Reorganize project folders (#6)

example/ck_tile/01_fmha/CMakeLists.txt

@@ -0,0 +1,123 @@
+# validate user-specified fmha_fwd API list
+set(FMHA_FWD_KNOWN_APIS "fwd;fwd_splitkv;fwd_appendkv")
+set(FMHA_FWD_ENABLE_APIS "fwd" CACHE STRING
+    "semicolon-separated list of APIs to generate (${FMHA_FWD_KNOWN_APIS}) & link, or \"all\".")
+if(FMHA_FWD_ENABLE_APIS STREQUAL "all")
+  set(FMHA_FWD_ENABLE_APIS ${FMHA_FWD_KNOWN_APIS})
+endif()
+
+foreach(api ${FMHA_FWD_ENABLE_APIS})
+  if(NOT "${api}" IN_LIST FMHA_FWD_KNOWN_APIS)
+    message(FATAL_ERROR "${api} isn't a known api: ${FMHA_FWD_KNOWN_APIS}.")
+  endif()
+endforeach()
+
+# "fwd" is a must-have api for the fmha_fwd example, add it if not specified
+if(NOT "fwd" IN_LIST FMHA_FWD_ENABLE_APIS)
+  list(APPEND FMHA_FWD_ENABLE_APIS "fwd")
+endif()
+
+string(REPLACE ";" "," FMHA_FWD_APIS "${FMHA_FWD_ENABLE_APIS}")
+# generate a list of kernels, but not actually emit files at config sta
+execute_process(
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${FMHA_FWD_APIS} --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/fwd_blob_list.txt
+  RESULT_VARIABLE ret
+)
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "CK Tile FMHA FAILED to genrate a list of FWD kernels via Python.")
+endif()
+
+execute_process(
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api bwd --list_blobs ${CMAKE_CURRENT_BINARY_DIR}/bwd_blob_list.txt --receipt 3
+  RESULT_VARIABLE ret
+)
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "CK Tile FMHA FAILED to genrate a list of BWD kernels via Python.")
+endif()
+
+# NOTE: for cmake, the FMHA_FWD_GEN_BLOBS/FMHA_BWD_GEN_BLOBS files must be in the same directory
+#       as current cmake list, otherwise will not figure out the dependency properly
+file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/fwd_blob_list.txt FMHA_FWD_GEN_BLOBS)
+file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/bwd_blob_list.txt FMHA_BWD_GEN_BLOBS)
+
+add_custom_command(
+  OUTPUT ${FMHA_FWD_GEN_BLOBS}
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${FMHA_FWD_APIS} --output_dir ${CMAKE_CURRENT_BINARY_DIR}
+)
+
+add_custom_command(
+  OUTPUT ${FMHA_BWD_GEN_BLOBS}
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api bwd --output_dir ${CMAKE_CURRENT_BINARY_DIR} --receipt 3
+)
+
+set(EXAMPLE_FMHA_FWD "tile_example_fmha_fwd")
+# 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"
+message("adding example ${EXAMPLE_FMHA_FWD}")
+add_executable(${EXAMPLE_FMHA_FWD} fmha_fwd.cpp)
+rocm_install(TARGETS ${EXAMPLE_FMHA_FWD} COMPONENT examples)
+target_include_directories(${EXAMPLE_FMHA_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+target_sources(${EXAMPLE_FMHA_FWD} PRIVATE ${FMHA_FWD_GEN_BLOBS})
+
+set(EXAMPLE_FMHA_BWD "tile_example_fmha_bwd")
+# 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"
+message("adding example ${EXAMPLE_FMHA_BWD}")
+add_executable(${EXAMPLE_FMHA_BWD} fmha_bwd.cpp)
+rocm_install(TARGETS ${EXAMPLE_FMHA_BWD} COMPONENT examples)
+target_include_directories(${EXAMPLE_FMHA_BWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+target_sources(${EXAMPLE_FMHA_BWD} PRIVATE ${FMHA_BWD_GEN_BLOBS})
+
+# NOTE: this is dangerous since will change the whole kernel to flush denormals
+#       WIP with compiler team for an exp2 intrinsic..., then remove this
+if(NOT DEFINED FMHA_FWD_FAST_EXP2)
+    set(FMHA_FWD_FAST_EXP2 true)
+endif()
+
+set(EXAMPLE_FMHA_FWD_COMPILE_OPTIONS)
+set(EXAMPLE_FMHA_BWD_COMPILE_OPTIONS)
+
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+#       ... because they are auto-generated
+if(FMHA_FWD_FAST_EXP2)
+	list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=1 -fgpu-flush-denormals-to-zero)
+else()
+	list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -DCK_TILE_FMHA_FWD_FAST_EXP2=0)
+endif()
+list(APPEND EXAMPLE_FMHA_BWD_COMPILE_OPTIONS -Wno-undefined-func-template -fgpu-flush-denormals-to-zero)
+
+# conditionally enable call to the fwd_splitkv API in fmha_fwd example
+if("fwd_splitkv" IN_LIST FMHA_FWD_ENABLE_APIS)
+  list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_SPLITKV_API=1)
+else()
+  list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_SPLITKV_API=0)
+endif()
+
+# conditionally enable call to the fwd_appendkv API in fmha_fwd example
+if("fwd_appendkv" IN_LIST FMHA_FWD_ENABLE_APIS)
+  list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_APPENDKV_API=1)
+else()
+  list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -DCK_TILE_FMHA_FWD_APPENDKV_API=0)
+endif()
+
+# conditionally specify the use of OCP_FP8
+if(CK_USE_OCP_FP8)
+  list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
+endif()
+
+# Allow comparing floating points directly in order to check sentinel values
+list(APPEND EXAMPLE_FMHA_FWD_COMPILE_OPTIONS -Wno-float-equal)
+list(APPEND EXAMPLE_FMHA_BWD_COMPILE_OPTIONS -Wno-float-equal)
+
+target_compile_options(${EXAMPLE_FMHA_FWD} PRIVATE ${EXAMPLE_FMHA_FWD_COMPILE_OPTIONS})
+target_compile_options(${EXAMPLE_FMHA_BWD} PRIVATE ${EXAMPLE_FMHA_BWD_COMPILE_OPTIONS})
+
+# TODO: we have to turn off this global prop, otherwise the progress bar generated
+# by cmake will print too many files, execvp: /bin/sh: Argument list too long
+# however, this property may affect global
+# TODO: consider codegen a makefile by us
+set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)

example/ck_tile/01_fmha/README.md

@@ -0,0 +1,131 @@
+# fused multi-head attention
+
+This folder contains example for fmha(fused multi-head attention) using ck_tile tile-programming implementation. It is a good example to demonstrate the usage of tile-programming API, as well as illustrate the new approach to construct a kernel template and instantiate it(them) while keeping compile time fast.
+
+## 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_fmha_fwd -j
+```
+This will result in an executable `build/bin/tile_example_fmha_fwd`
+
+## kernel
+The kernel template is `fmha_fwd_kernel.hpp`, this is the grid-wise op in old ck_tile's terminology. We put it here purposely, to demonstrate one can construct a kernel by using various internal component from ck_tile. We may still have an implementation under ck_tile's include path (in the future) for the kernel template.
+
+There are 2 template parameters for this kernel template.
+* `FmhaPipeline` is one of the block_tile_pipeline(under `include/ck_tile/tile_program/block_tile_pipeline`) which is a performance critical component. Indeed, we did a lot of optimization and trials to optimize the pipeline and may still workout more performance pipeline and update into that folder. People only need to replace this pipeline type and would be able to enjoy the benefit of different performant implementations (stay tuned for updated pipeline(s)).
+* `EpiloguePipeline` will modify and store out the result in the last phase. People usually will do lot of post-fusion at this stage, so we also abstract this concept. Currently we didn't do much thing at the epilogue stage but leave the room for future possible support.
+
+## codegen
+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 the arguments. Below is an example of the output (may subject to change)
+```
+args:
+          -v    weather do CPU validation or not (default:1)
+       -mode    kernel mode. 0:batch, 1:group (default:0)
+          -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
+          -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 (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
+    -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
+      -iperm    permute input (default:1)
+                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
+       -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)
+    -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
+       -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 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.
+
+### hdim
+Currently we support `32/64/128/256` hdim for `fp16`/`bf16`, within which `64`/`128` is better optimized. hdim should be multiple of 8, while seqlen_s can be arbitrary. For hdim be arbitrary number, it can be support through padding kernel of `qr` pipeline (we didn't generate this in generate.py by default)
+
+### group/batch mode
+Currently we support both `batch mode` and `group mode` (or `varlen`, in FA's term), by setting `-mode` = `0` or `1`. In `group mode` different kind of attention mask is also supported(see below)
+
+### MQA/GQA
+By setting `-h`(nhead for q) and `-h_k`(nhead for k/v) with different number, you can achieve MQA/GQA. Please pay attention that `h % h_K == 0` when you set different numbers.
+
+### input/output permute, and `b*s*3*h*d`
+If you look at the kernel argument inside `fmha_fwd_kernel.hpp`, we support providing arbitrary stride for seqlen(stride_q/k/v), nhead, batch of q/k/v matrix, hence it is very flexible to support `b*h*s*d` or `b*s*h*d` input/output permute. The `-iperm=0/1`, `-operm=0/1` is a convenient way to achieve this through the executable. We didn't provide a command-line arg to test `b*s*3*h*d` layout which is by default used by torch/FA, but it's trivial to achieve this if one set the proper `stride_q/k/v` value as `3*h*d`.
+
+### attention bias
+Attention bias is supported with the layout of `1*1*s*s`(similiar to input/output, different layout can be supported by changing the stride value for bias, or even extend to `b*h*s*s`) and bias value in float number.
+
+### alibi
+alibi is supported
+
+### lse
+For training kernels, "log sum exp" need to store out in forward and used in backward. We support this by setting `-lse=1`
+
+### vlayout
+We support v matrix in both row-major(`seqlen*hdim`) and col-major(`hdim*seqlen`). Since the accumulate(reduce) dimension for V is along `seqlen`, for current AMD's mfma layout which expect each thread to have contiguous register holding pixels along reduce dimension, it's easier to support col-major V layout. However, the performance of col-major is not necessarily faster than row-major, there are many factors that may affect the overall performance. We still provide the `-vlayout=r/c` here to switch/test between different layouts.
+
+### attention mask
+we support `causal mask` and `sliding window attention(swa)` mask in both batch and group mode, either from top-left or bottom-right.
+Underneath, we unify the mask expression into `generic attention mask coordinate`, providing an uniformed approach for each batch to locate the corresponding pixel need to be masked out.
+![](misc/gamc.png)
+
+Since FA/xformer style with window_size_left/right is more popular, we accept window_size as parameter and convert that internally to our generic coordinate(this coordinate can express more cases). Below shows some example of how to achieve different kind of mask through cmdline.
+
+| mask case|  cmdline    | FA style | xformer style |
+|----------|:-------------:|:-------------:|:-------------:|
+| no mask |  `-mask=0`(default) | | |
+| causal mask from top-left | `-mask=1` or `-mask=t` | `-mask=t:-1,0` | `-mask=xt:-1` |
+| causal mask from bottom-right | `-mask=2` or `-mask=b` | `-mask=b:-1,0` | `-mask=xb:-1` |
+| swa from top-left | | `-mask=t:3,5` | `-mask=xt:4` |
+| swa from bottom-right | |  `-mask=b:10,11` | `-mask=xb:16` |
+
+Note FA use bottom-right by default to express swa case, here we require you explicitly specify top-left/bottom-right.
+
+### dropout
+TBD
+
+## FP8 experimental support
+As described in [this blog](https://blog.hippoml.com/8bit-hippoattention-up-to-3x-faster-compared-to-flashattentionv2-8f9def90b482), we have an experimental support for fp8 fmha kernels, you can evaluate the performance by setting the arg `-prec=fp8` to the `tile_example_fmha_fwd`, on a gfx942 machine and ROCm 6.0+.
+
+Currently we only support `-vlayout=c`( `hdim*seqlen` for V matrix) and `-squant=1`(static quantization) with `hdim=128` for fp8 now. Full feature support will come later.

example/ck_tile/01_fmha/bias.hpp

@@ -0,0 +1,100 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <ostream>
+#include <string>
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/fmha.hpp"
+
+// keep sync with BlockAttentionBiasEnum
+enum class bias_enum
+{
+    no_bias          = 0,
+    elementwise_bias = 1,
+    alibi            = 2,
+};
+
+struct bias_info
+{
+    bias_enum type;
+    /*
+     * simple dispatch logic
+     *
+     * if type == elementwise_bias:
+     *      if rank_info == 0:
+     *           bias is 1*1*s*s
+     *      elif rank_info == 1:
+     *           bias is 1*h*s*s
+     *      elif rank_info == 2:
+     *           bias is b*h*s*s
+     *
+     * elif type == alibi:
+     *       if rank_info == 0:
+     *           alibi in 1*h
+     *       elif rank_info == 1:
+     *           alibi in b*h
+     */
+    int rank_info;
+
+    void serialize(std::ostream& os) const
+    {
+        if(type == bias_enum::no_bias)
+            os << "n";
+        else if(type == bias_enum::elementwise_bias)
+        {
+            os << "e";
+            if(rank_info != 0)
+            {
+                os << "[" << rank_info << "]";
+            }
+        }
+        else if(type == bias_enum::alibi)
+        {
+            os << "alibi";
+            if(rank_info != 0)
+            {
+                os << "[" << rank_info << "]";
+            }
+        }
+    }
+
+    static bias_info decode(std::string str)
+    {
+        bias_info info{bias_enum::no_bias, 0};
+        if(str == "0" || str == "n")
+        {
+            info.type = bias_enum::no_bias;
+        }
+        else if(str.compare(0, 1, "1") == 0 || str.compare(0, 1, "e") == 0 ||
+                str.compare(0, 11, "elementwise") == 0)
+        {
+            info.type    = bias_enum::elementwise_bias;
+            auto found_0 = str.find(':');
+            if(found_0 != std::string::npos)
+            {
+                std::string e  = str.substr(found_0 + 1);
+                info.rank_info = atoi(e.c_str());
+            }
+        }
+        else if(str.compare(0, 1, "2") == 0 || str.compare(0, 1, "a") == 0 ||
+                str.compare(0, 5, "alibi") == 0)
+        {
+            info.type    = bias_enum::alibi;
+            auto found_0 = str.find(':');
+            if(found_0 != std::string::npos)
+            {
+                std::string e  = str.substr(found_0 + 1);
+                info.rank_info = atoi(e.c_str());
+            }
+        }
+        return info;
+    }
+
+    friend std::ostream& operator<<(std::ostream& os, const bias_info& bi)
+    {
+        bi.serialize(os);
+        return os;
+    }
+};

example/ck_tile/01_fmha/codegen/cmake_config.py

@@ -0,0 +1,5 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+GEN_DIR = ""    # in Cmake, have to generate files in same folder

example/ck_tile/01_fmha/codegen/cpp_symbol_map.py

@@ -0,0 +1,128 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+FWD_DTYPE_MAP = {
+    "fp16"   : "FmhaFwdFp16",
+    "bf16"   : "FmhaFwdBf16",
+    "fp8"    : "FmhaFwdFp8",
+    "fp8fp16": "FmhaFwdFp8Fp16",
+    "fp8bf16": "FmhaFwdFp8Bf16"
+}
+
+BWD_DTYPE_MAP = {
+    "fp16": "FmhaBwdFp16",
+    "bf16": "FmhaBwdBf16"
+}
+
+MASK_IMPL = {
+    "generic" : "ck_tile::GenericAttentionMask",
+    "simplified"  : "ck_tile::SimplifiedGenericAttentionMask"
+}
+
+_MASK_SIMPLIFIED_MAP = {
+    "s_no" : "ck_tile::SimplifiedGenericAttentionMask<false>",
+    "s_mask" : "ck_tile::SimplifiedGenericAttentionMask<true>",
+}
+
+_MASK_MAP = {
+    "no" : "FmhaMasks::NoMask",
+    "causal" : "FmhaMasks::CausalMask",
+    "generic" : "FmhaMasks::GenericMask"
+}
+
+def get_mask_map(mask : str):
+    if mask == "generic":
+        return _MASK_MAP
+    elif mask == "simplified":
+        return _MASK_SIMPLIFIED_MAP
+    else:
+        assert False
+        return None
+
+_MASK_CHECK_MAP = {
+    "no" : "t.mask_type == mask_enum::no_mask",
+    "causal" : "t.mask_type == mask_enum::mask_top_left || t.mask_type == mask_enum::mask_bottom_right",
+    "generic" : "t.mask_type == mask_enum::window_generic",
+}
+
+_MASK_SIMPLIFIED_CHECK_MAP = {
+    "s_no" : "t.mask_type == mask_enum::no_mask",
+    "s_mask" : "t.mask_type != mask_enum::no_mask",
+}
+
+def get_mask_check_map(mask : str):
+    if mask == "generic":
+        return _MASK_CHECK_MAP
+    elif mask == "simplified":
+        return _MASK_SIMPLIFIED_CHECK_MAP
+    else:
+        assert False
+        return None
+
+BIAS_MAP = {
+    "no" : "ck_tile::BlockAttentionBiasEnum::NO_BIAS",
+    "bias"  : "ck_tile::BlockAttentionBiasEnum::ELEMENTWISE_BIAS",
+    "alibi" : "ck_tile::BlockAttentionBiasEnum::ALIBI"
+}
+
+# TODO: this is ugly
+BIAS_CHECK_MAP = {
+    "no" : "bias_enum::no_bias",
+    "bias"  : "bias_enum::elementwise_bias",
+    "alibi" : "bias_enum::alibi"
+}
+
+DROPOUT_MAP = {
+    "no"                        : "ck_tile::BlockDropoutBwd<false, true,  false>",
+    "dropout_wg32"              : "ck_tile::BlockDropoutBwd<true,  true,  false>",
+    "dropout_wg32_storerandval" : "ck_tile::BlockDropoutBwd<true,  true,  true >",
+    "dropout_wg16"              : "ck_tile::BlockDropoutBwd<true,  false, false>",
+    "dropout_wg16_storerandval" : "ck_tile::BlockDropoutBwd<true,  false, true >"
+}
+
+DROPOUT_CHECK_MAP = {
+    "no"                        : "t.has_dropout == false",
+    "dropout_wg32"              : "t.has_dropout == true && t.is_store_randval == false",
+    "dropout_wg32_storerandval" : "t.has_dropout == true && t.is_store_randval == true",
+    "dropout_wg16"              : "t.has_dropout == true && t.is_store_randval == false",
+    "dropout_wg16_storerandval" : "t.has_dropout == true && t.is_store_randval == true",
+}
+
+ROPE_MAP = {
+    "no" : "ck_tile::RotaryEmbeddingEnum::NONE",
+    "inter"  : "ck_tile::RotaryEmbeddingEnum::INTERLEAVED",
+    "half" : "ck_tile::RotaryEmbeddingEnum::HALF_ROTATED"
+}
+
+ROPE_CHECK_MAP = {
+    "no"    : "rope_enum::none",
+    "inter" : "rope_enum::interleaved",
+    "half"  : "rope_enum::half_rotated"
+}
+
+MODE_MAP = {
+    "batch" : "false",
+    "group" : "true"
+}
+
+LAYOUT_MAP = {
+    "row" : "true",
+    "col" : "false"
+}
+
+PIPELINE_MAP = {
+    "qr" : "ck_tile::BlockFmhaPipelineQRKSVS",
+    "qr_async" : "ck_tile::BlockFmhaPipelineQRKSVSAsync",
+}
+
+PIPELINE_ENUM_MAP = {
+    "qr" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS",
+    "qr_async" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS_ASYNC",
+    "qr_nwarp_sshuffle" : "ck_tile::BlockFmhaPipelineEnum::QRKSVS",
+}
+
+BOOL_MAP = {
+    "t" : "true",
+    "f" : "false"
+}

