transposeScale.IsSupportedArgument(argument.get()))
+        {
+            throw std::runtime_error(
+                "The runtime parameters seems not supported by the device instance, exiting!");
+        };
+
+        auto transposeScale_invoker_ptr = transposeScale.MakeInvokerPointer();
+        return transposeScale_invoker_ptr->Run(argument.get(), StreamConfig{nullptr, time_kernel});
+    };
+
+    auto launch_reduce = [&]() {
+        auto reduce = DeviceReduceInstance{};
+        auto reduce_argument_ptr =
+            reduce.MakeArgumentPointer(dims,
+                                       in_strides,
+                                       {1},    // Output Lengths
+                                       {1},    // Output Strides
+                                       {0, 1}, // Reduce Dims
+                                       static_cast<double>(1.f),
+                                       static_cast<double>(0.f),
+                                       output_scaled_casted_dev_buf.GetDeviceBuffer(),
+                                       nullptr,
+                                       output_amax_dev_buf.GetDeviceBuffer(),
+                                       nullptr,
+                                       UnaryAbs{},
+                                       PassThrough{});
+
+        if(!reduce.IsSupportedArgument(reduce_argument_ptr.get()))
+        {
+            throw std::runtime_error(
+                "The runtime parameters seems not supported by the device instance, exiting!");
+        };
+
+        auto invoker_ptr = reduce.MakeInvokerPointer();
+
+        return invoker_ptr->Run(reduce_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+    };
+
+    float ave_time = launch_transpose_scale();
+    ave_time += launch_reduce();
+    std::cout << "Perf: " << ave_time << " ms" << std::endl;
+    bool pass = true;
+
+    if(do_verification)
+    {
+        Tensor<OutputDataType> host_output_scaled_casted_transposed(dims, out_strides);
+        Tensor<OutputDataType> host_output_scaled_casted(dims, in_strides);
+        Tensor<OutputDataType> host_output_amax({1});
+
+        reference_scale_permute_amax(input,
+                                     host_output_scaled_casted_transposed,
+                                     host_output_scaled_casted,
+                                     host_output_amax,
+                                     scale);
+
+        output_scaled_casted_transposed_dev_buf.FromDevice(
+            output_scaled_casted_transposed.mData.data());
+        output_scaled_casted_dev_buf.FromDevice(output_scaled_casted.mData.data());
+        output_amax_dev_buf.FromDevice(output_amax.mData.data());
+
+        pass &= ck::utils::check_err(output_scaled_casted_transposed.mData,
+                                     host_output_scaled_casted_transposed.mData,
+                                     "Error: Incorrect results scaled transposed",
+                                     1e-3,
+                                     1e-3);
+        pass &= ck::utils::check_err(output_scaled_casted.mData,
+                                     host_output_scaled_casted.mData,
+                                     "Error: Incorrect results scaled",
+                                     1e-3,
+                                     1e-3);
+        pass &= ck::utils::check_err(
+            output_amax.mData, host_output_amax.mData, "Error: Incorrect results amax", 1e-3, 1e-3);
+    }
+
+    return pass ? 0 : 1;
+}

example/66_complex_contraction_bilinear/CMakeLists.txt

@@ -0,0 +1,3 @@
+add_example_executable(example_complex_contraction_bilinear_xdl_fp32 complex_contraction_bilinear_xdl_fp32.cpp)
+add_example_executable(example_complex_contraction_bilinear_xdl_fp64 complex_contraction_bilinear_xdl_fp64.cpp)
+

example/66_complex_contraction_bilinear/README.md

@@ -0,0 +1,11 @@
+# Instructions for ```example_complex_contraction_bilinear_xdl_fp32```
+
+## Run
+```bash
+#arg1: verification (0=no, 1=yes)
+#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
+#arg3: time kernel (0=no, 1=yes)
+./bin/example_contraction_bilinear_xdl_fp32 1 1 1
+```
+
+

example/66_complex_contraction_bilinear/common_instances.hpp

@@ -0,0 +1,196 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using F32  = float;
+using F64  = double;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+// Generic instances for fp32, fp16 and bf16 data types.
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp>
+// clang-format off
+using DeviceOpInstanceKK_Generic = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         Compute|
+        //#####################################|        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            Data|
+        //#####################################|        |        |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            Type|
+        //#####################################|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4, ComputeDataType>;
+// clang-format on
+
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp>
+// clang-format off
+using DeviceOpInstanceKN_Generic = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         Compute|
+        //#####################################|        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            Data|
+        //#####################################|        |        |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            Type|
+        //#####################################|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   4,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4, ComputeDataType>;
+// clang-format on
+
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp>
+// clang-format off
+using DeviceOpInstanceMK_Generic = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         Compute|
+        //#####################################|        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            Data|
+        //#####################################|        |        |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            Type|
+        //#####################################|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   4,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              4,              4,         1,           1,           1,              S<1, 16, 1, 16>,               4, ComputeDataType>;
+// clang-format on
+
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp>
+// clang-format off
+using DeviceOpInstanceMN_Generic = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         Compute|
+        //#####################################|        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            Data|
+        //#####################################|        |        |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            Type|
+        //#####################################|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   256,   128,    16,   1,   1,   32,   32,    4,    2,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              1,         0,           1,           1,              S<1, 16, 1, 16>,               4, ComputeDataType>;
+// clang-format on
+
+// Fp64 instances.
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp>
+// clang-format off
+using DeviceOpInstanceKK_FP64 = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         Compute|
+        //#####################################|        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            Data|
+        //#####################################|        |        |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            Type|
+        //#####################################|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1, ComputeDataType>;
+// clang-format on
+
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp>
+// clang-format off
+using DeviceOpInstanceKN_FP64 = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         Compute|
+        //#####################################|        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            Data|
+        //#####################################|        |        |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            Type|
+        //#####################################|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   2,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              2,              2,         1,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1, ComputeDataType>;
+// clang-format on
+
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp>
+// clang-format off
+using DeviceOpInstanceMK_FP64 = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         Compute|
+        //#####################################|        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            Data|
+        //#####################################|        |        |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            Type|
+        //#####################################|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   2,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              2,              2,         1,           1,           1,              S<1, 16, 1, 16>,               1, ComputeDataType>;
+// clang-format on
+
+template <ck::index_t NumDimM,
+          ck::index_t NumDimN,
+          ck::index_t NumDimK,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename ComputeDataType,
+          typename AElementOp,
+          typename BElementOp,
+          typename CDEElementOp>
+// clang-format off
+using DeviceOpInstanceMN_FP64 = ck::tensor_operation::device::
+        //#####################################| NumDimM| NumDimN| NumDimK|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|         Compute|
+        //#####################################|        |        |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|            Data|
+        //#####################################|        |        |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|            Type|
+        //#####################################|        |        |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |                |
+        DeviceContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    16,   1,   1,   16,   16,    4,    4,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              1,         0,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              1,         0,           1,           1,              S<1, 16, 1, 16>,               1, ComputeDataType>;
+// clang-format on

example/66_complex_contraction_bilinear/complex_contraction_bilinear_xdl_fp32.cpp

@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "common_instances.hpp"
+
+using ADataType        = F32;
+using BDataType        = F32;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F32;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F32;
+using ComputeDataType  = F32;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
+
+using DeviceOpInstanceKKNN = DeviceOpInstanceKK_Generic<NumDimM,
+                                                        NumDimN,
+                                                        NumDimK,
+                                                        ADataType,
+                                                        BDataType,
+                                                        AccDataType,
+                                                        CShuffleDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        ComputeDataType,
+                                                        AElementOp,
+                                                        BElementOp,
+                                                        CDEElementOp>;
+
+using DeviceOpInstanceKNNN = DeviceOpInstanceKN_Generic<NumDimM,
+                                                        NumDimN,
+                                                        NumDimK,
+                                                        ADataType,
+                                                        BDataType,
+                                                        AccDataType,
+                                                        CShuffleDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        ComputeDataType,
+                                                        AElementOp,
+                                                        BElementOp,
+                                                        CDEElementOp>;
+
+using DeviceOpInstanceMKNN = DeviceOpInstanceMK_Generic<NumDimM,
+                                                        NumDimN,
+                                                        NumDimK,
+                                                        ADataType,
+                                                        BDataType,
+                                                        AccDataType,
+                                                        CShuffleDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        ComputeDataType,
+                                                        AElementOp,
+                                                        BElementOp,
+                                                        CDEElementOp>;
+
+using DeviceOpInstanceMNNN = DeviceOpInstanceMN_Generic<NumDimM,
+                                                        NumDimN,
+                                                        NumDimK,
+                                                        ADataType,
+                                                        BDataType,
+                                                        AccDataType,
+                                                        CShuffleDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        ComputeDataType,
+                                                        AElementOp,
+                                                        BElementOp,
+                                                        CDEElementOp>;
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+#include "run_complex_contraction_bilinear_example.inc"
+
+int main(int argc, char* argv[]) { return run_complex_contraction_bilinear_example(argc, argv); }

example/66_complex_contraction_bilinear/complex_contraction_bilinear_xdl_fp64.cpp

@@ -0,0 +1,86 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "common_instances.hpp"
+
+using ADataType        = F64;
+using BDataType        = F64;
+using AccDataType      = F64;
+using CShuffleDataType = F64;
+using DDataType        = F64;
+using DsDataType       = ck::Tuple<DDataType>;
+using EDataType        = F64;
+using ComputeDataType  = F64;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+using AElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using BElementOp   = ck::tensor_operation::element_wise::PassThrough;
+using CDEElementOp = ck::tensor_operation::element_wise::Bilinear;
+
+using DeviceOpInstanceKKNN = DeviceOpInstanceKK_FP64<NumDimM,
+                                                     NumDimN,
+                                                     NumDimK,
+                                                     ADataType,
+                                                     BDataType,
+                                                     AccDataType,
+                                                     CShuffleDataType,
+                                                     DsDataType,
+                                                     EDataType,
+                                                     ComputeDataType,
+                                                     AElementOp,
+                                                     BElementOp,
+                                                     CDEElementOp>;
+
+using DeviceOpInstanceKNNN = DeviceOpInstanceKN_FP64<NumDimM,
+                                                     NumDimN,
+                                                     NumDimK,
+                                                     ADataType,
+                                                     BDataType,
+                                                     AccDataType,
+                                                     CShuffleDataType,
+                                                     DsDataType,
+                                                     EDataType,
+                                                     ComputeDataType,
+                                                     AElementOp,
+                                                     BElementOp,
+                                                     CDEElementOp>;
+
+using DeviceOpInstanceMKNN = DeviceOpInstanceMK_FP64<NumDimM,
+                                                     NumDimN,
+                                                     NumDimK,
+                                                     ADataType,
+                                                     BDataType,
+                                                     AccDataType,
+                                                     CShuffleDataType,
+                                                     DsDataType,
+                                                     EDataType,
+                                                     ComputeDataType,
+                                                     AElementOp,
+                                                     BElementOp,
+                                                     CDEElementOp>;
+
+using DeviceOpInstanceMNNN = DeviceOpInstanceMN_FP64<NumDimM,
+                                                     NumDimN,
+                                                     NumDimK,
+                                                     ADataType,
+                                                     BDataType,
+                                                     AccDataType,
+                                                     CShuffleDataType,
+                                                     DsDataType,
+                                                     EDataType,
+                                                     ComputeDataType,
+                                                     AElementOp,
+                                                     BElementOp,
+                                                     CDEElementOp>;
+
+using DeviceOpInstance = DeviceOpInstanceKKNN;
+
+#include "run_complex_contraction_bilinear_example.inc"
+
+int main(int argc, char* argv[]) { return run_complex_contraction_bilinear_example(argc, argv); }

example/66_complex_contraction_bilinear/run_complex_contraction_bilinear_example.inc

@@ -0,0 +1,484 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/numeric.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_contraction.hpp"
+
+int run_complex_contraction_bilinear_example(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // A[M0, M1, K0, K1]
+    std::vector<ck::index_t> a_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{524288, 4096, 128, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
+
+    float alpha = 1.f;
+    float beta  = 1.f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 28)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t M0 = std::stoi(argv[4]);
+        const ck::index_t M1 = std::stoi(argv[5]);
+
+        const ck::index_t N0 = std::stoi(argv[6]);
+        const ck::index_t N1 = std::stoi(argv[7]);
+
+        const ck::index_t K0 = std::stoi(argv[8]);
+        const ck::index_t K1 = std::stoi(argv[9]);
+
+        a_ms_ks_lengths = {M0, M1, K0, K1};
+        a_ms_ks_strides = {
+            std::stoi(argv[10]), std::stoi(argv[11]), std::stoi(argv[12]), std::stoi(argv[13])};
+
+        b_ns_ks_lengths = {N0, N1, K0, K1};
+        b_ns_ks_strides = {
+            std::stoi(argv[14]), std::stoi(argv[15]), std::stoi(argv[16]), std::stoi(argv[17])};
+
+        d_ms_ns_lengths = {M0, M1, N0, N1};
+        d_ms_ns_strides = {
+            std::stoi(argv[18]), std::stoi(argv[19]), std::stoi(argv[20]), std::stoi(argv[21])};
+
+        e_ms_ns_lengths = {M0, M1, N0, N1};
+        e_ms_ns_strides = {
+            std::stoi(argv[22]), std::stoi(argv[23]), std::stoi(argv[24]), std::stoi(argv[25])};
+
+        alpha = std::stof(argv[26]);
+        beta  = std::stof(argv[27]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M0, M1, N0, N1, K0, K1\n");
+        printf("arg10 to 13: Stride_A_M0, Stride_A_M1, Stride_A_K0, Stride_A_K1\n");
+        printf("arg14 to 17: Stride_B_N0, Stride_B_N1, Stride_B_K0, Stride_B_K1\n");
+        printf("arg18 to 21: Stride_D_M0, Stride_D_M1, Stride_D_N0, Stride_D_N1\n");
+        printf("arg22 to 25: Stride_E_M0, Stride_E_M1, Stride_E_N0, Stride_E_N1\n");
+        printf("arg26 to 27: alpha, beta\n");
+        exit(0);
+    }
+
+    // For Real Part of Complex Tensor
+    Tensor<ADataType> a_ms_ks_re(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks_re(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> d_ms_ns_re(d_ms_ns_lengths, d_ms_ns_strides);
+
+    Tensor<EDataType> e_ms_ns_host_result_re(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result_re(e_ms_ns_lengths, e_ms_ns_strides);
+
+    // For Imaginary Part of Complex Tensor
+    Tensor<ADataType> a_ms_ks_img(a_ms_ks_lengths, a_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks_img(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> d_ms_ns_img(d_ms_ns_lengths, d_ms_ns_strides);
+
+    Tensor<EDataType> e_ms_ns_host_result_img(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result_img(e_ms_ns_lengths, e_ms_ns_strides);
+
+    // Intermediate E tensor Definition
+    Tensor<EDataType> e_ms_ns_device_result_re1(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result_img1(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a_ms_ks_re: " << a_ms_ks_re.mDesc << std::endl;
+    std::cout << "b_ns_ks_re: " << b_ns_ks_re.mDesc << std::endl;
+    std::cout << "d_ms_ns_re: " << d_ms_ns_re.mDesc << std::endl;
+    std::cout << "e_ms_ns_re: " << e_ms_ns_host_result_re.mDesc << std::endl;
+
+    std::cout << "a_ms_ks_img: " << a_ms_ks_img.mDesc << std::endl;
+    std::cout << "b_ns_ks_img: " << b_ns_ks_img.mDesc << std::endl;
+    std::cout << "d_ms_ns_img: " << d_ms_ns_img.mDesc << std::endl;
+    std::cout << "e_ms_ns_img: " << e_ms_ns_host_result_img.mDesc << std::endl;
+
+    switch(init_method)
+    {
+        case 0: break;
+        case 1:
+
+            a_ms_ks_re.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_ns_ks_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            d_ms_ns_re.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+
+            a_ms_ks_img.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_ns_ks_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            d_ms_ns_img.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+
+        default:
+            a_ms_ks_re.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_ns_ks_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            d_ms_ns_re.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+
+            a_ms_ks_img.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_ns_ks_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            d_ms_ns_img.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+
+            break;
+    }
+
+    DeviceMem a_device_buf_re(sizeof(ADataType) * a_ms_ks_re.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf_re(sizeof(BDataType) * b_ns_ks_re.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf_re(sizeof(DDataType) * d_ms_ns_re.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf_re(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
+
+    DeviceMem a_device_buf_img(sizeof(ADataType) * a_ms_ks_img.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf_img(sizeof(BDataType) * b_ns_ks_img.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf_img(sizeof(DDataType) * d_ms_ns_img.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf_img(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
+
+    // Intermediate Value For E Real and Img
+    DeviceMem e_device_buf_re1(sizeof(EDataType) * e_ms_ns_device_result_re.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf_img1(sizeof(EDataType) * e_ms_ns_device_result_img.mDesc.GetElementSpaceSize());
+
+
+    a_device_buf_re.ToDevice(a_ms_ks_re.mData.data());
+    b_device_buf_re.ToDevice(b_ns_ks_re.mData.data());
+    d_device_buf_re.ToDevice(d_ms_ns_re.mData.data());
+
+    a_device_buf_img.ToDevice(a_ms_ks_img.mData.data());
+    b_device_buf_img.ToDevice(b_ns_ks_img.mData.data());
+    d_device_buf_img.ToDevice(d_ms_ns_img.mData.data());
+
+    // set zero
+    e_device_buf_re.SetZero();
+    e_device_buf_img.SetZero();
+
+    // set zero for intermediate values
+    e_device_buf_re1.SetZero();
+    e_device_buf_img1.SetZero();
+ 
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // device operation
+    // For real Intermediate Value re_1
+
+    auto op       = DeviceOpInstance{};
+    auto invoker  = op.MakeInvoker();
+    auto argument_re1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
+                                    b_device_buf_re.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{d_device_buf_re.GetDeviceBuffer()},
+                                    e_device_buf_re1.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument_re1))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time_re1 = invoker.Run(argument_re1, StreamConfig{nullptr, time_kernel});
+
+
+    alpha = -1.f;
+    beta  = 1.f;
+
+    a_element_op   = AElementOp{};
+    b_element_op   = BElementOp{};
+    cde_element_op = CDEElementOp{alpha, beta};
+
+    // device operation
+    // For real Intermediate Value re_2
+    // auto op       = DeviceOpInstance{};
+    // auto invoker  = op.MakeInvoker();
+    auto argument_re2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
+                                    b_device_buf_img.GetDeviceBuffer(),
+                                    std::array<const void*, 1>{e_device_buf_re1.GetDeviceBuffer()},
+                                    e_device_buf_re.GetDeviceBuffer(),
+                                    a_ms_ks_lengths,
+                                    a_ms_ks_strides,
+                                    b_ns_ks_lengths,
+                                    b_ns_ks_strides,
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                    std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                    e_ms_ns_lengths,
+                                    e_ms_ns_strides,
+                                    a_element_op,
+                                    b_element_op,
+                                    cde_element_op);
+
+    if(!op.IsSupportedArgument(argument_re2))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time_re2 = invoker.Run(argument_re2, StreamConfig{nullptr, time_kernel});
+
+    
+    alpha = 1.f;
+    beta  = 1.f;
+
+    a_element_op   = AElementOp{};
+    b_element_op   = BElementOp{};
+    cde_element_op = CDEElementOp{alpha, beta};
+
+    auto argument_img1 = op.MakeArgument(a_device_buf_re.GetDeviceBuffer(),
+                                b_device_buf_img.GetDeviceBuffer(),
+                                std::array<const void*, 1>{d_device_buf_img.GetDeviceBuffer()},
+                                e_device_buf_img1.GetDeviceBuffer(),
+                                a_ms_ks_lengths,
+                                a_ms_ks_strides,
+                                b_ns_ks_lengths,
+                                b_ns_ks_strides,
+                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                e_ms_ns_lengths,
+                                e_ms_ns_strides,
+                                a_element_op,
+                                b_element_op,
+                                cde_element_op);
+
+
+    if(!op.IsSupportedArgument(argument_img1))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time_img1 = invoker.Run(argument_img1, StreamConfig{nullptr, time_kernel});
+
+    alpha = 1.f;
+    beta  = 1.f;
+
+    auto argument_img2 = op.MakeArgument(a_device_buf_img.GetDeviceBuffer(),
+                                b_device_buf_re.GetDeviceBuffer(),
+                                std::array<const void*, 1>{e_device_buf_img1.GetDeviceBuffer()},
+                                e_device_buf_img.GetDeviceBuffer(),
+                                a_ms_ks_lengths,
+                                a_ms_ks_strides,
+                                b_ns_ks_lengths,
+                                b_ns_ks_strides,
+                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+                                std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+                                e_ms_ns_lengths,
+                                e_ms_ns_strides,
+                                a_element_op,
+                                b_element_op,
+                                cde_element_op);
+
+
+
+    if(!op.IsSupportedArgument(argument_img2))
+    {
+        std::cout << op.GetTypeString() << " does not support this problem" << std::endl;
+
+        return 0;
+    }
+
+    float ave_time_img2 = invoker.Run(argument_img2, StreamConfig{nullptr, time_kernel});
+
+
+    ck::index_t M =
+        ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+    ck::index_t N = ck::accumulate_n<ck::index_t>(
+        e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+    ck::index_t K = ck::accumulate_n<ck::index_t>(
+        a_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+    std::size_t flop      = std::size_t(2) * M * N * K * 2;
+    std::size_t num_btype = sizeof(ADataType) * M * K + sizeof(BDataType) * K * N +
+                            sizeof(DDataType) * M * N + sizeof(EDataType) * M * N * 2;
+
+    float ave_time = ave_time_img2 + ave_time_img1 + ave_time_re2 + ave_time_re1 ; 
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+    std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << op.GetTypeString() << std::endl;
+
+    e_device_buf_re.FromDevice(e_ms_ns_device_result_re.mData.data());
+    e_device_buf_img.FromDevice(e_ms_ns_device_result_img.mData.data());
+
+    auto isRealOk = 0;
+    auto isImgOk = 0;
+
+    if(do_verification)
+    {
+        // Real Part Verification
+        Tensor<CShuffleDataType> c_ms_ns_host_result_re(e_ms_ns_lengths, e_ms_ns_strides);
+        Tensor<CShuffleDataType> c_ms_ns_host_result_re1(e_ms_ns_lengths, e_ms_ns_strides);
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      ADataType,
+                                                                      BDataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      F32,
+                                                                      AElementOp,
+                                                                      BElementOp>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        auto ref_argument_re =
+            ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_re, c_ms_ns_host_result_re, a_element_op, b_element_op);
+
+        ref_invoker.Run(ref_argument_re);
+
+        alpha = 1.f;
+        beta  = 1.f;
+   
+        cde_element_op = CDEElementOp{alpha, beta};
+
+       
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result_re.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result_re.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result_re.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result_re.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result_re(m0, m1, n0, n1),
+                                       c_ms_ns_host_result_re(m0, m1, n0, n1),
+                                       d_ms_ns_re(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        alpha = 1.f;
+        beta  = -1.f;
+   
+        cde_element_op = CDEElementOp{alpha, beta};
+
+        auto ref_argument_re1 =
+            ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_img, c_ms_ns_host_result_re1, a_element_op, b_element_op);
+
+        ref_invoker.Run(ref_argument_re1);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result_re.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result_re.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result_re.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result_re.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result_re(m0, m1, n0, n1),
+                                       e_ms_ns_host_result_re(m0, m1, n0, n1),
+                                       c_ms_ns_host_result_re1(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        isRealOk =  ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
+
+        
+
+
+        // Img Part Verification
+        Tensor<CShuffleDataType> c_ms_ns_host_result_img(e_ms_ns_lengths, e_ms_ns_strides);
+        Tensor<CShuffleDataType> c_ms_ns_host_result_img1(e_ms_ns_lengths, e_ms_ns_strides);
+
+        auto ref_argument_img =
+            ref_op.MakeArgument(a_ms_ks_re, b_ns_ks_img, c_ms_ns_host_result_img, a_element_op, b_element_op);
+            
+        ref_invoker.Run(ref_argument_img);
+
+        alpha = 1.f;
+        beta  = 1.f;
+   
+        cde_element_op = CDEElementOp{alpha, beta};
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result_img.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result_img.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result_img.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result_img(m0, m1, n0, n1),
+                                       c_ms_ns_host_result_img(m0, m1, n0, n1),
+                                       d_ms_ns_img(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        auto ref_argument_img1 =
+            ref_op.MakeArgument(a_ms_ks_img, b_ns_ks_re, c_ms_ns_host_result_img1, a_element_op, b_element_op);
+            
+        ref_invoker.Run(ref_argument_img1);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result_img.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result_img.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result_img.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result_img.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result_img(m0, m1, n0, n1),
+                                       e_ms_ns_host_result_img(m0, m1, n0, n1),
+                                       c_ms_ns_host_result_img1(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        isImgOk =  ck::utils::check_err(e_ms_ns_device_result_re, e_ms_ns_host_result_re) ? 0 : 1;
+
+        return (isRealOk && isImgOk);
+    }
+
+    return 0;
+}

example/CMakeLists.txt

@@ -45,11 +45,7 @@ function(add_example_executable EXAMPLE_NAME FILE_NAME)
     endforeach()
     endif()
 
-    if(INSTANCES_ONLY)
-        set(EX_TARGETS ${DEFAULT_GPU_TARGETS})
-    else()
-        set(EX_TARGETS ${GPU_TARGETS})
-    endif()
+    set(EX_TARGETS ${SUPPORTED_GPU_TARGETS})
 
     #Do not build any DL examples if DL_KERNELS not set
     foreach(source IN LISTS FILE_NAME)
@@ -147,11 +143,8 @@ function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
     endforeach()
     endif()
 
-    if(INSTANCES_ONLY)
-        set(EX_TARGETS ${DEFAULT_GPU_TARGETS})
-    else()
-        set(EX_TARGETS ${GPU_TARGETS})
-    endif()
+    set(EX_TARGETS ${SUPPORTED_GPU_TARGETS})
+
     #Do not build any DL examples if DL_KERNELS not set
     foreach(source IN LISTS FILE_NAME)
         if(NOT DEFINED DL_KERNELS AND source MATCHES "_dl")

example/ck_tile/01_fmha/README.md

@@ -6,7 +6,8 @@ This folder contains example for fmha(fused multi-head attention) using ck_tile
 ```
 # in the root of ck_tile
 mkdir build && cd build
-sh ../script/cmake-ck-dev.sh  ../ <arch>  # you can replace this <arch> to gfx90a, gfx942...
+# you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
+sh ../script/cmake-ck-dev.sh  ../ <arch>
 make tile_example_fmha_fwd -j
 ```
 This will result in an executable `build/bin/tile_example_fmha_fwd`
@@ -23,7 +24,7 @@ There are 3 template parameters for this kernel template.
 To speed up compile time, we instantiate the kernels into separate file. In this way we can benefit from parallel building from CMake/Make system. This is achieved by `generate.py` script. Besides, you can look into this script to learn how to instantiate a kernel instance step by step, which is described in `FMHA_FWD_KERNEL_BODY` variable.
 