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

@@ -0,0 +1,902 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import copy
+from dataclasses import dataclass
+import fnmatch
+import itertools
+from pathlib import Path
+from typing import List, Optional, Tuple
+
+from codegen.cmake_config import *
+from codegen.cpp_symbol_map import *
+
+
+BWD_DQDKDV_PIPELINE_MAP = {
+    "kr_ktr_vr_iglp" : "ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVRIGLP",
+    "kr_ktr_vr"      : "ck_tile::BlockFmhaBwdDQDKDVPipelineKRKTRVR",
+}
+
+BWD_DQDKDV_PIPELINE_ENUM_MAP = {
+    "kr_ktr_vr_iglp" : "ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR_IGLP",
+    "kr_ktr_vr"      : "ck_tile::BlockFmhaBwdPipelineEnum::KRKTRVR",
+}
+
+FMHA_BWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.\n
+// auto generated by generate.py
+#include "fmha_bwd.hpp"
+"""
+
+FMHA_BWD_DQ_DK_DV_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+using fmha_block_tile_{F_idx} = ck_tile::
+    sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bk1}, {F_bk2}, {F_bk3}, {F_bk4}, {F_bhdq}, {F_bhdv}>;
+using fmha_block_warps0_{F_idx} = ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>;
+using fmha_block_warps1_{F_idx} = ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>;
+using fmha_block_warps2_{F_idx} = ck_tile::sequence<{F_rm2}, {F_rn2}, {F_rk2}>;
+using fmha_warp_tile0_{F_idx}   = ck_tile::sequence<{F_wm0}, {F_wn0}, {F_wk0}>;
+using fmha_warp_tile1_{F_idx}   = ck_tile::sequence<{F_wm1}, {F_wn1}, {F_wk1}>;
+
+// TODO: simplify Gemm0~4BlockWarps in TileFmhaBwdShape
+//       G0&G2 -> GSdP
+//       G1&G3 -> GdKV
+//       G4    -> GdQ
+using fmha_bwd_shape_{F_idx} = ck_tile::TileFmhaBwdShape<fmha_block_tile_{F_idx},
+                                                         fmha_block_warps0_{F_idx},
+                                                         fmha_warp_tile0_{F_idx},
+                                                         fmha_block_warps1_{F_idx},
+                                                         fmha_warp_tile1_{F_idx},
+                                                         fmha_block_warps0_{F_idx},
+                                                         fmha_warp_tile0_{F_idx},
+                                                         fmha_block_warps1_{F_idx},
+                                                         fmha_warp_tile1_{F_idx},
+                                                         fmha_block_warps2_{F_idx},
+                                                         fmha_warp_tile0_{F_idx}>;
+
+using fmha_bwd_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad},
+                                                       {F_skpad},
+                                                       {F_dpad},
+                                                       {F_dvpad},
+                                                       {F_bias},
+                                                       {F_dbias},
+                                                       false,
+                                                       false,
+                                                       false,
+                                                       {F_occupancy}>;
+using fmha_mask_{F_idx}      = {F_mask};
+using fmha_dropout_{F_idx}   = {F_dropout};
+
+using fmha_bwd_pipeline_problem_{F_idx} = ck_tile::BlockFmhaBwdPipelineProblem<
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::QDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::KDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::VDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::GemmDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::AccDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::DDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::BiasDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::RandValOutputDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::OGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::QGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::KGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::VGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::BiasGradDataType,
+    fmha_bwd_shape_{F_idx},
+    {F_mode},
+    {F_deterministic},
+    fmha_mask_{F_idx},
+    fmha_dropout_{F_idx},
+    fmha_bwd_trait_{F_idx}>;
+
+using fmha_bwd_pipeline_{F_idx} = {F_pipeline}<fmha_bwd_pipeline_problem_{F_idx}>;
+
+using fmha_bwd_dk_epilogue_{F_idx} = ck_tile::Default2DEpilogue<
+    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<{F_dtype}>::AccDataType,
+                                      typename FmhaBwdTypeConfig<{F_dtype}>::KGradDataType,
+                                      {F_skpad},
+                                      {F_dpad}>>;
+
+using fmha_bwd_dv_epilogue_{F_idx} = ck_tile::Default2DEpilogue<
+    ck_tile::Default2DEpilogueProblem<typename FmhaBwdTypeConfig<{F_dtype}>::AccDataType,
+                                      typename FmhaBwdTypeConfig<{F_dtype}>::VGradDataType,
+                                      {F_skpad},
+                                      {F_dvpad}>>;
+
+using fmha_bwd_dq_dk_dv_kernel_{F_idx} =
+    ck_tile::FmhaBwdDQDKDVKernel<fmha_bwd_pipeline_{F_idx},
+                                 fmha_bwd_dk_epilogue_{F_idx},
+                                 fmha_bwd_dv_epilogue_{F_idx}>;
+
+using dq_dk_dv_trait_{F_idx} = fmha_bwd_dq_dk_dv_traits_<{F_hdim},
+                                                         {F_dtype},
+                                                         {F_mode},
+                                                         {F_pipeline_enum},
+                                                         fmha_mask_{F_idx},
+                                                         fmha_dropout_{F_idx},
+                                                         {F_bias},
+                                                         {F_dbias},
+                                                         {F_spad},
+                                                         {F_skpad},
+                                                         {F_dpad},
+                                                         {F_dvpad},
+                                                         {F_deterministic}>;
+
+#include <iostream>
+
+template <>
+float fmha_bwd_dq_dk_dv_<dq_dk_dv_trait_{F_idx}>(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    using k_ = fmha_bwd_dq_dk_dv_kernel_{F_idx};
+    if(s.log_level_ > 0)
+        std::cout << ", " << k_::GetName() << std::flush;
+    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks                  = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    return ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
+}}
+
+template <>
+void fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_{F_idx}>(const ck_tile::stream_config& s,
+                                                        fmha_bwd_args a)
+{{
+    using k_                               = fmha_bwd_dq_dk_dv_kernel_{F_idx};
+    auto [kargs, grids]                    = fmha_bwd_dq_dk_dv_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks                  = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(
+        ck_tile::stream_config{{s.stream_id_}});
+}}
+
+template <>
+std::string fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_{F_idx}>()
+{{
+    using k_ = fmha_bwd_dq_dk_dv_kernel_{F_idx};
+    return k_::GetName();
+}}
+"""
+
+FMHA_BWD_API_FILENAME="fmha_bwd_api.cpp"
+FMHA_BWD_API="""
+#include <iostream>
+
+template <typename dot_do_o_trait_, typename dq_dk_dv_trait_, typename convert_dq_trait_>
+float fmha_bwd_(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    if(s.log_level_ > 0)
+        std::cout << ", " << fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_>() << ", " << fmha_bwd_dq_dk_dv_get_name_<dq_dk_dv_trait_>() << ", " << fmha_bwd_convert_dq_get_name_<convert_dq_trait_>() << std::flush;
+    return ck_tile::launch_kernel(s,
+        [=](const ck_tile::stream_config& s_){{ fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_>(s_, a); }},
+        [=](const ck_tile::stream_config& s_){{ fmha_bwd_dq_dk_dv_oneshot_<dq_dk_dv_trait_>(s_, a); }},
+        [=](const ck_tile::stream_config& s_){{ fmha_bwd_convert_dq_oneshot_<convert_dq_trait_>(s_, a); }}
+    );
+}}
+
+template <>
+float fmha_bwd<2>(fmha_bwd_traits t, fmha_bwd_args a, const ck_tile::stream_config& s){{
+    float r = -1;
+{F_dispatch}
+    return r;
+}}
+"""
+
+FMHA_BWD_API_PER_DTYPE="""    {F_if}(t.data_type.compare(\"{F_dtype}\") == 0){{
+{F_hdim_case}
+    }}
+"""
+FMHA_BWD_API_PER_HDIM_CASE="""        {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v <= {F_hdim}) {{
+{F_inner_dispatch}
+        }}
+"""
+
+FMHA_BWD_API_INNER_DISPATCH="""            {F_if}((t.is_group_mode == {F_mode}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_dbias == {F_dbias}) && ({F_dropout_check}) &&
+                        ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck}) && (t.is_deterministic == {F_deterministic})) {{
+                using dot_do_o_trait_ = fmha_bwd_dot_do_o_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_spad1}, {F_dvpad}>;
+                using dq_dk_dv_trait_ = fmha_bwd_dq_dk_dv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_pipeline_enum}, {F_mask}, {F_dropout}, {F_bias}, {F_dbias}, {F_spad0}, {F_skpad}, {F_dpad}, {F_dvpad}, {F_deterministic}>;
+                using convert_dq_trait_ = fmha_bwd_convert_dq_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_spad1}, {F_dpad}, {F_deterministic}>;
+                r = fmha_bwd_<dot_do_o_trait_, dq_dk_dv_trait_, convert_dq_trait_>(s, a);
+                return r;
+            }}
+"""
+
+@dataclass
+class FmhaBwdDQDKDVApiTrait:
+    pipeline      : str
+    # sync with fmha_bwd_traits<>, to generate fallback calls
+    hdim          : str
+    dtype         : str  # data type
+    mode          : str  # value from MODE_MAP
+    bm0           : int  # tile size along q seqlen (block size)
+    bn0           : int  # tile size along k seqlen
+    bhdq          : int  # q head_dim
+    bhdv          : int  # v head_dim
+    mask          : str
+    bias          : str
+    dbias         : str
+    dropout       : str
+    spad          : str
+    skpad         : str
+    dpad          : str
+    dvpad         : str
+    deterministic : str
+
+    def scheck(self, spad1 : str) -> str:
+        if self.mode == 'group':
+            return 'true' # always support
+        elif self.spad == 't' and spad1 == 't':
+            return f'a.seqlen_q % {self.bm0} != 0'
+        elif self.spad == 'f' and spad1 == 't':
+            return f'a.seqlen_q % {self.bm0} == 0 and a.seqlen_q % 64 != 0'
+        else: # self.skpad == 'f' and skpad1 == 'f'
+            return f'a.seqlen_q % 64 == 0'
+
+    @property
+    def skcheck(self) -> str:
+        if self.mode == 'group':
+            return 'true' # always support
+        elif self.skpad == 't':
+            return f'a.seqlen_k % {self.bn0} != 0'
+        else:
+            return f'a.seqlen_k % {self.bn0} == 0'
+
+    @property
+    def dcheck(self) -> str:
+        if self.dpad == 't': return f'a.hdim_q % {self.bhdq} != 0'
+        else :               return f'a.hdim_q % {self.bhdq} == 0'
+
+    @property
+    def dvcheck(self) -> str:
+        if self.dvpad == 't': return f'a.hdim_v % {self.bhdv} != 0'
+        else :                return f'a.hdim_v % {self.bhdv} == 0'
+
+class FmhaBwdApiPool:
+    def __init__(self, mask_impl):
+        self.dq_dk_dv_pool = dict()
+        self.mask_impl = mask_impl
+
+    def register_dq_dk_dv_traits(self, trait : FmhaBwdDQDKDVApiTrait) -> None:
+        # TODO: do we need to check duplication?
+        if trait.dtype not in self.dq_dk_dv_pool.keys():
+            self.dq_dk_dv_pool[trait.dtype] = dict()
+        if trait.hdim not in self.dq_dk_dv_pool[trait.dtype].keys():
+            self.dq_dk_dv_pool[trait.dtype][trait.hdim] = list()
+
+        self.dq_dk_dv_pool[trait.dtype][trait.hdim].append(copy.copy(trait))
+
+    @property
+    def api(self) -> str:
+        per_dtypes=str()
+        for i, dtype in enumerate(self.dq_dk_dv_pool.keys()):
+            per_hdim_case=str()
+            for j, hdim in enumerate(self.dq_dk_dv_pool[dtype].keys()):
+                traits=self.dq_dk_dv_pool[dtype][hdim]
+                hdim_int = int(hdim)
+                inners=str()
+                for k, trait in enumerate(traits):
+                    if_k = 'if' if k == 0 else 'else if'
+                    for spad1 in ["t", "f"]:
+                        if (spad1 == "f" and (trait.spad == "t" or trait.mode == "group")):
+                            continue
+                        inners = inners + FMHA_BWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_pipeline_enum=BWD_DQDKDV_PIPELINE_ENUM_MAP[trait.pipeline],
+                                    F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_mask=get_mask_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias],
+                                    F_bias=BIAS_MAP[trait.bias], F_dbias=BOOL_MAP[trait.dbias], F_dropout_check=DROPOUT_CHECK_MAP[trait.dropout], F_dropout=DROPOUT_MAP[trait.dropout],
+                                    F_scheck=trait.scheck(spad1=spad1), F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_hdim=hdim, F_dtype=BWD_DTYPE_MAP[dtype],
+                                    F_spad0=BOOL_MAP[trait.spad], F_spad1=BOOL_MAP[spad1], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
+                                    F_deterministic=BOOL_MAP[trait.deterministic])
+
+                if_j = 'if' if j == 0 else 'else if'
+                per_hdim_case = per_hdim_case + FMHA_BWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_inner_dispatch=inners)
+            if_i = 'if' if i == 0 else 'else if'
+            per_dtypes = per_dtypes + FMHA_BWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        if not per_dtypes:
+            # empty string we add some ignore to suppress warning in api
+            per_dtypes += '    (void)t ; (void)s ; (void)a;'
+        return FMHA_BWD_KERNEL_HEADER + FMHA_BWD_API.format(F_dispatch = per_dtypes)
+
+# GEMM0: Q@K=S^T
+# GEMM1: P^T@dO^T=dV(This was chosen as G1 to match fwd, but N1 must be equal to headdim_v)
+# GEMM2: dO@V=dP^T(This was chosen as G2 because of the calculation order)
+# GEMM3: dS^T@Q^T=dK(Similar to G1, but N3 must be equal to headdim_qk)
+# GEMM4: dS@K^T=dQ(N4 must be equal to headdim_qk)
+# Is it necessary to distinguish between K0~K4?
+@dataclass
+class FmhaBwdDQDKDVTileSize:
+    F_bm0       : int  # tile size along q seqlen (block size)
+    F_bn0       : int  # tile size along k seqlen
+    F_bk0       : int  # tile size along gemm0 unroll(F_bhdq)
+    F_bk1       : int  # tile size along gemm1 unroll(F_bm0)
+    F_bk2       : int  # tile size along gemm2 unroll(F_bhdv)
+    F_bk3       : int  # tile size along gemm3 unroll(F_bm0)
+    F_bk4       : int  # tile size along gemm4 unroll(F_bn0)
+    F_bhdq      : int  # q head_dim
+    F_bhdv      : int  # v head_dim
+    F_rm0       : int  # number of warps along q seqlen (block warps) in gemm0/gemm2
+    F_rn0       : int  # number of warps along k seqlen (block warps) in gemm0/gemm2
+    F_rk0       : int  # number of warps along headdim_qk/v (not used) in gemm0/gemm2
+    F_rm1       : int  # number of warps along k seqlen (block warps) in gemm1/gemm3
+    F_rn1       : int  # number of warps along headdim_qk/v (block warps) in gemm1/gemm3
+    F_rk1       : int  # number of warps along q seqlen (not used) in gemm1/gemm3
+    F_rm2       : int  # number of warps along q seqlen (block warps) in gemm4
+    F_rn2       : int  # number of warps along headdim_qk (block warps) in gemm4
+    F_rk2       : int  # number of warps along k seqlen (not used) in gemm4
+    F_wm0       : int  # warp size along m in gemm0/gemm2/gemm4
+    F_wn0       : int  # warp size along n in gemm0/gemm2/gemm4
+    F_wk0       : int  # warp size along k in gemm0/gemm2/gemm4
+    F_wm1       : int  # warp size along m in gemm1/gemm3
+    F_wn1       : int  # warp size along n in gemm1/gemm3
+    F_wk1       : int  # warp size along k in gemm1/gemm3
+    F_occupancy : int  # occupancy
+    @property
+    def name(self) -> str:
+        return f"b{self.F_bm0}x{self.F_bn0}x{self.F_bk0}x{self.F_bk1}x{self.F_bk2}x{self.F_bk3}x{self.F_bk4}x{self.F_bhdq}x{self.F_bhdv}" +\
+        f"_r{self.F_rm0}x{self.F_rn0}x{self.F_rk0}_r{self.F_rm1}x{self.F_rn1}x{self.F_rk1}_r{self.F_rm2}x{self.F_rn2}x{self.F_rk2}" +\
+        f"_w{self.F_wm0}x{self.F_wn0}x{self.F_wk0}_w{self.F_wm1}x{self.F_wn1}x{self.F_wk1}_o{self.F_occupancy}"
+
+@dataclass
+class FmhaBwdDQDKDVKernel:
+    F_idx           : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim          : int  # hdim
+    F_dtype         : str  # data type
+    F_tile          : FmhaBwdDQDKDVTileSize
+    F_spad          : str  # true/false
+    F_skpad         : str  #
+    F_dpad          : str  #
+    F_dvpad         : str  #
+    F_bias          : str  #
+    F_dbias         : str  #
+    F_dropout       : str  #
+    F_mask          : str  # value from MASK_MAP
+    F_mode          : str  # value from MODE_MAP
+    F_deterministic : str  #
+    F_pipeline      : str  #
+    mask_impl       : str  #
+
+    @property
+    def template(self) -> str:
+        return FMHA_BWD_KERNEL_HEADER + \
+            FMHA_BWD_DQ_DK_DV_KERNEL_BODY.format(
+                F_idx           = self.F_idx,
+                F_hdim          = self.F_hdim,
+                F_dtype         = BWD_DTYPE_MAP[self.F_dtype],
+                F_bm0           = self.F_tile.F_bm0,
+                F_bn0           = self.F_tile.F_bn0,
+                F_bk0           = self.F_tile.F_bk0,
+                F_bk1           = self.F_tile.F_bk1,
+                F_bk2           = self.F_tile.F_bk2,
+                F_bk3           = self.F_tile.F_bk3,
+                F_bk4           = self.F_tile.F_bk4,
+                F_bhdq          = self.F_tile.F_bhdq,
+                F_bhdv          = self.F_tile.F_bhdv,
+                F_rm0           = self.F_tile.F_rm0,
+                F_rn0           = self.F_tile.F_rn0,
+                F_rk0           = self.F_tile.F_rk0,
+                F_rm1           = self.F_tile.F_rm1,
+                F_rn1           = self.F_tile.F_rn1,
+                F_rk1           = self.F_tile.F_rk1,
+                F_rm2           = self.F_tile.F_rm2,
+                F_rn2           = self.F_tile.F_rn2,
+                F_rk2           = self.F_tile.F_rk2,
+                F_wm0           = self.F_tile.F_wm0,
+                F_wn0           = self.F_tile.F_wn0,
+                F_wk0           = self.F_tile.F_wk0,
+                F_wm1           = self.F_tile.F_wm1,
+                F_wn1           = self.F_tile.F_wn1,
+                F_wk1           = self.F_tile.F_wk1,
+                F_spad          = BOOL_MAP[self.F_spad],
+                F_skpad         = BOOL_MAP[self.F_skpad],
+                F_dpad          = BOOL_MAP[self.F_dpad],
+                F_dvpad         = BOOL_MAP[self.F_dvpad],
+                F_bias          = BIAS_MAP[self.F_bias],
+                F_dbias         = BOOL_MAP[self.F_dbias],
+                F_dropout       = DROPOUT_MAP[self.F_dropout],
+                F_occupancy     = self.F_tile.F_occupancy,
+                F_mask          = get_mask_map(self.mask_impl)[self.F_mask],
+                F_mode          = MODE_MAP[self.F_mode],
+                F_deterministic = BOOL_MAP[self.F_deterministic],
+                F_pipeline_enum = BWD_DQDKDV_PIPELINE_ENUM_MAP[self.F_pipeline],
+                F_pipeline      = BWD_DQDKDV_PIPELINE_MAP[self.F_pipeline])
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_skpad == 't' : n += 'sk'
+            if self.F_dpad == 't' : n += 'd'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f"fmha_bwd_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + self.F_tile.name + f'_{self.F_pipeline}'
+        if pn != '' : n += f'_{pn}'
+        else: n += '_npad'
+
+        if self.F_bias != 'no' : n += f'_{self.F_bias}'
+        else: n += '_nbias'
+
+        if self.F_dbias == 't' : n += '_dbias'
+        else: n += '_ndbias'
+
+        if self.F_mask[0:2] == 's_':
+            if self.F_mask == 's_mask': n += f'_mask'
+            else: n += '_nmask'
+        else:
+            if self.F_mask != 'no' : n += f'_m{self.F_mask[0]}'
+            else: n += '_nmask'
+
+        if self.F_dropout != 'no' : n += f'_{self.F_dropout}'
+        else: n += '_ndropout'
+
+        if self.F_deterministic == 't' : n += '_deterministic'
+        else: n += '_ndeterministic'
+        return n
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+    def api_trait(self) -> FmhaBwdDQDKDVApiTrait:
+        return FmhaBwdDQDKDVApiTrait(pipeline=self.F_pipeline,
+                hdim=str(self.F_hdim),
+                dtype=self.F_dtype,
+                mode=self.F_mode,
+                bm0=self.F_tile.F_bm0,
+                bn0=self.F_tile.F_bn0,
+                bhdq=self.F_tile.F_bhdq,
+                bhdv=self.F_tile.F_bhdv,
+                mask=self.F_mask,
+                bias=self.F_bias,
+                dbias=self.F_dbias,
+                dropout=self.F_dropout,
+                spad=self.F_spad,
+                skpad=self.F_skpad,
+                dpad=self.F_dpad,
+                dvpad=self.F_dvpad,
+                deterministic=self.F_deterministic
+                )
+
+# TODO: design a more practical way to do it
+# this is current supported tile size & pipeline.
+def get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype : str) -> Optional[dict]:
+    if dtype == 'fp16' or dtype == 'bf16':
+        return {
+            '32'  : [FmhaBwdDQDKDVTileSize( 32, 128,  32, 32,  32, 32, 64,  32,  32, 1, 4, 1, 4, 1, 1, 2, 2, 1, 16, 16, 32, 16, 16, 16, 1),
+                        "kr_ktr_vr_iglp", "kr_ktr_vr"],
+            '64'  : [FmhaBwdDQDKDVTileSize( 32, 128,  64, 32,  64, 32, 32,  64,  64, 1, 4, 1, 4, 1, 1, 1, 4, 1, 16, 16, 32, 16, 16, 16, 1),
+                        "kr_ktr_vr_iglp", "kr_ktr_vr"],
+            '128' : [FmhaBwdDQDKDVTileSize( 16, 128, 128, 16, 128, 16, 32, 128, 128, 1, 4, 1, 4, 1, 1, 1, 4, 1, 16, 16, 32, 16, 16, 16, 1),
+                        "kr_ktr_vr_iglp", "kr_ktr_vr"],
+            '256' : [FmhaBwdDQDKDVTileSize( 16,  64, 256, 16, 256, 16, 32, 256, 256, 1, 4, 1, 4, 1, 1, 1, 4, 1, 16, 16, 32, 16, 16, 16, 1),
+                        "kr_ktr_vr_iglp", "kr_ktr_vr"]
+        }
+    else:
+        return None
+
+def get_bwd_dq_dk_dv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[FmhaBwdApiPool, List[FmhaBwdDQDKDVKernel]]:
+    # TODO: we don't support tuning yet, so pick up one value for pad
+    #       support this in future
+    gen = list()
+    api_pool = FmhaBwdApiPool(mask_impl)
+
+    for dtype in BWD_DTYPE_MAP.keys():
+        d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        for hdim_str, mode, mask, bias, dbias, dropout, spad, skpad, dpad, dvpad, deterministic in itertools.product(d.keys(), MODE_MAP.keys(), get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], DROPOUT_MAP.keys(), ["t", "f"], ["t", "f"], ["t", "f"], ["t", "f"], ["t", "f"]):
+            tile = d[hdim_str][0]
+            ppl = d[hdim_str][1]
+            hdim = int(hdim_str)
+            if (mode == "group") and (spad == "f" or skpad == "f"):
+                continue
+            if ((bias == "no" or bias == "alibi") and dbias == "t"):
+                continue
+            if ("wg32" in dropout):
+                continue
+            if (dpad == "t" or dvpad == "t"):
+                ppl = d[hdim_str][2]
+            k = FmhaBwdDQDKDVKernel(F_idx=0, F_hdim=hdim, F_dtype=dtype, F_tile=tile,
+                                F_spad=spad, F_skpad=skpad, F_dpad=dpad, F_dvpad=dvpad,
+                                F_bias=bias, F_dbias=dbias, F_dropout=dropout, F_mask=mask, F_mode=mode,
+                                F_pipeline=ppl, mask_impl=mask_impl, F_deterministic=deterministic)
+            if kernel_filter != '':
+                if not fnmatch.fnmatch(k.name, kernel_filter):
+                    continue
+            # Flash attention integration
+            if receipt == 2:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= bias in ['no', 'alibi']
+                    cond &= dropout in ['no', 'dropout_wg32',  'dropout_wg16']
+                    cond &= dpad == dvpad
+                    if not cond:
+                        continue
+            elif receipt == 3:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= bias in ['no', 'alibi']
+                    cond &= dpad == dvpad
+                    cond &= deterministic == "f"
+                    if not cond:
+                        continue
+            # PyTorch integration
+            elif receipt == 4:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= bias in ['no', 'bias']
+                    cond &= dropout in ['no', 'dropout_wg32',  'dropout_wg16']
+                    cond &= dpad == dvpad
+                    cond &= deterministic == "f"
+                    if not cond:
+                        continue
+            # Aiter (mha_bwd) integration
+            elif receipt == 300:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == "batch"
+                    cond &= dropout in ['no', 'dropout_wg32',  'dropout_wg16']
+                    cond &= dpad == dvpad
+                    if not cond:
+                        continue
+            # Aiter (mha_varlen_bwd) integration
+            elif receipt == 400:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == "group"
+                    cond &= dropout in ['no', 'dropout_wg32',  'dropout_wg16']
+                    cond &= dpad == dvpad
+                    if not cond:
+                        continue
+            # aiter::mha_bwd C++ api integration
+            elif receipt == 600:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= dpad == dvpad
+                    if not cond:
+                        continue
+            api_pool.register_dq_dk_dv_traits(k.api_trait())
+            gen.append(k)
+
+    return (api_pool, gen)
+
+FMHA_BWD_DOT_DO_O_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+using fmha_bwd_dot_do_o_trait_{F_idx} =
+    ck_tile::TileFmhaBwdOGradDotOTraits<{F_spad}, {F_dvpad}, {F_occupancy}>;
+
+using fmha_bwd_dot_do_o_pipeline_problem_{F_idx} = ck_tile::BlockFmhaBwdOGradDotOPipelineProblem<
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::OGradDataType,
+    typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::DDataType,
+    /* BlockSize = */ 64,
+    {F_hdim},
+    {F_mode},
+    fmha_bwd_dot_do_o_trait_{F_idx}>;
+
+using fmha_bwd_dot_do_o_{F_idx} =
+    typename ck_tile::BlockFmhaBwdOGradDotO<fmha_bwd_dot_do_o_pipeline_problem_{F_idx}>;
+
+using fmha_bwd_dot_do_o_kernel_{F_idx} =
+    ck_tile::FmhaBwdOGradDotOKernel<fmha_bwd_dot_do_o_{F_idx}>;
+
+using dot_do_o_trait_{F_idx} =
+    fmha_bwd_dot_do_o_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_spad}, {F_dvpad}>;
+
+#include <iostream>
+
+template <>
+float fmha_bwd_dot_do_o_<dot_do_o_trait_{F_idx}>(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    using k_ = fmha_bwd_dot_do_o_kernel_{F_idx};
+    if(s.log_level_ > 0)
+        std::cout << ", " << k_::GetName() << std::flush;
+    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks                  = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    return ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
+}}
+
+template <>
+void fmha_bwd_dot_do_o_oneshot_<dot_do_o_trait_{F_idx}>(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    using k_                               = fmha_bwd_dot_do_o_kernel_{F_idx};
+    auto [kargs, grids]                    = fmha_bwd_dot_do_o_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks                  = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(
+        ck_tile::stream_config{{s.stream_id_}});
+}}
+
+template <>
+std::string fmha_bwd_dot_do_o_get_name_<dot_do_o_trait_{F_idx}>()
+{{
+    using k_ = fmha_bwd_dot_do_o_kernel_{F_idx};
+    return k_::GetName();
+}}
+"""
+
+@dataclass
+class FmhaBwdOGradDotOKernel:
+    F_idx       : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim      : int  # hdim
+    F_dtype     : str  # data type
+    F_spad      : str  # true/false
+    F_dvpad     : str  #
+    F_mode      : str  # value from MODE_MAP
+    F_occupancy : int
+
+    @property
+    def template(self) -> str:
+        return FMHA_BWD_KERNEL_HEADER + \
+            FMHA_BWD_DOT_DO_O_KERNEL_BODY.format(
+                F_idx       = self.F_idx,
+                F_hdim      = self.F_hdim,
+                F_dtype     = BWD_DTYPE_MAP[self.F_dtype],
+                F_spad      = BOOL_MAP[self.F_spad],
+                F_dvpad     = BOOL_MAP[self.F_dvpad],
+                F_mode      = MODE_MAP[self.F_mode],
+                F_occupancy = self.F_occupancy)
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f"fmha_bwd_dot_do_o_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_o{self.F_occupancy}"
+        if pn != '' : n += f'_{pn}'
+        else: n += '_npad'
+        return n
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+def get_bwd_dot_do_o_blobs(kernel_filter : Optional[str], receipt) -> List[FmhaBwdOGradDotOKernel]:
+    # TODO: we don't support tuning yet, so pick up one value for pad/occupancy
+    #       support this in future
+    def get_occupancy(dtype, hdim):
+        return 2
+
+    gen = list()
+
+    for dtype in BWD_DTYPE_MAP.keys():
+        d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        for hdim_str, mode, spad, dvpad in itertools.product(d.keys(), MODE_MAP.keys(), ["t", "f"], ["t", "f"]):
+            hdim = int(hdim_str)
+            if (mode == "group" and spad == "f"):
+                continue
+            k = FmhaBwdOGradDotOKernel(F_idx=0, F_hdim=hdim, F_dtype=dtype,
+                                F_spad=spad, F_dvpad=dvpad, F_mode=mode,
+                                F_occupancy=get_occupancy(dtype, hdim))
+            if kernel_filter != '':
+                if not fnmatch.fnmatch(k.name, kernel_filter):
+                    continue
+            # Aiter (mha_bwd) integration
+            if receipt == 300:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == "batch"
+                    if not cond:
+                        continue
+            # Aiter (mha_varlen_bwd) integration
+            elif receipt == 400:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == "group"
+                    if not cond:
+                        continue
+            # aiter::mha_bwd C++ api integration
+            elif receipt == 600:
+                    cond = dtype in ['fp16', 'bf16']
+                    if not cond:
+                        continue
+            gen.append(k)
+
+    return gen
+
+FMHA_BWD_CONVERT_DQ_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+using fmha_bwd_convert_dq_trait_{F_idx} =
+    ck_tile::TileFmhaBwdConvertQGradTraits<{F_spad}, {F_dpad}, {F_occupancy}>;
+
+using fmha_bwd_convert_dq_pipeline_problem_{F_idx} =
+    ck_tile::BlockFmhaBwdConvertQGradPipelineProblem<
+        typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::AccDataType,
+        typename FmhaBwdTypeConfig<fmha_dtype_{F_idx}>::QGradDataType,
+        /* BlockSize = */ 256,
+        {F_bm0},
+        {F_bn0},
+        {F_hdim},
+        {F_mode},
+        {F_deterministic},
+        fmha_bwd_convert_dq_trait_{F_idx}>;
+
+using fmha_bwd_convert_dq_{F_idx} =
+    typename ck_tile::BlockFmhaBwdConvertQGrad<fmha_bwd_convert_dq_pipeline_problem_{F_idx}>;
+
+using fmha_bwd_convert_dq_kernel_{F_idx} =
+    ck_tile::FmhaBwdConvertQGradKernel<fmha_bwd_convert_dq_{F_idx}>;
+
+using convert_dq_trait_{F_idx} = fmha_bwd_convert_dq_traits_<{F_hdim},
+                                                             {F_dtype},
+                                                             {F_mode},
+                                                             {F_spad},
+                                                             {F_dpad},
+                                                             {F_deterministic}>;
+
+#include <iostream>
+
+template <>
+float fmha_bwd_convert_dq_<convert_dq_trait_{F_idx}>(const ck_tile::stream_config& s, fmha_bwd_args a)
+{{
+    using k_ = fmha_bwd_convert_dq_kernel_{F_idx};
+    if(s.log_level_ > 0)
+        std::cout << ", " << k_::GetName() << std::flush;
+    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks                  = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    return ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
+}}
+
+template <>
+void fmha_bwd_convert_dq_oneshot_<convert_dq_trait_{F_idx}>(const ck_tile::stream_config& s,
+                                                            fmha_bwd_args a)
+{{
+    using k_                               = fmha_bwd_convert_dq_kernel_{F_idx};
+    auto [kargs, grids]                    = fmha_bwd_convert_dq_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks                  = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(
+        ck_tile::stream_config{{s.stream_id_}});
+}}
+
+template <>
+std::string fmha_bwd_convert_dq_get_name_<convert_dq_trait_{F_idx}>()
+{{
+    using k_ = fmha_bwd_convert_dq_kernel_{F_idx};
+    return k_::GetName();
+}}
+"""
+
+@dataclass
+class FmhaBwdConvertQGradKernel:
+    F_idx           : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim          : int  # hdim
+    F_dtype         : str  # data type
+    F_bm0           : int  # tile size along q seqlen (block size)
+    F_bn0           : int  # tile size along k seqlen
+    F_spad          : str  # true/false
+    F_dpad          : str  #
+    F_mode          : str  # value from MODE_MAP
+    F_occupancy     : int  #
+    F_deterministic : str  #
+
+    @property
+    def template(self) -> str:
+        return FMHA_BWD_KERNEL_HEADER + \
+            FMHA_BWD_CONVERT_DQ_KERNEL_BODY.format(
+                F_idx           = self.F_idx,
+                F_hdim          = self.F_hdim,
+                F_dtype         = BWD_DTYPE_MAP[self.F_dtype],
+                F_bm0           = self.F_bm0,
+                F_bn0           = self.F_bn0,
+                F_spad          = BOOL_MAP[self.F_spad],
+                F_dpad          = BOOL_MAP[self.F_dpad],
+                F_mode          = MODE_MAP[self.F_mode],
+                F_occupancy     = self.F_occupancy,
+                F_deterministic = BOOL_MAP[self.F_deterministic])
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_dpad == 't' : n += 'd'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f"fmha_bwd_convert_dq_d{self.F_hdim}_{self.F_dtype}_b{self.F_bm0}x{self.F_bn0}_{self.F_mode}_o{self.F_occupancy}"
+        if pn != '' : n += f'_{pn}'
+        else: n += '_npad'
+        if self.F_deterministic == 't' : n += '_deterministic'
+        else: n += '_ndeterministic'
+        return n
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+def get_bwd_convert_dq_blobs(kernel_filter : Optional[str], receipt) -> List[FmhaBwdConvertQGradKernel]:
+    # TODO: we don't support tuning yet, so pick up one value for pad/occupancy
+    #       support this in future
+    def get_occupancy(dtype, hdim):
+        return 2
+
+    gen = list()
+
+    for dtype in BWD_DTYPE_MAP.keys():
+        d = get_fmha_bwd_dq_dk_dv_tile_ppl_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        for hdim_str, mode, spad, dpad, deterministic in itertools.product(d.keys(), MODE_MAP.keys(), ["t", "f"], ["t", "f"], ["t", "f"]):
+            hdim = int(hdim_str)
+            tile = d[hdim_str][0]
+            if (mode == "group" and spad == "f"):
+                continue
+            k = FmhaBwdConvertQGradKernel(F_idx=0, F_hdim=hdim, F_dtype=dtype, F_bm0=64, F_bn0=tile.F_bn0,
+                                F_spad=spad, F_dpad=dpad, F_mode=mode, F_occupancy=get_occupancy(dtype, hdim), F_deterministic=deterministic)
+            if kernel_filter != '':
+                if not fnmatch.fnmatch(k.name, kernel_filter):
+                    continue
+            # Aiter (mha_bwd) integration
+            if receipt == 300:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == "batch"
+                    if not cond:
+                        continue
+            # Aiter (mha_varlen_bwd) integration
+            elif receipt == 400:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == "group"
+                    if not cond:
+                        continue
+            # aiter::mha_bwd C++ api integration
+            elif receipt == 600:
+                    cond = dtype in ['fp16', 'bf16']
+                    if not cond:
+                        continue
+            gen.append(k)
+
+    return gen
+
+def write_single_bwd_dq_dk_dv_kernel(kernel: FmhaBwdDQDKDVKernel, autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_single_bwd_dot_do_o_kernel(kernel: FmhaBwdOGradDotOKernel, autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_single_bwd_convert_dq_kernel(kernel: FmhaBwdConvertQGradKernel, autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_bwd_api(api_pool : FmhaBwdApiPool, autogen_dir: Path) -> None:
+    (autogen_dir / FMHA_BWD_API_FILENAME).write_text(api_pool.api)
+
+def write_blobs(output_dir : Path, filter_list : str, receipt, optdim_list, mask_impl) -> None:
+    filter_list = filter_list.split('@')
+    filter_list.extend([''] * (3 - len(filter_list)))
+    # TODO
+    assert optdim_list == [-1]
+
+    kernels = get_bwd_dot_do_o_blobs(filter_list[0], receipt)
+    for kernel in kernels:
+        write_single_bwd_dot_do_o_kernel(kernel, output_dir)
+    kernels = get_bwd_convert_dq_blobs(filter_list[1], receipt)
+    for kernel in kernels:
+        write_single_bwd_convert_dq_kernel(kernel, output_dir)
+    api_pool, kernels = get_bwd_dq_dk_dv_blobs(filter_list[2], receipt, mask_impl)
+    for kernel in kernels:
+        write_single_bwd_dq_dk_dv_kernel(kernel, output_dir)
+    write_bwd_api(api_pool, output_dir)
+
+def list_blobs(file_path : Path, filter_list : str, receipt, optdim_list, mask_impl) -> None:
+    filter_list = filter_list.split('@')
+    filter_list.extend([''] * (3 - len(filter_list)))
+    # TODO
+    assert optdim_list == [-1]
+
+    with file_path.open('a') as f:
+        kernels = get_bwd_dot_do_o_blobs(filter_list[0], receipt)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        kernels = get_bwd_convert_dq_blobs(filter_list[1], receipt)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        _, kernels = get_bwd_dq_dk_dv_blobs(filter_list[2], receipt, mask_impl)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        f.write(str(file_path.parent / GEN_DIR / FMHA_BWD_API_FILENAME) + "\n")

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

@@ -0,0 +1,574 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import copy
+from dataclasses import dataclass
+import fnmatch
+import itertools
+from pathlib import Path
+from typing import List, Optional, Tuple
+
+from codegen.cmake_config import *
+from codegen.cpp_symbol_map import *
+
+
+DTYPE_BITS = {
+    "fp32": 32,
+    "fp16": 16,
+    "bf16": 16,
+    "fp8" : 8,
+    "bf8" : 8
+}
+
+K0_MAX_SUBMAX_MAP = {
+    32 : 32,
+    64 : 64,
+    96 : 128,
+    128: 128,
+    256: 256
+}
+
+FMHA_FWD_KERNEL_HEADER = """// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.\n
+// auto generated by generate.py
+#include "fmha_fwd.hpp"
+"""
+
+FMHA_FWD_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+using fmha_block_tile_{F_idx} = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}>;
+
+using fmha_shape_{F_idx} = ck_tile::TileFmhaShape<fmha_block_tile_{F_idx},
+                                      ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>,
+                                      ck_tile::sequence<{F_wm0}, {F_wn0}, {F_wk0}>,
+                                      ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>,
+                                      ck_tile::sequence<{F_wm1}, {F_wn1}, {F_wk1}>,
+                                      {F_vlayout}>;
+
+using fmha_trait_{F_idx} = ck_tile::TileFmhaTraits<{F_spad},
+                                                    {F_skpad},
+                                                    {F_dpad},
+                                                    {F_dvpad},
+                                                    {F_bias},
+                                                    false,
+                                                    {F_lse},
+                                                    {F_dropout},
+                                                    {F_squant},
+                                                    {F_occupancy}>;
+using fmha_mask_{F_idx} = {F_mask};
+
+using fmha_pipeline_problem_{F_idx} = ck_tile::BlockFmhaPipelineProblem<
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::QDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::KDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::VDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SMPLComputeDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::BiasDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::RandValOutputDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::PDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
+    fmha_shape_{F_idx},
+    {F_mode},
+    fmha_mask_{F_idx},
+    fmha_trait_{F_idx}>;
+
+using fmha_pipeline_{F_idx} = {F_pipeline}<
+    fmha_pipeline_problem_{F_idx}>;
+
+using fmha_epilogue_{F_idx} =
+    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
+                                           typename FmhaFwdTypeConfig<{F_dtype}>::ODataType,
+                                           {F_spad}, {F_dvpad}>>;
+
+using fmha_kernel_{F_idx} =
+    ck_tile::FmhaFwdKernel<fmha_pipeline_{F_idx}, fmha_epilogue_{F_idx}>;
+
+using trait_{F_idx} = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode},{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout},
+                        {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+
+#include <iostream>
+
+template<>
+float fmha_fwd_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_args a)
+{{
+    using k_ = fmha_kernel_{F_idx};
+    if(s.log_level_ > 0)
+        std::cout << ", " << k_::GetName() << std::flush;
+    auto [kargs, grids] = fmha_fwd_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
+}}
+"""
+
+FMHA_FWD_API_FILENAME="fmha_fwd_api.cpp"
+FMHA_FWD_API="""
+float fmha_fwd(fmha_fwd_traits t, fmha_fwd_args a, const ck_tile::stream_config& s){{
+    float r = -1;
+{F_dispatch}
+    return r;
+}}
+"""
+
+FMHA_FWD_API_PER_DTYPE="""    {F_if}(t.data_type.compare(\"{F_dtype}\") == 0){{
+{F_hdim_case}
+    }}
+"""
+FMHA_FWD_API_PER_HDIM_CASE="""        {F_if} (t.hdim_q <= {F_hdim} && t.hdim_v <= {F_hdim_v}) {{
+{F_inner_dispatch}
+        }}
+"""
+
+FMHA_FWD_API_INNER_DISPATCH="""            {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.has_lse == {F_lse})  && (t.has_dropout == {F_dropout}) && (t.do_fp8_static_quant == {F_squant}) &&
+                        ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
+                using trait_ = fmha_fwd_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, {F_lse}, {F_dropout}, {F_squant}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+                return fmha_fwd_<trait_>(s, a);
+            }}
+"""
+
+@dataclass
+class FmhaFwdApiTrait:
+    pipeline_tag : str
+    # sync with fmha_fwd_traits<>, to generate fallback calls
+    hdim      : str
+    dtype     : str  # data type
+    mode      : str  # value from MODE_MAP
+    bm0       : int  # tile size along q seqlen (block size)
+    bn0       : int  # tile size along qk seqlen
+    bk0       : int  # tile size along qk gemm unroll
+    bn1       : int  # tile size along v head_dim
+    bk1       : int  # tile size along kv gemm unroll
+    bk0max    : int
+    vlayout   : str
+    mask      : str
+    bias      : str  #
+    lse       : str  #
+    dropout   : str
+    squant    : str  #
+    spad      : str
+    skpad     : str
+    dpad      : str
+    dvpad     : str
+
+    @property
+    def name(self) -> str:
+        return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0max}-'+\
+                    f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.dropout}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}'
+
+    @property
+    def scheck(self) -> str:
+        if self.mode == 'group': return 'true/*group mode spad always true*/'                  # group mode only generate spad/skpad == true
+        if self.pipeline_tag == 'qr_async':
+            if self.spad == 't' : return 'true' # always support
+            else :                return 'true'
+        elif self.pipeline_tag in ['qr']:
+            if self.spad == 't' : return f'true /*a.seqlen_q % {self.bm0} != 0*/'  # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.seqlen_q % {self.bm0} == 0'
+        else: assert False
+
+    @property
+    def skcheck(self) -> str:
+        if self.mode == 'group': return 'true/*group mode skpad always true*/'                  # group mode only generate spad/skpad == true
+        if self.pipeline_tag == 'qr_async':
+            if self.skpad == 't' : return f'a.seqlen_k == 0 || a.seqlen_k % {self.bn0} != 0'
+            else :                 return f'a.seqlen_k != 0 && a.seqlen_k % {self.bn0} == 0'
+        elif self.pipeline_tag in ['qr', 'qr_fp8']:
+            if self.skpad == 't' : return f'true /*a.seqlen_k % {self.bn0} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.seqlen_k % {self.bn0} == 0'
+        else: assert False
+
+    @property
+    def dcheck(self) -> str:
+        if self.pipeline_tag == 'qr_async':
+            vec = int((32 * 4) / DTYPE_BITS[self.dtype])
+            if self.dpad == 't': return f'a.hdim_q % {vec} == 0'
+            else :               assert False
+        elif self.pipeline_tag in ['qr']:
+            bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
+            if self.dpad == 't': return f'true /*a.hdim_q % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :               return f'a.hdim_q % {bk0submax} == 0'
+        else:   assert False
+
+    @property
+    def dvcheck(self) -> str:
+        if self.pipeline_tag == 'qr_async':
+            vec = int((32 * 4) / DTYPE_BITS[self.dtype])
+            if self.dvpad == 't': return f'a.hdim_v % {vec} == 0'
+            else :                assert False
+        elif self.pipeline_tag in ['qr']:
+            bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
+            if self.dvpad == 't': return f'true /*a.hdim_v % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.hdim_v % {bk0submax} == 0'
+        else:   assert False
+
+@dataclass
+class FmhaFwdPipeline:
+    tag : str
+
+    F_vlayout   : str  # row/col
+    F_spad      : str  # true/false
+    F_skpad     : str  #
+    F_dpad      : str  #
+    F_dvpad     : str  #
+    F_bias      : str  # true/false
+    F_lse       : str  #
+    F_dropout   : str  #
+    F_squant    : str  #
+    F_mask      : str  # value from MASK_MAP
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_skpad == 't' : n += 'sk'
+            if self.F_dpad == 't' : n += 'd'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f'{self.tag}_v{self.F_vlayout[0]}'
+        if pn != '' : n += f'_{pn}'
+        else: n += '_npad'
+
+        if self.F_bias != 'no' : n += f'_{self.F_bias}'
+        else: n += '_nbias'
+
+        if self.F_mask[0:2] == 's_':
+            if self.F_mask == 's_mask': n += f'_mask'
+            else: n += '_nmask'
+        else:
+            if self.F_mask != 'no' : n += f'_m{self.F_mask[0]}'
+            else: n += '_nmask'
+
+        if self.F_lse == 't' : n += '_lse'
+        else: n += '_nlse'
+
+        if self.F_dropout == 't' : n += '_dropout'
+        else: n += '_ndropout'
+
+        if self.F_squant == 't' : n += '_squant'
+        else: n += '_nsquant'
+        return n
+
+class FmhaFwdApiPool:
+    def __init__(self, mask_impl):
+        self.pool = dict()
+        self.mask_impl = mask_impl
+
+    def register_traits(self, trait : FmhaFwdApiTrait) -> None:
+        # TODO: do we need to check duplication?
+        if trait.dtype not in self.pool.keys():
+            self.pool[trait.dtype] = dict()
+        if trait.hdim not in self.pool[trait.dtype].keys():
+            self.pool[trait.dtype][trait.hdim] = list()
+
+        self.pool[trait.dtype][trait.hdim].append(copy.copy(trait))
+
+    @property
+    def api(self) -> str:
+        per_dtypes=str()
+        for i, dtype in enumerate(self.pool.keys()):
+            per_hdim_case=str()
+            for j, hdim in enumerate(self.pool[dtype].keys()):
+                traits=self.pool[dtype][hdim]
+                inners=str()
+                for k, trait in enumerate(traits):
+                    if_k = 'if' if k == 0 else 'else if'
+                    inners = inners + FMHA_FWD_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_vlayout=LAYOUT_MAP[trait.vlayout],
+                                   F_pipeline_enum=PIPELINE_ENUM_MAP[trait.pipeline_tag], F_mask=get_mask_map(self.mask_impl)[trait.mask],
+                                   F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias], F_bias=BIAS_MAP[trait.bias],
+                                   F_lse=BOOL_MAP[trait.lse], F_dropout=BOOL_MAP[trait.dropout] ,
+                                   F_squant=BOOL_MAP[trait.squant], F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
+                                   F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
+                                   F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
+                                   F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
+                if_j = 'if' if j == 0 else 'else if'
+                per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_hdim_v=trait.bn1, F_inner_dispatch=inners)
+            if_i = 'if' if i == 0 else 'else if'
+            per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        if not per_dtypes:
+            # empty string we add some ignore to suppress warning in api
+            per_dtypes += '    (void)t ; (void)s ; (void)a;'
+        return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_API.format(F_dispatch = per_dtypes)
+
+@dataclass
+class FmhaFwdTileSize:
+    F_bm0       : int  # tile size along q seqlen (block size)
+    F_bn0       : int  # tile size along k seqlen
+    F_bk0       : int  # tile size along qk gemm unroll
+    F_bn1       : int  # tile size along v head_dim
+    F_bk1       : int  # tile size along kv gemm unroll
+    F_bk0max    : int  # total length of K0, used for pipeline that need load Q at once (or repeately load Q as a whole tile)
+    F_rm0       : int  # number of warps for gemm0 along q seqlen
+    F_rn0       : int  # number of warps for gemm0 along k seqlen
+    F_rk0       : int  # number of warps for gemm0 along head dim q (not used)
+    F_rm1       : int  # number of warps for gemm1 along q seqlen
+    F_rn1       : int  # number of warps for gemm1 along head dim v
+    F_rk1       : int  # number of warps for gemm1 along k seqlen (not used)
+    F_wm0       : int  # gemm0 warp size along m
+    F_wn0       : int  # gemm0 warp size along n
+    F_wk0       : int  # gemm0 warp size along k
+    F_wm1       : int  # gemm1 warp size along m
+    F_wn1       : int  # gemm1 warp size along n
+    F_wk1       : int  # gemm1 warp size along k
+    F_occupancy : int  # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy
+    @property
+    def name(self) -> str:
+        return f"b{self.F_bm0}x{self.F_bn0}x{self.F_bk0}x{self.F_bn1}x{self.F_bk1}x{self.F_bk0max}" +\
+        f"_r{self.F_rm0}x{self.F_rn0}x{self.F_rk0}_r{self.F_rm1}x{self.F_rn1}x{self.F_rk1}" +\
+        f"_w{self.F_wm0}x{self.F_wn0}x{self.F_wk0}_w{self.F_wm1}x{self.F_wn1}x{self.F_wk1}" +\
+        ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
+
+@dataclass
+class FmhaFwdKernel:
+    F_idx           : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim          : int  # hdim
+    F_dtype         : str  # data type
+    F_mode          : str  # value from MODE_MAP
+    F_tile          : FmhaFwdTileSize
+    F_pipeline      : FmhaFwdPipeline
+    mask_impl       : str
+
+    @property
+    def template(self) -> str:
+        kernel_body = str()
+        return FMHA_FWD_KERNEL_HEADER + \
+            FMHA_FWD_KERNEL_BODY.format(
+                F_idx           = self.F_idx,
+                F_hdim          = self.F_hdim,
+                F_dtype         = FWD_DTYPE_MAP[self.F_dtype],
+                F_bm0           = self.F_tile.F_bm0,
+                F_bn0           = self.F_tile.F_bn0,
+                F_bk0           = self.F_tile.F_bk0,
+                F_bn1           = self.F_tile.F_bn1,
+                F_bk1           = self.F_tile.F_bk1,
+                F_bk0max        = self.F_tile.F_bk0max,
+                F_rm0           = self.F_tile.F_rm0,
+                F_rn0           = self.F_tile.F_rn0,
+                F_rk0           = self.F_tile.F_rk0,
+                F_rm1           = self.F_tile.F_rm1,
+                F_rn1           = self.F_tile.F_rn1,
+                F_rk1           = self.F_tile.F_rk1,
+                F_wm0           = self.F_tile.F_wm0,
+                F_wn0           = self.F_tile.F_wn0,
+                F_wk0           = self.F_tile.F_wk0,
+                F_wm1           = self.F_tile.F_wm1,
+                F_wn1           = self.F_tile.F_wn1,
+                F_wk1           = self.F_tile.F_wk1,
+                F_vlayout       = LAYOUT_MAP[self.F_pipeline.F_vlayout],
+                F_spad          = BOOL_MAP[self.F_pipeline.F_spad],
+                F_skpad         = BOOL_MAP[self.F_pipeline.F_skpad],
+                F_dpad          = BOOL_MAP[self.F_pipeline.F_dpad],
+                F_dvpad         = BOOL_MAP[self.F_pipeline.F_dvpad],
+                F_bias          = BIAS_MAP[self.F_pipeline.F_bias],
+                F_lse           = BOOL_MAP[self.F_pipeline.F_lse],
+                F_dropout       = BOOL_MAP[self.F_pipeline.F_dropout],
+                F_squant        = BOOL_MAP[self.F_pipeline.F_squant],
+                F_occupancy     = self.F_tile.F_occupancy,
+                F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag],
+                F_mask          = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask],
+                F_mode          = MODE_MAP[self.F_mode],
+                F_pipeline      = PIPELINE_MAP[self.F_pipeline.tag])
+
+    @property
+    def name(self) -> str:
+        # TODO: we don't encode idx here
+        return f"fmha_fwd_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + \
+                self.F_tile.name + '_' + self.F_pipeline.name
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+    def api_trait(self) -> FmhaFwdApiTrait:
+        return FmhaFwdApiTrait(
+                pipeline_tag=self.F_pipeline.tag,
+                hdim=str(self.F_hdim),
+                dtype=self.F_dtype,
+                mode=self.F_mode,
+                bm0=self.F_tile.F_bm0,
+                bn0=self.F_tile.F_bn0,
+                bk0=self.F_tile.F_bk0,
+                bn1=self.F_tile.F_bn1,
+                bk1=self.F_tile.F_bk1,
+                bk0max=self.F_tile.F_bk0max,
+                vlayout=self.F_pipeline.F_vlayout,
+                mask=self.F_pipeline.F_mask,
+                bias=self.F_pipeline.F_bias,
+                lse=self.F_pipeline.F_lse,
+                dropout=self.F_pipeline.F_dropout,
+                squant=self.F_pipeline.F_squant,
+                spad=self.F_pipeline.F_spad,
+                skpad=self.F_pipeline.F_skpad,
+                dpad=self.F_pipeline.F_dpad,
+                dvpad=self.F_pipeline.F_dvpad)
+
+# TODO: design a more practical way to do it
+# this is current supported tile size per hdim
+def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
+    if dtype == 'fp16' or dtype == 'bf16':
+        return {
+            '32'  : FmhaFwdTileSize(128, 64,  16, 32,  32,  32,   2, 1, 1,  2, 1, 1,  32, 32, 16,  32, 32, 16,  -1),
+            '64'  : FmhaFwdTileSize(128, 64,  32, 64,  32,  64,   4, 1, 1,  4, 1, 1,  32, 32, 16,  32, 32, 16,  -1),
+        ### '96'  : FmhaFwdTileSize(128, 128, 32, 128, 32,  96,   4, 1, 1,  4, 1, 1,  32, 32, 16,  32, 32, 16,  -1),
+            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32,  128,  4, 1, 1,  4, 1, 1,  32, 32, 16,  32, 32, 16,  -1),
+            '192' : FmhaFwdTileSize(128, 128, 32, 128, 32,  192,  4, 1, 1,  4, 1, 1,  32, 32, 16,  32, 32, 16,  -1),
+            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32,  256,  4, 1, 1,  4, 1, 1,  32, 32, 16,  32, 32, 16,  -1),
+        }
+    elif dtype == 'fp8' or dtype == 'bf8':
+        return {
+            '64'  : FmhaFwdTileSize(128, 64,  32, 64,  32,  64,   2, 1, 1,  2, 1, 1,  32, 32, 32,  32, 32, 32,  -1),
+            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32,  128,  4, 1, 1,  4, 1, 1,  32, 32, 32,  32, 32, 32,  -1),
+            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32,  256,  4, 1, 1,  4, 1, 1,  32, 32, 32,  32, 32, 32,  -1),
+        }
+    else:
+        return None
+
+def get_fwd_blobs(kernel_filter : Optional[str], receipt, optdim_list, mask_impl) -> Tuple[FmhaFwdApiPool, List[FmhaFwdKernel]]:
+    # TODO: we don't support tuning yet, so pick up one value for vlayout/pipeline/pad
+    #       support this in future
+    def get_pipelines(dtype, hdim) -> List[FmhaFwdPipeline]:
+        # this function will populate a list possible pipelines
+        # TODO: the order of List matters! the later in this list will be also be checked later
+        # TODO: currently for qr pipeline, let 't' padding to appear later!!
+        # TODO: how to design this more generic?
+        squant = 't' if dtype == 'fp8' else 'f'
+        pipelines = []
+        if dtype in ['fp16', 'bf16']:
+            for mask, bias, lse, dropout in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"], ["t", "f"]):
+                if hdim == 256:
+                # if True:
+                    pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                    pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                    # the below two is used for hdim vectorize load
+                    pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 'f', 'f', bias, lse, dropout, squant, mask))
+                    pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 'f', 'f', bias, lse, dropout, squant, mask))
+
+                    pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                else:
+                    if bias == "bias":
+                        # TODO: rocm 6.2 compiler problem if using qr_async for bias case
+                        pipelines.append(FmhaFwdPipeline('qr', 'row', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    else:
+                        pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr_async', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask))
+                        pipelines.append(FmhaFwdPipeline('qr_async', 'col', 't', 't', 't', 't', bias, lse, dropout, squant, mask))
+                    if receipt == 1 and bias != "bias":
+                        pipelines.append(FmhaFwdPipeline('qr', 'row', 't', 't', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
+                        pipelines.append(FmhaFwdPipeline('qr', 'col', 't', 'f', 't', 't', bias, lse, dropout, squant, mask)) # TODO: cover arbitraty hdim
+        elif dtype in ['fp8', 'bf8']:
+            # no need lse/dropout kernels
+            for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()):
+                pipelines.append(FmhaFwdPipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 'f', 'f', squant, mask))
+        elif dtype in ['fp8fp16', 'fp8bf16']:
+            # TODO
+            None
+        else:
+            assert False
+        return pipelines
+
+    gen = list()
+    api_pool = FmhaFwdApiPool(mask_impl)
+
+    for dtype in FWD_DTYPE_MAP.keys():
+        d = get_fmha_fwd_tile_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        #for hdim_str, mode, mask, bias, lse in itertools.product(d.keys(), MODE_MAP.keys(), MASK_MAP.keys(), ["t", "f"], ["t", "f"]):
+        for hdim_str, mode in itertools.product(d.keys(), MODE_MAP.keys()):
+            tile = d[hdim_str]
+            hdim = int(hdim_str)
+            for pipeline in get_pipelines(dtype, hdim):
+                if mode == "group":
+                    if pipeline.F_spad != 't' or pipeline.F_skpad != 't':
+                        # in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
+                        continue
+                if hdim == 192 and tile.F_bn1 == 128:
+                    # NOTE: this is used to speedup deepseek prefill case, we don't gen training
+                    if pipeline.F_bias != 'no' or pipeline.F_lse == 't' or pipeline.F_dropout == 't':
+                        continue
+                k = FmhaFwdKernel(F_idx=0,
+                                  F_hdim=hdim,
+                                  F_dtype=dtype,
+                                  F_mode=mode,
+                                  F_tile=tile,
+                                  F_pipeline=pipeline,
+                                  mask_impl=mask_impl)
+                if kernel_filter != '':
+                    if not fnmatch.fnmatch(k.name, kernel_filter):
+                        continue
+                if optdim_list != [-1]:
+                    if hdim not in optdim_list:
+                        continue
+                # 2 - Flash attention integration
+                if receipt in (2, 3):
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_bias in ['no', 'alibi']
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                # PyTorch integration
+                elif receipt == 4:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_bias in ['no', 'bias']
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                # Aiter(mha_fwd) integration
+                elif receipt == 100:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == 'batch'
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                # Aiter(mha_varlen_fwd) integration
+                elif receipt == 200:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == 'group'
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                # aiter::mha_fwd C++ api integration
+                elif receipt == 600:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                api_pool.register_traits(k.api_trait())
+                gen.append(k)
+
+    return (api_pool, gen)
+
+def write_single_fwd_kernel(kernel: FmhaFwdKernel, autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_fwd_api(api_pool : FmhaFwdApiPool, autogen_dir: Path) -> None:
+    (autogen_dir / FMHA_FWD_API_FILENAME).write_text(api_pool.api)
+
+def write_blobs(output_dir : Path, kernel_filter : str, receipt, optdim_list, mask_impl) -> None:
+    api_pool, kernels = get_fwd_blobs(kernel_filter, receipt, optdim_list, mask_impl)
+    for kernel in kernels:
+        write_single_fwd_kernel(kernel, output_dir)
+    write_fwd_api(api_pool, output_dir)
+
+def list_blobs(file_path : Path, kernel_filter : str, receipt, optdim_list, mask_impl) -> None:
+    with file_path.open('a') as f:
+        _, kernels = get_fwd_blobs(kernel_filter, receipt, optdim_list, mask_impl)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_API_FILENAME) + "\n")