 ## executable
-`tile_example_fmha_fwd` is the example executable, implemented in `fmha_fwd.cpp`. You can type `./bin/tile_example_fmha_fwd -?` to list all supported args. Below is an example of the output (may subject to change)
+`tile_example_fmha_fwd` is the example executable, implemented in `fmha_fwd.cpp`. You can type `./bin/tile_example_fmha_fwd -?` to list all the arguments. Below is an example of the output (may subject to change)
 ```
 args:
           -v    weather do CPU validation or not (default:1)
@@ -31,47 +32,52 @@ args:
           -b    batch size (default:2)
           -h    num of head, for q (default:8)
         -h_k    num of head, for k/v, -1 means equal to h (default:-1)
-                 if not equal to h, then this is GQA/MQA case
+                if not equal to h, then this is GQA/MQA case
           -s    seqlen_q. if group-mode, means the average value of seqlen_q (default:3328)
-                 total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary
-                 also with "-s=s0,s1,s2..." comma seperated int to set per batch seqlen(group-mode)
-        -s_k    seqlen_k, -1 means equal to s (default:-1)
+                total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary
+                also with "-s=s0,s1,s2..." comma seperated int to set per batch seqlen(group-mode)
+        -s_k    seqlen_k (including new key/value), -1 means equal to s (default:-1)
           -d    head dim for q, k (default:128)
         -d_v    head dim for v, -1 means equal to d (default:-1)
     -scale_s    scale factor of S. 0 means equal to 1/sqrt(hdim). (default:0)
-                 note when squant=1, this value will be modified by range_q/k
+                note when squant=1, this value will be modified by range_q/k
     -range_q    per-tensor quantization range of q. used if squant=1. (default:16)
     -range_k    per-tensor quantization range of k. used if squant=1. (default:16)
     -range_v    per-tensor quantization range of v. used if squant=1. (default:16)
     -range_p    per-tensor quantization range of p [e^(s-m)]. used if squant=1. (default:1)
     -range_o    per-tensor quantization range of o (p*v). used if squant=1. (default:16)
      -squant    if using static quantization fusion or not. auto: fp8 will default use squant, other will not (default:auto)
-                 0: no static quant(not implemented) 1: apply scale_p and scale_o with respect to P and O.
-                 calculate scale_s, scale_p, scale_o according to range_q, range_k, range_v, range_p, range_o
+                0: no static quant(not implemented) 1: apply scale_p and scale_o with respect to P and O.
+                calculate scale_s, scale_p, scale_o according to range_q, range_k, range_v, range_p, range_o
       -iperm    permute input (default:1)
-                 if true, will be b*h*s*d, else b*s*h*d
+                if true, will be b*h*s*d, else b*s*h*d
       -operm    permute output (default:1)
        -bias    n or 0, no bias (default:n)
-                 e(lementwise) or 1, elementwise bias with 1*1*s*s. e:1, 1*h*s*s. e:2, b*h*s*s
-                 a(libi) or 2, alibi with 1*h. a:1, b*h
+                e(lementwise) or 1, elementwise bias with 1*1*s*s. e:1, 1*h*s*s. e:2, b*h*s*s
+                a(libi) or 2, alibi with 1*h. a:1, b*h
        -prec    data type. fp16/bf16/fp8/bf8 (default:fp16)
        -mask    0: no mask, 1: top-left(same as 't'), 2:bottom-right(same as 'b') (default:0)
-                 't', top-left causal mask, 'b', bottom-r causal mask
-                 't:l,r', top-left sliding window attn(swa) with FA style left right size
-                 'b:l,r', bottom-r sliding window attn(swa) with FA style left right size
-                 'xt:window_size', xformer style masking from top-left, window_size negative is causal, positive is swa
-                 'xb:window_size', xformer style masking from bottom-r, window_size negative is causal, positive is swa
-                 'g:y,x', generic attention mask coordinate with y/x size (only debug purpose for now)
+                't', top-left causal mask, 'b', bottom-r causal mask
+                't:l,r', top-left sliding window attn(swa) with FA style left right size
+                'b:l,r', bottom-r sliding window attn(swa) with FA style left right size
+                'xt:window_size', xformer style masking from top-left, window_size negative is causal, positive is swa
+                'xb:window_size', xformer style masking from bottom-r, window_size negative is causal, positive is swa
+                'g:y,x', generic attention mask coordinate with y/x size (only debug purpose for now)
     -vlayout    r for row-major(seqlen*hdim), c for col-major(hdim*seqlen) (default:r)
         -lse    0 not store lse, 1 store lse (default:0)
       -kname    if set to 1 will print kernel name (default:0)
        -init    init method. ui, uniform random int, ni, normalized random int (default:uf)
-                 uf, uniform random float, nf, normalized random float, tf, trig float, uf:q, quantization
+                uf, uniform random float, nf, normalized random float, tf, trig float, uf:q, quantization
        -seed    random seed used for initializing input tensors. 0 for non-deterministic seed (default:11939)
+  -drop_seed    seed for random number generator (default:1)
+-drop_offset    offset for random number generator (default:0)
+ -drop_prefs    seed and offset values are present on GPU; 0 - host, 1 - device/GPU (default:0)
      -warmup    number of iterations before benchmark the kernel (default:5)
      -repeat    number of iterations to benchmark the kernel (default:20)
 ```
-Example: `./bin/tile_example_fmha_fwd -b=1 -h=16 -s=16384 -d=128` will run a fmha case with batch=1, nhead=16, sequence length=16384, hdim=128, fp16 case.
+Example 1: `./bin/tile_example_fmha_fwd -b=1 -h=16 -s=16384 -d=128` will run a fmha case with batch=1, nhead=16, sequence length=16384, hdim=128, fp16 case.
+Example 2: `./bin/tile_example_fmha_fwd -b=1 -h=8 -s=16384 -d=64 -drop_prefs=1 -drop_seed=10 -drop_offset=1234` will run a fmha case with 
+  batch=1, nhead=8, sequence length=16384, hdim=64, drop_seed=0 (in GPU memory), drop_offset=1234 (in GPU memory) fp16 case
 
 ## support features
 Currently we are still in rapid development stage, so more features/optimizations will be coming soon.

example/ck_tile/01_fmha/codegen/ops/fmha_fwd_splitkv.py

@@ -600,8 +600,8 @@ def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) ->
                 # TODO: use async pipeline when compiler is more stable 
                 if hdim == 256 or hdim in [32, 64, 128]:
                 # if True:
-                    pipelines.append(Pipeline('qr', 'row', 'f', 'f', 'f', 'f', bias, lse, squant, pagedkv, mask))
-                    pipelines.append(Pipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, lse, squant, pagedkv, mask))
+                    pipelines.append(Pipeline('qr', 'row', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
+                    pipelines.append(Pipeline('qr', 'col', 'f', 't', 'f', 'f', bias, lse, squant, pagedkv, mask))
 
                     pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask))
                     pipelines.append(Pipeline('qr', 'col', 't', 't', 't', 't', bias, lse, squant, pagedkv, mask))

example/ck_tile/01_fmha/fmha_bwd.cpp

@@ -85,6 +85,9 @@ auto create_args(int argc, char* argv[])
         .insert("p_drop", "0", "0~1 probability of dropout")
         .insert("drop_seed", "1", "seed for random number generator")
         .insert("drop_offset", "0", "offset for random number generator")
+        .insert("drop_prefs",
+                "0",
+                "seed and offset values are present on GPU; 0 - host, 1 - device/GPU")
         .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
         .insert("warmup", "5", "number of iterations before benchmark the kernel")
         .insert("repeat", "20", "number of iterations to benchmark the kernel")
@@ -110,13 +113,26 @@ auto create_args(int argc, char* argv[])
 
 // different threshold for different dtype
 template <typename DataType>
-auto get_elimit(int /*init_method*/)
+auto get_elimit(ck_tile::index_t /*hdim_q*/, ck_tile::index_t /*hdim_v*/)
 {
     double rtol = 1e-2;
     double atol = 1e-2;
     return ck_tile::make_tuple(rtol, atol);
 }
 
+template <>
+auto get_elimit<ck_tile::bf16_t>(ck_tile::index_t hdim_q, ck_tile::index_t hdim_v)
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    if(hdim_q > 128 && hdim_v > 128) // 3.2 for RTZ/1.5 for RTN
+    {
+        rtol = 3.2e-2;
+        atol = 3.2e-2;
+    }
+    return ck_tile::make_tuple(rtol, atol);
+}
+
 template <typename DataType>
 bool run(const ck_tile::ArgParser& arg_parser)
 {
@@ -156,6 +172,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     float p_drop         = arg_parser.get_float("p_drop");
     uint64_t drop_seed   = arg_parser.get_uint64("drop_seed");
     uint64_t drop_offset = arg_parser.get_uint64("drop_offset");
+    bool drop_prefs      = arg_parser.get_bool("drop_prefs");
+
     if(use_dbias && bias.type != bias_enum::elementwise_bias)
     {
         std::cerr << "dbias only exists when bias type is elementwise" << std::endl;
@@ -381,6 +399,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     ck_tile::DeviceMem dbias_buf(dbias_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem seqstart_q(seqstart_q_host.size() * sizeof(int32_t));
     ck_tile::DeviceMem seqstart_k(seqstart_k_host.size() * sizeof(int32_t));
+    ck_tile::DeviceMem drop_seed_buf(drop_prefs ? sizeof(uint64_t) : 0);
+    ck_tile::DeviceMem drop_offset_buf(drop_prefs ? sizeof(uint64_t) : 0);
     ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem dq_acc_buf(dq_acc_host.get_element_space_size_in_bytes());
 
@@ -391,6 +411,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     do_buf.ToDevice(do_host.data());
     seqstart_q.ToDevice(seqstart_q_host.data());
     seqstart_k.ToDevice(seqstart_k_host.data());
+    drop_seed_buf.ToDevice(drop_prefs ? &drop_seed : nullptr);
+    drop_offset_buf.ToDevice(drop_prefs ? &drop_offset : nullptr);
     alibi_slope_buf.ToDevice(alibi_slope_host.data());
 
     // clang-format off
@@ -478,6 +500,18 @@ bool run(const ck_tile::ArgParser& arg_parser)
         const ck_tile::index_t split_stride_dq_acc =
             (shape_batch * nhead * shape_seqlen_q * hdim_q);
 
+        const auto drop_seed_offset = [&]() -> decltype(fmha_bwd_args::drop_seed_offset) {
+            if(drop_prefs)
+            {
+                return std::make_pair(drop_seed_buf.GetDeviceBuffer(),
+                                      drop_offset_buf.GetDeviceBuffer());
+            }
+            else
+            {
+                return std::make_pair(drop_seed, drop_offset);
+            }
+        }();
+
         return fmha_bwd_args{q_buf.GetDeviceBuffer(),
                              k_buf.GetDeviceBuffer(),
                              v_buf.GetDeviceBuffer(),
@@ -551,7 +585,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
                              static_cast<ck_tile::index_t>(mask.type),
                              p_drop,
                              p_undrop,
-                             {drop_seed, drop_offset}};
+                             drop_seed_offset};
     }();
 
     float ave_time = fmha_bwd(fmha_traits, fmha_args, stream_config);
@@ -918,7 +952,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
         }
         // clang-format on
 
-        auto [rtol, atol] = get_elimit<DataType>(init_method);
+        auto [rtol, atol] = get_elimit<DataType>(hdim_q, hdim_v);
         bool dq_cur_pass  = ck_tile::check_err(dq_host_result,
                                               dq_host_ref,
                                               std::string("Error: QGrad Incorrect results!"),

example/ck_tile/01_fmha/fmha_bwd.hpp

@@ -9,7 +9,10 @@
 #include "ck_tile/ops/epilogue.hpp"
 #include "mask.hpp"
 #include "bias.hpp"
+
 #include <type_traits>
+#include <utility>
+#include <variant>
 
 template <typename DataType>
 struct FmhaBwdTypeConfig;
@@ -135,7 +138,8 @@ struct fmha_bwd_args
     ck_tile::index_t mask_type;
     float p_drop;
     float p_undrop;
-    std::tuple<uint64_t, uint64_t> drop_seed_offset;
+    std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+        drop_seed_offset;
 };
 
 template <typename FmhaBwdDQDKDVKernel>

example/ck_tile/01_fmha/fmha_fwd.cpp

@@ -122,6 +122,9 @@ auto create_args(int argc, char* argv[])
         .insert("p_drop", "0", "0~1 probability of dropout")
         .insert("drop_seed", "1", "seed for random number generator")
         .insert("drop_offset", "0", "offset for random number generator")
+        .insert("drop_prefs",
+                "0",
+                "seed and offset values are present on GPU; 0 - host, 1 - device/GPU")
         .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
         .insert(
             "rotary_dim", "0", "RoPE rotary dimension. rotary_dim <= 0 means not apply RoPE at all")
@@ -442,6 +445,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     float p_drop         = arg_parser.get_float("p_drop");
     uint64_t drop_seed   = arg_parser.get_uint64("drop_seed");
     uint64_t drop_offset = arg_parser.get_uint64("drop_offset");
+    bool drop_prefs      = arg_parser.get_bool("drop_prefs");
+
     if(p_drop < 0.0f || p_drop > 1.0f)
     {
         std::cerr << "The value of p_drop should be 0~1" << std::endl;
@@ -552,16 +557,33 @@ bool run(const ck_tile::ArgParser& arg_parser)
     }
 #endif
 
-    auto get_lengths = [&](bool permute,
-                           ck_tile::index_t b /*batch*/,
-                           ck_tile::index_t h /*nhead*/,
-                           ck_tile::index_t s /*seqlen*/,
-                           ck_tile::index_t d /*hdim*/) {
-        if(permute)
-            return std::array<ck_tile::index_t, 4>{b, h, s, d};
-        else
-            return std::array<ck_tile::index_t, 4>{b, s, h, d};
-    };
+    struct
+    {
+        auto operator()(bool permute,
+                        ck_tile::index_t b /*batch*/,
+                        ck_tile::index_t h /*nhead*/,
+                        ck_tile::index_t s /*seqlen*/,
+                        ck_tile::index_t d /*hdim*/)
+        {
+            if(permute)
+                return std::array<ck_tile::index_t, 4>{b, h, s, d};
+            else
+                return std::array<ck_tile::index_t, 4>{b, s, h, d};
+        }
+
+        auto operator()(bool permute,
+                        ck_tile::index_t ns /*num_splits*/,
+                        ck_tile::index_t b /*batch*/,
+                        ck_tile::index_t h /*nhead*/,
+                        ck_tile::index_t s /*seqlen*/,
+                        ck_tile::index_t d /*hdim*/)
+        {
+            if(permute)
+                return std::array<ck_tile::index_t, 5>{ns, b, h, s, d};
+            else
+                return std::array<ck_tile::index_t, 5>{ns, b, s, h, d};
+        }
+    } get_lengths;
 
     bool is_v_rowmajor = vlayout == std::string("r");
 
@@ -617,7 +639,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
             : std::array<ck_tile::index_t, 4>{1, 1, 1, 1});
     ck_tile::HostTensor<OaccDataType> o_acc_host(
         1 < num_splits || use_kvcache
-            ? std::array<ck_tile::index_t, 5>{num_splits, batch, nhead, max_seqlen_q, hdim_v}
+            ? get_lengths(o_perm, num_splits, shape_batch, nhead, shape_seqlen_q, hdim_v)
             : std::array<ck_tile::index_t, 5>{1, 1, 1, 1, 1});
 
     // batch mode of lse data layout is [batch, nhead, seqlen_q]
@@ -739,6 +761,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
         need_append_kvcache ? cache_seqlen_ks.size() * sizeof(int32_t) : 0);
     ck_tile::DeviceMem rotary_cos_buf(rotary_cos_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem rotary_sin_buf(rotary_sin_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem drop_seed_buf(drop_prefs ? sizeof(uint64_t) : 0);
+    ck_tile::DeviceMem drop_offset_buf(drop_prefs ? sizeof(uint64_t) : 0);
     ck_tile::DeviceMem randval_buf(randval_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem alibi_slope_buf(alibi_slope_host.get_element_space_size_in_bytes());
     ck_tile::DeviceMem block_table_buf(block_table_host.get_element_space_size_in_bytes());
@@ -757,6 +781,8 @@ bool run(const ck_tile::ArgParser& arg_parser)
     cache_seqlen_k_buf.ToDevice(need_append_kvcache ? cache_seqlen_ks.data() : nullptr);
     rotary_cos_buf.ToDevice(rotary_cos_host.data());
     rotary_sin_buf.ToDevice(rotary_sin_host.data());
+    drop_seed_buf.ToDevice(drop_prefs ? &drop_seed : nullptr);
+    drop_offset_buf.ToDevice(drop_prefs ? &drop_offset : nullptr);
     alibi_slope_buf.ToDevice(alibi_slope_host.data());
     block_table_buf.ToDevice(block_table_host.data());
     cache_batch_idx_buf.ToDevice(cache_batch_idx_host.data());
@@ -854,7 +880,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
         }();
         const ck_tile::index_t stride_bias    = (i_perm ? shape_seqlen_k : 1 * shape_seqlen_k);
         const ck_tile::index_t stride_randval = (max_seqlen_k);
-        const ck_tile::index_t stride_o_acc   = hdim_v;
+        const ck_tile::index_t stride_o_acc   = (o_perm ? hdim_v : nhead * hdim_v);
         const ck_tile::index_t stride_o       = (o_perm ? hdim_v : nhead * hdim_v);
         // setup nhead_stride_* arguments
         const ck_tile::index_t nhead_stride_q = (i_perm ? shape_seqlen_q * hdim_q : hdim_q);
@@ -881,7 +907,7 @@ bool run(const ck_tile::ArgParser& arg_parser)
         const ck_tile::index_t nhead_stride_randval = (shape_seqlen_q * max_seqlen_k);
         const ck_tile::index_t nhead_stride_lse     = shape_seqlen_q;
         const ck_tile::index_t nhead_stride_lse_acc = shape_seqlen_q;
-        const ck_tile::index_t nhead_stride_o_acc   = (max_seqlen_q * hdim_v);
+        const ck_tile::index_t nhead_stride_o_acc   = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
         const ck_tile::index_t nhead_stride_o       = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
         // setup batch_stride_* arguments
         const ck_tile::index_t batch_stride_q = (nhead * shape_seqlen_q * hdim_q);
@@ -897,12 +923,12 @@ bool run(const ck_tile::ArgParser& arg_parser)
         const ck_tile::index_t batch_stride_randval = (nhead * shape_seqlen_q * max_seqlen_k);
         const ck_tile::index_t batch_stride_lse     = (nhead * shape_seqlen_q);
         const ck_tile::index_t batch_stride_lse_acc = (nhead * shape_seqlen_q);
-        const ck_tile::index_t batch_stride_o_acc   = (nhead * max_seqlen_q * hdim_v);
+        const ck_tile::index_t batch_stride_o_acc   = (nhead * shape_seqlen_q * hdim_v);
         const ck_tile::index_t batch_stride_o       = (nhead * shape_seqlen_q * hdim_v);
         const ck_tile::index_t batch_stride_block_table = (max_num_page_blocks / batch);
         // setup split_stride_* arguments (only used in split-kv kernel)
         const ck_tile::index_t split_stride_lse_acc = (shape_batch * nhead * shape_seqlen_q);
-        const ck_tile::index_t split_stride_o_acc   = (batch * nhead * max_seqlen_q * hdim_v);
+        const ck_tile::index_t split_stride_o_acc = (shape_batch * nhead * shape_seqlen_q * hdim_v);
 
         args.q_ptr = q_buf.GetDeviceBuffer();
         args.k_ptr = k_buf.GetDeviceBuffer();
@@ -996,9 +1022,17 @@ bool run(const ck_tile::ArgParser& arg_parser)
                 args.nhead_stride_randval = nhead_stride_randval;
                 args.batch_stride_randval = batch_stride_randval;
 
-                args.p_drop           = p_drop;
-                args.s_randval        = s_randval;
-                args.drop_seed_offset = std::tie(drop_seed, drop_offset);
+                args.p_drop    = p_drop;
+                args.s_randval = s_randval;
+                if(drop_prefs)
+                {
+                    args.drop_seed_offset = std::make_pair(drop_seed_buf.GetDeviceBuffer(),
+                                                           drop_offset_buf.GetDeviceBuffer());
+                }
+                else
+                {
+                    args.drop_seed_offset = std::make_pair(drop_seed, drop_offset);
+                }
             }
             else if constexpr(std::is_same_v<fmha_fwd_splitkv_args, std::decay_t<decltype(args)>>)
             {

example/ck_tile/01_fmha/fmha_fwd.hpp

@@ -13,6 +13,8 @@
 #include "rotary.hpp"
 