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

@@ -0,0 +1,359 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import copy
+from dataclasses import dataclass
+import fnmatch
+import itertools
+from pathlib import Path
+from typing import List, Optional, Tuple
+
+from codegen.cmake_config import *
+from codegen.cpp_symbol_map import *
+
+from codegen.ops.fmha_fwd import (
+    FmhaFwdApiTrait,
+    DTYPE_BITS,
+    FMHA_FWD_KERNEL_HEADER,
+    FMHA_FWD_API_PER_DTYPE,
+    FMHA_FWD_API_PER_HDIM_CASE,
+)
+
+
+FMHA_FWD_APPENDKV_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+using fmha_trait_{F_idx} = ck_tile::TileFmhaFwdAppendKVTraits<{F_spad},
+                                                    {F_skpad},
+                                                    {F_dpad},
+                                                    {F_dvpad},
+                                                    {F_occupancy}>;
+
+using fmha_pipeline_problem_{F_idx} = ck_tile::BlockFmhaFwdAppendKVPipelineProblem<
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::QDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::KDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::VDataType,
+    {F_bs},
+    {F_bsk},
+    {F_bd},
+    {F_bdv},
+    {F_vlayout},
+    {F_rope},
+    {F_pagedkv},
+    fmha_trait_{F_idx}>;
+
+using fmha_pipeline_{F_idx} = ck_tile::BlockFmhaFwdAppendKVPipeline<
+    fmha_pipeline_problem_{F_idx}>;
+
+using fmha_kernel_{F_idx} = ck_tile::FmhaFwdAppendKVKernel<fmha_pipeline_{F_idx}>;
+
+using trait_{F_idx} = fmha_fwd_appendkv_traits_<{F_hdim}, {F_dtype}, {F_bs}, {F_bsk}, {F_bd}, {F_bdv}, {F_vlayout},
+                        {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}, {F_rope}, {F_pagedkv}>;
+
+#include <iostream>
+
+template<>
+float fmha_fwd_appendkv_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_appendkv_args a)
+{{
+    using k_ = fmha_kernel_{F_idx};
+    if(s.log_level_ > 0)
+        std::cout << ", " << k_::GetName() << std::flush;
+    auto [kargs, grids] = fmha_fwd_appendkv_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    return ck_tile::launch_kernel(s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs));
+}}
+"""
+
+FMHA_FWD_APPENDKV_API_FILENAME="fmha_fwd_appendkv_api.cpp"
+FMHA_FWD_APPENDKV_API="""
+float fmha_fwd_appendkv(fmha_fwd_appendkv_traits t, fmha_fwd_appendkv_args a, const ck_tile::stream_config& s){{
+    float r = -1;
+{F_dispatch}
+    return r;
+}}
+"""
+
+FMHA_FWD_APPENDKV_API_INNER_DISPATCH="""            {F_if}((t.is_v_rowmajor == {F_vlayout}) &&
+                        ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck}) && (t.rope_type == {F_rope_check}) &&
+                        ((a.block_table_ptr != nullptr) == {F_pagedkv})) {{
+                using trait_ = fmha_fwd_appendkv_traits_<{F_hdim}, {F_dtype}, {F_bs}, {F_bsk}, {F_bd}, {F_bdv}, {F_vlayout}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}, {F_rope}, {F_pagedkv}>;
+                return fmha_fwd_appendkv_<trait_>(s, a);
+            }}
+"""
+
+@dataclass
+class FmhaFwdAppendKVApiTrait:
+    # sync with fmha_fwd_traits<>, to generate fallback calls
+    hdim      : str
+    dtype     : str  # data type
+    bs        : int  # tile size along q seqlen
+    bsk       : int  # tile size along k seqlen
+    bd        : int  # tile size along qk gemm unroll
+    bdv       : int  # tile size along kv gemm unroll
+    vlayout   : str
+    spad      : str
+    skpad     : str
+    dpad      : str
+    dvpad     : str
+    rope      : str # key from ROPE_MAP
+    pagedkv   : str
+
+    @property
+    def name(self) -> str:
+        return f'{self.hdim}-{self.dtype}-{self.bs}-{self.bsk}-{self.bd}-{self.bdv}-{self.vlayout}-'+\
+               f'{self.spad}-{self.skpad}-{self.dpad}-{self.dvpad}-{self.rope}-{self.pagedkv}'
+
+    @property
+    def scheck(self) -> str:
+        if self.spad == 't' : return f'true /*a.seqlen_q % {self.bs} != 0*/'
+        else :                return f'a.seqlen_q % {self.bs} == 0'
+
+    @property
+    def skcheck(self) -> str:
+        # we do not check all the values in a.seqlen_k_ptr
+        return 'true'
+
+    @property
+    def dcheck(self) -> str:
+        if self.dpad == 't': return f'true /*a.hdim_q % {self.bd} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+        else :               return f'a.hdim_q % {self.bd} == 0'
+
+    @property
+    def dvcheck(self) -> str:
+        if self.dvpad == 't': return f'true /*a.hdim_v % {self.bdv} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+        else :                return f'a.hdim_v % {self.bdv} == 0'
+
+@dataclass
+class FmhaFwdAppendKVPipeline:
+    F_vlayout   : str  # row/col
+    F_spad      : str  # true/false
+    F_skpad     : str  #
+    F_dpad      : str  #
+    F_dvpad     : str  #
+    F_rope      : str  # key from ROPE_MAP
+    F_pagedkv   : str  # t/f
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_skpad == 't' : n += 'sk'
+            if self.F_dpad == 't' : n += 'd'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f'v{self.F_vlayout[0]}'
+        if pn != '' : n += f'_{pn}'
+        if self.F_rope != 'no': n += f'_{self.F_rope}'
+        if self.F_pagedkv == 't': n += '_pagedkv'
+        return n
+
+class FmhaFwdAppendKVApiPool:
+    def __init__(self, mask_impl):
+        self.pool = dict()
+        self.mask_impl = mask_impl
+
+    def register_traits(self, trait : FmhaFwdApiTrait) -> None:
+        # TODO: do we need to check duplication?
+        if trait.dtype not in self.pool.keys():
+            self.pool[trait.dtype] = dict()
+        if trait.hdim not in self.pool[trait.dtype].keys():
+            self.pool[trait.dtype][trait.hdim] = list()
+
+        self.pool[trait.dtype][trait.hdim].append(copy.copy(trait))
+
+    @property
+    def api(self) -> str:
+        per_dtypes=str()
+        for i, dtype in enumerate(self.pool.keys()):
+            per_hdim_case=str()
+            for j, hdim in enumerate(self.pool[dtype].keys()):
+                traits=self.pool[dtype][hdim]
+                inners=str()
+                for k, trait in enumerate(traits):
+                    if_k = 'if' if k == 0 else 'else if'
+                    inners = inners + FMHA_FWD_APPENDKV_API_INNER_DISPATCH.format(F_if=if_k, F_vlayout=LAYOUT_MAP[trait.vlayout],
+                                   F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck, F_rope_check=ROPE_CHECK_MAP[trait.rope],
+                                   F_pagedkv=BOOL_MAP[trait.pagedkv], F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
+                                   F_rope=ROPE_MAP[trait.rope], F_bs=trait.bs, F_bsk=trait.bsk, F_bd=trait.bd, F_bdv=trait.bdv, F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
+                if_j = 'if' if j == 0 else 'else if'
+                per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_hdim_v=hdim, F_inner_dispatch=inners)
+            if_i = 'if' if i == 0 else 'else if'
+            per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_APPENDKV_API.format(F_dispatch = per_dtypes)
+
+@dataclass
+class FmhaFwdAppendKVTileSize:
+    F_bs        : int  # tile size along q seqlen
+    F_bsk       : int  # tile size along k seqlen
+    F_bd        : int  # tile size along qk gemm unroll
+    F_bdv       : int  # tile size along kv gemm unroll
+    F_occupancy : int  # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy
+    @property
+    def name(self) -> str:
+        return f"b{self.F_bs}x{self.F_bsk}x{self.F_bd}x{self.F_bdv}" +\
+            ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
+
+@dataclass
+class FmhaFwdAppendKVKernel:
+    F_idx           : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim          : int  # hdim
+    F_dtype         : str  # data type
+    F_tile          : FmhaFwdAppendKVTileSize
+    F_pipeline      : FmhaFwdAppendKVPipeline
+    mask_impl       : str
+
+    @property
+    def template(self) -> str:
+        kernel_body = str()
+        return FMHA_FWD_KERNEL_HEADER + \
+            FMHA_FWD_APPENDKV_KERNEL_BODY.format(
+                F_idx           = self.F_idx,
+                F_hdim          = self.F_hdim,
+                F_dtype         = FWD_DTYPE_MAP[self.F_dtype],
+                F_bs            = self.F_tile.F_bs,
+                F_bsk           = self.F_tile.F_bsk,
+                F_bd            = self.F_tile.F_bd,
+                F_bdv           = self.F_tile.F_bdv,
+                F_vlayout       = LAYOUT_MAP[self.F_pipeline.F_vlayout],
+                F_spad          = BOOL_MAP[self.F_pipeline.F_spad],
+                F_skpad         = BOOL_MAP[self.F_pipeline.F_skpad],
+                F_dpad          = BOOL_MAP[self.F_pipeline.F_dpad],
+                F_dvpad         = BOOL_MAP[self.F_pipeline.F_dvpad],
+                F_rope          = ROPE_MAP[self.F_pipeline.F_rope],
+                F_pagedkv       = BOOL_MAP[self.F_pipeline.F_pagedkv],
+                F_occupancy     = self.F_tile.F_occupancy)
+
+    @property
+    def name(self) -> str:
+        # TODO: we don't encode idx here
+        return f"fmha_fwd_appendkv_d{self.F_hdim}_{self.F_dtype}_" + \
+                self.F_tile.name + '_' + self.F_pipeline.name
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+    def api_trait(self) -> FmhaFwdAppendKVApiTrait:
+        return FmhaFwdAppendKVApiTrait(
+                hdim=str(self.F_hdim),
+                dtype=self.F_dtype,
+                bs=self.F_tile.F_bs,
+                bsk=self.F_tile.F_bsk,
+                bd=self.F_tile.F_bd,
+                bdv=self.F_tile.F_bdv,
+                vlayout=self.F_pipeline.F_vlayout,
+                spad=self.F_pipeline.F_spad,
+                skpad=self.F_pipeline.F_skpad,
+                dpad=self.F_pipeline.F_dpad,
+                dvpad=self.F_pipeline.F_dvpad,
+                rope=self.F_pipeline.F_rope,
+                pagedkv=self.F_pipeline.F_pagedkv)
+
+# TODO: design a more practical way to do it
+# this is current supported tile size per hdim
+def get_fmha_fwd_appendkv_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
+    if dtype == 'fp16' or dtype == 'bf16':
+        return {
+            '32'  : FmhaFwdAppendKVTileSize(64, 64,  32,  32, -1),
+            '64'  : FmhaFwdAppendKVTileSize(64, 64,  64,  64, -1),
+            '128' : FmhaFwdAppendKVTileSize(64, 64, 128, 128, -1),
+            '256' : FmhaFwdAppendKVTileSize(64, 64, 256, 256, -1),
+        }
+    elif dtype == 'fp8' or dtype == 'bf8':
+        return {
+            '64'  : FmhaFwdAppendKVTileSize(64, 64,  64,  64, -1),
+            '128' : FmhaFwdAppendKVTileSize(64, 64, 128, 128, -1),
+            '256' : FmhaFwdAppendKVTileSize(64, 64, 256, 256, -1)
+        }
+    else:
+        return None
+
+def get_fwd_appendkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[FmhaFwdAppendKVApiPool, List[FmhaFwdAppendKVKernel]]:
+    # TODO: we don't support tuning yet, so pick up one value for vlayout/pipeline/pad
+    #       support this in future
+    def get_pipelines(dtype, hdim) -> List[FmhaFwdAppendKVPipeline]:
+        # this function will populate a list possible pipelines
+        # TODO: the order of List matters! the later in this list will be also be checked later
+        # TODO: currently for qr pipeline, let 't' padding to appear later!!
+        # TODO: how to design this more generic?
+        squant = 't' if dtype == 'fp8' else 'f'
+        pipelines = []
+        if dtype in ['fp16', 'bf16']:
+            # NOTICE: it will be very complicated if we consider all the hdim_q padding cases while
+            #         applying rotary embedding, so I just use 't' in inter/half pipelines
+            for vlayout in ['row', 'col']:
+                for pagedkv in ["t", "f"]:
+                    pipelines.append(FmhaFwdAppendKVPipeline(vlayout, 'f', 't', 'f', 'f', 'no', pagedkv))
+                    pipelines.append(FmhaFwdAppendKVPipeline(vlayout, 't', 't', 't', 't', 'no', pagedkv))
+
+                    pipelines.append(FmhaFwdAppendKVPipeline(vlayout, 'f', 't', 't', 'f', 'inter', pagedkv))
+                    pipelines.append(FmhaFwdAppendKVPipeline(vlayout, 't', 't', 't', 't', 'inter', pagedkv))
+
+                    pipelines.append(FmhaFwdAppendKVPipeline(vlayout, 'f', 't', 't', 'f', 'half', pagedkv))
+                    pipelines.append(FmhaFwdAppendKVPipeline(vlayout, 't', 't', 't', 't', 'half', pagedkv))
+        elif dtype in ['fp8', 'bf8']:
+            # rope/paged-kv is not supported
+            pipelines.append(FmhaFwdAppendKVPipeline('col', 't', 't', 't', 't', 'no', 'f'))
+        elif dtype in ['fp8fp16', 'fp8bf16']:
+            # TODO
+            None
+        else:
+            assert False
+        return pipelines
+
+    gen = list()
+    api_pool = FmhaFwdAppendKVApiPool(mask_impl)
+
+    for dtype in FWD_DTYPE_MAP.keys():
+        d = get_fmha_fwd_appendkv_tile_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        for hdim_str in d.keys():
+            tile = d[hdim_str]
+            hdim = int(hdim_str)
+            for pipeline in get_pipelines(dtype, hdim):
+                k = FmhaFwdAppendKVKernel(F_idx=0,
+                                  F_hdim=hdim,
+                                  F_dtype=dtype,
+                                  F_tile=tile,
+                                  F_pipeline=pipeline,
+                                  mask_impl=mask_impl)
+                if kernel_filter != '':
+                    if not fnmatch.fnmatch(k.name, kernel_filter):
+                        continue
+                # 2 - Flash attention integration
+                if receipt == 2:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= pipeline.F_vlayout == 'row'
+                    if not cond:
+                        continue
+                api_pool.register_traits(k.api_trait())
+                gen.append(k)
+
+    return (api_pool, gen)
+
+def write_single_kernel(kernel: FmhaFwdAppendKVKernel, autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_fwd_appendkv_api(api_pool : FmhaFwdAppendKVApiPool, autogen_dir: Path) -> None:
+    (autogen_dir / FMHA_FWD_APPENDKV_API_FILENAME).write_text(api_pool.api)
+
+def write_blobs(output_dir : Path, kernel_filter : Optional[str], receipt, optdim_list, mask_impl) -> None:
+    assert optdim_list == [-1]
+    api_pool, kernels = get_fwd_appendkv_blobs(kernel_filter, receipt, mask_impl)
+    for kernel in kernels:
+        write_single_kernel(kernel, output_dir)
+    write_fwd_appendkv_api(api_pool, output_dir)
+
+def list_blobs(file_path : Path, kernel_filter : Optional[str], receipt, optdim_list, mask_impl) -> None:
+    assert optdim_list == [-1]
+    with file_path.open('a') as f:
+        _, kernels = get_fwd_appendkv_blobs(kernel_filter, receipt, mask_impl)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_APPENDKV_API_FILENAME) + "\n")