 #include <type_traits>
+#include <utility>
+#include <variant>
 
 template <typename DataType>
 struct FmhaFwdTypeConfig;
@@ -144,7 +146,9 @@ struct fmha_fwd_args
 
     float p_drop;
     bool s_randval;
-    std::tuple<uint64_t, uint64_t> drop_seed_offset;
+
+    std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+        drop_seed_offset;
 };
 
 struct fmha_fwd_splitkv_args
@@ -398,10 +402,8 @@ auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_splitkv_args args)
                                      args.nhead_stride_bias,
                                      args.nhead_stride_lse_acc,
                                      args.nhead_stride_o_acc,
-                                     args.batch_stride_k,
-                                     args.batch_stride_v,
-                                     args.batch_stride_lse_acc,
-                                     args.batch_stride_o_acc,
+                                     args.batch_stride_k, // only used for paged-kvcache
+                                     args.batch_stride_v, // only used for paged-kvcache
                                      args.split_stride_lse_acc,
                                      args.split_stride_o_acc,
                                      args.window_size_left,
@@ -475,7 +477,6 @@ auto fmha_fwd_splitkv_combine_create_kargs_and_grids(fmha_fwd_splitkv_args args)
                                      args.lse_ptr,
                                      args.o_ptr,
                                      args.batch,
-                                     args.max_seqlen_q,
                                      args.seqstart_q_ptr,
                                      args.hdim_v,
                                      args.num_splits,
@@ -486,7 +487,6 @@ auto fmha_fwd_splitkv_combine_create_kargs_and_grids(fmha_fwd_splitkv_args args)
                                      args.nhead_stride_o_acc,
                                      args.nhead_stride_lse,
                                      args.nhead_stride_o,
-                                     args.batch_stride_o_acc,
                                      args.split_stride_lse_acc,
                                      args.split_stride_o_acc);
         }
@@ -497,7 +497,6 @@ auto fmha_fwd_splitkv_combine_create_kargs_and_grids(fmha_fwd_splitkv_args args)
                                      args.lse_ptr,
                                      args.o_ptr,
                                      args.batch,
-                                     args.max_seqlen_q,
                                      args.seqlen_q,
                                      args.hdim_v,
                                      args.num_splits,

example/ck_tile/02_layernorm2d/README.md

@@ -6,7 +6,8 @@ This folder contains example for Layernorm2D forward using ck_tile tile-programm
 ```
 # in the root of ck_tile
 mkdir build && cd build
-sh ../script/cmake-ck-dev.sh  ../ <arch>  # you can replace this <arch> to gfx90a, gfx942...
+# you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
+sh ../script/cmake-ck-dev.sh  ../ <arch>
 make tile_example_layernorm2d_fwd -j
 ```
 This will result in an executable `build/bin/tile_example_layernorm2d_fwd`

example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp

@@ -35,7 +35,9 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t,
                                                                     YDataType,
                                                                     MeanDataType,
                                                                     InvStdDataType,
-                                                                    Shape>;
+                                                                    Shape,
+                                                                    true,
+                                                                    true>;
 
         using Kernel = ck_tile::Layernorm2dFwd<PipelineProblem>;
 

example/ck_tile/03_gemm/README.md

@@ -6,7 +6,8 @@ This folder contains example for GEMM using ck_tile tile-programming implementat
 ```
 # in the root of ck_tile
 mkdir build && cd build
-sh ../script/cmake-ck-dev.sh  ../ <arch>  # you can replace this <arch> to gfx90a, gfx942...
+# you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
+sh ../script/cmake-ck-dev.sh  ../ <arch>
 make tile_example_gemm_basic -j
 ```
 This will result in an executable `build/bin/tile_example_gemm_basic`
@@ -14,10 +15,17 @@ This will result in an executable `build/bin/tile_example_gemm_basic`
 ## example
 ```
 args:
-          -m    m dimension (default:3328)
-          -n    m dimension (default:4096)
+          -b    batch size (default:1)
+          -m    m dimension (default:1024)
+          -n    n dimension (default:2048)
           -k    k dimension (default:64)
-          -e    epsilon (default:1e-5)
-          -v    cpu validation or not (default:1)
-       -prec    precision (default:fp16)
+   -stride_a    Tensor A stride (default:0)
+   -stride_b    Tensor B stride (default:0)
+   -stride_c    Tensor C stride (default:0)
+          -v    0. No validation, 1. Validation on CPU, 2. Validation on GPU (default:2)
+          -e    Absolute error tolerance (default:1e-5)
+       -prec    data type. fp16/bf16/fp8/bf8 (default:fp16)
+     -warmup    number of iterations before benchmark the kernel (default:10)
+     -repeat    number of iterations to benchmark the kernel (default:100)
+      -timer    gpu:gpu timer, cpu:cpu timer (default:gpu)
 ```

example/ck_tile/03_gemm/gemm_basic.cpp

@@ -1,4 +1,3 @@
-
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
 
@@ -41,18 +40,39 @@ template <typename LayoutA,
 float gemm_calc(const gemm_basic_args& args, const ck_tile::stream_config& s)
 {
     // The kPadA, kPadB, kPadC & kBlockPerCu should also come from the Codegen part.
-    constexpr bool kPadA = true;
-    constexpr bool kPadB = true;
+    constexpr bool kPadA        = true;
+    constexpr bool kPadB        = true;
+    constexpr bool kTilePermute = false;
 
     constexpr int kBlockPerCu = 1;
 
     using TilePartitioner = ck_tile::GemmTilePartitioner<GemmShape>;
-    using GemmEpilogue    = ck_tile::Default2DEpilogue<
-        ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadA, kPadB>>;
+
+    // The rank and permutation will also be generate out by the CodeGen part.
+    constexpr ck_tile::index_t kOutputRank = 2;
+
+    // Whether doing the CShuffle (transpose before the global memory), depending on the output
+    // layout.
+    constexpr bool CShuffleEpilogue =
+        std::is_same_v<LayoutC, ck_tile::tensor_layout::gemm::ColumnMajor>;
+
+    using GemmEpilogue = std::conditional_t<
+        CShuffleEpilogue,
+        ck_tile::CShuffleEpilogue<ck_tile::CShuffleEpilogueProblem<AccDataType,
+                                                                   CDataType,
+                                                                   kPadA,
+                                                                   kPadB,
+                                                                   kTilePermute,
+                                                                   kOutputRank,
+                                                                   1,
+                                                                   0,
+                                                                   TilePartitioner::kM,
+                                                                   TilePartitioner::kN>>,
+        ck_tile::Default2DEpilogue<
+            ck_tile::Default2DEpilogueProblem<AccDataType, CDataType, kPadA, kPadB>>>;
     // ToDo: Will add the codegen part to test different pipeline policies in GEMM.
     // Now we only use the BlockGemmASmemBSmemCRegV1DefaultPolicy.
-    using Kernel =
-        ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue, LayoutA, LayoutB, LayoutC>;
+    using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
 
     auto kargs = Kernel::MakeKargs(args.p_a,
                                    args.p_b,
@@ -255,15 +275,17 @@ int main(int argc, char* argv[])
                                ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
                                ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
 
-    using CodegenPipelineProblem = ck_tile::BlockGemmPipelineProblem<ADataType,
-                                                                     BDataType,
-                                                                     AccDataType,
-                                                                     CodegenGemmShape,
-                                                                     kPadA,
-                                                                     kPadB,
-                                                                     kPadC>;
+    using CodegenGemmTraits = ck_tile::
+        TileGemmTraits<kPadA, kPadB, kPadC, matrix_a_layout, matrix_b_layout, matrix_c_layout>;
 
-    using CodegenGemmPipeline = ck_tile::BlockGemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
+    using CodegenPipelineProblem = ck_tile::
+        GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
+
+    using CodegenGemmPolicy = ck_tile::
+        UniversalGemmPipelineAgBgCrPolicy<matrix_a_layout, matrix_b_layout, matrix_c_layout>;
+
+    using CodegenGemmPipeline =
+        ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenGemmPolicy>;
 
     invoke_gemm<ck_tile::half_t,
                 matrix_a_layout,
@@ -341,7 +363,13 @@ int main(int argc, char* argv[])
         ck_tile::HostTensor<CDataType> c_host_gpu_ref(c_dimensions);
         ck_tile::DeviceMem c_gpu_buf(c_host_gpu_ref.get_element_space_size_in_bytes());
 
-        ck_tile::reference_gemm_gpu<ADataType, BDataType, AccDataType, CDataType>(
+        ck_tile::reference_gemm_gpu<ADataType,
+                                    BDataType,
+                                    AccDataType,
+                                    CDataType,
+                                    matrix_a_layout,
+                                    matrix_b_layout,
+                                    matrix_c_layout>(
             a_buf, b_buf, c_gpu_buf, M, N, K, stride_a, stride_b, stride_c);
 
         c_buf.FromDevice(c_host_gpu_ref.data());

example/ck_tile/04_img2col/CMakeLists.txt

@@ -0,0 +1,3 @@
+# not using add_example_executable() to add this target, since we don't want this to have
+# to be included in "make all/install/check"
+add_executable(tile_example_img2col EXCLUDE_FROM_ALL image_to_column.cpp)

example/ck_tile/04_img2col/README.md

@@ -0,0 +1,13 @@
+# Image to Column
+
+This folder contains example for Image to Column using ck_tile tile-programming implementation.
+
+## build
+```
+# in the root of ck_tile
+mkdir build && cd build
+# you can replace <arch> with the appropriate architecture (for example gfx90a or gfx942) or leave it blank
+sh ../script/cmake-ck-dev.sh  ../ <arch>
+make tile_example_img2col -j
+```
+This will result in an executable `build/bin/tile_example_img2col`

example/ck_tile/04_img2col/image_to_column.cpp

@@ -0,0 +1,170 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <algorithm>
+#include <cstring>
+
+#include "ck_tile/host.hpp"
+#include "image_to_column.hpp"
+
+// Host API implementation
+template <>
+float image_to_column(const image_to_column_traits& traits,
+                      const image_to_column_args<2>& args,
+                      const ck_tile::stream_config& stream_conf)
+{
+    if(traits.data_type.compare("fp16") == 0)
+    {
+        constexpr ck_tile::index_t NDimSpatial = 2;
+        constexpr ck_tile::index_t VectorSize  = 8;
+
+        using thread_tile = ck_tile::sequence<8, 8>;
+        using warp_tile   = ck_tile::sequence<64, 64>;
+        using block_tile  = ck_tile::sequence<128, 128>;
+
+        using Shape = ck_tile::TileImageToColumnShape<thread_tile, warp_tile, block_tile>;
+
+        using InDataType  = ck_tile::half_t;
+        using OutDataType = ck_tile::half_t;
+
+        using PipelineProblem = ck_tile::BlockImageToColumnProblem<InDataType,
+                                                                   OutDataType,
+                                                                   Shape,
+                                                                   NDimSpatial,
+                                                                   VectorSize,
+                                                                   VectorSize>;
+
+        using Kernel = ck_tile::ImageToColumn<PipelineProblem>;
+
+        auto kargs = Kernel::MakeKargs(args.p_in,
+                                       args.p_out,
+                                       args.G,
+                                       args.N,
+                                       args.C,
+                                       args.input_spatial_lengths,
+                                       args.filter_spatial_lengths,
+                                       args.output_spatial_lengths,
+                                       args.image_g_n_c_wis_strides,
+                                       args.gemm_g_m_k_strides,
+                                       args.conv_filter_strides,
+                                       args.conv_filter_dilations,
+                                       args.input_left_pads,
+                                       args.input_right_pads);
+
+        const dim3 grids = Kernel::GridSize(
+            args.N * args.output_spatial_lengths[0] * args.output_spatial_lengths[1],
+            args.filter_spatial_lengths[0] * args.filter_spatial_lengths[1] * args.C,
+            args.G);
+        constexpr dim3 blocks = Kernel::BlockSize();
+
+        constexpr ck_tile::index_t kBlockPerCu = 2;
+
+        float ave_time = ck_tile::launch_kernel(
+            stream_conf,
+            ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+
+        return ave_time;
+    }
+
+    return 0;
+}
+
+int main(int argc, char* argv[])
+{
+    constexpr ck_tile::index_t NDimSpatial = 2;
+
+    ExecutionConfig config;
+    ck_tile::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatial dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    using InDataType  = ck_tile::half_t;
+    using OutDataType = ck_tile::half_t;
+    using ImLayout    = ck_tile::tensor_layout::convolution::NHWGC;
+
+    const auto G = conv_params.G_;
+    const auto N = conv_params.N_;
+    const auto C = conv_params.C_;
+
+    const ck_tile::long_index_t NHoWo =
+        N * std::accumulate(conv_params.output_spatial_lengths_.begin(),
+                            std::next(conv_params.output_spatial_lengths_.begin(), NDimSpatial),
+                            1,
+                            std::multiplies<>());
+
+    const ck_tile::long_index_t CYX =
+        C * std::accumulate(conv_params.filter_spatial_lengths_.begin(),
+                            std::next(conv_params.filter_spatial_lengths_.begin(), NDimSpatial),
+                            1,
+                            std::multiplies<>());
+
+    const auto in_desc =
+        ck_tile::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(conv_params);
+    const auto out_desc = ck_tile::HostTensorDescriptor({G, NHoWo, CYX});
+
+    // host verify
+    ck_tile::HostTensor<InDataType> in(in_desc);
+    ck_tile::HostTensor<OutDataType> out_device(out_desc);
+    ck_tile::HostTensor<OutDataType> out_host(out_desc);
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1: ck_tile::FillUniformDistributionIntegerValue<InDataType>{-5.f, 5.f}(in); break;
+    default: ck_tile::FillUniformDistribution<InDataType>{-0.5, 0.5}(in); break;
+    }
+
+    ck_tile::DeviceMem in_device_buf(in.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem out_device_buf(out_device.get_element_space_size_in_bytes());
+
+    in_device_buf.ToDevice(in.data());
+
+    image_to_column_traits traits{"fp16"};
+
+    image_to_column_args<NDimSpatial> args{
+        in_device_buf.GetDeviceBuffer(),
+        out_device_buf.GetDeviceBuffer(),
+        G,
+        N,
+        C,
+        ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_spatial_lengths_),
+        ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.filter_spatial_lengths_),
+        ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.output_spatial_lengths_),
+        ck_tile::to_array<ck_tile::long_index_t, NDimSpatial + 3>(in_desc.get_strides()),
+        ck_tile::to_array<ck_tile::long_index_t, 3>(out_desc.get_strides()),
+        ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.conv_filter_strides_),
+        ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.conv_filter_dilations_),
+        ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_left_pads_),
+        ck_tile::to_array<ck_tile::long_index_t, NDimSpatial>(conv_params.input_right_pads_)};
+
+    float ave_time =
+        image_to_column(traits, args, ck_tile::stream_config{nullptr, config.time_kernel});
+
+    std::size_t num_btype = G * NHoWo * CYX * (sizeof(OutDataType) + sizeof(InDataType));
+    float gb_per_sec      = num_btype / 1.E6 / ave_time;
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    bool pass = true;
+
+    if(config.do_verification)
+    {
+        // reference
+        ck_tile::reference_im2col<InDataType, OutDataType, NDimSpatial>(in, out_host, conv_params);
+
+        out_device_buf.FromDevice(out_device.data());
+        pass = ck_tile::check_err(out_device, out_host);
+
+        std::cout << "valid:" << (pass ? "y" : "n") << std::endl;
+    }
+
+    return !pass;
+}

example/ck_tile/04_img2col/image_to_column.hpp

@@ -0,0 +1,105 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/kernel_launch.hpp"
+#include "ck_tile/ops/image_to_column.hpp"
+#include <string>
+
+#define DefaultConvParams                                                    \
+    ck_tile::conv::ConvParam                                                 \
+    {                                                                        \
+        2, 2, 32, 32, 32, {4, 4}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, { 0, 0 } \
+    }
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck_tile::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck_tile::conv::ConvParam& conv_params)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+        config = ExecutionConfig{};
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck_tile::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_params =
+            ck_tile::conv::parse_conv_param(num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}
+
+struct image_to_column_traits
+{
+    std::string data_type;
+};
+
+template <ck_tile::index_t NDimSpatial>
+struct image_to_column_args
+{
+    const void* p_in;
+    void* p_out;
+    const ck_tile::long_index_t G;
+    const ck_tile::long_index_t N;
+    const ck_tile::long_index_t C;
+    const ck_tile::array<ck_tile::long_index_t, NDimSpatial> input_spatial_lengths;
+    const ck_tile::array<ck_tile::long_index_t, NDimSpatial> filter_spatial_lengths;
+    const ck_tile::array<ck_tile::long_index_t, NDimSpatial> output_spatial_lengths;
+    const ck_tile::array<ck_tile::long_index_t, NDimSpatial + 3> image_g_n_c_wis_strides;
+    const ck_tile::array<ck_tile::long_index_t, 3> gemm_g_m_k_strides;
+    const ck_tile::array<ck_tile::long_index_t, NDimSpatial> conv_filter_strides;
+    const ck_tile::array<ck_tile::long_index_t, NDimSpatial> conv_filter_dilations;
+    const ck_tile::array<ck_tile::long_index_t, NDimSpatial> input_left_pads;
+    const ck_tile::array<ck_tile::long_index_t, NDimSpatial> input_right_pads;
+};
+
+// host API
+template <ck_tile::index_t NDimSpatial>
+float image_to_column(const image_to_column_traits&,
+                      const image_to_column_args<NDimSpatial>&,
+                      const ck_tile::stream_config&);

example/ck_tile/CMakeLists.txt

@@ -5,3 +5,4 @@ include_directories(AFTER
 add_subdirectory(01_fmha)
 add_subdirectory(02_layernorm2d)
 add_subdirectory(03_gemm)
+add_subdirectory(04_img2col)

include/ck/config.h.in

@@ -97,13 +97,6 @@
 #cmakedefine CK_ENABLE_DL_KERNELS @CK_ENABLE_DL_KERNELS@
 #endif
 
-//
-// Instances supports in the current CK build
-//
-#ifndef CK_ENABLE_INSTANCES_ONLY
-#cmakedefine CK_ENABLE_INSTANCES_ONLY @CK_ENABLE_INSTANCES_ONLY@
-#endif
-
 //
 // CK kernels which support XDL (MI series)
 //

include/ck/host_utility/kernel_launch.hpp

@@ -66,6 +66,9 @@ float launch_and_time_kernel(const StreamConfig& stream_config,
 
         hip_check_error(hipEventElapsedTime(&total_time, start, stop));
 
+        hip_check_error(hipEventDestroy(start));
+        hip_check_error(hipEventDestroy(stop));
+
         return total_time / nrepeat;
     }
     else
@@ -143,6 +146,9 @@ float launch_and_time_kernel_with_preprocess(const StreamConfig& stream_config,
 
         hip_check_error(hipEventElapsedTime(&total_time, start, stop));
 
+        hip_check_error(hipEventDestroy(start));
+        hip_check_error(hipEventDestroy(stop));
+
         return total_time / nrepeat;
     }
     else

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops.hpp

@@ -406,7 +406,7 @@ struct BlockwiseGemmXdlops_pipeline_v4
     }
 
     template <>
-    __device__ static constexpr auto TailScheduler<1>()
+    __device__ constexpr auto TailScheduler<1>()
     {
         // schedule
         constexpr auto num_ds_read_inst =
@@ -433,7 +433,7 @@ struct BlockwiseGemmXdlops_pipeline_v4
     }
 
     template <>
-    __device__ static constexpr auto TailScheduler<2>()
+    __device__ constexpr auto TailScheduler<2>()
     {
         // schedule
         constexpr auto num_ds_read_inst =

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v1_ab_scale.hpp

@@ -308,7 +308,7 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
                                 typename vector_type<ComputeDataType,
                                                      xdlops_gemm.K1PerXdlops>::type;
 
-                            xdlops_gemm.template Run(
+                            xdlops_gemm.template Run<>(
                                 a_thread_vec.template AsType<mfma_input_type>(),
                                 b_thread_vec.template AsType<mfma_input_type>(),
                                 c_thread_buf_per_scale.GetVectorTypeReference(I0));
@@ -390,9 +390,10 @@ struct BlockwiseGemmXdlops_pipeline_v1_ab_scale<BlockGemmPipelineScheduler::Intr
                         using mfma_input_type =
                             typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
 
-                        xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
-                                                 b_thread_vec.template AsType<mfma_input_type>(),
-                                                 c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf_per_scale.GetVectorTypeReference(I0));
                     });
                     static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
                         constexpr index_t c_offset =

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v2_ab_scale.hpp

@@ -350,7 +350,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
                                     typename vector_type<ComputeDataType,
                                                          xdlops_gemm.K1PerXdlops>::type;
 
-                                xdlops_gemm.template Run(
+                                xdlops_gemm.template Run<>(
                                     a_thread_vec.template AsType<mfma_input_type>(),
                                     b_thread_vec.template AsType<mfma_input_type>(),
                                     c_thread_buf_per_scale.GetVectorTypeReference(I0));
@@ -443,7 +443,7 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
                                 typename vector_type<ComputeDataType,
                                                      xdlops_gemm.K1PerXdlops>::type;
 
-                            xdlops_gemm.template Run(
+                            xdlops_gemm.template Run<>(
                                 a_thread_vec.template AsType<mfma_input_type>(),
                                 b_thread_vec.template AsType<mfma_input_type>(),
                                 c_thread_buf_per_scale.GetVectorTypeReference(I0));
@@ -518,9 +518,10 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
                         using mfma_input_type =
                             typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
 
-                        xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
-                                                 b_thread_vec.template AsType<mfma_input_type>(),
-                                                 c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf_per_scale.GetVectorTypeReference(I0));
                     });
                     static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
                         constexpr index_t c_offset =
@@ -575,9 +576,10 @@ struct BlockwiseGemmXdlops_pipeline_v2_ab_scale<BlockGemmPipelineScheduler::Intr
                         using mfma_input_type =
                             typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
 
-                        xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
-                                                 b_thread_vec.template AsType<mfma_input_type>(),
-                                                 c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf_per_scale.GetVectorTypeReference(I0));
                     });
                     static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
                         constexpr index_t c_offset =

include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_v3_ab_scale.hpp

@@ -427,7 +427,7 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
                                 typename vector_type<ComputeDataType,
                                                      xdlops_gemm.K1PerXdlops>::type;
 
-                            xdlops_gemm.template Run(
+                            xdlops_gemm.template Run<>(
                                 a_thread_vec.template AsType<mfma_input_type>(),
                                 b_thread_vec.template AsType<mfma_input_type>(),
                                 c_thread_buf_per_scale.GetVectorTypeReference(I0));
@@ -504,9 +504,10 @@ struct BlockwiseGemmXdlops_pipeline_v3_ab_scale<BlockGemmPipelineScheduler::Intr
                         using mfma_input_type =
                             typename vector_type<ComputeDataType, xdlops_gemm.K1PerXdlops>::type;
 
-                        xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
-                                                 b_thread_vec.template AsType<mfma_input_type>(),
-                                                 c_thread_buf_per_scale.GetVectorTypeReference(I0));
+                        xdlops_gemm.template Run<>(
+                            a_thread_vec.template AsType<mfma_input_type>(),
+                            b_thread_vec.template AsType<mfma_input_type>(),
+                            c_thread_buf_per_scale.GetVectorTypeReference(I0));
                     });
                     static_for<0, xdlops_gemm.GetRegSizePerXdlops(), 1>{}([&](auto t) {
                         constexpr index_t c_offset =

include/ck/tensor_operation/gpu/device/device_cgemm.hpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 #include "device_base.hpp"
@@ -31,13 +31,13 @@ struct DeviceCGemm : public BaseOperator
                                                               CElementwiseOperation c_element_op,
                                                               ck::index_t KBatch = 1) = 0;
 
-    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer()   = 0;
     virtual std::size_t GetWorkspaceSize(index_t MRaw,
                                          index_t NRaw,
                                          index_t KRaw,
                                          index_t StrideA,
                                          index_t StrideB,
-                                         index_t StrideC)     = 0;
+                                         index_t StrideC) const = 0;
 };
 
 template <typename AElementwiseOperation,

include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp

@@ -598,10 +598,26 @@ struct DeviceCGemm_4Gemm_Xdl_CShuffle
                                  [[maybe_unused]] index_t K,
                                  [[maybe_unused]] index_t StrideA,
                                  [[maybe_unused]] index_t StrideB,
-                                 index_t StrideC) override
+                                 index_t StrideC) const override
     {
         return 2 * sizeof(CDataType) * GetCElementSpaceSize(M, N, StrideC);
     }
+
+    std::size_t GetWorkSpaceSize(const BaseArgument* base_arg) const override
+    {
+        const auto* parg = dynamic_cast<const Argument*>(base_arg);
+
+        if(!parg)
+        {
+            std::ostringstream err;
+            err << "Provided argument pointer is not of an Argument class!"
+                << " In " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+            throw std::runtime_error(err.str());
+        }
+
+        return GetWorkspaceSize(
+            parg->M, parg->N, parg->K, parg->StrideA, parg->StrideB, parg->StrideC);
+    }
 };
 
 } // namespace device

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multi_abd_xdl_fixed_nk.hpp

@@ -64,7 +64,7 @@ __global__ void
     const index_t N = gemm_desc_ptr[group_id].N;
     const index_t K = gemm_desc_ptr[group_id].K;
 
-    if(M * N * K == 0)
+    if(M == 0 || N == 0 || K == 0)
         return;
 
     const auto StrideAs = gemm_desc_ptr[group_id].StrideAs;

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_splitk_xdl_cshuffle_two_stage.hpp

@@ -345,7 +345,7 @@ struct DeviceGroupedGemmMultipleDSplitKXdlCShuffleTwoStage
                 const index_t N = gemm_descs[i].N_;
                 const index_t K = gemm_descs[i].K_;
 
-                if(M * N * K == 0)
+                if(M == 0 || N == 0 || K == 0)
                 {
                     skipped_group_count_++;
                     continue;

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_xdl_cshuffle_tile_loop.hpp

@@ -109,7 +109,7 @@ __global__ void
             N = gemm_desc_ptr[group_id].N;
             K = gemm_desc_ptr[group_id].K;
 
-            if(M * N * K == 0)
+            if(M == 0 || N == 0 || K == 0)
             {
                 grid_size_grp = 0;
                 continue;

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp

@@ -68,7 +68,7 @@ __global__ void
     const index_t N = gemm_desc_ptr[group_id].N;
     const index_t K = gemm_desc_ptr[group_id].K;
 
-    if(M * N * K == 0)
+    if(M == 0 || N == 0 || K == 0)
         return;
 
     const auto StrideA  = gemm_desc_ptr[group_id].StrideA;

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -419,6 +419,12 @@ struct UnaryAbs
 
         y = ck::math::abs(x);
     };
+
+    template <>
+    __host__ __device__ void operator()(f8_t& y, const f8_t& x) const
+    {
+        y = ck::type_convert<f8_t>(ck::math::abs(ck::type_convert<float>(x)));
+    };
 };
 
 struct UnarySqrt

include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp

@@ -324,55 +324,55 @@ struct DppSelector
     static constexpr auto GetDpp();
 
     template <>
-    static constexpr auto GetDpp<half_t, 8, 32>()
+    constexpr auto GetDpp<half_t, 8, 32>()
     {
         return DppInstr::dpp8_f16_8x32x2;
     }
 
     template <>
-    static constexpr auto GetDpp<half_t, 8, 16>()
+    constexpr auto GetDpp<half_t, 8, 16>()
     {
         return DppInstr::dpp8_f16_8x16x2;
     }
 
     template <>
-    static constexpr auto GetDpp<half_t, 16, 16>()
+    constexpr auto GetDpp<half_t, 16, 16>()
     {
         return DppInstr::dpp8_f16_16x16x2;
     }
 
     template <>
-    static constexpr auto GetDpp<half_t, 32, 8>()
+    constexpr auto GetDpp<half_t, 32, 8>()
     {
         return DppInstr::dpp8_f16_32x8x2;
     }
 
     template <>
-    static constexpr auto GetDpp<half_t, 1, 32>()
+    constexpr auto GetDpp<half_t, 1, 32>()
     {
         return DppInstr::dpp8_f16_1x32x2;
     }
 
     template <>
-    static constexpr auto GetDpp<half_t, 2, 32>()
+    constexpr auto GetDpp<half_t, 2, 32>()
     {
         return DppInstr::dpp8_f16_2x32x2;
     }
 
     template <>
-    static constexpr auto GetDpp<half_t, 2, 16>()
+    constexpr auto GetDpp<half_t, 2, 16>()
     {
         return DppInstr::dpp8_f16_2x16x2;
     }
 
     template <>
-    static constexpr auto GetDpp<half_t, 4, 16>()
+    constexpr auto GetDpp<half_t, 4, 16>()
     {
         return DppInstr::dpp8_f16_4x16x2;
     }
 
     template <>
-    static constexpr auto GetDpp<half_t, 4, 32>()
+    constexpr auto GetDpp<half_t, 4, 32>()
     {
         return DppInstr::dpp8_f16_4x32x2;
     }

include/ck/tensor_operation/gpu/warp/wmma_gemm.hpp

@@ -415,7 +415,7 @@ struct WmmaSelector
     static constexpr auto GetWmma();
 
     template <>
-    static constexpr auto GetWmma<half_t, half_t, float, 16, 16>()
+    constexpr auto GetWmma<half_t, half_t, float, 16, 16>()
     {
 #ifdef __gfx12__
         return WmmaInstr::wmma_f32_16x16x16_f16_gfx12;
@@ -425,7 +425,7 @@ struct WmmaSelector
     }
 
     template <>
-    static constexpr auto GetWmma<bhalf_t, bhalf_t, float, 16, 16>()
+    constexpr auto GetWmma<bhalf_t, bhalf_t, float, 16, 16>()
     {
 #ifdef __gfx12__
         return WmmaInstr::wmma_f32_16x16x16_bf16_gfx12;
@@ -435,19 +435,19 @@ struct WmmaSelector
     }
 
     template <>
-    static constexpr auto GetWmma<half_t, half_t, half_t, 16, 16>()
+    constexpr auto GetWmma<half_t, half_t, half_t, 16, 16>()
     {
         return WmmaInstr::wmma_f16_16x16x16_f16;
     }
 
     template <>
-    static constexpr auto GetWmma<bhalf_t, bhalf_t, bhalf_t, 16, 16>()
+    constexpr auto GetWmma<bhalf_t, bhalf_t, bhalf_t, 16, 16>()
     {
         return WmmaInstr::wmma_bf16_16x16x16_bf16;
     }
 
     template <>
-    static constexpr auto GetWmma<int8_t, int8_t, int, 16, 16>()
+    constexpr auto GetWmma<int8_t, int8_t, int, 16, 16>()
     {
 #ifdef __gfx12__
         return WmmaInstr::wmma_i32_16x16x16_iu8_gfx12;
@@ -458,7 +458,7 @@ struct WmmaSelector
 
 #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
     template <>
-    static constexpr auto GetWmma<int4_t, int4_t, int, 16, 16>()
+    constexpr auto GetWmma<int4_t, int4_t, int, 16, 16>()
     {
         return WmmaInstr::wmma_i32_16x16x16_iu4;
     }

include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp

@@ -651,97 +651,97 @@ struct MfmaSelector
     static constexpr auto GetMfma();
 
     template <>
-    static constexpr auto GetMfma<double, 16, 16>()
+    constexpr auto GetMfma<double, 16, 16>()
     {
         return MfmaInstr::mfma_f64_16x16x4f64;
     }
 
     template <>
-    static constexpr auto GetMfma<float, 64, 64>()
+    constexpr auto GetMfma<float, 64, 64>()
     {
         return MfmaInstr::mfma_f32_32x32x1xf32;
     }
 
     template <>
-    static constexpr auto GetMfma<float, 32, 64>()
+    constexpr auto GetMfma<float, 32, 64>()
     {
         return MfmaInstr::mfma_f32_32x32x1xf32;
     }
 
     template <>
-    static constexpr auto GetMfma<float, 16, 64>()
+    constexpr auto GetMfma<float, 16, 64>()
     {
         return MfmaInstr::mfma_f32_16x16x1xf32;
     }
 
     template <>
-    static constexpr auto GetMfma<float, 8, 64>()
+    constexpr auto GetMfma<float, 8, 64>()
     {
         return MfmaInstr::mfma_f32_4x4x1xf32;
     }
 
     template <>
-    static constexpr auto GetMfma<float, 4, 64>()
+    constexpr auto GetMfma<float, 4, 64>()
     {
         return MfmaInstr::mfma_f32_4x4x1xf32;
     }
 
     template <>
-    static constexpr auto GetMfma<float, 32, 32>()
+    constexpr auto GetMfma<float, 32, 32>()
     {
         return MfmaInstr::mfma_f32_32x32x2xf32;
     }
 
     template <>
-    static constexpr auto GetMfma<float, 16, 16>()
+    constexpr auto GetMfma<float, 16, 16>()
     {
         return MfmaInstr::mfma_f32_16x16x4xf32;
     }
 
     template <>
-    static constexpr auto GetMfma<half_t, 64, 64>()
+    constexpr auto GetMfma<half_t, 64, 64>()
     {
         return MfmaInstr::mfma_f32_32x32x4f16;
     }
 
     template <>
-    static constexpr auto GetMfma<half_t, 32, 64>()
+    constexpr auto GetMfma<half_t, 32, 64>()
     {
         return MfmaInstr::mfma_f32_32x32x4f16;
     }
 
     template <>
-    static constexpr auto GetMfma<half_t, 32, 32>()
+    constexpr auto GetMfma<half_t, 32, 32>()
     {
         return MfmaInstr::mfma_f32_32x32x8f16;
     }
 
     template <>
-    static constexpr auto GetMfma<half_t, 16, 16>()
+    constexpr auto GetMfma<half_t, 16, 16>()
     {
         return MfmaInstr::mfma_f32_16x16x16f16;
     }
 
     template <>
-    static constexpr auto GetMfma<half_t, 16, 64>()
+    constexpr auto GetMfma<half_t, 16, 64>()
     {
         return MfmaInstr::mfma_f32_16x16x4f16;
     }
 
     template <>
-    static constexpr auto GetMfma<half_t, 8, 64>()
+    constexpr auto GetMfma<half_t, 8, 64>()
     {
         return MfmaInstr::mfma_f32_4x4x4f16;
     }
 
     template <>
-    static constexpr auto GetMfma<half_t, 4, 64>()
+    constexpr auto GetMfma<half_t, 4, 64>()
     {
         return MfmaInstr::mfma_f32_4x4x4f16;
     }
 
     template <>
-    static constexpr auto GetMfma<bhalf_t, 32, 32>()
+    constexpr auto GetMfma<bhalf_t, 32, 32>()
     {
 #if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
         return MfmaInstr::mfma_f32_32x32x8bf16_1k;
@@ -751,7 +751,7 @@ struct MfmaSelector
     }
 
     template <>
-    static constexpr auto GetMfma<bhalf_t, 16, 16>()
+    constexpr auto GetMfma<bhalf_t, 16, 16>()
     {
 #if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
         return MfmaInstr::mfma_f32_16x16x16bf16_1k;
@@ -762,72 +762,72 @@ struct MfmaSelector
 
 #if defined(CK_USE_AMD_MFMA_GFX940)
     template <>
-    static constexpr auto GetMfma<int8_t, 32, 32>()
+    constexpr auto GetMfma<int8_t, 32, 32>()
     {
         return MfmaInstr::mfma_i32_32x32x16i8;
     }
     template <>
-    static constexpr auto GetMfma<int8_t, 16, 16>()
+    constexpr auto GetMfma<int8_t, 16, 16>()
     {
         return MfmaInstr::mfma_i32_16x16x32i8;
     }
 #else
     template <>
-    static constexpr auto GetMfma<int8_t, 32, 32>()
+    constexpr auto GetMfma<int8_t, 32, 32>()
     {
         return MfmaInstr::mfma_i32_32x32x8i8;
     }
     template <>
-    static constexpr auto GetMfma<int8_t, 16, 16>()
+    constexpr auto GetMfma<int8_t, 16, 16>()
     {
         return MfmaInstr::mfma_i32_16x16x16i8;
     }
 #endif
 
     template <>
-    static constexpr auto GetMfma<f8_t, 32, 32>()
+    constexpr auto GetMfma<f8_t, 32, 32>()
     {
         return MfmaInstr::mfma_f32_32x32x16f8f8;
     }
 
     template <>
-    static constexpr auto GetMfma<f8_t, 16, 16>()
+    constexpr auto GetMfma<f8_t, 16, 16>()
     {
         return MfmaInstr::mfma_f32_16x16x32f8f8;
     }
 
     template <>
-    static constexpr auto GetMfma<bf8_t, 32, 32>()
+    constexpr auto GetMfma<bf8_t, 32, 32>()
     {
         return MfmaInstr::mfma_f32_32x32x16bf8bf8;
     }
 
     template <>
-    static constexpr auto GetMfma<bf8_t, 16, 16>()
+    constexpr auto GetMfma<bf8_t, 16, 16>()
     {
         return MfmaInstr::mfma_f32_16x16x32bf8bf8;
     }
 
     template <>
-    static constexpr auto GetMfma<f8_t, 32, 32, bf8_t>()
+    constexpr auto GetMfma<f8_t, 32, 32, bf8_t>()
     {
         return MfmaInstr::mfma_f32_32x32x16f8bf8;
     }
 
     template <>
-    static constexpr auto GetMfma<f8_t, 16, 16, bf8_t>()
+    constexpr auto GetMfma<f8_t, 16, 16, bf8_t>()
     {
         return MfmaInstr::mfma_f32_16x16x32f8bf8;
     }
 
     template <>
-    static constexpr auto GetMfma<bf8_t, 32, 32, f8_t>()
+    constexpr auto GetMfma<bf8_t, 32, 32, f8_t>()
     {
         return MfmaInstr::mfma_f32_32x32x16bf8f8;
     }
 
     template <>
-    static constexpr auto GetMfma<bf8_t, 16, 16, f8_t>()
+    constexpr auto GetMfma<bf8_t, 16, 16, f8_t>()
     {
         return MfmaInstr::mfma_f32_16x16x32bf8f8;
     }

include/ck/utility/data_type.hpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -13,8 +13,24 @@ using int4_t  = _BitInt(4);
 using f8_t    = _BitInt(8);
 using bf8_t   = unsigned _BitInt(8);
 
+inline constexpr auto next_pow2(uint32_t x)
+{
+    // Precondition: x > 1.
+    return x > 1u ? (1u << (32u - __builtin_clz(x - 1u))) : x;
+}
+
+// native types: double, float, _Float16, ushort, int32_t, int8_t, uint8_t, f8_t, bf8_t, bool
+template <typename T>
+inline constexpr bool is_native_type()
+{
+    return is_same<T, double>::value || is_same<T, float>::value || is_same<T, half_t>::value ||
+           is_same<T, bhalf_t>::value || is_same<T, int32_t>::value || is_same<T, int8_t>::value ||
+           is_same<T, uint8_t>::value || is_same<T, f8_t>::value || is_same<T, bf8_t>::value ||
+           is_same<T, bool>::value;
+}
+
 // vector_type
-template <typename T, index_t N>
+template <typename T, index_t N, typename Enable = void>
 struct vector_type;
 
 // Caution: DO NOT REMOVE
@@ -171,7 +187,7 @@ struct scalar_type<bool>
 };
 
 template <typename T>
-struct vector_type<T, 1>
+struct vector_type<T, 1, typename std::enable_if_t<is_native_type<T>()>>
 {
     using d1_t = T;
     using type = d1_t;
@@ -189,7 +205,8 @@ struct vector_type<T, 1>
     template <typename X>
     __host__ __device__ constexpr const auto& AsType() const
     {
-        static_assert(is_same<X, d1_t>::value, "wrong!");
+        static_assert(is_same<X, d1_t>::value,
+                      "Something went wrong, please check src and dst types.");
 
         return data_.d1x1_;
     }
@@ -197,7 +214,8 @@ struct vector_type<T, 1>
     template <typename X>
     __host__ __device__ constexpr auto& AsType()
     {
-        static_assert(is_same<X, d1_t>::value, "wrong!");
+        static_assert(is_same<X, d1_t>::value,
+                      "Something went wrong, please check src and dst types.");
 
         return data_.d1x1_;
     }
@@ -205,7 +223,7 @@ struct vector_type<T, 1>
 
 __device__ int static err = 0;
 template <typename T>
-struct vector_type<T, 2>
+struct vector_type<T, 2, typename std::enable_if_t<is_native_type<T>()>>
 {
     using d1_t = T;
     typedef T d2_t __attribute__((ext_vector_type(2)));
@@ -226,7 +244,8 @@ struct vector_type<T, 2>
     template <typename X>
     __host__ __device__ constexpr const auto& AsType() const
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value, "wrong!");
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value,
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -245,7 +264,8 @@ struct vector_type<T, 2>
     template <typename X>
     __host__ __device__ constexpr auto& AsType()
     {
-        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value, "wrong!");
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value,
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -263,7 +283,7 @@ struct vector_type<T, 2>
 };
 
 template <typename T>
-struct vector_type<T, 4>
+struct vector_type<T, 4, typename std::enable_if_t<is_native_type<T>()>>
 {
     using d1_t = T;
     typedef T d2_t __attribute__((ext_vector_type(2)));
@@ -287,7 +307,7 @@ struct vector_type<T, 4>
     __host__ __device__ constexpr const auto& AsType() const
     {
         static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -311,7 +331,7 @@ struct vector_type<T, 4>
     __host__ __device__ constexpr auto& AsType()
     {
         static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -333,7 +353,7 @@ struct vector_type<T, 4>
 };
 
 template <typename T>
-struct vector_type<T, 8>
+struct vector_type<T, 8, typename std::enable_if_t<is_native_type<T>()>>
 {
     using d1_t = T;
     typedef T d2_t __attribute__((ext_vector_type(2)));
@@ -360,7 +380,7 @@ struct vector_type<T, 8>
     {
         static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -389,7 +409,7 @@ struct vector_type<T, 8>
     {
         static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -415,7 +435,7 @@ struct vector_type<T, 8>
 };
 
 template <typename T>
-struct vector_type<T, 16>
+struct vector_type<T, 16, typename std::enable_if_t<is_native_type<T>()>>
 {
     using d1_t = T;
     typedef T d2_t __attribute__((ext_vector_type(2)));
@@ -445,7 +465,7 @@ struct vector_type<T, 16>
         static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
                           is_same<X, d16_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -479,7 +499,7 @@ struct vector_type<T, 16>
         static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
                           is_same<X, d16_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -509,7 +529,7 @@ struct vector_type<T, 16>
 };
 
 template <typename T>
-struct vector_type<T, 32>
+struct vector_type<T, 32, typename std::enable_if_t<is_native_type<T>()>>
 {
     using d1_t = T;
     typedef T d2_t __attribute__((ext_vector_type(2)));
@@ -541,7 +561,7 @@ struct vector_type<T, 32>
         static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
                           is_same<X, d16_t>::value || is_same<X, d32_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -579,7 +599,7 @@ struct vector_type<T, 32>
         static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
                           is_same<X, d16_t>::value || is_same<X, d32_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -613,7 +633,7 @@ struct vector_type<T, 32>
 };
 
 template <typename T>
-struct vector_type<T, 64>
+struct vector_type<T, 64, typename std::enable_if_t<is_native_type<T>()>>
 {
     using d1_t = T;
     typedef T d2_t __attribute__((ext_vector_type(2)));
@@ -648,7 +668,7 @@ struct vector_type<T, 64>
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
                           is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
                           is_same<X, d64_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -691,7 +711,7 @@ struct vector_type<T, 64>
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
                           is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
                           is_same<X, d64_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -729,7 +749,7 @@ struct vector_type<T, 64>
 };
 
 template <typename T>
-struct vector_type<T, 128>
+struct vector_type<T, 128, typename std::enable_if_t<is_native_type<T>()>>
 {
     using d1_t = T;
     typedef T d2_t __attribute__((ext_vector_type(2)));
@@ -766,7 +786,7 @@ struct vector_type<T, 128>
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
                           is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
                           is_same<X, d64_t>::value || is_same<X, d128_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -813,7 +833,7 @@ struct vector_type<T, 128>
                           is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
                           is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
                           is_same<X, d64_t>::value || is_same<X, d128_t>::value,
-                      "wrong!");
+                      "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -855,7 +875,7 @@ struct vector_type<T, 128>
 };
 
 template <typename T>
-struct vector_type<T, 256>
+struct vector_type<T, 256, typename std::enable_if_t<is_native_type<T>()>>
 {
     using d1_t = T;
     typedef T d2_t __attribute__((ext_vector_type(2)));
@@ -894,7 +914,7 @@ struct vector_type<T, 256>
             is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
                 is_same<X, d8_t>::value || is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
                 is_same<X, d64_t>::value || is_same<X, d128_t>::value || is_same<X, d256_t>::value,
-            "wrong!");
+            "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -945,7 +965,7 @@ struct vector_type<T, 256>
             is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value ||
                 is_same<X, d8_t>::value || is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
                 is_same<X, d64_t>::value || is_same<X, d128_t>::value || is_same<X, d256_t>::value,
-            "wrong!");
+            "Something went wrong, please check src and dst types.");
 
         if constexpr(is_same<X, d1_t>::value)
         {
@@ -990,6 +1010,581 @@ struct vector_type<T, 256>
     }
 };
 