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

@@ -0,0 +1,855 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import copy
+from dataclasses import dataclass
+import fnmatch
+import itertools
+from pathlib import Path
+from typing import List, Optional, Tuple, Union
+
+from codegen.cmake_config import *
+from codegen.cpp_symbol_map import *
+
+from codegen.ops.fmha_fwd import (
+    FmhaFwdTileSize,
+    FmhaFwdApiTrait,
+    FMHA_FWD_KERNEL_HEADER,
+    FMHA_FWD_API_PER_DTYPE,
+    FMHA_FWD_API_PER_HDIM_CASE,
+)
+
+
+DTYPE_BITS = {
+    "fp32": 32,
+    "fp16": 16,
+    "bf16": 16,
+    "fp8" : 8,
+    "bf8" : 8
+}
+
+K0_MAX_SUBMAX_MAP = {
+    32 : 32,
+    64 : 64,
+    96 : 128,
+    128: 128,
+    256: 256
+}
+
+FMHA_FWD_SPLITKV_PIPELINE_MAP = {
+    "qr" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVS",
+    "qr_nwarp_sshuffle" : "ck_tile::BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS",
+    "qr_async" : "ck_tile::BlockFmhaFwdSplitKVPipelineQRKSVSAsync",
+}
+
+FMHA_FWD_SPLITKV_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+using fmha_mask_{F_idx} = {F_mask};
+
+namespace {{
+template <bool kHasUnevenSplits, bool kMergeNumHeadGroupsSeqLenQ = false>
+struct instance {{
+using fmha_block_tile = ck_tile::sequence<{F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}>;
+
+using fmha_shape = ck_tile::TileFmhaShape<fmha_block_tile,
+                                          ck_tile::sequence<{F_rm0}, {F_rn0}, {F_rk0}>,
+                                          ck_tile::sequence<{F_wm0}, {F_wn0}, {F_wk0}>,
+                                          ck_tile::sequence<{F_rm1}, {F_rn1}, {F_rk1}>,
+                                          ck_tile::sequence<{F_wm1}, {F_wn1}, {F_wk1}>,
+                                          {F_vlayout}>;
+
+using fmha_trait = ck_tile::TileFmhaFwdSplitKVTraits<{F_spad},
+                                                     {F_skpad},
+                                                     {F_dpad},
+                                                     {F_dvpad},
+                                                     {F_bias},
+                                                     /*kHasBiasGrad=*/false,
+                                                     {F_lse},
+                                                     {F_squant},
+                                                     {F_pagedkv},
+                                                     kHasUnevenSplits,
+                                                     kMergeNumHeadGroupsSeqLenQ,
+                                                     {F_occupancy}>;
+
+using fmha_pipeline_problem = ck_tile::BlockFmhaFwdSplitKVPipelineProblem<
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::QDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::KDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::VDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::SMPLComputeDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::BiasDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::PDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
+    fmha_shape,
+    {F_mode},
+    fmha_mask_{F_idx},
+    fmha_trait>;
+
+using fmha_pipeline = {F_pipeline}<
+    fmha_pipeline_problem>;
+
+/// FIXME: use {F_spad}/{F_dvpad} as kPadM/kPadN parameters after solving
+///        store_tile_raw() data corruption issue
+using fmha_epilogue =
+    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
+                                           typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
+                                           false, false>>;
+
+using fmha_kernel =
+    ck_tile::FmhaFwdSplitKVKernel<fmha_pipeline, fmha_epilogue>;
+
+static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
+{{
+    using k_ = fmha_kernel;
+    auto [kargs, grids] = fmha_fwd_splitkv_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
+}}
+}};
+}}
+
+using trait_{F_idx} = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout},
+                        {F_pipeline_enum}, fmha_mask_{F_idx}, {F_bias}, {F_lse}, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad},
+                        {F_dvpad}>;
+
+#include <iostream>
+
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wtautological-compare"
+
+namespace {{
+template <bool kHasUnevenSplits>
+void run_instance(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a) {{
+    if constexpr ({F_hdim} == 128 && {F_bias} == ck_tile::BlockAttentionBiasEnum::NO_BIAS
+                  && (std::is_same_v<{F_mask}, ck_tile::SimplifiedGenericAttentionMask<false>>
+                      || std::is_same_v<{F_mask}, FmhaMasks::NoMask>)) {{
+        if (a.max_seqlen_q == 1 && a.nhead_k < a.nhead_q) {{
+            instance<kHasUnevenSplits, /*kMergeNumHeadGroupsSeqLenQ=*/true>::run(s, a);
+        }} else {{
+            instance<kHasUnevenSplits>::run(s, a);
+        }}
+    }} else {{
+        instance<kHasUnevenSplits>::run(s, a);
+    }}
+}}
+}} // anonymous namespace
+
+#pragma clang diagnostic pop
+
+template<>
+void fmha_fwd_splitkv_oneshot_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
+{{
+    if constexpr({F_mode} == false) {{ // batch mode
+        // we don't check every seqlen_k values for kvcache
+        if (a.seqlen_k_ptr != nullptr) {{
+            run_instance</*kHasUnevenSplits=*/true>(s, a);
+        // make sure F_bn0 is divisible by F_bk1
+        }} else if (a.seqlen_k % (a.num_splits * {F_bn0}) == 0) {{
+            run_instance</*kHasUnevenSplits=*/false>(s, a);
+        }} else {{
+            run_instance</*kHasUnevenSplits=*/true>(s, a);
+        }}
+    }} else {{
+        run_instance</*kHasUnevenSplits=*/true>(s, a);
+    }}
+}}
+
+template<>
+std::string fmha_fwd_splitkv_get_name_<trait_{F_idx}>()
+{{
+    using k_ = instance<true>::fmha_kernel; /// FIXME: choose real kernel type
+    return k_::GetName();
+}}
+"""
+
+FMHA_FWD_SPLITKV_COMBINE_KERNEL_BODY="""
+using fmha_dtype_{F_idx} = {F_dtype};
+
+namespace {{
+template <ck_tile::index_t kLogMaxSplits>
+struct instance {{
+using fmha_trait = ck_tile::TileFmhaFwdSplitKVCombineTraits<{F_spad},
+                                                    {F_dvpad},
+                                                    {F_lse},
+                                                    {F_squant},
+                                                    kLogMaxSplits,
+                                                    {F_occupancy}>;
+
+using fmha_pipeline_problem = ck_tile::BlockFmhaSplitKVCombinePipelineProblem<
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::LSEDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::OaccDataType,
+    typename FmhaFwdTypeConfig<fmha_dtype_{F_idx}>::ODataType,
+    {F_hdim},
+    {F_mode},
+    {F_bn1},
+    fmha_trait>;
+
+using fmha_pipeline = ck_tile::BlockFmhaFwdSplitKVCombinePipeline<
+    fmha_pipeline_problem>;
+
+/// FIXME: use {F_spad}/{F_dvpad} as kPadM/kPadN parameters after solving
+///        store_tile_raw() data corruption issue
+using fmha_epilogue =
+    ck_tile::Default2DEpilogue<ck_tile::Default2DEpilogueProblem<typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType,
+                                           typename FmhaFwdTypeConfig<{F_dtype}>::ODataType,
+                                           false, false>>;
+
+using fmha_kernel =
+    ck_tile::FmhaFwdSplitKVCombineKernel<fmha_pipeline, fmha_epilogue>;
+
+static void run(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
+{{
+    using k_ = fmha_kernel;
+    auto [kargs, grids] = fmha_fwd_splitkv_combine_create_kargs_and_grids<k_>(a);
+    constexpr dim3 blocks             = k_::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = k_::kBlockPerCu;
+    ck_tile::make_kernel<blocks.x, kBlockPerCu>(k_{{}}, grids, blocks, 0, kargs)(ck_tile::stream_config{{s.stream_id_}});
+}}
+}};
+}}
+
+using trait_{F_idx} = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bn1},
+                        {F_lse}, {F_squant}, {F_spad}, {F_dvpad}>;
+
+#include <iostream>
+
+template<>
+void fmha_fwd_splitkv_combine_oneshot_<trait_{F_idx}>(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
+{{
+    if (a.num_splits <= 8) {{
+        instance<3>::run(s, a);
+    }} else if (a.num_splits <= 16) {{
+        instance<4>::run(s, a);
+    }} else if (a.num_splits <= 32) {{
+        instance<5>::run(s, a);
+    }} else if (a.num_splits <= 64) {{
+        instance<6>::run(s, a);
+    }} else if (a.num_splits <= 128) {{
+        instance<7>::run(s, a);
+    }}
+}}
+
+template<>
+std::string fmha_fwd_splitkv_combine_get_name_<trait_{F_idx}>()
+{{
+    using k_ = instance<6>::fmha_kernel; /// FIXME: choose real kernel type
+    return k_::GetName();
+}}
+"""
+
+FMHA_FWD_SPLITKV_API_FILENAME="fmha_fwd_splitkv_api.cpp"
+FMHA_FWD_SPLITKV_API="""
+#include <iostream>
+
+template<typename fmha_fwd_splitkv_traits_, typename fmha_fwd_splitkv_combine_traits_>
+float fmha_fwd_splitkv_(const ck_tile::stream_config& s, fmha_fwd_splitkv_args a)
+{{
+    if(s.log_level_ > 0)
+    std::cout
+    << ", " << fmha_fwd_splitkv_get_name_<fmha_fwd_splitkv_traits_>()
+    << ", " << fmha_fwd_splitkv_combine_get_name_<fmha_fwd_splitkv_combine_traits_>()
+    << std::flush;
+
+    return ck_tile::launch_kernel(s,
+        [=](const ck_tile::stream_config& s_){{ fmha_fwd_splitkv_oneshot_<fmha_fwd_splitkv_traits_>(s_, a); }},
+        [=](const ck_tile::stream_config& s_){{ fmha_fwd_splitkv_combine_oneshot_<fmha_fwd_splitkv_combine_traits_>(s_, a); }}
+    );
+}}
+
+float fmha_fwd_splitkv(fmha_fwd_splitkv_traits t, fmha_fwd_splitkv_args a, const ck_tile::stream_config& s){{
+    float r = -1;
+{F_dispatch}
+    return r;
+}}
+"""
+
+FMHA_FWD_SPLITKV_API_INNER_DISPATCH="""            {F_if}((t.is_group_mode == {F_mode}) && (t.is_v_rowmajor == {F_vlayout}) && ({F_mask_check}) && (t.bias_type == {F_bias_check}) && (t.do_fp8_static_quant == {F_squant}) &&
+                        ((a.block_table_ptr != nullptr) == {F_pagedkv}) && ({F_scheck}) && ({F_skcheck}) && ({F_dcheck}) && ({F_dvcheck})) {{
+                using traits_ = fmha_fwd_splitkv_traits_<{F_hdim}, {F_dtype}, {F_mode}, {F_bm0}, {F_bn0}, {F_bk0}, {F_bn1}, {F_bk1}, {F_bk0max}, {F_vlayout}, {F_pipeline_enum}, {F_mask}, {F_bias}, true, {F_squant}, {F_pagedkv}, {F_spad}, {F_skpad}, {F_dpad}, {F_dvpad}>;
+
+                // get combine kernel tile sizes
+                using OaccDataType = typename FmhaFwdTypeConfig<{F_dtype}>::OaccDataType;
+                constexpr ck_tile::index_t kM0 = ck_tile::BlockFmhaSplitKVCombinePipelineTileSizes<OaccDataType, /*F_bn1=*/32>::kM0;
+
+                // make sure we can reuse the padding flags in combine kernels
+                static_assert({F_bm0} % kM0 == 0);
+                static_assert({F_bn1} % 32 == 0);
+
+                if (t.has_lse) {{
+                    if constexpr (std::is_same_v<{F_dtype}, FmhaFwdFp8>) {{
+                        return -1;
+                    }} else {{
+                        using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, /*F_bn1=*/32, true, {F_squant}, {F_spad}, {F_dvpad}>;
+
+                        return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
+                    }}
+                }} else {{
+                    using traits2_ = fmha_fwd_splitkv_combine_traits_<{F_hdim}, {F_dtype}, {F_mode}, /*F_bn1=*/32, false, {F_squant}, {F_spad}, {F_dvpad}>;
+
+                    return fmha_fwd_splitkv_<traits_, traits2_>(s, a);
+                }}
+            }}
+"""
+
+@dataclass
+class FmhaFwdSplitKVApiTrait:
+    pipeline_tag : str
+    # sync with fmha_fwd_traits<>, to generate fallback calls
+    hdim      : str
+    dtype     : str  # data type
+    mode      : str  # value from MODE_MAP
+    bm0       : int  # tile size along q seqlen (block size)
+    bn0       : int  # tile size along qk seqlen
+    bk0       : int  # tile size along qk gemm unroll
+    bn1       : int  # tile size along v head_dim
+    bk1       : int  # tile size along kv gemm unroll
+    bk0max    : int
+    vlayout   : str
+    mask      : str
+    bias      : str  #
+    lse       : str  #
+    squant    : str  #
+    spad      : str
+    skpad     : str
+    dpad      : str
+    dvpad     : str
+    pagedkv   : str
+
+    @property
+    def name(self) -> str:
+        return f'{self.hdim}-{self.dtype}-{self.mode}-{self.bm0}-{self.bn0}-{self.bk0}-{self.bn0}-{self.bk1}-{self.bk0max}-'+\
+                    f'{self.vlayout}-{self.mask}-{self.bias}-{self.lse}-{self.squant}-{self.spad}-{self.skpad}-{self.dpad}-'+\
+                    f'{self.dvpad}-{self.pagedkv}'
+
+    @property
+    def scheck(self) -> str:
+        if self.mode == 'group': return 'true/*group mode spad always true*/'                  # group mode only generate spad/skpad == true
+        if self.pipeline_tag == 'qr_async':
+            if self.spad == 't' : return 'true' # always support
+            else :                return 'true'
+        elif self.pipeline_tag in ['qr', 'qr_nwarp_sshuffle']:
+            if self.spad == 't' : return f'true /*a.seqlen_q % {self.bm0} != 0*/'  # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.seqlen_q % {self.bm0} == 0'
+        else: assert False
+
+    @property
+    def skcheck(self) -> str:
+        if self.mode == 'group': return 'true/*group mode skpad always true*/'                  # group mode only generate spad/skpad == true
+        if self.pipeline_tag == 'qr_async':
+            if self.skpad == 't' : return f'a.seqlen_k == 0 || a.seqlen_k % {self.bn0} != 0'
+            else :                 return f'a.seqlen_k != 0 && a.seqlen_k % {self.bn0} == 0'
+        elif self.pipeline_tag in ['qr', 'qr_nwarp_sshuffle']:
+            if self.skpad == 't' : return f'true /*a.seqlen_k % {self.bn0} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.seqlen_k % {self.bn0} == 0'
+        else: assert False
+
+    @property
+    def dcheck(self) -> str:
+        if self.pipeline_tag == 'qr_async':
+            vec = int((32 * 4) / DTYPE_BITS[self.dtype])
+            if self.dpad == 't': return f'a.hdim_q % {vec} == 0'
+            else :               assert False
+        elif self.pipeline_tag in ['qr', 'qr_nwarp_sshuffle']:
+            bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
+            if self.dpad == 't': return f'true /*a.hdim_q % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :               return f'a.hdim_q % {bk0submax} == 0'
+        else:   assert False
+
+    @property
+    def dvcheck(self) -> str:
+        if self.pipeline_tag == 'qr_async':
+            vec = int((32 * 4) / DTYPE_BITS[self.dtype])
+            if self.dvpad == 't': return f'a.hdim_v % {vec} == 0'
+            else :                assert False
+        elif self.pipeline_tag in ['qr', 'qr_nwarp_sshuffle']:
+            bk0submax = K0_MAX_SUBMAX_MAP[self.bk0max]
+            if self.dvpad == 't': return f'true /*a.hdim_v % {bk0submax} != 0*/' # TODO: order of get_pipelines() matters! (ugly)
+            else :                return f'a.hdim_v % {bk0submax} == 0'
+        else:   assert False
+
+@dataclass
+class FmhaFwdSplitKVPipeline:
+    tag : str
+
+    F_vlayout   : str  # row/col
+    F_spad      : str  # true/false
+    F_skpad     : str  #
+    F_dpad      : str  #
+    F_dvpad     : str  #
+    F_bias      : str  # true/false
+    F_lse       : str  #
+    F_squant    : str  #
+    F_pagedkv   : str  # t/f
+    F_mask      : str  # value from MASK_MAP
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_skpad == 't' : n += 'sk'
+            if self.F_dpad == 't' : n += 'd'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f'{self.tag}_v{self.F_vlayout[0]}'
+        if pn != '' : n += f'_{pn}'
+        else: n += '_npad'
+
+        if self.F_bias != 'no' : n += f'_{self.F_bias}'
+        else: n += '_nbias'
+
+        if self.F_mask[0:2] == 's_':
+            if self.F_mask == 's_mask': n += f'_mask'
+            else: n += '_nmask'
+        else:
+            if self.F_mask != 'no' : n += f'_m{self.F_mask[0]}'
+            else: n += '_nmask'
+
+        if self.F_lse == 't' : n += '_lse'
+        else: n += '_nlse'
+
+        if self.F_squant == 't' : n += '_squant'
+        else: n += '_nsquant'
+
+        if self.F_pagedkv == 't' : n += '_pagedkv'
+        else: n += '_npagedkv'
+        return n
+
+@dataclass
+class FmhaFwdSplitKVCombinePipeline:
+    tag : str
+
+    F_spad      : str  # true/false
+    F_dvpad     : str  #
+    F_lse       : str  #
+    F_squant    : str  #
+
+    @property
+    def name(self) -> str:
+        def pad_name() -> str:
+            n = ''
+            if self.F_spad == 't': n += 's'
+            if self.F_dvpad == 't' : n += 'dv'
+            if n != '' : n = 'p' + n
+            return n
+        pn = pad_name()
+        n = f'{self.tag}'
+        if pn != '' : n += f'_{pn}'
+        else: n += '_npad'
+
+        if self.F_lse == 't' : n += '_lse'
+        else: n += '_nlse'
+
+        if self.F_squant == 't' : n += '_squant'
+        else: n += '_nsquant'
+        return n
+
+class FmhaFwdSplitKVApiPool:
+    def __init__(self, mask_impl):
+        self.pool = dict()
+        self.mask_impl = mask_impl
+
+    def register_traits(self, trait : FmhaFwdSplitKVApiTrait) -> None:
+        # TODO: do we need to check duplication?
+        if trait.dtype not in self.pool.keys():
+            self.pool[trait.dtype] = dict()
+        if trait.hdim not in self.pool[trait.dtype].keys():
+            self.pool[trait.dtype][trait.hdim] = list()
+
+        self.pool[trait.dtype][trait.hdim].append(copy.copy(trait))
+
+    @property
+    def api(self) -> str:
+        per_dtypes=str()
+        for i, dtype in enumerate(self.pool.keys()):
+            per_hdim_case=str()
+            for j, hdim in enumerate(self.pool[dtype].keys()):
+                traits=self.pool[dtype][hdim]
+                inners=str()
+                for k, trait in enumerate(traits):
+                    if_k = 'if' if k == 0 else 'else if'
+                    inners = inners + FMHA_FWD_SPLITKV_API_INNER_DISPATCH.format(F_if=if_k, F_mode=MODE_MAP[trait.mode], F_vlayout=LAYOUT_MAP[trait.vlayout],
+                                   F_pipeline_enum=PIPELINE_ENUM_MAP[trait.pipeline_tag], F_mask=get_mask_map(self.mask_impl)[trait.mask],
+                                   F_mask_check=get_mask_check_map(self.mask_impl)[trait.mask], F_bias_check=BIAS_CHECK_MAP[trait.bias], F_bias=BIAS_MAP[trait.bias],
+                                   F_lse=BOOL_MAP[trait.lse], F_squant=BOOL_MAP[trait.squant], F_pagedkv=BOOL_MAP[trait.pagedkv],
+                                   F_scheck=trait.scheck, F_skcheck=trait.skcheck, F_dcheck=trait.dcheck, F_dvcheck=trait.dvcheck,
+                                   F_spad=BOOL_MAP[trait.spad], F_skpad=BOOL_MAP[trait.skpad], F_dpad=BOOL_MAP[trait.dpad], F_dvpad=BOOL_MAP[trait.dvpad],
+                                   F_bm0=trait.bm0, F_bn0=trait.bn0, F_bk0=trait.bk0, F_bn1=trait.bn1, F_bk1=trait.bk1, F_bk0max=trait.bk0max,
+                                   F_hdim=hdim, F_dtype=FWD_DTYPE_MAP[dtype])
+                if_j = 'if' if j == 0 else 'else if'
+                per_hdim_case = per_hdim_case + FMHA_FWD_API_PER_HDIM_CASE.format(F_if=if_j, F_hdim=hdim, F_hdim_v=hdim, F_inner_dispatch=inners)
+            if_i = 'if' if i == 0 else 'else if'
+            per_dtypes = per_dtypes + FMHA_FWD_API_PER_DTYPE.format(F_if=if_i, F_dtype=dtype, F_hdim_case=per_hdim_case)
+        if not per_dtypes:
+            # empty string we add some ignore to suppress warning in api
+            per_dtypes += '    (void)t ; (void)s ; (void)a;'
+        return FMHA_FWD_KERNEL_HEADER + FMHA_FWD_SPLITKV_API.format(F_dispatch = per_dtypes)
+
+@dataclass
+class FmhaFwdSplitKVCombineTileSize:
+    F_bn1       : int  # tile size along v head_dim
+    F_occupancy : int  # occupancy, -1 will let pipeline decide the occupancy, other value will overwrite occupancy
+    @property
+    def name(self) -> str:
+        return f"b{self.F_bn1}" +\
+            ("" if self.F_occupancy == -1 else f"_o{self.F_occupancy}")
+
+@dataclass
+class FmhaFwdSplitKVKernel:
+    F_idx           : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim          : int  # hdim
+    F_dtype         : str  # data type
+    F_mode          : str  # value from MODE_MAP
+    F_tile          : FmhaFwdTileSize
+    F_pipeline      : FmhaFwdSplitKVPipeline
+    mask_impl       : str
+
+    @property
+    def template(self) -> str:
+        kernel_body = str()
+        return FMHA_FWD_KERNEL_HEADER + \
+            FMHA_FWD_SPLITKV_KERNEL_BODY.format(
+                F_idx           = self.F_idx,
+                F_hdim          = self.F_hdim,
+                F_dtype         = FWD_DTYPE_MAP[self.F_dtype],
+                F_bm0           = self.F_tile.F_bm0,
+                F_bn0           = self.F_tile.F_bn0,
+                F_bk0           = self.F_tile.F_bk0,
+                F_bn1           = self.F_tile.F_bn1,
+                F_bk1           = self.F_tile.F_bk1,
+                F_bk0max        = self.F_tile.F_bk0max,
+                F_rm0           = self.F_tile.F_rm0,
+                F_rn0           = self.F_tile.F_rn0,
+                F_rk0           = self.F_tile.F_rk0,
+                F_rm1           = self.F_tile.F_rm1,
+                F_rn1           = self.F_tile.F_rn1,
+                F_rk1           = self.F_tile.F_rk1,
+                F_wm0           = self.F_tile.F_wm0,
+                F_wn0           = self.F_tile.F_wn0,
+                F_wk0           = self.F_tile.F_wk0,
+                F_wm1           = self.F_tile.F_wm1,
+                F_wn1           = self.F_tile.F_wn1,
+                F_wk1           = self.F_tile.F_wk1,
+                F_vlayout       = LAYOUT_MAP[self.F_pipeline.F_vlayout],
+                F_spad          = BOOL_MAP[self.F_pipeline.F_spad],
+                F_skpad         = BOOL_MAP[self.F_pipeline.F_skpad],
+                F_dpad          = BOOL_MAP[self.F_pipeline.F_dpad],
+                F_dvpad         = BOOL_MAP[self.F_pipeline.F_dvpad],
+                F_bias          = BIAS_MAP[self.F_pipeline.F_bias],
+                F_lse           = BOOL_MAP[self.F_pipeline.F_lse],
+                F_squant        = BOOL_MAP[self.F_pipeline.F_squant],
+                F_pagedkv       = BOOL_MAP[self.F_pipeline.F_pagedkv],
+                F_occupancy     = self.F_tile.F_occupancy,
+                F_pipeline_enum = PIPELINE_ENUM_MAP[self.F_pipeline.tag],
+                F_mask          = get_mask_map(self.mask_impl)[self.F_pipeline.F_mask],
+                F_mode          = MODE_MAP[self.F_mode],
+                F_pipeline      = FMHA_FWD_SPLITKV_PIPELINE_MAP[self.F_pipeline.tag])
+
+    @property
+    def name(self) -> str:
+        # TODO: we don't encode idx here
+        return f"fmha_fwd_splitkv_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + \
+                self.F_tile.name + '_' + self.F_pipeline.name
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+    def api_trait(self) -> FmhaFwdSplitKVApiTrait:
+        return FmhaFwdSplitKVApiTrait(
+                pipeline_tag=self.F_pipeline.tag,
+                hdim=str(self.F_hdim),
+                dtype=self.F_dtype,
+                mode=self.F_mode,
+                bm0=self.F_tile.F_bm0,
+                bn0=self.F_tile.F_bn0,
+                bk0=self.F_tile.F_bk0,
+                bn1=self.F_tile.F_bn1,
+                bk1=self.F_tile.F_bk1,
+                bk0max=self.F_tile.F_bk0max,
+                vlayout=self.F_pipeline.F_vlayout,
+                mask=self.F_pipeline.F_mask,
+                bias=self.F_pipeline.F_bias,
+                lse=self.F_pipeline.F_lse,
+                squant=self.F_pipeline.F_squant,
+                pagedkv=self.F_pipeline.F_pagedkv,
+                spad=self.F_pipeline.F_spad,
+                skpad=self.F_pipeline.F_skpad,
+                dpad=self.F_pipeline.F_dpad,
+                dvpad=self.F_pipeline.F_dvpad)
+
+@dataclass
+class FmhaFwdSplitKVCombineKernel:
+    F_idx           : int  # this is not a tunable, but a counter to differentiate symbol
+    F_hdim          : int  # hdim
+    F_dtype         : str  # data type
+    F_mode          : str  # value from MODE_MAP
+    F_tile          : FmhaFwdSplitKVCombineTileSize
+    F_pipeline      : FmhaFwdSplitKVCombinePipeline
+
+    @property
+    def template(self) -> str:
+        kernel_body = str()
+        return FMHA_FWD_KERNEL_HEADER + \
+            FMHA_FWD_SPLITKV_COMBINE_KERNEL_BODY.format(
+                F_idx           = self.F_idx,
+                F_hdim          = self.F_hdim,
+                F_dtype         = FWD_DTYPE_MAP[self.F_dtype],
+                F_bn1           = self.F_tile.F_bn1,
+                F_spad          = BOOL_MAP[self.F_pipeline.F_spad],
+                F_dvpad         = BOOL_MAP[self.F_pipeline.F_dvpad],
+                F_lse           = BOOL_MAP[self.F_pipeline.F_lse],
+                F_squant        = BOOL_MAP[self.F_pipeline.F_squant],
+                F_occupancy     = self.F_tile.F_occupancy,
+                F_mode          = MODE_MAP[self.F_mode])
+
+    @property
+    def name(self) -> str:
+        # TODO: we don't encode idx here
+        return f"fmha_fwd_splitkv_combine_d{self.F_hdim}_{self.F_dtype}_{self.F_mode}_" + \
+                self.F_tile.name + '_' + self.F_pipeline.name
+
+    @property
+    def filename(self) -> str:
+        return self.name + ".cpp"
+
+# TODO: design a more practical way to do it
+# this is current supported tile size per hdim
+def get_fmha_fwd_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
+    if dtype == 'fp16' or dtype == 'bf16':
+        return {
+            '32'  : FmhaFwdTileSize(32, 64,  16, 32,  32,  32,   2, 1, 1,  2, 1, 1,  16, 16, 16,  16, 16, 16,  -1),
+            '64'  : FmhaFwdTileSize(64, 64,  32, 64,  32,  64,   4, 1, 1,  4, 1, 1,  16, 16, 16,  16, 16, 16,  -1),
+        ### '96'  : FmhaFwdTileSize(64, 128, 32, 128, 32,  96,   4, 1, 1,  4, 1, 1,  16, 16, 16,  16, 16, 16,  -1),
+            '128' : FmhaFwdTileSize(64, 128, 32, 128, 32,  128,  4, 1, 1,  4, 1, 1,  16, 16, 16,  16, 16, 16,  -1),
+            '256' : FmhaFwdTileSize(64, 128, 32, 256, 32,  256,  4, 1, 1,  4, 1, 1,  16, 16, 16,  16, 16, 16,  -1),
+        }
+    elif dtype == 'fp8' or dtype == 'bf8':
+        return {
+            '64'  : FmhaFwdTileSize(128, 64,  32, 64,  32,  64,   2, 1, 1,  2, 1, 1,  32, 32, 32,  32, 32, 32,  -1),
+            '128' : FmhaFwdTileSize(128, 128, 32, 128, 32,  128,  4, 1, 1,  4, 1, 1,  32, 32, 32,  32, 32, 32,  -1),
+            '256' : FmhaFwdTileSize(128, 128, 32, 256, 32,  256,  4, 1, 1,  4, 1, 1,  32, 32, 32,  32, 32, 32,  -1),
+        }
+    else:
+        return None
+
+def get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype : str) -> Optional[dict]:
+    if dtype == 'fp16' or dtype == 'bf16':
+        return {
+            '32'  : FmhaFwdSplitKVCombineTileSize(32,  -1),
+            '64'  : FmhaFwdSplitKVCombineTileSize(32,  -1),
+        ### '96'  : FmhaFwdSplitKVCombineTileSize(32,  -1),
+            '128' : FmhaFwdSplitKVCombineTileSize(32,  -1),
+            '256' : FmhaFwdSplitKVCombineTileSize(32,  -1),
+    }
+    elif dtype == 'fp8' or dtype == 'bf8':
+        return {
+            '64'  : FmhaFwdSplitKVCombineTileSize(32,  -1),
+            '128' : FmhaFwdSplitKVCombineTileSize(32,  -1),
+            '256' : FmhaFwdSplitKVCombineTileSize(32,  -1),
+        }
+    else:
+        return None
+
+def get_fwd_splitkv_blobs(kernel_filter : Optional[str], receipt, mask_impl) -> Tuple[FmhaFwdSplitKVApiPool, List[FmhaFwdSplitKVKernel]]:
+    Pipeline = FmhaFwdSplitKVPipeline
+    Kernel = FmhaFwdSplitKVKernel
+
+    # TODO: we don't support tuning yet, so pick up one value for vlayout/pipeline/pad
+    #       support this in future
+    def get_pipelines(dtype, hdim) -> List[FmhaFwdSplitKVPipeline]:
+        # this function will populate a list possible pipelines
+        # TODO: the order of List matters! the later in this list will be also be checked later
+        # TODO: currently for qr pipeline, let 't' padding to appear later!!
+        # TODO: how to design this more generic?
+        squant = 't' if dtype == 'fp8' else 'f'
+        pipelines = []
+        if dtype in ['fp16', 'bf16']:
+            for mask, bias, pagedkv in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys(), ["t", "f"]):
+                # TODO: use async pipeline when compiler is more stable
+                if hdim == 256 or hdim in [32, 64, 128]:         ### [32, 64, 96, 128]:
+                # if True:
+                    pipelines.append(Pipeline('qr', 'row', 'f', 't', 'f', 'f', bias, 't', squant, pagedkv, mask))
+                    pipelines.append(Pipeline('qr', 'col', 'f', 't', 'f', 'f', bias, 't', squant, pagedkv, mask))
+
+                    pipelines.append(Pipeline('qr', 'row', 't', 'f', 'f', 'f', bias, 't', squant, pagedkv, mask))
+                    pipelines.append(Pipeline('qr', 'col', 't', 'f', 'f', 'f', bias, 't', squant, pagedkv, mask))
+
+                    pipelines.append(Pipeline('qr', 'row', 't', 't', 'f', 'f', bias, 't', squant, pagedkv, mask))
+                    pipelines.append(Pipeline('qr', 'col', 't', 't', 'f', 'f', bias, 't', squant, pagedkv, mask))
+
+                    pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask))
+                    pipelines.append(Pipeline('qr', 'col', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask))
+                else:
+                    pipelines.append(Pipeline('qr_async', 'row', 't', 'f', 't', 't', bias, 't', squant, pagedkv, mask))
+                    pipelines.append(Pipeline('qr_async', 'row', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask))
+                    pipelines.append(Pipeline('qr_async', 'col', 't', 'f', 't', 't', bias, 't', squant, pagedkv, mask))
+                    pipelines.append(Pipeline('qr_async', 'col', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask))
+                    if receipt == 1:
+                        pipelines.append(Pipeline('qr', 'row', 't', 't', 't', 't', bias, 't', squant, pagedkv, mask)) # TODO: cover arbitraty hdim
+                        pipelines.append(Pipeline('qr', 'col', 't', 'f', 't', 't', bias, 't', squant, pagedkv, mask)) # TODO: cover arbitraty hdim
+        elif dtype in ['fp8', 'bf8']:
+            for mask, bias in itertools.product(get_mask_map(mask_impl).keys(), BIAS_MAP.keys()):
+                pipelines.append(Pipeline('qr', 'col', 'f', 'f', 'f', 'f', bias, 't', squant, 'f', mask))
+        elif dtype in ['fp8fp16', 'fp8bf16']:
+            # TODO
+            None
+        else:
+            assert False
+        return pipelines
+
+    gen = list()
+    api_pool = FmhaFwdSplitKVApiPool(mask_impl)
+
+    for dtype in FWD_DTYPE_MAP.keys():
+        d = get_fmha_fwd_tile_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        #for hdim_str, mode, mask, bias, lse in itertools.product(d.keys(), MODE_MAP.keys(), MASK_MAP.keys(), ["t", "f"], ["t", "f"]):
+        for hdim_str, mode in itertools.product(d.keys(), MODE_MAP.keys()):
+            tile = d[hdim_str]
+            hdim = int(hdim_str)
+            for pipeline in get_pipelines(dtype, hdim):
+                if mode == "group":
+                    if pipeline.F_spad != 't' or pipeline.F_skpad != 't':
+                        # in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
+                        continue
+                k = Kernel(F_idx=0,
+                           F_hdim=hdim,
+                           F_dtype=dtype,
+                           F_mode=mode,
+                           F_tile=tile,
+                           F_pipeline=pipeline,
+                           mask_impl=mask_impl)
+                if kernel_filter != '':
+                    if not fnmatch.fnmatch(k.name, kernel_filter):
+                        continue
+                # Flash attention integration
+                if receipt == 2:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_bias in ['no', 'alibi']
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                # Aiter(mha_varlen_fwd) integration
+                elif receipt == 200:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == "group"
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                # aiter::mha_fwd_splikv C++ api integration
+                elif receipt == 600:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= pipeline.F_vlayout == 'row'
+                    cond &= pipeline.F_squant == 'f'
+                    if not cond:
+                        continue
+                api_pool.register_traits(k.api_trait())
+                gen.append(k)
+
+    return (api_pool, gen)
+
+def get_fwd_splitkv_combine_blobs(kernel_filter : Optional[str], receipt) -> List[FmhaFwdSplitKVCombineKernel]:
+    Pipeline = FmhaFwdSplitKVCombinePipeline
+    Kernel = FmhaFwdSplitKVCombineKernel
+
+    # TODO: we don't support tuning yet, so pick up one value for vlayout/pipeline/pad
+    #       support this in future
+    def get_pipelines(dtype, hdim) -> List[FmhaFwdSplitKVCombinePipeline]:
+        # this function will populate a list possible pipelines
+        # TODO: the order of List matters! the later in this list will be also be checked later
+        # TODO: currently for qr pipeline, let 't' padding to appear later!!
+        # TODO: how to design this more generic?
+        squant = 't' if dtype == 'fp8' else 'f'
+        pipelines = []
+        if dtype in ['fp16', 'bf16']:
+            for spad, dvpad, lse in itertools.product(["t", "f"], ["t", "f"], ["t", "f"]):
+                pipelines.append(Pipeline('unused', spad, dvpad, lse, squant))
+        elif dtype in ['fp8', 'bf8']:
+            # no need lse kernels
+            pipelines.append(Pipeline('unused', 'f', 'f', 'f', squant))
+        else:
+            assert False
+        return pipelines
+
+    gen = list()
+
+    for dtype in FWD_DTYPE_MAP.keys():
+        d = get_fmha_fwd_splitkv_combine_tile_dict_from_dtype(dtype)
+        if d == None:
+            continue
+        #for hdim_str, mode, mask, bias, lse in itertools.product(d.keys(), MODE_MAP.keys(), MASK_MAP.keys(), ["t", "f"], ["t", "f"]):
+        for hdim_str, mode in itertools.product(d.keys(), MODE_MAP.keys()):
+            tile = d[hdim_str]
+            hdim = int(hdim_str)
+            for pipeline in get_pipelines(dtype, hdim):
+                if mode == "group":
+                    if pipeline.F_spad != 't':
+                        # in group mode, spad/skpad must be true, since we can't predict if seqlen of current batch need pad or not
+                        continue
+                k = Kernel(F_idx=0,
+                           F_hdim=hdim,
+                           F_dtype=dtype,
+                           F_mode=mode,
+                           F_tile=tile,
+                           F_pipeline=pipeline)
+                if kernel_filter != '':
+                    if not fnmatch.fnmatch(k.name, kernel_filter):
+                        continue
+                # Aiter(mha_varlen_fwd) integration
+                if receipt == 200:
+                    cond = dtype in ['fp16', 'bf16']
+                    cond &= mode == "group"
+                    if not cond:
+                        continue
+                # aiter::mha_fwd_splikv C++ api integration
+                elif receipt == 600:
+                    cond = dtype in ['fp16', 'bf16']
+                    if not cond:
+                        continue
+                gen.append(k)
+
+    return gen
+
+def write_single_kernel(kernel: Union[FmhaFwdSplitKVKernel, FmhaFwdSplitKVCombineKernel], autogen_dir: Path) -> None:
+    (autogen_dir / kernel.filename).write_text(kernel.template)
+
+def write_fwd_splitkv_api(api_pool : FmhaFwdSplitKVApiPool, autogen_dir: Path) -> None:
+    file_path = autogen_dir / FMHA_FWD_SPLITKV_API_FILENAME
+    file_path.write_text(api_pool.api)
+
+def write_blobs(output_dir : Path, filter_list : str, receipt, optdim_list, mask_impl) -> None:
+    filter_list = filter_list.split('@')
+    filter_list.extend([''] * (2 - len(filter_list)))
+    assert optdim_list == [-1]
+
+    kernels = get_fwd_splitkv_combine_blobs(filter_list[0], receipt)
+    for kernel in kernels:
+        write_single_kernel(kernel, output_dir)
+    api_pool, kernels = get_fwd_splitkv_blobs(filter_list[1], receipt, mask_impl)
+    for kernel in kernels:
+        write_single_kernel(kernel, output_dir)
+    write_fwd_splitkv_api(api_pool, output_dir)
+
+def list_blobs(file_path : Path, filter_list : str, receipt, optdim_list, mask_impl) -> None:
+    filter_list = filter_list.split('@')
+    filter_list.extend([''] * (2 - len(filter_list)))
+    assert optdim_list == [-1]
+
+    with file_path.open('a') as f:
+        kernels = get_fwd_splitkv_combine_blobs(filter_list[0], receipt)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        _, kernels = get_fwd_splitkv_blobs(filter_list[1], receipt, mask_impl)
+        for kernel in kernels:
+            f.write(str(file_path.parent / GEN_DIR / kernel.filename) + "\n")
+        f.write(str(file_path.parent / GEN_DIR / FMHA_FWD_SPLITKV_API_FILENAME) + "\n")