+template <typename T, index_t N>
+struct non_native_vector_base
+{
+    using type = non_native_vector_base<T, N>;
+
+    __host__ __device__ non_native_vector_base()            = default;
+    __host__ __device__ non_native_vector_base(const type&) = default;
+    __host__ __device__ non_native_vector_base(type&&)      = default;
+    __host__ __device__ ~non_native_vector_base()           = default;
+
+    T d[N];
+};
+
+// non-native vector_type implementation
+template <typename T>
+struct vector_type<T, 1, typename std::enable_if_t<!is_native_type<T>()>>
+{
+    using d1_t = T;
+    using type = d1_t;
+
+    union alignas(next_pow2(1 * sizeof(T)))
+    {
+        d1_t d1_;
+        StaticallyIndexedArray<d1_t, 1> d1x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        return data_.d1x1_;
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        return data_.d1x1_;
+    }
+};
+
+template <typename T>
+struct vector_type<T, 2, typename std::enable_if_t<!is_native_type<T>()>>
+{
+    using d1_t = T;
+    using d2_t = non_native_vector_base<T, 2>;
+
+    using type = d2_t;
+
+    union alignas(next_pow2(2 * sizeof(T)))
+    {
+        d2_t d2_;
+        StaticallyIndexedArray<d1_t, 2> d1x2_;
+        StaticallyIndexedArray<d2_t, 1> d2x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x2_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x2_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 4, typename std::enable_if_t<!is_native_type<T>()>>
+{
+    using d1_t = T;
+    using d2_t = non_native_vector_base<T, 2>;
+    using d4_t = non_native_vector_base<T, 4>;
+
+    using type = d4_t;
+
+    union alignas(next_pow2(4 * sizeof(T)))
+    {
+        d4_t d4_;
+        StaticallyIndexedArray<d1_t, 4> d1x4_;
+        StaticallyIndexedArray<d2_t, 2> d2x2_;
+        StaticallyIndexedArray<d4_t, 1> d4x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x4_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x2_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value || is_same<X, d4_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x4_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x2_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 8, typename std::enable_if_t<!is_native_type<T>()>>
+{
+    using d1_t = T;
+    using d2_t = non_native_vector_base<T, 2>;
+    using d4_t = non_native_vector_base<T, 4>;
+    using d8_t = non_native_vector_base<T, 8>;
+
+    using type = d8_t;
+
+    union alignas(next_pow2(8 * sizeof(T)))
+    {
+        d8_t d8_;
+        StaticallyIndexedArray<d1_t, 8> d1x8_;
+        StaticallyIndexedArray<d2_t, 4> d2x4_;
+        StaticallyIndexedArray<d4_t, 2> d4x2_;
+        StaticallyIndexedArray<d8_t, 1> d8x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x8_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x4_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x2_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x8_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x4_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x2_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 16, typename std::enable_if_t<!is_native_type<T>()>>
+{
+    using d1_t  = T;
+    using d2_t  = non_native_vector_base<T, 2>;
+    using d4_t  = non_native_vector_base<T, 4>;
+    using d8_t  = non_native_vector_base<T, 8>;
+    using d16_t = non_native_vector_base<T, 16>;
+
+    using type = d16_t;
+
+    union alignas(next_pow2(16 * sizeof(T)))
+    {
+        d16_t d16_;
+        StaticallyIndexedArray<d1_t, 16> d1x16_;
+        StaticallyIndexedArray<d2_t, 8> d2x8_;
+        StaticallyIndexedArray<d4_t, 4> d4x4_;
+        StaticallyIndexedArray<d8_t, 2> d8x2_;
+        StaticallyIndexedArray<d16_t, 1> d16x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x16_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x8_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x4_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x2_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x16_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x8_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x4_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x2_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 32, typename std::enable_if_t<!is_native_type<T>()>>
+{
+    using d1_t  = T;
+    using d2_t  = non_native_vector_base<T, 2>;
+    using d4_t  = non_native_vector_base<T, 4>;
+    using d8_t  = non_native_vector_base<T, 8>;
+    using d16_t = non_native_vector_base<T, 16>;
+    using d32_t = non_native_vector_base<T, 32>;
+
+    using type = d32_t;
+
+    union alignas(next_pow2(32 * sizeof(T)))
+    {
+        d32_t d32_;
+        StaticallyIndexedArray<d1_t, 32> d1x32_;
+        StaticallyIndexedArray<d2_t, 16> d2x16_;
+        StaticallyIndexedArray<d4_t, 8> d4x8_;
+        StaticallyIndexedArray<d8_t, 4> d8x4_;
+        StaticallyIndexedArray<d16_t, 2> d16x2_;
+        StaticallyIndexedArray<d32_t, 1> d32x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x32_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x16_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x8_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x4_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x2_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x32_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x16_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x8_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x4_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x2_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
+template <typename T>
+struct vector_type<T, 64, typename std::enable_if_t<!is_native_type<T>()>>
+{
+    using d1_t  = T;
+    using d2_t  = non_native_vector_base<T, 2>;
+    using d4_t  = non_native_vector_base<T, 4>;
+    using d8_t  = non_native_vector_base<T, 8>;
+    using d16_t = non_native_vector_base<T, 16>;
+    using d32_t = non_native_vector_base<T, 32>;
+    using d64_t = non_native_vector_base<T, 64>;
+
+    using type = d64_t;
+
+    union alignas(next_pow2(64 * sizeof(T)))
+    {
+        d64_t d64_;
+        StaticallyIndexedArray<d1_t, 64> d1x64_;
+        StaticallyIndexedArray<d2_t, 32> d2x32_;
+        StaticallyIndexedArray<d4_t, 16> d4x16_;
+        StaticallyIndexedArray<d8_t, 8> d8x8_;
+        StaticallyIndexedArray<d16_t, 4> d16x4_;
+        StaticallyIndexedArray<d32_t, 2> d32x2_;
+        StaticallyIndexedArray<d64_t, 1> d64x1_;
+    } data_;
+
+    __host__ __device__ constexpr vector_type() : data_{type{}} {}
+
+    __host__ __device__ constexpr vector_type(type v) : data_{v} {}
+
+    template <typename X>
+    __host__ __device__ constexpr const auto& AsType() const
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x64_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x32_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x16_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x8_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x4_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x2_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+
+    template <typename X>
+    __host__ __device__ constexpr auto& AsType()
+    {
+        static_assert(is_same<X, d1_t>::value || is_same<X, d2_t>::value ||
+                          is_same<X, d4_t>::value || is_same<X, d8_t>::value ||
+                          is_same<X, d16_t>::value || is_same<X, d32_t>::value ||
+                          is_same<X, d64_t>::value,
+                      "Something went wrong, please check src and dst types.");
+
+        if constexpr(is_same<X, d1_t>::value)
+        {
+            return data_.d1x64_;
+        }
+        else if constexpr(is_same<X, d2_t>::value)
+        {
+            return data_.d2x32_;
+        }
+        else if constexpr(is_same<X, d4_t>::value)
+        {
+            return data_.d4x16_;
+        }
+        else if constexpr(is_same<X, d8_t>::value)
+        {
+            return data_.d8x8_;
+        }
+        else if constexpr(is_same<X, d16_t>::value)
+        {
+            return data_.d16x4_;
+        }
+        else if constexpr(is_same<X, d32_t>::value)
+        {
+            return data_.d32x2_;
+        }
+        else if constexpr(is_same<X, d64_t>::value)
+        {
+            return data_.d64x1_;
+        }
+        else
+        {
+            return err;
+        }
+    }
+};
+
 using int64_t = long;
 
 // fp64
@@ -1051,8 +1646,8 @@ using bf8x8_t  = typename vector_type<bf8_t, 8>::type;
 using bf8x16_t = typename vector_type<bf8_t, 16>::type;
 using bf8x32_t = typename vector_type<bf8_t, 32>::type;
 using bf8x64_t = typename vector_type<bf8_t, 64>::type;
+
 // u8
-// i8
 using uint8x2_t  = typename vector_type<uint8_t, 2>::type;
 using uint8x4_t  = typename vector_type<uint8_t, 4>::type;
 using uint8x8_t  = typename vector_type<uint8_t, 8>::type;

include/ck/utility/math_v2.hpp

@@ -80,6 +80,8 @@ static inline __host__ bool isnan(half_t x)
     return (xx & 0x7FFF) > 0x7C00;
 };
 
+static inline __host__ bool isnan(f8_t x) { return (x & 0x80); };
+
 #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
 static inline __host__ bool isnan(int4_t x)
 {
@@ -529,6 +531,8 @@ static inline __device__ bool isnan(half_t x)
     return (xx & 0x7FFF) > 0x7C00;
 };
 
+static inline __device__ bool isnan(f8_t x) { return (x & 0x80); };
+
 static inline __device__ half_t sqrt(half_t x)
 {
     return static_cast<half_t>(__builtin_amdgcn_sqrtf(static_cast<float>(x)));

include/ck_tile/core/config.hpp

@@ -157,8 +157,11 @@
 #endif
 #endif
 
+// workaround for ROCm 6.2 and later
 #ifndef CK_TILE_WORKAROUND_ROCM_6_2_SCRATCH_MEMORY_ISSUE
-#if HIP_VERSION_MAJOR == 6 && HIP_VERSION_MINOR == 2 && HIP_VERSION_PATCH >= 41133
+#if(HIP_VERSION_MAJOR == 6 && HIP_VERSION_MINOR == 2 && HIP_VERSION_PATCH >= 41133) ||  \
+    (HIP_VERSION_MAJOR == 6 && HIP_VERSION_MINOR == 3 && HIP_VERSION_PATCH >= 42131) || \
+    (HIP_VERSION_MAJOR == 6 && HIP_VERSION_MINOR > 3)
 #define CK_TILE_WORKAROUND_ROCM_6_2_SCRATCH_MEMORY_ISSUE 1
 #else
 #define CK_TILE_WORKAROUND_ROCM_6_2_SCRATCH_MEMORY_ISSUE 0

include/ck_tile/core/container/array.hpp

@@ -1,9 +1,10 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
 #include <initializer_list>
+#include <vector>
 
 #include "ck_tile/core/config.hpp"
 #include "ck_tile/core/numeric/integer.hpp"
@@ -236,6 +237,16 @@ CK_TILE_HOST_DEVICE constexpr bool operator!=(const array<T, Size>& a, const arr
     return !(a == b);
 }
 
+template <typename T, index_t N, typename X>
+CK_TILE_HOST_DEVICE constexpr auto to_array(const std::vector<X>& x)
+{
+    array<T, N> arr;
+
+    static_for<0, N, 1>{}([&x, &arr](auto i) { arr(i) = x[i]; });
+
+    return arr;
+}
+
 template <typename T, index_t N, typename X>
 CK_TILE_HOST_DEVICE constexpr auto to_array(const X& x)
 {

include/ck_tile/core/container/thread_buffer.hpp

@@ -58,7 +58,7 @@ struct thread_buffer {
     template <index_t I> CK_TILE_HOST_DEVICE constexpr const auto& at() const           { return get(I); }
     template <index_t I> CK_TILE_HOST_DEVICE constexpr auto& at(number<I>)              { return get(I); }
     template <index_t I> CK_TILE_HOST_DEVICE constexpr const auto& at(number<I>) const  { return get(I); }
-
+    
     template <typename X_,
               typename std::enable_if<has_same_scalar_type<value_type, X_>::value, bool>::type = false>
     CK_TILE_HOST_DEVICE constexpr auto _get_as() const

include/ck_tile/host.hpp

@@ -5,6 +5,8 @@
 
 #include "ck_tile/host/arg_parser.hpp"
 #include "ck_tile/host/check_err.hpp"
+#include "ck_tile/host/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck_tile/host/convolution_parameter.hpp"
 #include "ck_tile/host/device_memory.hpp"
 #include "ck_tile/host/fill.hpp"
 #include "ck_tile/host/hip_check_error.hpp"

include/ck_tile/host/arg_parser.hpp

@@ -50,12 +50,22 @@ class ArgParser
         }
         return *this;
     }
-    void print()
+    void print() const
     {
+        // find max key length
+        std::string::size_type max_key_length = 11;
+        for(auto& key : keys)
+        {
+            if(max_key_length < key.length())
+            {
+                max_key_length = key.length();
+            }
+        }
+
         printf("args:\n");
         for(auto& key : keys)
         {
-            auto value = input_map[key];
+            auto value = input_map.at(key);
             std::vector<std::string> help_text_lines;
             size_t pos = 0;
             for(size_t next_pos = value.help_text.find('\n', pos); next_pos != std::string::npos;)
@@ -69,8 +79,7 @@ class ArgParser
                 std::string(value.help_text.begin() + pos, value.help_text.end()));
 
             std::string default_value = std::string("(default:") + value.value + std::string(")");
-
-            std::cout << std::setw(2) << std::setw(12 - value.name.length()) << "-" << key
+            std::cout << std::setw(1 + max_key_length - value.name.length()) << "-" << key
                       << std::setw(4) << " " << help_text_lines[0] << " " << default_value
                       << std::endl;
 
@@ -78,7 +87,8 @@ class ArgParser
                 help_next_line != help_text_lines.end();
                 ++help_next_line)
             {
-                std::cout << std::setw(17) << " " << *help_next_line << std::endl;
+                std::cout << std::setw(1 + max_key_length + 4) << " " << *help_next_line
+                          << std::endl;
             }
         }
     }

include/ck_tile/host/convolution_host_tensor_descriptor_helper.hpp

@@ -0,0 +1,266 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/ops/common/tensor_layout.hpp"
+#include "ck_tile/host/convolution_parameter.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+
+namespace ck_tile {
+namespace conv {
+namespace detail {
+
+template <typename OldLayout>
+CK_TILE_HOST std::vector<std::size_t> get_layout_transpose_gnchw_to_old()
+{
+    if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNCW> ||
+                 std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKCX> ||
+                 std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNKW>)
+    {
+        return {0, 1, 2, 3};
+    }
+    else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNCHW> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKCYX> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNKHW>)
+    {
+        return {0, 1, 2, 3, 4};
+    }
+    else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNCDHW> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKCZYX> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNKDHW>)
+    {
+        return {0, 1, 2, 3, 4, 5};
+    }
+    if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNWC> ||
+                 std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKXC> ||
+                 std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNWK>)
+    {
+        return {0, 1, 3, 2};
+    }
+    else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNHWC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKYXC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNHWK>)
+    {
+        return {0, 1, 4, 2, 3};
+    }
+    else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNDHWC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GKZYXC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::GNDHWK>)
+    {
+        return {0, 1, 5, 2, 3, 4};
+    }
+    else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NWGC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::KXGC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NWGK>)
+    {
+        return {2, 0, 3, 1};
+    }
+    else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NHWGC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::KYXGC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NHWGK>)
+    {
+        return {3, 0, 4, 1, 2};
+    }
+    else if constexpr(std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NDHWGC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::KZYXGC> ||
+                      std::is_same_v<OldLayout, ck_tile::tensor_layout::convolution::NDHWGK>)
+    {
+        return {4, 0, 5, 1, 2, 3};
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+}
+
+} // namespace detail
+
+// make tensor descriptor for packed input tensor, and order the dimension in the order of GNCHW
+// regardless of physical layout
+template <typename InLayout>
+CK_TILE_HOST HostTensorDescriptor
+make_input_host_tensor_descriptor_g_n_c_wis_packed(const ck_tile::conv::ConvParam& param)
+{
+    std::vector<std::size_t> physical_lengths;
+
+    if constexpr(std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNCW> ||
+                 std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNCHW> ||
+                 std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNCDHW>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNWC> ||
+                      std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNHWC> ||
+                      std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::GNDHWC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::NWGC> ||
+                      std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::NHWGC> ||
+                      std::is_same_v<InLayout, ck_tile::tensor_layout::convolution::NDHWGC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 1,
+                                param.input_spatial_lengths_.begin(),
+                                param.input_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        printf("%s\n", InLayout::name);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+
+    return transpose_host_tensor_descriptor_given_new2old(
+        HostTensorDescriptor(physical_lengths),
+        detail::get_layout_transpose_gnchw_to_old<InLayout>());
+}
+
+// make tensor descriptor for packed weight tensor, and order the dimension in the order of GKCYX
+// regardless of physical layout
+template <typename WeiLayout>
+CK_TILE_HOST HostTensorDescriptor
+make_weight_host_tensor_descriptor_g_k_c_xs_packed(const ck_tile::conv::ConvParam& param)
+{
+    std::vector<std::size_t> physical_lengths;
+
+    if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KXC> ||
+                 std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KYXC> ||
+                 std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KZYXC>)
+    {
+        if(param.G_ != 1)
+        {
+            throw std::runtime_error("wrong! G != 1");
+        }
+
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKCX> ||
+                      std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKCYX> ||
+                      std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKCZYX>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKXC> ||
+                      std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKYXC> ||
+                      std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::GKZYXC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KXGC> ||
+                      std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KYXGC> ||
+                      std::is_same_v<WeiLayout, ck_tile::tensor_layout::convolution::KZYXGC>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.K_),
+                                                    static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.C_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 1,
+                                param.filter_spatial_lengths_.begin(),
+                                param.filter_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        printf("%s\n", WeiLayout::name);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+
+    return transpose_host_tensor_descriptor_given_new2old(
+        HostTensorDescriptor(physical_lengths),
+        detail::get_layout_transpose_gnchw_to_old<WeiLayout>());
+}
+
+// make tensor descriptor for packed output tensor, and order the dimension in the order of GNKHW
+// regardless of physical layout
+template <typename OutLayout>
+CK_TILE_HOST HostTensorDescriptor
+make_output_host_tensor_descriptor_g_n_k_wos_packed(const ck_tile::conv::ConvParam& param)
+{
+    std::vector<std::size_t> physical_lengths;
+
+    if constexpr(std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNKW> ||
+                 std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNKHW> ||
+                 std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNKDHW>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.end(),
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    // separate from legacy code above
+    else if constexpr(std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNWK> ||
+                      std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNHWK> ||
+                      std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::GNDHWK>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 2,
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else if constexpr(std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::NWGK> ||
+                      std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::NHWGK> ||
+                      std::is_same_v<OutLayout, ck_tile::tensor_layout::convolution::NDHWGK>)
+    {
+        physical_lengths = std::vector<std::size_t>{static_cast<std::size_t>(param.N_),
+                                                    static_cast<std::size_t>(param.G_),
+                                                    static_cast<std::size_t>(param.K_)};
+
+        physical_lengths.insert(physical_lengths.begin() + 1,
+                                param.output_spatial_lengths_.begin(),
+                                param.output_spatial_lengths_.begin() + param.num_dim_spatial_);
+    }
+    else
+    {
+        printf("%s\n", __func__);
+        printf("%s\n", OutLayout::name);
+        throw std::runtime_error("wrong! unsupported layout");
+    }
+
+    return transpose_host_tensor_descriptor_given_new2old(
+        HostTensorDescriptor(physical_lengths),
+        detail::get_layout_transpose_gnchw_to_old<OutLayout>());
+}
+
+} // namespace conv
+} // namespace ck_tile

include/ck_tile/host/convolution_parameter.hpp

@@ -0,0 +1,277 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <numeric>
+#include <iterator>
+#include <vector>
+
+namespace ck_tile {
+namespace conv {
+
+struct ConvParam
+{
+    ConvParam(ck_tile::index_t n_dim,
+              ck_tile::index_t group_count,
+              ck_tile::index_t n_batch,
+              ck_tile::index_t n_out_channels,
+              ck_tile::index_t n_in_channels,
+              const std::vector<ck_tile::index_t>& filters_len,
+              const std::vector<ck_tile::index_t>& input_len,
+              const std::vector<ck_tile::index_t>& strides,
+              const std::vector<ck_tile::index_t>& dilations,
+              const std::vector<ck_tile::index_t>& left_pads,
+              const std::vector<ck_tile::index_t>& right_pads)
+        : num_dim_spatial_(static_cast<ck_tile::long_index_t>(n_dim)),
+          G_(static_cast<ck_tile::long_index_t>(group_count)),
+          N_(static_cast<ck_tile::long_index_t>(n_batch)),
+          K_(static_cast<ck_tile::long_index_t>(n_out_channels)),
+          C_(static_cast<ck_tile::long_index_t>(n_in_channels)),
+          filter_spatial_lengths_(num_dim_spatial_),
+          input_spatial_lengths_(num_dim_spatial_),
+          output_spatial_lengths_(num_dim_spatial_),
+          conv_filter_strides_(num_dim_spatial_),
+          conv_filter_dilations_(num_dim_spatial_),
+          input_left_pads_(num_dim_spatial_),
+          input_right_pads_(num_dim_spatial_)
+    {
+        if(static_cast<ck_tile::index_t>(filter_spatial_lengths_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(input_spatial_lengths_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(conv_filter_strides_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(conv_filter_dilations_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(input_left_pads_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(input_right_pads_.size()) != num_dim_spatial_)
+        {
+            throw(std::runtime_error(
+                "ConvParam::ConvParam: "
+                "parameter size is different from number of declared dimensions!"));
+        }
+
+        for(ck_tile::index_t i = 0; i < num_dim_spatial_; ++i)
+        {
+            filter_spatial_lengths_[i] = static_cast<ck_tile::long_index_t>(filters_len[i]);
+            input_spatial_lengths_[i]  = static_cast<ck_tile::long_index_t>(input_len[i]);
+            conv_filter_strides_[i]    = static_cast<ck_tile::long_index_t>(strides[i]);
+            conv_filter_dilations_[i]  = static_cast<ck_tile::long_index_t>(dilations[i]);
+            input_left_pads_[i]        = static_cast<ck_tile::long_index_t>(left_pads[i]);
+            input_right_pads_[i]       = static_cast<ck_tile::long_index_t>(right_pads[i]);
+
+            // XEff = (X - 1) * conv_dilation_w + 1;
+            // Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+            const ck_tile::long_index_t x_eff =
+                (filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
+
+            output_spatial_lengths_[i] =
+                (input_spatial_lengths_[i] + input_left_pads_[i] + input_right_pads_[i] - x_eff) /
+                    conv_filter_strides_[i] +
+                1;
+        }
+    }
+
+    ConvParam(ck_tile::long_index_t n_dim,
+              ck_tile::long_index_t group_count,
+              ck_tile::long_index_t n_batch,
+              ck_tile::long_index_t n_out_channels,
+              ck_tile::long_index_t n_in_channels,
+              const std::vector<ck_tile::long_index_t>& filters_len,
+              const std::vector<ck_tile::long_index_t>& input_len,
+              const std::vector<ck_tile::long_index_t>& strides,
+              const std::vector<ck_tile::long_index_t>& dilations,
+              const std::vector<ck_tile::long_index_t>& left_pads,
+              const std::vector<ck_tile::long_index_t>& right_pads)
+        : num_dim_spatial_(n_dim),
+          G_(group_count),
+          N_(n_batch),
+          K_(n_out_channels),
+          C_(n_in_channels),
+          filter_spatial_lengths_(filters_len),
+          input_spatial_lengths_(input_len),
+          output_spatial_lengths_(num_dim_spatial_),
+          conv_filter_strides_(strides),
+          conv_filter_dilations_(dilations),
+          input_left_pads_(left_pads),
+          input_right_pads_(right_pads)
+    {
+        if(static_cast<ck_tile::index_t>(filter_spatial_lengths_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(input_spatial_lengths_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(conv_filter_strides_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(conv_filter_dilations_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(input_left_pads_.size()) != num_dim_spatial_ ||
+           static_cast<ck_tile::index_t>(input_right_pads_.size()) != num_dim_spatial_)
+        {
+            throw(std::runtime_error(
+                "ConvParam::ConvParam: "
+                "parameter size is different from number of declared dimensions!"));
+        }
+
+        for(ck_tile::index_t i = 0; i < num_dim_spatial_; ++i)
+        {
+            // XEff = (X - 1) * conv_dilation_w + 1;
+            // Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+            const ck_tile::long_index_t x_eff =
+                (filter_spatial_lengths_[i] - 1) * conv_filter_dilations_[i] + 1;
+
+            output_spatial_lengths_[i] =
+                (input_spatial_lengths_[i] + input_left_pads_[i] + input_right_pads_[i] - x_eff) /
+                    conv_filter_strides_[i] +
+                1;
+        }
+    }
+
+    ck_tile::long_index_t num_dim_spatial_;
+    ck_tile::long_index_t G_;
+    ck_tile::long_index_t N_;
+    ck_tile::long_index_t K_;
+    ck_tile::long_index_t C_;
+
+    std::vector<ck_tile::long_index_t> filter_spatial_lengths_;
+    std::vector<ck_tile::long_index_t> input_spatial_lengths_;
+    std::vector<ck_tile::long_index_t> output_spatial_lengths_;
+
+    std::vector<ck_tile::long_index_t> conv_filter_strides_;
+    std::vector<ck_tile::long_index_t> conv_filter_dilations_;
+
+    std::vector<ck_tile::long_index_t> input_left_pads_;
+    std::vector<ck_tile::long_index_t> input_right_pads_;
+
+    std::vector<ck_tile::long_index_t> GetOutputSpatialLengths() const
+    {
+        return output_spatial_lengths_;
+    }
+
+    std::size_t GetFlops() const
+    {
+        // 2 * G * N * K * C * <output spatial lengths product> * <filter spatial lengths product>
+        return static_cast<std::size_t>(2) * G_ * N_ * K_ * C_ *
+               std::accumulate(std::begin(output_spatial_lengths_),
+                               std::next(std::begin(output_spatial_lengths_), num_dim_spatial_),
+                               1,
+                               std::multiplies<>()) *
+               std::accumulate(std::begin(filter_spatial_lengths_),
+                               std::next(std::begin(filter_spatial_lengths_), num_dim_spatial_),
+                               1,
+                               std::multiplies<>());
+    }
+
+    template <typename InDataType>
+    std::size_t GetInputByte() const
+    {
+        // sizeof(InDataType) * (G * N * C * <input spatial lengths product>) +
+        return sizeof(InDataType) *
+               (G_ * N_ * C_ *
+                std::accumulate(std::begin(input_spatial_lengths_),
+                                std::next(std::begin(input_spatial_lengths_), num_dim_spatial_),
+                                1,
+                                std::multiplies<>()));
+    }
+
+    template <typename WeiDataType>
+    std::size_t GetWeightByte() const
+    {
+        // sizeof(WeiDataType) * (G * K * C * <filter spatial lengths product>) +
+        return sizeof(WeiDataType) *
+               (G_ * K_ * C_ *
+                std::accumulate(std::begin(filter_spatial_lengths_),
+                                std::next(std::begin(filter_spatial_lengths_), num_dim_spatial_),
+                                1,
+                                std::multiplies<>()));
+    }
+
+    template <typename OutDataType>
+    std::size_t GetOutputByte() const
+    {
+        // sizeof(OutDataType) * (G * N * K * <output spatial lengths product>);
+        return sizeof(OutDataType) * (G_ * N_ * K_ *
+                                      std::accumulate(std::begin(output_spatial_lengths_),
+                                                      std::end(output_spatial_lengths_),
+                                                      static_cast<std::size_t>(1),
+                                                      std::multiplies<std::size_t>()));
+    }
+
+    template <typename InDataType, typename WeiDataType, typename OutDataType>
+    std::size_t GetByte() const
+    {
+        return GetInputByte<InDataType>() + GetWeightByte<WeiDataType>() +
+               GetOutputByte<OutDataType>();
+    }
+};
+
+CK_TILE_HOST std::string get_conv_param_parser_helper_msg()
+{
+    std::string msg;
+
+    msg += "Following arguments (depending on number of spatial dims):\n"
+           " Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)\n"
+           " G, N, K, C, \n"
+           " <filter spatial dimensions>, (ie Y, X for 2D)\n"
+           " <input image spatial dimensions>, (ie Hi, Wi for 2D)\n"
+           " <strides>, (ie Sy, Sx for 2D)\n"
+           " <dilations>, (ie Dy, Dx for 2D)\n"
+           " <left padding>, (ie LeftPy, LeftPx for 2D)\n"
+           " <right padding>, (ie RightPy, RightPx for 2D)\n";
+
+    return msg;
+}
+
+CK_TILE_HOST ck_tile::conv::ConvParam
+parse_conv_param(int num_dim_spatial, int arg_idx, char* const argv[])
+{
+    const ck_tile::long_index_t G = std::stol(argv[arg_idx++]);
+    const ck_tile::long_index_t N = std::stol(argv[arg_idx++]);
+    const ck_tile::long_index_t K = std::stol(argv[arg_idx++]);
+    const ck_tile::long_index_t C = std::stol(argv[arg_idx++]);
+
+    std::vector<ck_tile::long_index_t> filter_spatial_lengths(num_dim_spatial);
+    std::vector<ck_tile::long_index_t> input_spatial_lengths(num_dim_spatial);
+    std::vector<ck_tile::long_index_t> conv_filter_strides(num_dim_spatial);
+    std::vector<ck_tile::long_index_t> conv_filter_dilations(num_dim_spatial);
+    std::vector<ck_tile::long_index_t> input_left_pads(num_dim_spatial);
+    std::vector<ck_tile::long_index_t> input_right_pads(num_dim_spatial);
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        filter_spatial_lengths[i] = std::stol(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        input_spatial_lengths[i] = std::stol(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        conv_filter_strides[i] = std::stol(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        conv_filter_dilations[i] = std::stol(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        input_left_pads[i] = std::stol(argv[arg_idx++]);
+    }
+
+    for(int i = 0; i < num_dim_spatial; ++i)
+    {
+        input_right_pads[i] = std::stol(argv[arg_idx++]);
+    }
+
+    return ck_tile::conv::ConvParam{num_dim_spatial,
+                                    G,
+                                    N,
+                                    K,
+                                    C,
+                                    filter_spatial_lengths,
+                                    input_spatial_lengths,
+                                    conv_filter_strides,
+                                    conv_filter_dilations,
+                                    input_left_pads,
+                                    input_right_pads};
+}
+
+} // namespace conv
+} // namespace ck_tile

include/ck_tile/host/host_tensor.hpp

@@ -176,7 +176,20 @@ struct HostTensorDescriptor
         return std::inner_product(iss.begin(), iss.end(), mStrides.begin(), std::size_t{0});
     }
 
-    friend std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc);
+    friend std::ostream& operator<<(std::ostream& os, const HostTensorDescriptor& desc)
+    {
+        os << "dim " << desc.get_num_of_dimension() << ", ";
+
+        os << "lengths {";
+        LogRange(os, desc.get_lengths(), ", ");
+        os << "}, ";
+
+        os << "strides {";
+        LogRange(os, desc.get_strides(), ", ");
+        os << "}";
+
+        return os;
+    }
 
     private:
     std::vector<std::size_t> mLens;

include/ck_tile/host/reference/reference_gemm.hpp

@@ -27,7 +27,9 @@ CK_TILE_HOST void reference_gemm(const HostTensor<ADataType>& a_m_k,
                                  const BElementOp& b_element_op     = {},
                                  const ACCElementOp& acc_element_op = {})
 {
-    const int N = b_n_k.mDesc.get_lengths()[0];
+    const int N = (std::is_same_v<LayoutB, tensor_layout::gemm::ColumnMajor>)
+                      ? b_n_k.mDesc.get_lengths()[0]
+                      : b_n_k.mDesc.get_lengths()[1];
     const int K = (std::is_same_v<LayoutA, tensor_layout::gemm::RowMajor>)
                       ? a_m_k.mDesc.get_lengths()[1]
                       : a_m_k.mDesc.get_lengths()[0];
@@ -45,20 +47,31 @@ CK_TILE_HOST void reference_gemm(const HostTensor<ADataType>& a_m_k,
                 ADataType v_a = (std::is_same_v<LayoutA, tensor_layout::gemm::RowMajor>)
                                     ? a_element_op(a_m_k(m, k))
                                     : a_element_op(a_m_k(k, m));
-                BDataType v_b = b_element_op(b_n_k(n, k));
+                BDataType v_b = (std::is_same_v<LayoutB, tensor_layout::gemm::ColumnMajor>)
+                                    ? b_element_op(b_n_k(n, k))
+                                    : b_element_op(b_n_k(k, n));
 
                 v_acc += ck_tile::type_convert<AccDataType>(v_a) *
                          ck_tile::type_convert<AccDataType>(v_b);
             }
 
-            c_m_n(m, n) = ck_tile::type_convert<CDataType>(acc_element_op(v_acc));
+            CDataType& c_ref = (std::is_same_v<LayoutC, tensor_layout::gemm::RowMajor>)
+                                   ? c_m_n(m, n)
+                                   : c_m_n(n, m);
+            c_ref            = ck_tile::type_convert<CDataType>(acc_element_op(v_acc));
         }
     };
 
     make_ParallelTensorFunctor(f, M)(std::thread::hardware_concurrency());
 }
 
-template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename LayoutA,
+          typename LayoutB,
+          typename LayoutC>
 __global__ void naive_gemm_kernel(ADataType* A,
                                   BDataType* B,
                                   CDataType* C,
@@ -76,18 +89,32 @@ __global__ void naive_gemm_kernel(ADataType* A,
     if(row < M && col < N)
     {
         AccDataType acc = 0.0;
-
         for(int k = 0; k < K; ++k)
         {
-            acc += static_cast<AccDataType>(A[row * strideA + k]) *
-                   static_cast<AccDataType>(B[col * strideB + k]);
+            // Adjust indexing based on matrix layout
+            int a_index = (std::is_same_v<LayoutA, tensor_layout::gemm::RowMajor>)
+                              ? row * strideA + k
+                              : k * strideA + row;
+            int b_index = (std::is_same_v<LayoutB, tensor_layout::gemm::ColumnMajor>)
+                              ? col * strideB + k
+                              : k * strideB + col;
+            acc += static_cast<AccDataType>(A[a_index]) * static_cast<AccDataType>(B[b_index]);
         }
 
-        C[row * strideC + col] = acc; // Store as AccDataType
+        int c_index = (std::is_same_v<LayoutC, tensor_layout::gemm::RowMajor>)
+                          ? row * strideC + col
+                          : col * strideC + row;
+        C[c_index]  = acc;
     }
 }
 
-template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename LayoutA,
+          typename LayoutB,
+          typename LayoutC>
 void reference_gemm_gpu(DeviceMem& a_device,
                         DeviceMem& b_device,
                         DeviceMem& c_device,
@@ -145,7 +172,7 @@ void reference_gemm_gpu(DeviceMem& a_device,
     int numThreadsPerBlock = 256; // Common choice for threads per block
     int numBlocks          = (totalElements + numThreadsPerBlock - 1) / numThreadsPerBlock;
 
-    naive_gemm_kernel<ADataType, BDataType, AccDataType, CDataType>
+    naive_gemm_kernel<ADataType, BDataType, AccDataType, CDataType, LayoutA, LayoutB, LayoutC>
         <<<numBlocks, numThreadsPerBlock>>>(d_A, d_B, d_C, M, N, K, stride_a, stride_b, stride_c);
     errC = hipMemcpy(
         c_device.GetDeviceBuffer(), d_C, M * N * sizeof(CDataType), hipMemcpyDeviceToHost);

include/ck_tile/host/reference/reference_im2col.hpp

@@ -1,5 +1,5 @@
 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -9,53 +9,125 @@
 
 namespace ck_tile {
 
-template <typename T>
-CK_TILE_HOST void reference_im2col(HostTensor<T>& in_mtx_host_ref,
-                                   const HostTensor<T>& in_host,
-                                   int /*N*/,
-                                   int /*K*/,
-                                   int C,
-                                   int /*Y*/,
-                                   int X,
-                                   int Hi,
-                                   int Wi,
-                                   int Ho,
-                                   int Wo,
-                                   int ConvStrideH,
-                                   int ConvStrideW,
-                                   int ConvDilationH,
-                                   int ConvDilationW,
-                                   int InLeftPadH,
-                                   int InLeftPadW,
-                                   int /*InRightPadH*/,
-                                   int /*InRightPadW*/)
+template <typename InDataType, typename OutDataType, index_t NDimSpatial>
+CK_TILE_HOST void reference_im2col(const HostTensor<InDataType>& in_host,
+                                   HostTensor<OutDataType>& out_host,
+                                   const ck_tile::conv::ConvParam& conv_params)
 {
-    int GemmM = in_mtx_host_ref.get_lengths()[0];
-    int GemmK = in_mtx_host_ref.get_lengths()[1];
+    const long_index_t G = in_host.get_lengths()[0];
+    const long_index_t N = in_host.get_lengths()[1];
+    const long_index_t C = in_host.get_lengths()[2];
 
-    for(int gemm_m = 0; gemm_m < GemmM; ++gemm_m)
+    if constexpr(NDimSpatial == 1)
     {
-        int mtmp = gemm_m;
-        int n    = mtmp / (Ho * Wo);
-        mtmp -= n * Ho * Wo;
-        int ho = mtmp / Wo;
-        int wo = mtmp - ho * Wo;
+        const long_index_t Wo = conv_params.output_spatial_lengths_[0];
+        auto func             = [&](auto g, auto n, auto wo) {
+            long_index_t row    = n * Wo + wo;
+            long_index_t column = 0;
 
-        for(int gemm_k = 0; gemm_k < GemmK; ++gemm_k)
-        {
-            int ktmp = gemm_k;
-            int y    = ktmp / (X * C);
-            ktmp -= y * X * C;
-            int x = ktmp / C;
-            int c = ktmp - x * C;
+            for(long_index_t x = 0; x < conv_params.filter_spatial_lengths_[0]; ++x)
+            {
+                auto wi = static_cast<long_index_t>(wo * conv_params.conv_filter_strides_[0]) +
+                          static_cast<long_index_t>(x * conv_params.conv_filter_dilations_[0]) -
+                          static_cast<long_index_t>(conv_params.input_left_pads_[0]);
 
-            int hi = y * ConvDilationH + ho * ConvStrideH - InLeftPadH;
-            int wi = x * ConvDilationW + wo * ConvStrideW - InLeftPadW;
+                for(long_index_t c = 0; c < C; ++c)
+                {
+                    if(wi >= 0 && type_convert<std::size_t>(wi) < in_host.get_lengths()[3])
+                    {
+                        InDataType v_in          = in_host(g, n, c, wi);
+                        out_host(g, row, column) = type_convert<OutDataType>(v_in);
+                    }
+                    column++;
+                }
+            }
+        };
 
-            bool inbound = (hi >= 0 && hi < Hi && wi >= 0 && wi < Wi);
+        make_ParallelTensorFunctor(func, G, N, Wo)(std::thread::hardware_concurrency());
+    }
+    else if constexpr(NDimSpatial == 2)
+    {
+        const long_index_t Ho = conv_params.output_spatial_lengths_[0];
+        const long_index_t Wo = conv_params.output_spatial_lengths_[1];
 
-            in_mtx_host_ref(gemm_m, gemm_k) = inbound ? in_host(n, hi, wi, c) : 0;
-        }
+        auto func = [&](auto g, auto n, auto ho, auto wo) {
+            long_index_t row    = n * Ho * Wo + ho * Wo + wo;
+            long_index_t column = 0;
+
+            for(long_index_t y = 0; y < conv_params.filter_spatial_lengths_[0]; ++y)
+            {
+                auto hi = static_cast<long_index_t>(ho * conv_params.conv_filter_strides_[0]) +
+                          static_cast<long_index_t>(y * conv_params.conv_filter_dilations_[0]) -
+                          static_cast<long_index_t>(conv_params.input_left_pads_[0]);
+
+                for(long_index_t x = 0; x < conv_params.filter_spatial_lengths_[1]; ++x)
+                {
+                    auto wi = static_cast<long_index_t>(wo * conv_params.conv_filter_strides_[1]) +
+                              static_cast<long_index_t>(x * conv_params.conv_filter_dilations_[1]) -
+                              static_cast<long_index_t>(conv_params.input_left_pads_[1]);
+
+                    for(long_index_t c = 0; c < C; ++c)
+                    {
+
+                        if(hi >= 0 && type_convert<std::size_t>(hi) < in_host.get_lengths()[3] &&
+                           wi >= 0 && type_convert<std::size_t>(wi) < in_host.get_lengths()[4])
+                        {
+                            InDataType v_in          = in_host(g, n, c, hi, wi);
+                            out_host(g, row, column) = type_convert<OutDataType>(v_in);
+                        }
+                        column++;
+                    }
+                }
+            }
+        };
+
+        make_ParallelTensorFunctor(func, G, N, Ho, Wo)(std::thread::hardware_concurrency());
+    }
+    else if constexpr(NDimSpatial == 3)
+    {
+        const long_index_t Do = conv_params.output_spatial_lengths_[0];
+        const long_index_t Ho = conv_params.output_spatial_lengths_[1];
+        const long_index_t Wo = conv_params.output_spatial_lengths_[2];
+
+        auto func = [&](auto g, auto n, auto d_o, auto ho, auto wo) {
+            long_index_t row    = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo;
+            long_index_t column = 0;
+
+            for(long_index_t z = 0; z < conv_params.filter_spatial_lengths_[0]; ++z)
+            {
+                auto di = static_cast<long_index_t>(d_o * conv_params.conv_filter_strides_[0]) +
+                          static_cast<long_index_t>(z * conv_params.conv_filter_dilations_[0]) -
+                          static_cast<long_index_t>(conv_params.input_left_pads_[0]);
+                for(long_index_t y = 0; y < conv_params.filter_spatial_lengths_[1]; ++y)
+                {
+                    auto hi = static_cast<long_index_t>(ho * conv_params.conv_filter_strides_[1]) +
+                              static_cast<long_index_t>(y * conv_params.conv_filter_dilations_[1]) -
+                              static_cast<long_index_t>(conv_params.input_left_pads_[1]);
+                    for(long_index_t x = 0; x < conv_params.filter_spatial_lengths_[2]; ++x)
+                    {
+                        auto wi =
+                            static_cast<long_index_t>(wo * conv_params.conv_filter_strides_[2]) +
+                            static_cast<long_index_t>(x * conv_params.conv_filter_dilations_[2]) -
+                            static_cast<long_index_t>(conv_params.input_left_pads_[2]);
+                        for(long_index_t c = 0; c < C; ++c)
+                        {
+                            if(di >= 0 &&
+                               type_convert<std::size_t>(di) < in_host.get_lengths()[3] &&
+                               hi >= 0 &&
+                               type_convert<std::size_t>(hi) < in_host.get_lengths()[4] &&
+                               wi >= 0 && type_convert<std::size_t>(wi) < in_host.get_lengths()[5])
+                            {
+                                InDataType v_in          = in_host(g, n, c, di, hi, wi);
+                                out_host(g, row, column) = type_convert<OutDataType>(v_in);
+                            }
+                            column++;
+                        }
+                    }
+                }
+            }
+        };
+
+        make_ParallelTensorFunctor(func, G, N, Do, Ho, Wo)(std::thread::hardware_concurrency());
     }
 }
 } // namespace ck_tile

include/ck_tile/ops/epilogue.hpp

@@ -3,5 +3,6 @@
 
 #pragma once
 
+#include "ck_tile/ops/epilogue/cshuffle_epilogue.hpp"
 #include "ck_tile/ops/epilogue/default_2d_epilogue.hpp"
 #include "ck_tile/ops/common/tensor_layout.hpp"

include/ck_tile/ops/epilogue/cshuffle_epilogue.hpp

@@ -0,0 +1,171 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+
+#define CK_TILE_MAX_RANK 5
+
+namespace ck_tile {
+
+// this epilogue aiming to store a matrix with different layout from the shared memory to the global
+// memory.
+template <typename AccDataType_,
+          typename ODataType_,
+          bool kPadM_,
+          bool kPadN_,
+          bool kTilePermute_,
+          index_t kRank_,
+          index_t kPerm0,
+          index_t kPerm1,
+          index_t TileSize0,
+          index_t TileSize1,
+          index_t kPerm2    = 0,
+          index_t kPerm3    = 0,
+          index_t kPerm4    = 0,
+          index_t TileSize2 = 0,
+          index_t TileSize3 = 0,
+          index_t TileSize4 = 0>
+struct CShuffleEpilogueProblem
+{
+    using AccDataType                                = remove_cvref_t<AccDataType_>;
+    using ODataType                                  = remove_cvref_t<ODataType_>;
+    static constexpr bool kPadM                      = kPadM_;
+    static constexpr bool kPadN                      = kPadN_;
+    static constexpr bool kTilePermute               = kTilePermute_;
+    static constexpr index_t kRank                   = kRank_;
+    static constexpr index_t kPerm[CK_TILE_MAX_RANK] = {kPerm0, kPerm1, kPerm2, kPerm3, kPerm4};
+    static constexpr index_t tile_sizes[CK_TILE_MAX_RANK] = {
+        TileSize0, TileSize1, TileSize2, TileSize3, TileSize4};
+};
+
+template <typename Problem_, typename Policy_ = void>
+struct CShuffleEpilogue
+{
+    using Problem                      = remove_cvref_t<Problem_>;
+    using AccDataType                  = remove_cvref_t<typename Problem::AccDataType>;
+    using ODataType                    = remove_cvref_t<typename Problem::ODataType>;
+    static constexpr bool kPadM        = Problem::kPadM;
+    static constexpr bool kPadN        = Problem::kPadN;
+    const index_t* kPerm               = Problem::kPerm;
+    static constexpr bool kTilePermute = Problem::kTilePermute;
+    static constexpr index_t kRank     = Problem::kRank;
+    const index_t* tile_sizes          = Problem::tile_sizes;
+
+    // No additional shared memory needed
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize() { return 0; }
+
+    template <typename OAccTile>
+    CK_TILE_DEVICE void permute_tile_data(OAccTile& o_acc_tile)
+    {
+        using DataType = typename OAccTile::DataType;
+
+        // Get thread buffer
+        auto& thread_buf = o_acc_tile.get_thread_buffer();
+
+        // Create a temporary buffer to hold the permuted data
+        thread_buffer<DataType, OAccTile::kThreadElementSpaceSize> permuted_thread_buf;
+
+        // Get the lengths of each dimension
+        auto thread_tensor_lengths = o_acc_tile.get_lengths();
+
+        // Total number of elements
+        index_t total_elements = OAccTile::kThreadElementSpaceSize;
+
+        // Iterate over all elements
+        for(index_t linear_idx = 0; linear_idx < total_elements; ++linear_idx)
+        {
+            // Convert linear index to multi-dimensional indices
+            array<index_t, kRank> indices;
+            index_t remaining = linear_idx;
+            static_for<0, kRank, 1>{}([&](auto i) {
+                constexpr auto rev_i = kRank - 1 - i;
+                indices(rev_i)       = remaining % thread_tensor_lengths.get(number<rev_i>{});
+                remaining /= thread_tensor_lengths.get(number<rev_i>{});
+            });
+
+            // Apply the permutation
+            array<index_t, kRank> permuted_indices;
+            static_for<0, kRank, 1>{}(
+                [&](auto i) { permuted_indices(i) = indices.get(number<Problem::kPerm[i]>{}); });
+
+            // Compute offsets
+            index_t dst_offset = 0;
+            index_t stride     = 1;
+
+            static_for<0, kRank, 1>{}([&](auto i) {
+                constexpr auto rev_i = kRank - 1 - i;
+                dst_offset += permuted_indices[rev_i] * stride;
+                stride *= thread_tensor_lengths.get(number<rev_i>{});
+            });
+
+            // Move the data
+            permuted_thread_buf(dst_offset) = thread_buf[linear_idx];
+        }
+
+        // Copy the permuted data back to the original thread buffer
+        for(index_t i = 0; i < total_elements; ++i)
+        {
+            thread_buf.set_as(i, permuted_thread_buf.get(i));
+        }
+    }
+
+    template <typename ODramWindowTmp, typename OAccTile>
+    CK_TILE_DEVICE auto operator()(ODramWindowTmp& o_dram_window_tmp, OAccTile& o_acc_tile)
+    {
+        const auto& current_window_origin = o_dram_window_tmp.get_window_origin();
+
+        // Compute the tile coordinates by dividing the window origin by the tile sizes
+        index_t tile_coords[CK_TILE_MAX_RANK] = {0};
+        for(index_t i = 0; i < kRank; ++i)
+        {
+            tile_coords[i] = current_window_origin[i] / tile_sizes[i];
+            // printf("The tile_coord is: %d", tile_coords[i]);
+        }
+
+        // Apply the permutation to the tile coordinates
+        index_t permuted_tile_coords[CK_TILE_MAX_RANK];
+        for(index_t i = 0; i < kRank; ++i)
+        {
+            permuted_tile_coords[i] = tile_coords[kPerm[i]];
+            // printf("The new permuted_tile_coords is: %d", permuted_tile_coords[i]);
+        }
+
+        // Compute the permuted window origin
+        index_t permuted_window_origin[CK_TILE_MAX_RANK] = {0};
+        for(index_t i = 0; i < kRank; ++i)
+        {
+            permuted_window_origin[i] = permuted_tile_coords[i] * tile_sizes[i];
+            // printf("The new permuted_window_origin is: %d", permuted_window_origin[i]);
+        }
+
+        typename ODramWindowTmp::BottomTensorIndex step = {};
+        for(index_t i = 0; i < kRank; ++i)
+        {
+            step[i] = permuted_window_origin[i] - current_window_origin[i];
+        }
+
+        // Move the window
+        move_tile_window(o_dram_window_tmp, step);
+
+        // Permute the data within the tile if necessary
+        if constexpr(kTilePermute)
+        {
+            permute_tile_data(o_acc_tile);
+        }
+
+        // Store the tile data to the permuted location
+        if constexpr(kPadM || kPadN)
+        {
+            store_tile_raw(o_dram_window_tmp, cast_tile<ODataType>(o_acc_tile));
+            buffer_store_fence();
+        }
+        else
+        {
+            store_tile(o_dram_window_tmp, cast_tile<ODataType>(o_acc_tile));
+        }
+    }
+};
+
+} // namespace ck_tile

include/ck_tile/ops/fmha/block/block_masking.hpp

@@ -308,9 +308,9 @@ struct SimplifiedGenericAttentionMask
     {
         auto [origin_start, origin_end] = GetTileRangeAlongX(i_y, height, width);
 