example/ck_tile/01_fmha/fmha_bwd.cpp

@@ -0,0 +1,997 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "fmha_bwd.hpp"
+#include "ck_tile/host.hpp"
+#include "mask.hpp"
+#include "utils.hpp"
+
+#include <array>
+#include <cstring>
+#include <functional>
+#include <numeric>
+#include <ostream>
+#include <string>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+template <typename T>
+std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
+{
+    using size_type = typename std::vector<T>::size_type;
+
+    os << "[";
+    for(size_type idx = 0; idx < v.size(); ++idx)
+    {
+        if(0 < idx)
+        {
+            os << ", ";
+        }
+        os << v[idx];
+    }
+    return os << "]";
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("v", "1", "weather do CPU validation or not")
+        .insert("mode", "0", "kernel mode. 0:batch, 1:group")
+        .insert("b", "2", "batch size")
+        .insert("h", "8", "num of head, for q")
+        .insert("h_k",
+                "-1",
+                "num of head, for k/v, -1 means equal to h\n"
+                "if not equal to h, then this is GQA/MQA case")
+        .insert("s",
+                "3328",
+                "seqlen_q. if group-mode, means the average value of seqlen_q\n"
+                "total_seqlen_q = seqlen_q * batch, and seqlen_q per batch may vary")
+        .insert("s_k", "-1", "seqlen_k, -1 means equal to s")
+        .insert("d", "128", "head dim for q, k")
+        .insert("d_v", "-1", "head dim for v, -1 means equal to d")
+        .insert("scale", "0", "scale factor. 0 means equal to 1/sqrt(hdim)")
+        .insert("iperm",
+                "1",
+                "permute input\n"
+                "if true, will be b*h*s*d, else b*s*h*d")
+        .insert("operm", "1", "permute output")
+        .insert("bias",
+                "n",
+                "n or 0, no bias\n"
+                "e(lementwise) or 1, elementwise bias with 1*1*s*s. e:1, 1*h*s*s. e:2, b*h*s*s\n"
+                "a(libi) or 2, alibi with 1*h. a:1, b*h")
+        .insert("dbias", "0", "output bias gradient or not")
+        .insert("prec", "fp16", "data type. fp16 or bf16")
+        .insert("mask",
+                "0",
+                "0: no mask, 1: top-left(same as 't'), 2:bottom-right(same as 'b')\n"
+                "'t', top-left causal mask, 'b', bottom-r causal mask\n"
+                "'t:l,r', top-left sliding window attn(swa) with FA style left right size\n"
+                "'b:l,r', bottom-r sliding window attn(swa) with FA style left right size\n"
+                "'xt:window_size', xformer style masking from top-left, window_size negative is "
+                "causal, positive is swa\n"
+                "'xb:window_size', xformer style masking from bottom-r, window_size negative is "
+                "causal, positive is swa\n"
+                "'g:y,x', generic attention mask coordinate with y/x size (only debug purpose for "
+                "now)")
+        .insert("kname", "0", "if set to 1 will print kernel name")
+        .insert("init", "1", "init method. 0:random int, 1:random float, 2:trig float")
+        .insert("seed",
+                "11939",
+                "random seed used for initializing input tensors. 0 for "
+                "non-deterministic seed")
+        .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")
+        .insert("deterministic",
+                "0",
+                "if set to 1 will use multi-buffer reduction strategy for dq, atomic opeartion "
+                "will not be used");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+// different threshold for different dtype
+template <typename DataTypeConfig>
+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<FmhaBwdBf16>(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 DataTypeConfig>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    std::string data_type    = arg_parser.get_str("prec");
+    int do_validation        = arg_parser.get_int("v");
+    auto mode                = static_cast<mode_enum>(arg_parser.get_uint32("mode"));
+    ck_tile::index_t batch   = arg_parser.get_int("b");
+    ck_tile::index_t nhead   = arg_parser.get_int("h");
+    ck_tile::index_t nhead_k = arg_parser.get_int("h_k");
+    if(nhead_k < 0)
+        nhead_k = nhead;
+
+    if(nhead % nhead_k != 0)
+    {
+        std::cerr << "nhead:" << nhead << " must be multiple of nhead_k:" << nhead_k << std::endl;
+        return false;
+    }
+
+    ck_tile::index_t seqlen_q = arg_parser.get_int("s");
+    ck_tile::index_t seqlen_k = arg_parser.get_int("s_k");
+    if(seqlen_k < 0)
+        seqlen_k = seqlen_q;
+    ck_tile::index_t hdim_q = arg_parser.get_int("d");
+    ck_tile::index_t hdim_v = arg_parser.get_int("d_v");
+    if(hdim_v < 0)
+        hdim_v = hdim_q;
+
+    bool i_perm = arg_parser.get_bool("iperm"); // if true, will be batch * nhead * seqlen * hdim
+    bool o_perm = arg_parser.get_bool("operm"); // if false, will be batch * seqlen * nhead * hdim
+
+    float scale = arg_parser.get_float("scale");
+    if(scale == .0f)
+        scale = 1.0 / ck_tile::sqrt(static_cast<float>(hdim_q));
+
+    bias_info bias       = bias_info::decode(arg_parser.get_str("bias"));
+    bool use_dbias       = arg_parser.get_bool("dbias");
+    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;
+        return false;
+    }
+
+    if(p_drop < 0.0f || p_drop > 1.0f)
+    {
+        std::cerr << "The value of p_drop should be 0~1" << std::endl;
+        return false;
+    }
+    float p_undrop = 1.0 - p_drop;
+    uint8_t p_undrop_in_uint8_t =
+        uint8_t(std::floor(p_undrop * std::numeric_limits<uint8_t>::max()));
+    float rp_undrop = 1.0 / p_undrop;
+
+    bool s_randval = false;
+    if(p_drop > 0.0f && do_validation)
+    {
+        s_randval = true;
+    }
+
+    mask_info mask = mask_info::decode(arg_parser.get_str("mask"), seqlen_q, seqlen_k);
+
+    int init_method              = arg_parser.get_int("init");
+    std::optional<uint32_t> seed = arg_parser.get_uint32("seed");
+    if(*seed == 0)
+    {
+        seed.reset();
+    }
+
+    int stream_warmup  = arg_parser.get_int("warmup");
+    int stream_repeat  = arg_parser.get_int("repeat");
+    bool kname         = arg_parser.get_bool("kname");
+    bool deterministic = arg_parser.get_bool("deterministic");
+
+    ck_tile::stream_config stream_config{nullptr,
+                                         true,
+                                         /* log_level = */ (kname ? 1 : 0),
+                                         stream_warmup,
+                                         stream_repeat,
+                                         arg_parser.get_str("timer") == std::string("gpu")};
+
+    const auto seqstart_q_host = generate_seqstarts(mode, batch, seqlen_q);
+    const auto seqstart_k_host = generate_seqstarts(mode, batch, seqlen_k);
+
+    using TypeConfig = FmhaBwdTypeConfig<DataTypeConfig>;
+
+    using QDataType             = typename TypeConfig::QDataType;
+    using KDataType             = typename TypeConfig::KDataType;
+    using VDataType             = typename TypeConfig::VDataType;
+    using GemmDataType          = typename TypeConfig::GemmDataType;
+    using BiasDataType          = typename TypeConfig::BiasDataType;
+    using LSEDataType           = typename TypeConfig::LSEDataType;
+    using AccDataType           = typename TypeConfig::AccDataType;
+    using DDataType             = typename TypeConfig::DDataType;
+    using RandValOutputDataType = typename TypeConfig::RandValOutputDataType;
+    using ODataType             = typename TypeConfig::ODataType;
+    using OGradDataType         = typename TypeConfig::OGradDataType;
+    using QGradDataType         = typename TypeConfig::QGradDataType;
+    using KGradDataType         = typename TypeConfig::KGradDataType;
+    using VGradDataType         = typename TypeConfig::VGradDataType;
+    using BiasGradDataType      = typename TypeConfig::BiasGradDataType;
+
+    // accumulation numbers for performance evaluation
+    std::size_t flop = 0, num_byte = 0;
+    auto max_seqlen_q =
+        std::numeric_limits<int32_t>::min(); // we will use max seqlen to decide grid size
+    auto max_seqlen_k =
+        std::numeric_limits<int32_t>::min(); // we will use max seqlen to decide grid size
+    {
+        for(ck_tile::index_t wb = 0; wb < batch; ++wb)
+        {
+            const int32_t real_seqlen_q = seqstart_q_host[wb + 1] - seqstart_q_host[wb];
+            const int32_t real_seqlen_k = seqstart_k_host[wb + 1] - seqstart_k_host[wb];
+
+            if(max_seqlen_q < real_seqlen_q)
+            {
+                max_seqlen_q = real_seqlen_q;
+            }
+
+            if(max_seqlen_k < real_seqlen_k)
+            {
+                max_seqlen_k = real_seqlen_k;
+            }
+
+            flop += nhead * (static_cast<std::size_t>(3) * static_cast<std::size_t>(2) *
+                                 real_seqlen_q * real_seqlen_k * hdim_q + // Q@K/dS^T@Q^T/dS@K^T
+                             static_cast<std::size_t>(2) * static_cast<std::size_t>(2) *
+                                 real_seqlen_q * real_seqlen_k * hdim_v); // dO@V/P^T@dO^T
+
+            num_byte += nhead * (sizeof(QDataType) * real_seqlen_q * hdim_q +
+                                 sizeof(KDataType) * real_seqlen_k * hdim_q +
+                                 sizeof(VDataType) * real_seqlen_k * hdim_v +
+                                 sizeof(ODataType) * real_seqlen_q * hdim_v +
+                                 sizeof(OGradDataType) * real_seqlen_q * hdim_v +
+                                 sizeof(QGradDataType) * real_seqlen_q * hdim_q +
+                                 sizeof(KGradDataType) * real_seqlen_k * hdim_q +
+                                 sizeof(VGradDataType) * real_seqlen_k * hdim_v +
+                                 sizeof(LSEDataType) * real_seqlen_q);
+        }
+    }
+
+    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};
+    };
+
+    // host memory for storing all the tensor elements
+    const ck_tile::index_t shape_batch = (mode == mode_enum::batch ? batch : 1);
+    const ck_tile::index_t shape_seqlen_q =
+        (mode == mode_enum::batch ? seqlen_q : seqstart_q_host.back());
+    const ck_tile::index_t shape_seqlen_k =
+        (mode == mode_enum::batch ? seqlen_k : seqstart_k_host.back());
+    const ck_tile::index_t kN0 = (hdim_q <= 128) ? 128 : 64;
+    const ck_tile::index_t nsplits =
+        deterministic ? ck_tile::integer_divide_ceil(max_seqlen_k, kN0) : 1;
+
+    ck_tile::HostTensor<QDataType> q_host(
+        get_lengths(i_perm, shape_batch, nhead, shape_seqlen_q, hdim_q));
+    ck_tile::HostTensor<KDataType> k_host(
+        get_lengths(i_perm, shape_batch, nhead_k, shape_seqlen_k, hdim_q));
+    ck_tile::HostTensor<VDataType> v_host(
+        get_lengths(i_perm, shape_batch, nhead_k, shape_seqlen_k, hdim_v));
+    ck_tile::HostTensor<BiasDataType> bias_host(
+        bias.type == bias_enum::elementwise_bias
+            ? get_lengths(i_perm, 1, 1, shape_seqlen_q, max_seqlen_k)
+            : std::array<ck_tile::index_t, 4>{1, 1, 1, 1} /* dummy shape for simplifying code */);
+    ck_tile::HostTensor<AccDataType> alibi_slope_host(
+        bias.type == bias_enum::alibi
+            ? (bias.rank_info == 0 ? std::array<ck_tile::index_t, 2>{1, nhead}
+                                   : std::array<ck_tile::index_t, 2>{batch, nhead})
+            : std::array<ck_tile::index_t, 2>{1, 1});
+    ck_tile::HostTensor<ODataType> o_host(
+        get_lengths(o_perm, shape_batch, nhead, shape_seqlen_q, hdim_v));
+    ck_tile::HostTensor<LSEDataType> lse_host(
+        std::array<ck_tile::index_t, 3>{shape_batch, nhead, shape_seqlen_q});
+    ck_tile::HostTensor<DDataType> d_host(
+        std::array<ck_tile::index_t, 3>{shape_batch, nhead, shape_seqlen_q});
+    ck_tile::HostTensor<RandValOutputDataType> randval_host(
+        p_drop > 0 ? get_lengths(true, shape_batch, nhead, shape_seqlen_q, max_seqlen_k)
+                   : std::array<ck_tile::index_t, 4>{1, 1, 1, 1});
+    ck_tile::HostTensor<QGradDataType> dq_host(
+        get_lengths(i_perm, shape_batch, nhead, shape_seqlen_q, hdim_q));
+    ck_tile::HostTensor<KGradDataType> dk_host(
+        get_lengths(i_perm, shape_batch, nhead, shape_seqlen_k, hdim_q));
+    ck_tile::HostTensor<VGradDataType> dv_host(
+        get_lengths(i_perm, shape_batch, nhead, shape_seqlen_k, hdim_v));
+    ck_tile::HostTensor<OGradDataType> do_host(
+        get_lengths(o_perm, shape_batch, nhead, shape_seqlen_q, hdim_v));
+    ck_tile::HostTensor<BiasGradDataType> dbias_host(
+        use_dbias
+            ? get_lengths(i_perm, shape_batch, nhead, shape_seqlen_q, max_seqlen_k)
+            : std::array<ck_tile::index_t, 4>{1, 1, 1, 1} /* dummy shape for simplifying code */);
+    ck_tile::HostTensor<AccDataType> dq_acc_host(
+        i_perm
+            ? std::array<ck_tile::index_t, 5>{nsplits, shape_batch, nhead, shape_seqlen_q, hdim_q}
+            : std::array<ck_tile::index_t, 5>{nsplits, shape_batch, shape_seqlen_q, nhead, hdim_q});
+
+    if(init_method == 0)
+    {
+        ck_tile::FillUniformDistributionIntegerValue<QDataType>{-2.f, 2.f, seed}(q_host);
+        ck_tile::FillUniformDistributionIntegerValue<KDataType>{-2.f, 2.f, seed}(k_host);
+        ck_tile::FillUniformDistributionIntegerValue<VDataType>{-2.f, 2.f, seed}(v_host);
+        ck_tile::FillUniformDistributionIntegerValue<BiasDataType>{-2.f, 2.f, seed}(bias_host);
+        ck_tile::FillUniformDistributionIntegerValue<OGradDataType>{-2.f, 2.f, seed}(do_host);
+    }
+    else if(init_method == 1)
+    {
+        ck_tile::FillUniformDistribution<QDataType>{0.f, 1.f, seed}(q_host);
+        ck_tile::FillUniformDistribution<KDataType>{0.f, 1.f, seed}(k_host);
+        ck_tile::FillUniformDistribution<VDataType>{0.f, 1.f, seed}(v_host);
+        ck_tile::FillUniformDistribution<BiasDataType>{0.f, 1.f, seed}(bias_host);
+        ck_tile::FillUniformDistribution<OGradDataType>{0.f, 1.f, seed}(do_host);
+    }
+    else if(init_method == 2)
+    {
+        ck_tile::FillTrigValue<QDataType>{}(q_host);
+        ck_tile::FillTrigValue<KDataType>{}(k_host);
+        ck_tile::FillTrigValue<VDataType>{}(v_host);
+        ck_tile::FillTrigValue<BiasDataType>{}(bias_host);
+        ck_tile::FillTrigValue<OGradDataType>{}(do_host);
+    }
+    if(bias.type == bias_enum::alibi)
+    {
+        auto slopes = ck_tile::get_alibi_slopes<AccDataType>(nhead);
+        assert(slopes.size() == nhead);
+        if(bias.rank_info == 0)
+        {
+            // alibi in 1*h
+            std::copy(slopes.begin(), slopes.end(), alibi_slope_host.begin());
+        }
+        else
+        {
+            // alibi in b*h
+            for(auto i_b = 0; i_b < batch; i_b++)
+            {
+                std::copy(slopes.begin(), slopes.end(), alibi_slope_host.begin() + i_b * nhead);
+            }
+        }
+    }
+
+    ck_tile::DeviceMem q_buf(q_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem k_buf(k_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem v_buf(v_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem bias_buf(bias_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem o_buf(o_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem lse_buf(lse_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem d_buf(d_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem randval_buf(randval_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem dq_buf(dq_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem dk_buf(dk_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem dv_buf(dv_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem do_buf(do_host.get_element_space_size_in_bytes());
+    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());
+
+    q_buf.ToDevice(q_host.data());
+    k_buf.ToDevice(k_host.data());
+    v_buf.ToDevice(v_host.data());
+    bias_buf.ToDevice(bias_host.data());
+    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
+    auto layout_str = [&](bool permute){
+        if (permute) return std::string("bhsd");
+        else return std::string("bshd");
+    };
+    auto io_layout = [&](bool iperm_, bool operm_) {
+        if (iperm_ == operm_) return layout_str(iperm_);
+        else return layout_str(iperm_) + std::string("-") + layout_str(operm_);
+    };
+    // clang-format on
+    const std::string prec = arg_parser.get_str("prec");
+
+    std::cout << "[" << prec << "|" << mode << "|" << io_layout(i_perm, o_perm) << "] b:" << batch
+              << ", h:" << nhead << "/" << nhead_k << ", s:" << seqlen_q << "/" << seqlen_k
+              << ", d:" << hdim_q << "/" << hdim_v << ", scale:" << scale << ", bias:" << bias
+              << ", dbias:" << use_dbias << ", p_drop:" << p_drop << ", s_randval:" << s_randval
+              << ", deterministic:" << deterministic << ", mask:" << mask << std::flush;
+
+    std::size_t workspace_size =
+        dq_acc_host.get_element_space_size_in_bytes() * sizeof(AccDataType) / (1024 * 1024);
+
+    if(deterministic == 1)
+    {
+        std::cout << "\nDeterministic mode ON: " << workspace_size
+                  << " MByte memory workspace allocated" << std::endl;
+    }
+
+    auto fmha_traits = fmha_bwd_traits{hdim_q,
+                                       hdim_v,
+                                       data_type,
+                                       mode == mode_enum::group,
+                                       mask.type,
+                                       bias.type,
+                                       use_dbias,
+                                       p_drop > 0.0f,
+                                       s_randval,
+                                       deterministic};
+    auto fmha_args   = [&]() {
+        assert(nhead % nhead_k == 0);
+        /// NOTE: we broadcast bias from [1, 1, seqlen_q, seqlen_k] to [batch, nhead, seqlen_q,
+        ///       seqlen_k] in this example, hence both the 'batch_stride_bias' &
+        ///       'nhead_stride_bias' are 0.
+        // setup stride_* arguments
+        const ck_tile::index_t stride_q       = (i_perm ? hdim_q : nhead * hdim_q);
+        const ck_tile::index_t stride_k       = (i_perm ? hdim_q : nhead_k * hdim_q);
+        const ck_tile::index_t stride_v       = (i_perm ? hdim_v : nhead_k * hdim_v);
+        const ck_tile::index_t stride_bias    = (max_seqlen_k);
+        const ck_tile::index_t stride_o       = (o_perm ? hdim_v : nhead * hdim_v);
+        const ck_tile::index_t stride_randval = (max_seqlen_k);
+        const ck_tile::index_t stride_do      = (o_perm ? hdim_v : nhead * hdim_v);
+        const ck_tile::index_t stride_dk      = (i_perm ? hdim_q : nhead * hdim_q);
+        const ck_tile::index_t stride_dv      = (i_perm ? hdim_v : nhead * hdim_v);
+        const ck_tile::index_t stride_dbias   = (i_perm ? max_seqlen_k : nhead * max_seqlen_k);
+        // setup nhead_stride_* arguments
+        const ck_tile::index_t nhead_stride_q       = (i_perm ? shape_seqlen_q * hdim_q : hdim_q);
+        const ck_tile::index_t nhead_stride_k       = (i_perm ? shape_seqlen_k * hdim_q : hdim_q);
+        const ck_tile::index_t nhead_stride_v       = (i_perm ? shape_seqlen_k * hdim_v : hdim_v);
+        const ck_tile::index_t nhead_stride_bias    = 0;
+        const ck_tile::index_t nhead_stride_o       = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
+        const ck_tile::index_t nhead_stride_randval = (shape_seqlen_q * max_seqlen_k);
+        const ck_tile::index_t nhead_stride_do      = (o_perm ? shape_seqlen_q * hdim_v : hdim_v);
+        const ck_tile::index_t nhead_stride_lsed    = shape_seqlen_q;
+        const ck_tile::index_t nhead_stride_dbias =
+            (i_perm ? shape_seqlen_q * max_seqlen_k : max_seqlen_k);
+        // setup batch_stride_* arguments
+        const ck_tile::index_t batch_stride_q       = (nhead * shape_seqlen_q * hdim_q);
+        const ck_tile::index_t batch_stride_k       = (nhead_k * shape_seqlen_k * hdim_q);
+        const ck_tile::index_t batch_stride_v       = (nhead_k * shape_seqlen_k * hdim_v);
+        const ck_tile::index_t batch_stride_bias    = 0;
+        const ck_tile::index_t batch_stride_o       = (nhead * shape_seqlen_q * hdim_v);
+        const ck_tile::index_t batch_stride_randval = (nhead * shape_seqlen_q * max_seqlen_k);
+        const ck_tile::index_t batch_stride_do      = (nhead * shape_seqlen_q * hdim_v);
+        const ck_tile::index_t batch_stride_lsed    = (nhead * shape_seqlen_q);
+        const ck_tile::index_t batch_stride_dk      = (nhead * shape_seqlen_k * hdim_q);
+        const ck_tile::index_t batch_stride_dv      = (nhead * shape_seqlen_k * hdim_v);
+        const ck_tile::index_t batch_stride_dbias   = (nhead * shape_seqlen_q * max_seqlen_k);
+        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(),
+                             bias.type == bias_enum::alibi ? alibi_slope_buf.GetDeviceBuffer()
+                                                             : bias_buf.GetDeviceBuffer(),
+                             o_buf.GetDeviceBuffer(),
+                             lse_buf.GetDeviceBuffer(),
+                             do_buf.GetDeviceBuffer(),
+                             d_buf.GetDeviceBuffer(),
+                             randval_buf.GetDeviceBuffer(),
+                             dq_buf.GetDeviceBuffer(),
+                             dk_buf.GetDeviceBuffer(),
+                             dv_buf.GetDeviceBuffer(),
+                             dbias_buf.GetDeviceBuffer(),
+                             dq_acc_buf.GetDeviceBuffer(),
+                             seqstart_q.GetDeviceBuffer(),
+                             seqstart_k.GetDeviceBuffer(),
+                             nullptr,
+                             shape_seqlen_q,
+                             shape_seqlen_k,
+                             batch,
+                             max_seqlen_q,
+                             max_seqlen_k,
+                             hdim_q,
+                             hdim_v,
+                             nhead,
+                             nhead_k,
+                             scale,
+                             stride_q,
+                             stride_k,
+                             stride_v,
+                             bias.type == bias_enum::alibi ? (bias.rank_info == 0 ? 0 : nhead)