-        const index_t x_per_split = ck_tile::max(1, x_total / num_splits);
+        const index_t x_per_split = ck_tile::max(1, integer_divide_ceil(x_total, num_splits));
         const index_t split_start = x_per_split * i_split;
-        const index_t split_end = (i_split == num_splits - 1 ? x_total : split_start + x_per_split);
+        const index_t split_end   = split_start + x_per_split;
 
         return ck_tile::make_tuple(ck_tile::max(origin_start, split_start),
                                    ck_tile::min(origin_end, split_end));

include/ck_tile/ops/fmha/kernel/fmha_bwd_kernel.hpp

@@ -6,8 +6,11 @@
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/common.hpp"
 #include "ck_tile/ops/fmha/block/block_attention_bias_enum.hpp"
+
 #include <string>
 #include <type_traits>
+#include <utility>
+#include <variant>
 
 // S[seqlen_q, seqlen_k] = Q[seqlen_q, hdim_q] @ K[seqlen_k, hdim_q]
 // S'[seqlen_q, seqlen_k] = S[seqlen_q, seqlen_k] * Scale[1]
@@ -194,11 +197,23 @@ struct FmhaBwdDQDKDVKernel
         ck_tile::GenericAttentionMaskEnum mask_type;
     };
 
-    struct FmhaBwdCommonDropoutKargs
+    struct FmhaBwdDropoutSeedOffset
     {
-        void init_dropout(const float p_drop,
-                          const std::tuple<uint64_t, uint64_t>& drop_seed_offset,
-                          const float raw_scale)
+        template <typename T>
+        union ValueOrPointer
+        {
+            T val;
+            const T* ptr;
+        };
+
+        ValueOrPointer<uint64_t> drop_seed;
+        ValueOrPointer<uint64_t> drop_offset;
+        bool is_drop_seed_offset_from_host;
+    };
+
+    struct FmhaBwdCommonDropoutKargs : FmhaBwdDropoutSeedOffset
+    {
+        void init_dropout(float p_drop, uint64_t seed, uint64_t offset, float raw_scale)
         {
             float p_undrop = 1.0 - p_drop;
             p_undrop_in_uint8_t =
@@ -206,23 +221,41 @@ struct FmhaBwdDQDKDVKernel
             rp_undrop       = 1.0 / p_undrop;
             scale_rp_undrop = rp_undrop * raw_scale;
 
-            drop_seed   = std::get<0>(drop_seed_offset);
-            drop_offset = std::get<1>(drop_seed_offset);
+            this->drop_seed.val                 = seed;
+            this->drop_offset.val               = offset;
+            this->is_drop_seed_offset_from_host = true;
         }
+
+        void init_dropout(float p_drop,
+                          const uint64_t* seed_ptr,
+                          const uint64_t* offset_ptr,
+                          float raw_scale)
+        {
+            float p_undrop = 1.0 - p_drop;
+            p_undrop_in_uint8_t =
+                uint8_t(std::floor(p_undrop * std::numeric_limits<uint8_t>::max()));
+            rp_undrop       = 1.0 / p_undrop;
+            scale_rp_undrop = rp_undrop * raw_scale;
+
+            this->drop_seed.ptr                 = seed_ptr;
+            this->drop_offset.ptr               = offset_ptr;
+            this->is_drop_seed_offset_from_host = false;
+        }
+
         float rp_undrop             = 1;
         float scale_rp_undrop       = 1;
         uint8_t p_undrop_in_uint8_t = std::numeric_limits<uint8_t>::max();
-        uint64_t drop_seed          = 1;
-        uint64_t drop_offset        = 0;
         void* rand_val_ptr          = nullptr;
 
         ck_tile::index_t stride_randval       = 0;
         ck_tile::index_t nhead_stride_randval = 0;
     };
+
     struct FmhaBwdBatchModeDropoutKargs : FmhaBwdCommonDropoutKargs
     {
         ck_tile::index_t batch_stride_randval = 0;
     };
+
     struct FmhaBwdDeterministicKargs
     {
         ck_tile::index_t split_stride_dq_acc = 0;
@@ -327,7 +360,8 @@ struct FmhaBwdDQDKDVKernel
               ck_tile::index_t window_size_right,
               ck_tile::index_t mask_type,
               float p_drop,
-              const std::tuple<uint64_t, uint64_t>& drop_seed_offset)
+              std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+                  drop_seed_offset)
     {
         Kargs kargs{{q_ptr,
                      k_ptr,
@@ -405,7 +439,20 @@ struct FmhaBwdDQDKDVKernel
 
         if constexpr(kHasDropout)
         {
-            kargs.init_dropout(p_drop, drop_seed_offset, scale);
+            if(drop_seed_offset.index() == 0) // seed & offset come from host
+            {
+                const auto& [seed, offset] = std::get<0>(drop_seed_offset);
+                kargs.init_dropout(p_drop, seed, offset, scale);
+            }
+            else // seed & offset come from device
+            {
+                const auto& [seed_ptr, offset_ptr] = std::get<1>(drop_seed_offset);
+                kargs.init_dropout(p_drop,
+                                   reinterpret_cast<const uint64_t*>(seed_ptr),
+                                   reinterpret_cast<const uint64_t*>(offset_ptr),
+                                   scale);
+            }
+
             if constexpr(kIsStoreRandval)
             {
                 kargs.rand_val_ptr         = rand_val_ptr;
@@ -471,7 +518,8 @@ struct FmhaBwdDQDKDVKernel
               ck_tile::index_t window_size_right,
               ck_tile::index_t mask_type,
               float p_drop,
-              const std::tuple<uint64_t, uint64_t>& drop_seed_offset)
+              std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+                  drop_seed_offset)
     {
         Kargs kargs{{q_ptr,
                      k_ptr,
@@ -539,7 +587,20 @@ struct FmhaBwdDQDKDVKernel
         }
         if constexpr(kHasDropout)
         {
-            kargs.init_dropout(p_drop, drop_seed_offset, scale);
+            if(drop_seed_offset.index() == 0) // seed & offset come from host
+            {
+                const auto& [seed, offset] = std::get<0>(drop_seed_offset);
+                kargs.init_dropout(p_drop, seed, offset, scale);
+            }
+            else // seed & offset come from device
+            {
+                const auto& [seed_ptr, offset_ptr] = std::get<1>(drop_seed_offset);
+                kargs.init_dropout(p_drop,
+                                   reinterpret_cast<const uint64_t*>(seed_ptr),
+                                   reinterpret_cast<const uint64_t*>(offset_ptr),
+                                   scale);
+            }
+
             if constexpr(kIsStoreRandval)
             {
                 kargs.rand_val_ptr         = rand_val_ptr;
@@ -958,8 +1019,10 @@ struct FmhaBwdDQDKDVKernel
                 return FmhaDropout{i_batch_,
                                    i_nhead_,
                                    kargs.num_head_q,
-                                   kargs.drop_seed,
-                                   kargs.drop_offset,
+                                   kargs.is_drop_seed_offset_from_host ? kargs.drop_seed.val
+                                                                       : *kargs.drop_seed.ptr,
+                                   kargs.is_drop_seed_offset_from_host ? kargs.drop_offset.val
+                                                                       : *kargs.drop_offset.ptr,
                                    kargs.rp_undrop,
                                    kargs.p_undrop_in_uint8_t};
             }

include/ck_tile/ops/fmha/kernel/fmha_fwd_kernel.hpp

@@ -6,8 +6,11 @@
 #include "ck_tile/core.hpp"
 #include "ck_tile/ops/common.hpp"
 #include "ck_tile/ops/fmha/block/block_attention_bias_enum.hpp"
+
 #include <string>
 #include <type_traits>
+#include <utility>
+#include <variant>
 
 // S[seqlen_q, seqlen_k] = Q[seqlen_q, hdim_q] @ K[seqlen_k, hdim_q]
 // S'[seqlen_q, seqlen_k] = S[seqlen_q, seqlen_k] * Scale[1]
@@ -170,29 +173,55 @@ struct FmhaFwdKernel
         ck_tile::index_t batch_stride_lse = 0;
     };
 
-    struct FmhaFwdCommonDropoutKargs
+    struct FmhaFwdDropoutSeedOffset
     {
-        void init_dropout(const float p_drop,
-                          const std::tuple<uint64_t, uint64_t>& drop_seed_offset)
+        template <typename T>
+        union ValueOrPointer
+        {
+            T val;
+            const T* ptr;
+        };
+
+        ValueOrPointer<uint64_t> drop_seed;
+        ValueOrPointer<uint64_t> drop_offset;
+        bool is_drop_seed_offset_from_host;
+    };
+
+    struct FmhaFwdCommonDropoutKargs : FmhaFwdDropoutSeedOffset
+    {
+        void init_dropout(float p_drop, uint64_t seed, uint64_t offset)
         {
             float p_undrop = 1.0 - p_drop;
             p_undrop_in_uint8_t =
                 uint8_t(std::floor(p_undrop * std::numeric_limits<uint8_t>::max()));
             rp_undrop = 1.0 / p_undrop;
 
-            drop_seed   = std::get<0>(drop_seed_offset);
-            drop_offset = std::get<1>(drop_seed_offset);
+            this->drop_seed.val                 = seed;
+            this->drop_offset.val               = offset;
+            this->is_drop_seed_offset_from_host = true;
         }
+
+        void init_dropout(float p_drop, const uint64_t* seed_ptr, const uint64_t* offset_ptr)
+        {
+            float p_undrop = 1.0 - p_drop;
+            p_undrop_in_uint8_t =
+                uint8_t(std::floor(p_undrop * std::numeric_limits<uint8_t>::max()));
+            rp_undrop = 1.0 / p_undrop;
+
+            this->drop_seed.ptr                 = seed_ptr;
+            this->drop_offset.ptr               = offset_ptr;
+            this->is_drop_seed_offset_from_host = false;
+        }
+
         float rp_undrop             = 1;
         uint8_t p_undrop_in_uint8_t = std::numeric_limits<uint8_t>::max();
         bool is_store_randval       = false;
-        uint64_t drop_seed          = 1;
-        uint64_t drop_offset        = 0;
         void* rand_val_ptr          = nullptr;
 
         ck_tile::index_t stride_randval       = 0;
         ck_tile::index_t nhead_stride_randval = 0;
     };
+
     struct FmhaFwdBatchModeDropoutKargs : FmhaFwdCommonDropoutKargs
     {
         ck_tile::index_t batch_stride_randval = 0;
@@ -278,7 +307,8 @@ struct FmhaFwdKernel
               ck_tile::index_t mask_type,
               float p_drop,
               bool s_randval,
-              const std::tuple<uint64_t, uint64_t>& drop_seed_offset)
+              std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+                  drop_seed_offset)
     {
         Kargs kargs{{q_ptr,
                      k_ptr,
@@ -344,7 +374,19 @@ struct FmhaFwdKernel
         }
         if constexpr(kHasDropout)
         {
-            kargs.init_dropout(p_drop, drop_seed_offset);
+            if(drop_seed_offset.index() == 0) // seed & offset come from host
+            {
+                const auto& [seed, offset] = std::get<0>(drop_seed_offset);
+                kargs.init_dropout(p_drop, seed, offset);
+            }
+            else // seed & offset come from device
+            {
+                const auto& [seed_ptr, offset_ptr] = std::get<1>(drop_seed_offset);
+                kargs.init_dropout(p_drop,
+                                   reinterpret_cast<const uint64_t*>(seed_ptr),
+                                   reinterpret_cast<const uint64_t*>(offset_ptr));
+            }
+
             kargs.rand_val_ptr         = rand_val_ptr;
             kargs.stride_randval       = stride_randval;
             kargs.nhead_stride_randval = nhead_stride_randval;
@@ -392,7 +434,8 @@ struct FmhaFwdKernel
               ck_tile::index_t mask_type,
               float p_drop,
               bool s_randval,
-              const std::tuple<uint64_t, uint64_t>& drop_seed_offset)
+              std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+                  drop_seed_offset)
     {
         Kargs kargs{{q_ptr,
                      k_ptr,
@@ -455,7 +498,19 @@ struct FmhaFwdKernel
         }
         if constexpr(kHasDropout)
         {
-            kargs.init_dropout(p_drop, drop_seed_offset);
+            if(drop_seed_offset.index() == 0) // seed & offset come from host
+            {
+                const auto& [seed, offset] = std::get<0>(drop_seed_offset);
+                kargs.init_dropout(p_drop, seed, offset);
+            }
+            else // seed & offset come from device
+            {
+                const auto& [seed_ptr, offset_ptr] = std::get<1>(drop_seed_offset);
+                kargs.init_dropout(p_drop,
+                                   reinterpret_cast<const uint64_t*>(seed_ptr),
+                                   reinterpret_cast<const uint64_t*>(offset_ptr));
+            }
+
             kargs.rand_val_ptr         = rand_val_ptr;
             kargs.stride_randval       = stride_randval;
             kargs.nhead_stride_randval = nhead_stride_randval;
@@ -748,8 +803,10 @@ struct FmhaFwdKernel
                 return BlockDropout{i_batch_,
                                     i_nhead_,
                                     kargs.num_head_q,
-                                    kargs.drop_seed,
-                                    kargs.drop_offset,
+                                    kargs.is_drop_seed_offset_from_host ? kargs.drop_seed.val
+                                                                        : *kargs.drop_seed.ptr,
+                                    kargs.is_drop_seed_offset_from_host ? kargs.drop_offset.val
+                                                                        : *kargs.drop_offset.ptr,
                                     kargs.rp_undrop,
                                     kargs.p_undrop_in_uint8_t,
                                     kargs.is_store_randval};

include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_kernel.hpp

@@ -78,8 +78,6 @@ struct FmhaFwdSplitKVCombineKernel
         void* o_ptr;
 
         ck_tile::index_t batch;
-        ck_tile::index_t max_seqlen_q;
-
         ck_tile::index_t seqlen_q;
         ck_tile::index_t hdim_v;
         ck_tile::index_t num_splits;
@@ -91,8 +89,6 @@ struct FmhaFwdSplitKVCombineKernel
         ck_tile::index_t nhead_stride_o_acc;
         ck_tile::index_t nhead_stride_o;
 
-        ck_tile::index_t batch_stride_o_acc;
-
         ck_tile::index_t split_stride_lse_acc;
         ck_tile::index_t split_stride_o_acc;
     };
@@ -114,8 +110,9 @@ struct FmhaFwdSplitKVCombineKernel
           std::conditional_t<kStoreLSE, CommonLSEKargs, EmptyKargs<0>>,
           std::conditional_t<kDoFp8StaticQuant, Fp8StaticQuantKargs, EmptyKargs<1>>
     {
-        ck_tile::index_t batch_stride_o;
         ck_tile::index_t batch_stride_lse_acc;
+        ck_tile::index_t batch_stride_o_acc;
+        ck_tile::index_t batch_stride_o;
     };
 
     struct GroupModeKargs
@@ -135,7 +132,6 @@ struct FmhaFwdSplitKVCombineKernel
               void* lse_ptr,
               void* o_ptr,
               ck_tile::index_t batch,
-              ck_tile::index_t max_seqlen_q,
               ck_tile::index_t seqlen_q,
               ck_tile::index_t hdim_v,
               ck_tile::index_t num_splits,
@@ -157,7 +153,6 @@ struct FmhaFwdSplitKVCombineKernel
                      o_acc_ptr,
                      o_ptr,
                      batch,
-                     max_seqlen_q,
                      seqlen_q,
                      hdim_v,
                      num_splits,
@@ -166,13 +161,13 @@ struct FmhaFwdSplitKVCombineKernel
                      nhead_stride_lse_acc,
                      nhead_stride_o_acc,
                      nhead_stride_o,
-                     batch_stride_o_acc,
                      split_stride_lse_acc,
                      split_stride_o_acc}, // args for common karg
                     {},                   // placeholder for lse
                     {},                   // placeholder for fp8_static_quant args
-                    batch_stride_o,
-                    batch_stride_lse_acc};
+                    batch_stride_lse_acc,
+                    batch_stride_o_acc,
+                    batch_stride_o};
 
         if constexpr(kStoreLSE)
         {
@@ -195,7 +190,6 @@ struct FmhaFwdSplitKVCombineKernel
               void* lse_ptr,
               void* o_ptr,
               ck_tile::index_t batch,
-              ck_tile::index_t max_seqlen_q,
               const void* seqstart_q_ptr,
               ck_tile::index_t hdim_v,
               ck_tile::index_t num_splits,
@@ -206,7 +200,6 @@ struct FmhaFwdSplitKVCombineKernel
               ck_tile::index_t nhead_stride_o_acc,
               ck_tile::index_t nhead_stride_lse,
               ck_tile::index_t nhead_stride_o,
-              ck_tile::index_t batch_stride_o_acc,
               ck_tile::index_t split_stride_lse_acc,
               ck_tile::index_t split_stride_o_acc)
     {
@@ -214,7 +207,6 @@ struct FmhaFwdSplitKVCombineKernel
                      o_acc_ptr,
                      o_ptr,
                      batch,
-                     max_seqlen_q,
                      -1, // seqlen will be updated by another pointer
                      hdim_v,
                      num_splits,
@@ -223,7 +215,6 @@ struct FmhaFwdSplitKVCombineKernel
                      nhead_stride_lse_acc,
                      nhead_stride_o_acc,
                      nhead_stride_o,
-                     batch_stride_o_acc,
                      split_stride_lse_acc,
                      split_stride_o_acc}, // args for common karg
                     {},                   // placeholder for lse
@@ -243,12 +234,12 @@ struct FmhaFwdSplitKVCombineKernel
         return kargs;
     }
 
-    __host__ static constexpr auto GridSize(ck_tile::index_t batch_size_,
-                                            ck_tile::index_t nhead_,
-                                            ck_tile::index_t seqlen_q_,
-                                            ck_tile::index_t hdim_v_)
+    __host__ static constexpr auto GridSize(ck_tile::index_t batch_size,
+                                            ck_tile::index_t nhead,
+                                            ck_tile::index_t max_seqlen_q,
+                                            ck_tile::index_t hdim_v)
     {
-        return TilePartitioner::GridSize(batch_size_, nhead_, seqlen_q_, hdim_v_);
+        return TilePartitioner::GridSize(batch_size, nhead, max_seqlen_q, hdim_v);
     }
 
     __host__ static constexpr auto BlockSize() { return dim3(kBlockSize); }
@@ -270,10 +261,8 @@ struct FmhaFwdSplitKVCombineKernel
         const index_t i_m0 = __builtin_amdgcn_readfirstlane(i_tile_m * FmhaPipeline::kM0);
         const index_t i_n1 = __builtin_amdgcn_readfirstlane(i_tile_n * FmhaPipeline::kN1);
 
-        const long_index_t batch_offset_o_acc =
-            static_cast<long_index_t>(i_batch) * kargs.batch_stride_o_acc;
-
         long_index_t batch_offset_lse_acc = 0;
+        long_index_t batch_offset_o_acc   = 0;
         long_index_t batch_offset_lse     = 0;
         long_index_t batch_offset_o       = 0;
 
@@ -282,14 +271,16 @@ struct FmhaFwdSplitKVCombineKernel
             // get starting offset for each batch
             const long_index_t query_start = kargs.seqstart_q_ptr[i_batch];
 
-            batch_offset_o       = query_start * kargs.row_stride_o;
             batch_offset_lse_acc = query_start;
+            batch_offset_o_acc   = query_start * kargs.row_stride_o_acc;
 
             if constexpr(kStoreLSE)
             {
                 batch_offset_lse = query_start;
             }
 
+            batch_offset_o = query_start * kargs.row_stride_o;
+
             // get real # queries & # keys under group mode
             const auto adjusted_seqstart_q_ptr = kargs.seqstart_q_ptr + i_batch;
             kargs.seqlen_q = adjusted_seqstart_q_ptr[1] - adjusted_seqstart_q_ptr[0];
@@ -303,13 +294,15 @@ struct FmhaFwdSplitKVCombineKernel
         }
         else
         {
-            batch_offset_o       = static_cast<long_index_t>(i_batch) * kargs.batch_stride_o;
             batch_offset_lse_acc = static_cast<long_index_t>(i_batch) * kargs.batch_stride_lse_acc;
+            batch_offset_o_acc   = static_cast<long_index_t>(i_batch) * kargs.batch_stride_o_acc;
 
             if constexpr(kStoreLSE)
             {
                 batch_offset_lse = static_cast<long_index_t>(i_batch) * kargs.batch_stride_lse;
             }
+
+            batch_offset_o = static_cast<long_index_t>(i_batch) * kargs.batch_stride_o;
         }
 
         // for simplicity, batch stride we just modify the pointer
@@ -341,7 +334,7 @@ struct FmhaFwdSplitKVCombineKernel
         auto o_acc_dram = [&]() {
             const auto o_acc_dram_naive = make_naive_tensor_view<address_space_enum::global>(
                 o_acc_ptr,
-                make_tuple(kargs.num_splits, kargs.max_seqlen_q, kargs.hdim_v),
+                make_tuple(kargs.num_splits, kargs.seqlen_q, kargs.hdim_v),
                 make_tuple(kargs.split_stride_o_acc, kargs.row_stride_o_acc, 1),
                 number<FmhaPipeline::kAlignmentOacc>{},
                 number<1>{});
@@ -351,14 +344,14 @@ struct FmhaFwdSplitKVCombineKernel
                 make_tuple(number<1>{}, number<FmhaPipeline::kM0>{}, number<FmhaPipeline::kN1>{}),
                 sequence<false, kPadSeqLenQ, kPadHeadDimV>{});
 