+                                                             : stride_bias,
+                             stride_o,
+                             stride_randval,
+                             stride_do,
+                             stride_q, // stride_dq_acc
+                             stride_q, // stride_dq
+                             stride_dk,
+                             stride_dv,
+                             stride_dbias,
+                             nhead_stride_q,
+                             nhead_stride_k,
+                             nhead_stride_v,
+                             nhead_stride_bias,
+                             nhead_stride_o,
+                             nhead_stride_randval,
+                             nhead_stride_do,
+                             nhead_stride_lsed,
+                             nhead_stride_q, // nhead_stride_dq_acc
+                             nhead_stride_q, // nhead_stride_dq
+                             nhead_stride_k, // nhead_stride_dk
+                             nhead_stride_v, // nhead_stride_dv
+                             nhead_stride_dbias,
+                             batch_stride_q,
+                             batch_stride_k,
+                             batch_stride_v,
+                             batch_stride_bias,
+                             batch_stride_o,
+                             batch_stride_randval,
+                             batch_stride_do,
+                             batch_stride_lsed,
+                             batch_stride_q, // batch_stride_dq_acc
+                             batch_stride_q, // batch_stride_dq
+                             batch_stride_dk,
+                             batch_stride_dv,
+                             batch_stride_dbias,
+                             split_stride_dq_acc,
+                             mask.left,
+                             mask.right,
+                             static_cast<ck_tile::index_t>(mask.type),
+                             p_drop,
+                             p_undrop,
+                             drop_seed_offset};
+    }();
+
+    float ave_time = fmha_bwd(fmha_traits, fmha_args, stream_config);
+    if(ave_time < 0)
+    {
+        std::cout << ", not supported yet" << std::flush << std::endl;
+        return false;
+    }
+
+    float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
+
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+    std::cout << std::fixed << ", " << std::setprecision(3) << ave_time << " ms, "
+              << std::setprecision(2) << tflops << " TFlops, " << std::setprecision(2) << gb_per_sec
+              << " GB/s" << std::flush;
+
+    if(!do_validation)
+    {
+        std::cout << std::flush << std::endl;
+        return true;
+    }
+
+    bool pass = true;
+
+    std::vector<ck_tile::HostTensor<QDataType>> q_host_refs;
+    std::vector<ck_tile::HostTensor<KDataType>> k_host_refs;
+    std::vector<ck_tile::HostTensor<VDataType>> v_host_refs;
+    std::vector<ck_tile::HostTensor<ODataType>> o_host_refs;
+    std::vector<ck_tile::HostTensor<RandValOutputDataType>> randval_host_refs;
+    std::vector<ck_tile::HostTensor<AccDataType>> p_hp_host_refs;
+    std::vector<ck_tile::HostTensor<GemmDataType>> p_lp_host_refs;
+
+    randval_buf.FromDevice(randval_host.data());
+
+    for(ck_tile::index_t wb = 0; wb < batch; ++wb)
+    {
+        const ck_tile::index_t real_seqlen_q = seqstart_q_host[wb + 1] - seqstart_q_host[wb];
+        const ck_tile::index_t real_seqlen_k = seqstart_k_host[wb + 1] - seqstart_k_host[wb];
+
+        // adjust matrix index according to the mode
+        const ck_tile::index_t b            = (mode == mode_enum::batch ? wb : 0);
+        const ck_tile::index_t query_offset = (mode == mode_enum::batch ? 0 : seqstart_q_host[wb]);
+        const ck_tile::index_t key_offset   = (mode == mode_enum::batch ? 0 : seqstart_k_host[wb]);
+
+        ck_tile::HostTensor<QDataType> q_host_ref({nhead, real_seqlen_q, hdim_q}); // q_g_m_k
+        ck_tile::HostTensor<KDataType> k_host_ref({nhead, real_seqlen_k, hdim_q}); // k_g_n_k
+        ck_tile::HostTensor<VDataType> v_host_ref({nhead, hdim_v, real_seqlen_k}); // v_g_o_n
+        ck_tile::HostTensor<ODataType> o_host_ref({nhead, real_seqlen_q, hdim_v}); // o_g_m_o
+        ck_tile::HostTensor<LSEDataType> lse_host_ref({nhead, real_seqlen_q});     // lse_g_m
+        ck_tile::HostTensor<RandValOutputDataType> randval_host_ref(
+            {nhead, real_seqlen_q, real_seqlen_k}); // randval_g_m_n
+        ck_tile::HostTensor<AccDataType> s_host_ref(
+            {nhead, real_seqlen_q, real_seqlen_k}); // s_g_m_n
+        ck_tile::HostTensor<AccDataType> p_hp_host_ref(
+            {nhead, real_seqlen_q, real_seqlen_k}); // p_hp_g_m_n high precision
+        ck_tile::HostTensor<AccDataType> p_dropped_hp_host_ref(
+            {nhead, real_seqlen_q, real_seqlen_k}); // p_dropped_hp_g_m_n high precision
+        ck_tile::HostTensor<GemmDataType> p_lp_host_ref(
+            {nhead, real_seqlen_q, real_seqlen_k}); // p_lp_g_m_n low precision
+
+        ck_tile::index_t nr = nhead / nhead_k;
+
+        // clang-format off
+        // permute
+        if(i_perm) q_host_ref.ForEach([&](auto& self, auto i) { self(i) = q_host(b, i[0], i[1] + query_offset, i[2]); });
+        else       q_host_ref.ForEach([&](auto& self, auto i) { self(i) = q_host(b, i[1] + query_offset, i[0], i[2]); });
+
+        if(i_perm) k_host_ref.ForEach([&](auto& self, auto i) { self(i) = k_host(b, i[0] / nr, i[1] + key_offset, i[2]); });
+        else       k_host_ref.ForEach([&](auto& self, auto i) { self(i) = k_host(b, i[1] + key_offset, i[0] / nr, i[2]); });
+
+        // v_host_ref: [nhead, hdim, seq], v_host: [b, h_k, s, d]
+        if(i_perm) v_host_ref.ForEach([&](auto& self, auto i) { self(i) = v_host(b, i[0] / nr, i[2] + key_offset, i[1]); });
+        // v_host_ref: [nhead, hdim, seq], v_host: [b, s, h_k, d]
+        else       v_host_ref.ForEach([&](auto& self, auto i) { self(i) = v_host(b, i[2] + key_offset, i[0] / nr, i[1]); });
+        // clang-format on
+
+        // reference
+        // S = scale * Q * K^T
+        ck_tile::reference_batched_gemm<QDataType, KDataType, AccDataType, AccDataType>(
+            q_host_ref,
+            k_host_ref,
+            s_host_ref,
+            ck_tile::identity{},
+            ck_tile::identity{},
+            ck_tile::scales(scale)); // s_g_m_n = scale * q_g_m_k@k_g_n_k
+
+        if(bias.type == bias_enum::elementwise_bias)
+        {
+            // elementwise bias
+            ck_tile::HostTensor<BiasDataType> bias_host_ref({1, real_seqlen_q, real_seqlen_k});
+            // clang-format off
+            if(i_perm)
+                bias_host_ref.ForEach([&](auto& self, auto i) { self(i) = bias_host(0, 0, i[1] + query_offset, i[2]); });
+            else
+                bias_host_ref.ForEach([&](auto& self, auto i) { self(i) = bias_host(0, i[1] + query_offset, 0, i[2]); });
+            // clang-format on
+
+            // broadcast from [1, real_seqlen_q, real_seqlen_k] to [nhead, real_seqlen_q,
+            // real_seqlen_k]
+            ck_tile::
+                reference_batched_elementwise<AccDataType, BiasDataType, AccDataType, AccDataType>(
+                    s_host_ref, bias_host_ref, s_host_ref);
+        }
+        else if(bias.type == bias_enum::alibi)
+        {
+            // alibi construct elementwise bias to verify
+            auto alibi_host = [&]() {
+                if(mask.type != mask_enum::no_mask)
+                {
+                    return ck_tile::make_alibi_from_lr_mask<AccDataType, false>(
+                        0,
+                        mask.left,
+                        mask.right,
+                        real_seqlen_q,
+                        real_seqlen_k,
+                        static_cast<ck_tile::GenericAttentionMaskEnum>(mask.type));
+                }
+                else
+                {
+                    return ck_tile::Alibi<AccDataType, false>{
+                        0, real_seqlen_q, real_seqlen_k, ck_tile::AlibiMode::FROM_BOTTOM_RIGHT};
+                }
+            }();
+
+            ck_tile::HostTensor<AccDataType> alibi_bias_host_ref(
+                {nhead, real_seqlen_q, real_seqlen_k});
+            auto i_b_slope = bias.rank_info == 0 ? 0 : wb;
+            for(auto i_h = 0; i_h < nhead; i_h++)
+            {
+                AccDataType current_slope = alibi_slope_host(i_b_slope, i_h);
+                alibi_host.slope = alibi_host.mode == ck_tile::AlibiMode::VERTICAL ? current_slope
+                                                                                   : -current_slope;
+                for(auto i_r = 0; i_r < real_seqlen_q; i_r++)
+                {
+                    for(auto i_c = 0; i_c < real_seqlen_k; i_c++)
+                    {
+                        AccDataType pixel = 0;
+                        alibi_host.update(pixel, i_r, i_c);
+                        alibi_bias_host_ref(i_h, i_r, i_c) = pixel;
+                    }
+                }
+            }
+            // [nhead, real_seqlen_q, real_seqlen_k]
+            ck_tile::
+                reference_batched_elementwise<AccDataType, AccDataType, AccDataType, AccDataType>(
+                    s_host_ref, alibi_bias_host_ref, s_host_ref);
+        }
+
+        if(mask.type == mask_enum::no_mask)
+        {
+            ck_tile::reference_batched_masking<AccDataType>(
+                s_host_ref, FmhaMasks::NoMask{real_seqlen_q, real_seqlen_k});
+        }
+        else if(mask.type == mask_enum::window_generic)
+        {
+            ck_tile::reference_batched_masking<AccDataType>(
+                s_host_ref,
+                ck_tile::make_generic_attention_mask_from_lr_window<FmhaMasks::GenericMask>(
+                    mask.left, mask.right, real_seqlen_q, real_seqlen_k));
+        }
+        else
+        {
+            // if left window size is negative, means causal
+            // else means generic (for current batch)
+            if(mask.left < 0)
+                ck_tile::reference_batched_masking<AccDataType>(
+                    s_host_ref,
+                    ck_tile::make_generic_attention_mask_from_lr_window<FmhaMasks::CausalMask>(
+                        mask.left,
+                        mask.right,
+                        real_seqlen_q,
+                        real_seqlen_k,
+                        mask.type == mask_enum::mask_top_left));
+            else
+                ck_tile::reference_batched_masking<AccDataType>(
+                    s_host_ref,
+                    ck_tile::make_generic_attention_mask_from_lr_window<FmhaMasks::GenericMask>(
+                        mask.left,
+                        mask.right,
+                        real_seqlen_q,
+                        real_seqlen_k,
+                        mask.type == mask_enum::mask_top_left));
+        }
+        ck_tile::reference_batched_softmax<AccDataType, LSEDataType, AccDataType>(
+            s_host_ref, p_hp_host_ref, ck_tile::identity{}, lse_host_ref);
+
+        if(p_drop > 0)
+        {
+            p_hp_host_ref.ForEach(
+                [&](auto& self, auto idx) { p_dropped_hp_host_ref(idx) = self(idx); });
+            randval_host_ref.ForEach([&](auto& self, auto idx) {
+                self(idx) = randval_host(b, idx[0], idx[1] + query_offset, idx[2]);
+            });
+            ck_tile::reference_batched_dropout(
+                p_dropped_hp_host_ref, randval_host_ref, p_undrop_in_uint8_t, rp_undrop);
+            p_dropped_hp_host_ref.ForEach([&](auto& self, auto idx) {
+                p_lp_host_ref(idx) = ck_tile::type_convert<GemmDataType>(self(idx));
+            });
+        }
+        else
+        {
+            p_hp_host_ref.ForEach([&](auto& self, auto idx) {
+                p_lp_host_ref(idx) = ck_tile::type_convert<GemmDataType>(self(idx));
+            });
+        }
+
+        // O = P * V
+        ck_tile::reference_batched_gemm<GemmDataType, VDataType, AccDataType, ODataType>(
+            p_lp_host_ref, v_host_ref, o_host_ref); // o_g_m_o = p_lp_g_m_n@v_g_o_n
+
+        // clang-format off
+        // permute
+        if(o_perm) o_host_ref.ForEach([&](auto& self, auto idx) { o_host(b, idx[0], idx[1] + query_offset, idx[2]) = self(idx); });
+        else       o_host_ref.ForEach([&](auto& self, auto idx) { o_host(b, idx[1] + query_offset, idx[0], idx[2]) = self(idx); });
+
+        lse_host_ref.ForEach([&](auto& self, auto idx) { lse_host(b, idx[0], idx[1] + query_offset) = self(idx); });
+        // clang-format on
+
+        q_host_refs.push_back(q_host_ref);
+        k_host_refs.push_back(k_host_ref);
+        v_host_refs.push_back(v_host_ref);
+        o_host_refs.push_back(o_host_ref);
+        p_hp_host_refs.push_back(p_hp_host_ref);
+        p_lp_host_refs.push_back(p_lp_host_ref);
+        if(p_drop > 0)
+        {
+            randval_host_refs.push_back(randval_host_ref);
+        }
+    }
+
+    o_buf.ToDevice(o_host.data());
+    lse_buf.ToDevice(lse_host.data());
+    dq_buf.SetZero();
+    dbias_buf.SetZero();
+    dq_acc_buf.SetZero();
+
+    ck_tile::stream_config stream_config_v{
+        nullptr, true, 0, 0, 1, arg_parser.get_str("timer") == std::string("gpu")};
+    fmha_bwd(fmha_traits, fmha_args, stream_config_v);
+
+    dq_buf.FromDevice(dq_host.data());
+    dk_buf.FromDevice(dk_host.data());
+    dv_buf.FromDevice(dv_host.data());
+    dbias_buf.FromDevice(dbias_host.data());
+
+    for(ck_tile::index_t wb = 0; wb < batch; ++wb)
+    {
+        const ck_tile::index_t real_seqlen_q = seqstart_q_host[wb + 1] - seqstart_q_host[wb];
+        const ck_tile::index_t real_seqlen_k = seqstart_k_host[wb + 1] - seqstart_k_host[wb];
+
+        // adjust matrix index according to the mode
+        const ck_tile::index_t b            = (mode == mode_enum::batch ? wb : 0);
+        const ck_tile::index_t query_offset = (mode == mode_enum::batch ? 0 : seqstart_q_host[wb]);
+        const ck_tile::index_t key_offset   = (mode == mode_enum::batch ? 0 : seqstart_k_host[wb]);
+
+        ck_tile::HostTensor<OGradDataType> do_host_ref({nhead, real_seqlen_q, hdim_v}); // do_g_m_o
+        ck_tile::HostTensor<AccDataType> ds_hp_host_ref(
+            {nhead, real_seqlen_q, real_seqlen_k}); // ds_g_m_n high precision
+        ck_tile::HostTensor<GemmDataType> ds_lp_host_ref(
+            {nhead, real_seqlen_q, real_seqlen_k}); // ds_g_m_n low precision
+        ck_tile::HostTensor<AccDataType> dp_hp_host_ref(
+            {nhead, real_seqlen_q, real_seqlen_k}); // dp_g_m_n high precision
+        ck_tile::HostTensor<BiasGradDataType> dbias_host_ref(
+            {nhead, real_seqlen_q, real_seqlen_k}); // dbias_g_m_n
+        ck_tile::HostTensor<QGradDataType> dq_host_ref({nhead, real_seqlen_q, hdim_q}); // dq_g_m_k
+        ck_tile::HostTensor<KGradDataType> dk_host_ref({nhead, real_seqlen_k, hdim_q}); // dk_g_n_k
+        ck_tile::HostTensor<VGradDataType> dv_host_ref({nhead, real_seqlen_k, hdim_v}); // dv_g_n_o
+
+        // clang-format off
+        if(o_perm) do_host_ref.ForEach([&](auto& self, auto i) { self(i) = do_host(b, i[0], i[1] + query_offset, i[2]); });
+        else       do_host_ref.ForEach([&](auto& self, auto i) { self(i) = do_host(b, i[1] + query_offset, i[0], i[2]); });
+        // clang-format on
+
+        // dP = dO@V x Z w/  dropout
+        // dP = dO@V     w/o dropout
+        auto v_t_host_ref = v_host_refs[wb].transpose({0, 2, 1}); // v_g_o_n -> v_g_n_o
+        ck_tile::reference_batched_gemm<OGradDataType, VDataType, AccDataType, AccDataType>(
+            do_host_ref, v_t_host_ref, dp_hp_host_ref); // dp_g_m_n = do_g_m_o@v_g_n_o
+
+        if(p_drop > 0)
+        {
+            ck_tile::reference_batched_dropout(
+                dp_hp_host_ref, randval_host_refs[wb], p_undrop_in_uint8_t, rp_undrop);
+        }
+
+        // dS_i_j = P_i_j .* (dP_i_j - dO_i dot O_i)
+        ds_hp_host_ref.ForEach([&](auto& self, auto idx_gmn) {
+            AccDataType do_dot_o = 0;
+            for(int o = 0; o < hdim_v; o++)
+            {
+                auto idx_gmo = idx_gmn;
+                idx_gmo[2]   = o;
+                do_dot_o += ck_tile::type_convert<AccDataType>(do_host_ref(idx_gmo)) *
+                            ck_tile::type_convert<AccDataType>(o_host_refs[wb](idx_gmo));
+            }
+            self(idx_gmn) = ck_tile::type_convert<AccDataType>(
+                p_hp_host_refs[wb](idx_gmn) * (dp_hp_host_ref(idx_gmn) - do_dot_o));
+        });
+
+        if(use_dbias)
+        {
+            ds_hp_host_ref.ForEach([&](auto& self, auto idx) {
+                dbias_host_ref(idx) = ck_tile::type_convert<BiasGradDataType>(self(idx));
+            });
+        }
+
+        ds_hp_host_ref.ForEach([&](auto& self, auto idx) {
+            ds_lp_host_ref(idx) = ck_tile::type_convert<GemmDataType>(self(idx));
+        });
+
+        // dV = P_drop^T@dO^T
+        // dV = P^T@dO^T w/o dropout
+        auto p_t_lp_host_ref = p_lp_host_refs[wb].transpose({0, 2, 1}); // p_lp_g_m_n -> p_lp_g_n_m
+        auto do_t_host_ref   = do_host_ref.transpose({0, 2, 1});        // do_g_m_o -> do_g_o_m
+        ck_tile::reference_batched_gemm<GemmDataType, OGradDataType, AccDataType, VGradDataType>(
+            p_t_lp_host_ref, do_t_host_ref, dv_host_ref); // dv_g_n_o = p_lp_g_n_m@do_g_o_m
+
+        // dQ = scale * dS@K^T
+        auto k_t_host_ref = k_host_refs[wb].transpose({0, 2, 1}); // k_g_n_k -> k_g_k_n
+        ck_tile::reference_batched_gemm<GemmDataType, KDataType, AccDataType, QGradDataType>(
+            ds_lp_host_ref,
+            k_t_host_ref,
+            dq_host_ref,
+            ck_tile::identity{},
+            ck_tile::identity{},
+            ck_tile::scales(scale)); // dq_g_m_k = ds_g_m_n@k_g_k_n
+
+        // dK = scale * dS^T@Q^T
+        auto ds_t_lp_host_ref = ds_lp_host_ref.transpose({0, 2, 1});  // ds_g_m_n -> ds_g_n_m
+        auto q_t_host_ref     = q_host_refs[wb].transpose({0, 2, 1}); // q_g_m_k -> q_g_k_m
+        ck_tile::reference_batched_gemm<GemmDataType, QDataType, AccDataType, KGradDataType>(
+            ds_t_lp_host_ref,
+            q_t_host_ref,
+            dk_host_ref,
+            ck_tile::identity{},
+            ck_tile::identity{},
+            ck_tile::scales(scale)); // dk_g_n_k = ds_g_n_m@q_g_k_m
+
+        ck_tile::HostTensor<QGradDataType> dq_host_result(
+            {nhead, real_seqlen_q, hdim_q}); // dq_g_m_k
+        ck_tile::HostTensor<KGradDataType> dk_host_result(
+            {nhead, real_seqlen_k, hdim_q}); // dk_g_n_k
+        ck_tile::HostTensor<VGradDataType> dv_host_result(
+            {nhead, real_seqlen_k, hdim_v}); // dv_g_n_o
+        ck_tile::HostTensor<BiasGradDataType> dbias_host_result(
+            {nhead, real_seqlen_q, real_seqlen_k}); // dbias_g_m_n
+
+        // clang-format off
+        // permute
+        if(i_perm) dq_host_result.ForEach([&](auto& self, auto idx) {self(idx) = dq_host(b, idx[0], idx[1] + query_offset, idx[2]); });
+        else       dq_host_result.ForEach([&](auto& self, auto idx) {self(idx) = dq_host(b, idx[1] + query_offset, idx[0], idx[2]); });
+
+        if(i_perm) dk_host_result.ForEach([&](auto& self, auto idx) {self(idx) = dk_host(b, idx[0], idx[1] + key_offset, idx[2]); });
+        else       dk_host_result.ForEach([&](auto& self, auto idx) {self(idx) = dk_host(b, idx[1] + key_offset, idx[0], idx[2]); });
+
+        if(i_perm) dv_host_result.ForEach([&](auto& self, auto idx) {self(idx) = dv_host(b, idx[0], idx[1] + key_offset, idx[2]); });
+        else       dv_host_result.ForEach([&](auto& self, auto idx) {self(idx) = dv_host(b, idx[1] + key_offset, idx[0], idx[2]); });
+
+        if(use_dbias)
+        {
+            if(i_perm) dbias_host_result.ForEach([&](auto& self, auto idx) {self(idx) = dbias_host(b, idx[0], idx[1] + query_offset, idx[2]); });
+            else       dbias_host_result.ForEach([&](auto& self, auto idx) {self(idx) = dbias_host(b, idx[1] + query_offset, idx[0], idx[2]); });
+        }
+        // clang-format on
+
+        auto [rtol, atol] = get_elimit<DataTypeConfig>(hdim_q, hdim_v);
+        bool dq_cur_pass  = ck_tile::check_err(dq_host_result,
+                                              dq_host_ref,
+                                              std::string("Error: QGrad Incorrect results!"),
+                                              rtol,
+                                              atol);
+        bool dk_cur_pass  = ck_tile::check_err(dk_host_result,
+                                              dk_host_ref,
+                                              std::string("Error: KGrad Incorrect results!"),
+                                              rtol,
+                                              atol);
+        bool dv_cur_pass  = ck_tile::check_err(dv_host_result,
+                                              dv_host_ref,
+                                              std::string("Error: VGrad Incorrect results!"),
+                                              rtol,
+                                              atol);
+
+        bool dbias_cur_pass = true;
+        if(use_dbias)
+        {
+            dbias_cur_pass = ck_tile::check_err(dbias_host_result,
+                                                dbias_host_ref,
+                                                std::string("Error: BiasGrad Incorrect results!"),
+                                                rtol,
+                                                atol);
+        }
+        pass &= (dq_cur_pass & dk_cur_pass & dv_cur_pass & dbias_cur_pass);
+        if(!(dq_cur_pass & dk_cur_pass & dv_cur_pass & dbias_cur_pass))
+        {
+            std::cerr << "mismatch found at batch: " << wb << std::endl
+                      << "\tseqlen_q: " << real_seqlen_q << std::endl
+                      << "\tseqlen_k: " << real_seqlen_k << std::endl
+                      << "\tseqstart_q: " << seqstart_q_host << std::endl
+                      << "\tseqstart_k: " << seqstart_k_host << std::endl;
+
+            break;
+        }
+    }
+
+    std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    const std::string data_type = arg_parser.get_str("prec");
+    if(data_type == "fp16")
+    {
+        return run<FmhaBwdFp16>(arg_parser) ? 0 : -2;
+    }
+    else if(data_type == "bf16")
+    {
+        return run<FmhaBwdBf16>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/01_fmha/fmha_bwd.hpp