-            const index_t padded_max_seqlen_q =
+            const index_t padded_seqlen_q =
                 o_acc_dram_view.get_tensor_descriptor().get_lengths()[number<1>{}];
             const index_t padded_hdim_v =
                 o_acc_dram_view.get_tensor_descriptor().get_lengths()[number<2>{}];
 
             return transform_tensor_view(
                 o_acc_dram_view,
-                make_tuple(make_merge_transform(make_tuple(kargs.num_splits, padded_max_seqlen_q)),
+                make_tuple(make_merge_transform(make_tuple(kargs.num_splits, padded_seqlen_q)),
                            make_pass_through_transform(padded_hdim_v)),
                 make_tuple(sequence<0, 1>{}, sequence<2>{}),
                 make_tuple(sequence<0>{}, sequence<1>{}));
@@ -417,7 +410,7 @@ struct FmhaFwdSplitKVCombineKernel
                     identity{},                                          // lse_element_func
                     composes(saturates<fp8_t>{}, scales{kargs.scale_o}), // o_acc_element_func
                     kargs.num_splits,
-                    kargs.max_seqlen_q,
+                    kargs.seqlen_q,
                     smem_ptr);
             }
             else
@@ -426,7 +419,7 @@ struct FmhaFwdSplitKVCombineKernel
                                       o_acc_dram_window,
                                       lse_dram_window,
                                       kargs.num_splits,
-                                      kargs.max_seqlen_q,
+                                      kargs.seqlen_q,
                                       smem_ptr);
             }
         }();

include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_combine_tile_partitioner.hpp

@@ -13,21 +13,20 @@ struct FmhaFwdSplitKVCombineTilePartitioner
     static constexpr ck_tile::index_t kM0 = kM0_;
     static constexpr ck_tile::index_t kN1 = kN1_;
 
-    CK_TILE_HOST static constexpr auto GridSize(ck_tile::index_t batch_size_,
-                                                ck_tile::index_t nhead_,
-                                                ck_tile::index_t seqlen_q_,
-                                                ck_tile::index_t hdim_v_)
+    CK_TILE_HOST static constexpr auto GridSize(ck_tile::index_t batch_size,
+                                                ck_tile::index_t nhead,
+                                                ck_tile::index_t max_seqlen_q,
+                                                ck_tile::index_t hdim_v)
     {
         // TODO: this may need tuning
-        return dim3(ck_tile::integer_divide_ceil(seqlen_q_, kM0) *
-                        ck_tile::integer_divide_ceil(hdim_v_, kN1),
-                    nhead_,
-                    batch_size_);
+        return dim3(ck_tile::integer_divide_ceil(max_seqlen_q, kM0) *
+                        ck_tile::integer_divide_ceil(hdim_v, kN1),
+                    nhead,
+                    batch_size);
     }
 
     CK_TILE_DEVICE auto operator()(ck_tile::index_t /*seqlen_q*/, ck_tile::index_t hdim_v)
     {
-        // const index_t num_tile_m0 = seqlen_q / kM0;
         const index_t num_tile_n1 = ck_tile::integer_divide_ceil(hdim_v, kN1);
 
         const index_t i_block = blockIdx.x;

include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp

@@ -135,9 +135,6 @@ struct FmhaFwdSplitKVKernel
         ck_tile::index_t nhead_stride_lse_acc;
         ck_tile::index_t nhead_stride_o_acc;
 
-        ck_tile::index_t batch_stride_lse_acc;
-        ck_tile::index_t batch_stride_o_acc;
-
         ck_tile::index_t split_stride_lse_acc;
         ck_tile::index_t split_stride_o_acc;
     };
@@ -201,6 +198,8 @@ struct FmhaFwdSplitKVKernel
         ck_tile::index_t batch_stride_q;
         ck_tile::index_t batch_stride_k;
         ck_tile::index_t batch_stride_v;
+        ck_tile::index_t batch_stride_lse_acc;
+        ck_tile::index_t batch_stride_o_acc;
     };
 
     struct GroupModeKargs
@@ -217,8 +216,8 @@ struct FmhaFwdSplitKVKernel
         const int32_t* seqstart_k_ptr;
         const int32_t* seqlen_k_ptr;
 
-        ck_tile::index_t batch_stride_k;
-        ck_tile::index_t batch_stride_v;
+        ck_tile::index_t batch_stride_k; // only used for paged-kvcache
+        ck_tile::index_t batch_stride_v; // only used for paged-kvcache
     };
 
     using Kargs = std::conditional_t<kIsGroupMode, GroupModeKargs, BatchModeKargs>;
@@ -296,8 +295,6 @@ struct FmhaFwdSplitKVKernel
                      nhead_stride_v,
                      nhead_stride_lse_acc,
                      nhead_stride_o_acc,
-                     batch_stride_lse_acc,
-                     batch_stride_o_acc,
                      split_stride_lse_acc,
                      split_stride_o_acc}, // args for common karg
                     {},                   // placeholder for bias
@@ -307,7 +304,9 @@ struct FmhaFwdSplitKVKernel
                     reinterpret_cast<const int32_t*>(seqlen_k_ptr),
                     batch_stride_q,
                     batch_stride_k,
-                    batch_stride_v};
+                    batch_stride_v,
+                    batch_stride_lse_acc,
+                    batch_stride_o_acc};
 
         if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
         {
@@ -375,10 +374,8 @@ struct FmhaFwdSplitKVKernel
               ck_tile::index_t nhead_stride_bias,
               ck_tile::index_t nhead_stride_lse_acc,
               ck_tile::index_t nhead_stride_o_acc,
-              ck_tile::index_t batch_stride_k,
-              ck_tile::index_t batch_stride_v,
-              ck_tile::index_t batch_stride_lse_acc,
-              ck_tile::index_t batch_stride_o_acc,
+              ck_tile::index_t batch_stride_k, // only used for paged-kvcache
+              ck_tile::index_t batch_stride_v, // only used for paged-kvcache
               ck_tile::index_t split_stride_lse_acc,
               ck_tile::index_t split_stride_o_acc,
               ck_tile::index_t window_size_left,
@@ -412,8 +409,6 @@ struct FmhaFwdSplitKVKernel
                      nhead_stride_v,
                      nhead_stride_lse_acc,
                      nhead_stride_o_acc,
-                     batch_stride_lse_acc,
-                     batch_stride_o_acc,
                      split_stride_lse_acc,
                      split_stride_o_acc}, // args for common karg
                     {},                   // placeholder for bias
@@ -452,11 +447,11 @@ struct FmhaFwdSplitKVKernel
 
     __host__ static constexpr auto GridSize(ck_tile::index_t batch_size,
                                             ck_tile::index_t nhead,
-                                            ck_tile::index_t seqlen_q,
+                                            ck_tile::index_t max_seqlen_q,
                                             ck_tile::index_t hdim_v,
                                             ck_tile::index_t num_splits)
     {
-        return TilePartitioner::GridSize(batch_size, nhead, seqlen_q, hdim_v, num_splits);
+        return TilePartitioner::GridSize(batch_size, nhead, max_seqlen_q, hdim_v, num_splits);
     }
 
     __host__ static constexpr auto BlockSize() { return dim3(kBlockSize); }
@@ -483,8 +478,7 @@ struct FmhaFwdSplitKVKernel
         long_index_t batch_offset_v       = 0;
         long_index_t batch_offset_bias    = 0;
         long_index_t batch_offset_lse_acc = 0;
-        const long_index_t batch_offset_o_acc =
-            static_cast<long_index_t>(i_batch) * kargs.batch_stride_o_acc;
+        long_index_t batch_offset_o_acc   = 0;
 
         if constexpr(kIsGroupMode)
         {
@@ -492,9 +486,9 @@ struct FmhaFwdSplitKVKernel
             const long_index_t query_start = kargs.seqstart_q_ptr[i_batch];
             const long_index_t key_start   = kargs.seqstart_k_ptr[i_batch];
 
-            batch_offset_q       = query_start * kargs.stride_q;
-            batch_offset_k       = key_start * kargs.stride_k;
-            batch_offset_lse_acc = query_start;
+            batch_offset_q = query_start * kargs.stride_q;
+            batch_offset_k = key_start * kargs.stride_k;
+
             if constexpr(std::is_same_v<VLayout, ck_tile::tensor_layout::gemm::RowMajor>)
             {
                 batch_offset_v = key_start * kargs.stride_v;
@@ -508,6 +502,9 @@ struct FmhaFwdSplitKVKernel
                 batch_offset_bias = query_start * kargs.stride_bias + key_start;
             }
 
+            batch_offset_lse_acc = query_start;
+            batch_offset_o_acc   = query_start * kargs.stride_o_acc;
+
             // get real # queries & # keys under group mode
             kargs.seqlen_q = kargs.seqstart_q_ptr[i_batch + 1] - kargs.seqstart_q_ptr[i_batch];
 
@@ -545,6 +542,7 @@ struct FmhaFwdSplitKVKernel
             batch_offset_k       = static_cast<long_index_t>(i_cache_batch) * kargs.batch_stride_k;
             batch_offset_v       = static_cast<long_index_t>(i_cache_batch) * kargs.batch_stride_v;
             batch_offset_lse_acc = static_cast<long_index_t>(i_batch) * kargs.batch_stride_lse_acc;
+            batch_offset_o_acc   = static_cast<long_index_t>(i_batch) * kargs.batch_stride_o_acc;
 
             if constexpr(BiasEnum == BlockAttentionBiasEnum::ELEMENTWISE_BIAS)
             {
@@ -895,8 +893,8 @@ struct FmhaFwdSplitKVKernel
             const auto o_acc_dram_naive = make_naive_tensor_view<address_space_enum::global>(
                 o_acc_ptr,
                 make_tuple(kargs.seqlen_q, kargs.hdim_v),
-                make_tuple(kargs.hdim_v, 1),
-                number<FmhaPipeline::kAlignmentO>{},
+                make_tuple(kargs.stride_o_acc, 1),
+                number<1>{},
                 number<1>{});
 
             return pad_tensor_view(

include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_tile_partitioner.hpp

@@ -20,12 +20,12 @@ struct FmhaFwdSplitKVTilePartitioner
 
     __host__ static constexpr auto GridSize(ck_tile::index_t batch_size,
                                             ck_tile::index_t nhead,
-                                            ck_tile::index_t seqlen_q,
+                                            ck_tile::index_t max_seqlen_q,
                                             ck_tile::index_t hdim_v,
                                             ck_tile::index_t num_splits)
     {
         // TODO: this may need tuning
-        return dim3(ck_tile::integer_divide_ceil(seqlen_q, kM0) *
+        return dim3(ck_tile::integer_divide_ceil(max_seqlen_q, kM0) *
                         ck_tile::integer_divide_ceil(hdim_v, kN1),
                     nhead * num_splits,
                     batch_size);

include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr.hpp

@@ -178,13 +178,13 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
             k_lds_ptr, Policy::template MakeKLdsWriteBlockDescriptor<Problem>());
 
         auto k_lds_write_window =
-            make_tile_window(k_lds, make_tuple(number<kN0>{}, number<kK0>{}), {0, 0});
+            make_tile_window(k_lds, make_tuple(number<kN0>{}, number<kQKHeaddim>{}), {0, 0});
 
         auto k_lds_read_window =
             make_tile_window(k_lds_write_window.get_bottom_tensor_view(),
                              make_tuple(number<kN0>{}, number<kK0>{}),
                              k_lds_write_window.get_window_origin(),
-                             Policy::template MakeKRegSliceBlockDescriptor<Problem>());
+                             Policy::template MakeKRegBlockDescriptor<Problem>());
 
         auto k_reg_tensor = make_static_distributed_tensor<KDataType>(
             Policy::template MakeKRegBlockDescriptor<Problem>());
@@ -204,16 +204,13 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
             v_lds_ptr, Policy::template MakeVLdsWriteBlockDescriptor<Problem>());
 
         auto v_lds_write_window =
-            make_tile_window(v_lds, make_tuple(number<kN0>{}, number<kK2>{}), {0, 0});
+            make_tile_window(v_lds, make_tuple(number<kN0>{}, number<kVHeaddim>{}), {0, 0});
 
         auto v_lds_read_window =
             make_tile_window(v_lds_write_window.get_bottom_tensor_view(),
                              make_tuple(number<kN0>{}, number<kK2>{}),
                              v_lds_write_window.get_window_origin(),
-                             Policy::template MakeVRegSliceBlockDescriptor<Problem>());
-
-        auto v_reg_tensor = make_static_distributed_tensor<VDataType>(
-            Policy::template MakeVRegBlockDescriptor<Problem>());
+                             Policy::template MakeVRegBlockDescriptor<Problem>());
 
         //------------------------------------------------------------------
         // KT, Reg ->LDS ->Reg
@@ -227,7 +224,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
             kt_lds_ptr, Policy::template MakeShuffledKLdsWriteBlockDescriptor<Problem>());
 
         auto shuffled_k_lds_write_window = make_tile_window(
-            shuffled_k_lds_write, make_tuple(number<kN0>{}, number<kK0>{}), {0, 0});
+            shuffled_k_lds_write, make_tuple(number<kN0>{}, number<kQKHeaddim>{}), {0, 0});
 
         auto kt_lds_read = make_tensor_view<address_space_enum::lds>(
             kt_lds_ptr, Policy::template MakeKTLdsReadBlockDescriptor<Problem>());
@@ -257,7 +254,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
 
         block_sync_lds();
 
-        v_reg_tensor = load_tile(v_lds_read_window);
+        auto v_reg_tensor = load_tile(v_lds_read_window);
         block_sync_lds();
         //---------------------------- Loop Load in ----------------------------//
         // Q: HBM ->Reg ->LDS
@@ -276,7 +273,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
             q_lds_ptr, Policy::template MakeQLdsBlockDescriptor<Problem>());
 
         auto q_lds_window =
-            make_tile_window(q_lds, make_tuple(number<kM0>{}, number<kK0>{}), {0, 0});
+            make_tile_window(q_lds, make_tuple(number<kM0>{}, number<kQKHeaddim>{}), {0, 0});
 
         auto q_lds_read_window =
             make_tile_window(q_lds_window.get_bottom_tensor_view(),
@@ -297,7 +294,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
             qt_lds_ptr, Policy::template MakeShuffledQLdsWriteBlockDescriptor<Problem>());
 
         auto shuffled_q_lds_write_window = make_tile_window(
-            shuffled_q_lds_write, make_tuple(number<kM0>{}, number<kK0>{}), {0, 0});
+            shuffled_q_lds_write, make_tuple(number<kM0>{}, number<kQKHeaddim>{}), {0, 0});
 
         auto qt_lds_read = make_tensor_view<address_space_enum::lds>(
             qt_lds_ptr, Policy::template MakeQTLdsReadBlockDescriptor<Problem>());
@@ -322,7 +319,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
             do_lds_ptr, Policy::template MakeOGradLdsBlockDescriptor<Problem>());
 
         auto do_lds_window =
-            make_tile_window(do_lds, make_tuple(number<kM0>{}, number<kK2>{}), {0, 0});
+            make_tile_window(do_lds, make_tuple(number<kM0>{}, number<kVHeaddim>{}), {0, 0});
 
         auto do_lds_read_window =
             make_tile_window(do_lds_window.get_bottom_tensor_view(),
@@ -341,7 +338,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
             dot_lds_ptr, Policy::template MakeShuffledOGradLdsWriteBlockDescriptor<Problem>());
 
         auto shuffled_do_lds_write_window = make_tile_window(
-            shuffled_do_lds_write, make_tuple(number<kM0>{}, number<kK2>{}), {0, 0});
+            shuffled_do_lds_write, make_tuple(number<kM0>{}, number<kVHeaddim>{}), {0, 0});
 
         auto dot_read_lds = make_tensor_view<address_space_enum::lds>(
             dot_lds_ptr, Policy::template MakeOGradTLdsReadBlockDescriptor<Problem>());
@@ -483,9 +480,9 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVR
 
         index_t i_total_loops = 0;
         index_t seqlen_q_step = seqlen_q_start;
-        static_assert(kQKHeaddim == kK0, "kQKHeaddim should equal to kK0");
+        static_assert(kQKHeaddim >= kK0, "kQKHeaddim should be equal or greater than kK0");
         static_assert(kM0 == kK1, "kM0 should equal to kK1");
-        static_assert(kVHeaddim == kK2, "kVHeaddim should equal to kK2");
+        static_assert(kVHeaddim >= kK2, "kVHeaddim should be equal or greater than kK2");
         static_assert(kM0 == kK3, "kM0 should equal to kK3");
         constexpr index_t k4_loops = kN0 / kK4;
 

include/ck_tile/ops/fmha/pipeline/block_fmha_bwd_dq_dk_dv_pipeline_kr_ktr_vr_iglp.hpp

@@ -178,13 +178,13 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
             k_lds_ptr, Policy::template MakeKLdsWriteBlockDescriptor<Problem>());
 
         auto k_lds_write_window =
-            make_tile_window(k_lds, make_tuple(number<kN0>{}, number<kK0>{}), {0, 0});
+            make_tile_window(k_lds, make_tuple(number<kN0>{}, number<kQKHeaddim>{}), {0, 0});
 
         auto k_lds_read_window =
             make_tile_window(k_lds_write_window.get_bottom_tensor_view(),
                              make_tuple(number<kN0>{}, number<kK0>{}),
                              k_lds_write_window.get_window_origin(),
-                             Policy::template MakeKRegSliceBlockDescriptor<Problem>());
+                             Policy::template MakeKRegBlockDescriptor<Problem>());
 
         auto k_reg_tensor = make_static_distributed_tensor<KDataType>(
             Policy::template MakeKRegBlockDescriptor<Problem>());
@@ -204,16 +204,13 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
             v_lds_ptr, Policy::template MakeVLdsWriteBlockDescriptor<Problem>());
 
         auto v_lds_write_window =
-            make_tile_window(v_lds, make_tuple(number<kN0>{}, number<kK2>{}), {0, 0});
+            make_tile_window(v_lds, make_tuple(number<kN0>{}, number<kVHeaddim>{}), {0, 0});
 
         auto v_lds_read_window =
             make_tile_window(v_lds_write_window.get_bottom_tensor_view(),
                              make_tuple(number<kN0>{}, number<kK2>{}),
                              v_lds_write_window.get_window_origin(),
-                             Policy::template MakeVRegSliceBlockDescriptor<Problem>());
-
-        auto v_reg_tensor = make_static_distributed_tensor<VDataType>(
-            Policy::template MakeVRegBlockDescriptor<Problem>());
+                             Policy::template MakeVRegBlockDescriptor<Problem>());
 
         //------------------------------------------------------------------
         // KT, Reg ->LDS ->Reg
@@ -227,7 +224,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
             kt_lds_ptr, Policy::template MakeShuffledKLdsWriteBlockDescriptor<Problem>());
 
         auto shuffled_k_lds_write_window = make_tile_window(
-            shuffled_k_lds_write, make_tuple(number<kN0>{}, number<kK0>{}), {0, 0});
+            shuffled_k_lds_write, make_tuple(number<kN0>{}, number<kQKHeaddim>{}), {0, 0});
 
         auto kt_lds_read = make_tensor_view<address_space_enum::lds>(
             kt_lds_ptr, Policy::template MakeKTLdsReadBlockDescriptor<Problem>());
@@ -257,7 +254,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
 
         block_sync_lds();
 
-        v_reg_tensor = load_tile(v_lds_read_window);
+        auto v_reg_tensor = load_tile(v_lds_read_window);
         //---------------------------- Loop Load in ----------------------------//
         // Q: HBM ->Reg ->LDS
         auto q_dram_window =
@@ -275,7 +272,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
             q_lds_ptr, Policy::template MakeQLdsBlockDescriptor<Problem>());
 
         auto q_lds_window =
-            make_tile_window(q_lds, make_tuple(number<kM0>{}, number<kK0>{}), {0, 0});
+            make_tile_window(q_lds, make_tuple(number<kM0>{}, number<kQKHeaddim>{}), {0, 0});
 
         auto q_lds_read_window =
             make_tile_window(q_lds_window.get_bottom_tensor_view(),
@@ -296,7 +293,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
             qt_lds_ptr, Policy::template MakeShuffledQLdsWriteBlockDescriptor<Problem>());
 
         auto shuffled_q_lds_write_window = make_tile_window(
-            shuffled_q_lds_write, make_tuple(number<kM0>{}, number<kK0>{}), {0, 0});
+            shuffled_q_lds_write, make_tuple(number<kM0>{}, number<kQKHeaddim>{}), {0, 0});
 
         auto qt_lds_read = make_tensor_view<address_space_enum::lds>(
             qt_lds_ptr, Policy::template MakeQTLdsReadBlockDescriptor<Problem>());
@@ -321,7 +318,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
             do_lds_ptr, Policy::template MakeOGradLdsBlockDescriptor<Problem>());
 
         auto do_lds_window =
-            make_tile_window(do_lds, make_tuple(number<kM0>{}, number<kK2>{}), {0, 0});
+            make_tile_window(do_lds, make_tuple(number<kM0>{}, number<kVHeaddim>{}), {0, 0});
 
         auto do_lds_read_window =
             make_tile_window(do_lds_window.get_bottom_tensor_view(),
@@ -340,7 +337,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
             dot_lds_ptr, Policy::template MakeShuffledOGradLdsWriteBlockDescriptor<Problem>());
 
         auto shuffled_do_lds_write_window = make_tile_window(
-            shuffled_do_lds_write, make_tuple(number<kM0>{}, number<kK2>{}), {0, 0});
+            shuffled_do_lds_write, make_tuple(number<kM0>{}, number<kVHeaddim>{}), {0, 0});
 
         auto dot_read_lds = make_tensor_view<address_space_enum::lds>(
             dot_lds_ptr, Policy::template MakeOGradTLdsReadBlockDescriptor<Problem>());
@@ -482,9 +479,9 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
 
         index_t i_total_loops = 0;
         index_t seqlen_q_step = seqlen_q_start;
-        static_assert(kQKHeaddim == kK0, "kQKHeaddim should equal to kK0");
+        static_assert(kQKHeaddim >= kK0, "kQKHeaddim should be equal or greater than kK0");
         static_assert(kM0 == kK1, "kM0 should equal to kK1");
-        static_assert(kVHeaddim == kK2, "kVHeaddim should equal to kK2");
+        static_assert(kVHeaddim >= kK2, "kVHeaddim should be equal or greater than kK2");
         static_assert(kM0 == kK3, "kM0 should equal to kK3");
         constexpr index_t k4_loops = kN0 / kK4;
 
@@ -827,6 +824,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
                 },
                 s_acc,
                 bias_s_tile);
+            __builtin_amdgcn_sched_barrier(0);
         }
         else if constexpr(BiasEnum == BlockAttentionBiasEnum::ALIBI)
         {
@@ -918,6 +916,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
         gemm_1(dv_acc, pt_reg_tensor, dot_reg_tensor);
 
         HotLoopScheduler::template GemmStagedScheduler<1>();
+        __builtin_amdgcn_sched_barrier(0);
 
         // STAGE 4, OGrad@V Gemm2
         auto dp_acc = SPGradBlockTileType{};
@@ -927,6 +926,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
         dp_acc = gemm_2(do_reg_tensor, v_reg_tensor);
 
         HotLoopScheduler::template GemmStagedScheduler<2>();
+        __builtin_amdgcn_sched_barrier(0);
 
         // STAGE 5, P^T(PGrad^T - D)
         auto ds                 = SPGradBlockTileType{};
@@ -965,6 +965,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
                 Policy::template MakeBiasTileDistribution<Problem>());
             shuffle_tile(dbias_tile, shuffled_dbias_tile);
             store_tile(dbias_dram_window, dbias_tile);
+            __builtin_amdgcn_sched_barrier(0);
         }
 
         // STAGE 6, SGrad^T@Q^T Gemm3
@@ -984,6 +985,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
         move_tile_window(ds_lds_read_window, {0, kK4});
 
         HotLoopScheduler::template GemmStagedScheduler<3>();
+        __builtin_amdgcn_sched_barrier(0);
         // STAGE 7, SGrad@K^T Gemm4
         auto dq_acc = QGradBlockTileType{};
         clear_tile(dq_acc);
@@ -1005,6 +1007,7 @@ struct BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP
         });
 
         HotLoopScheduler::template GemmStagedScheduler<4>();
+        __builtin_amdgcn_sched_barrier(0);
 
         // Results Scale
         if constexpr(FmhaDropout::IsDropout)