@@ -0,0 +1,456 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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/fmha.hpp"
+#include "ck_tile/ops/epilogue.hpp"
+#include "mask.hpp"
+#include "bias.hpp"
+
+#include <type_traits>
+#include <utility>
+#include <variant>
+
+struct FmhaBwdFp16
+{
+};
+
+struct FmhaBwdBf16
+{
+};
+
+template <typename DataType>
+struct FmhaBwdTypeConfig;
+
+template <>
+struct FmhaBwdTypeConfig<FmhaBwdFp16>
+{
+    using QDataType             = ck_tile::half_t;
+    using KDataType             = ck_tile::half_t;
+    using VDataType             = ck_tile::half_t;
+    using GemmDataType          = ck_tile::half_t;
+    using BiasDataType          = ck_tile::half_t;
+    using LSEDataType           = float;
+    using AccDataType           = float; // data type for gemm accumulation
+    using DDataType             = float;
+    using RandValOutputDataType = uint8_t;
+    using ODataType             = ck_tile::half_t;
+    using OGradDataType         = ck_tile::half_t;
+    using QGradDataType         = ck_tile::half_t;
+    using KGradDataType         = ck_tile::half_t;
+    using VGradDataType         = ck_tile::half_t;
+    using BiasGradDataType      = ck_tile::half_t;
+};
+
+template <>
+struct FmhaBwdTypeConfig<FmhaBwdBf16>
+{
+    using QDataType             = ck_tile::bf16_t;
+    using KDataType             = ck_tile::bf16_t;
+    using VDataType             = ck_tile::bf16_t;
+    using GemmDataType          = ck_tile::bf16_t;
+    using BiasDataType          = ck_tile::bf16_t;
+    using LSEDataType           = float;
+    using AccDataType           = float; // data type for gemm accumulation
+    using DDataType             = float;
+    using RandValOutputDataType = uint8_t;
+    using ODataType             = ck_tile::bf16_t;
+    using OGradDataType         = ck_tile::bf16_t;
+    using QGradDataType         = ck_tile::bf16_t;
+    using KGradDataType         = ck_tile::bf16_t;
+    using VGradDataType         = ck_tile::bf16_t;
+    using BiasGradDataType      = ck_tile::bf16_t;
+};
+
+struct FmhaMasks
+{
+    using NoMask      = ck_tile::GenericAttentionMask<false>;
+    using GenericMask = ck_tile::GenericAttentionMask<true, true>;
+    using CausalMask  = ck_tile::GenericAttentionMask<true, false>;
+};
+
+// runtime args, some will passed to karg, some will used to compute grids/blocks
+struct fmha_bwd_args
+{
+    const void* q_ptr;
+    const void* k_ptr;
+    const void* v_ptr;
+    const void* bias_ptr; // bias or alibi_slope pointer
+    const void* o_ptr;
+    const void* lse_ptr;
+    const void* do_ptr;
+    void* d_ptr;
+    void* rand_val_ptr;
+    void* dq_ptr;
+    void* dk_ptr;
+    void* dv_ptr;
+    void* dbias_ptr;
+    void* dq_acc_ptr;
+    const void* seqstart_q_ptr;
+    const void* seqstart_k_ptr;
+    const void* seqlen_k_ptr;
+    ck_tile::index_t seqlen_q;
+    ck_tile::index_t seqlen_k;
+    ck_tile::index_t batch;
+    ck_tile::index_t max_seqlen_q;
+    ck_tile::index_t max_seqlen_k;
+    ck_tile::index_t hdim_q;
+    ck_tile::index_t hdim_v;
+    ck_tile::index_t nhead_q;
+    ck_tile::index_t nhead_k;
+    float scale;
+    ck_tile::index_t stride_q;
+    ck_tile::index_t stride_k;
+    ck_tile::index_t stride_v;
+    ck_tile::index_t stride_bias; // if alibi, b*h need set this to h, 1*h need set this to 0
+    ck_tile::index_t stride_o;
+    ck_tile::index_t stride_randval;
+    ck_tile::index_t stride_do;
+    ck_tile::index_t stride_dq_acc;
+    ck_tile::index_t stride_dq;
+    ck_tile::index_t stride_dk;
+    ck_tile::index_t stride_dv;
+    ck_tile::index_t stride_dbias;
+    ck_tile::index_t nhead_stride_q;
+    ck_tile::index_t nhead_stride_k;
+    ck_tile::index_t nhead_stride_v;
+    ck_tile::index_t nhead_stride_bias;
+    ck_tile::index_t nhead_stride_o;
+    ck_tile::index_t nhead_stride_randval;
+    ck_tile::index_t nhead_stride_do;
+    ck_tile::index_t nhead_stride_lsed;
+    ck_tile::index_t nhead_stride_dq_acc;
+    ck_tile::index_t nhead_stride_dq;
+    ck_tile::index_t nhead_stride_dk;
+    ck_tile::index_t nhead_stride_dv;
+    ck_tile::index_t nhead_stride_dbias;
+    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_bias;
+    ck_tile::index_t batch_stride_o;
+    ck_tile::index_t batch_stride_randval;
+    ck_tile::index_t batch_stride_do;
+    ck_tile::index_t batch_stride_lsed;
+    ck_tile::index_t batch_stride_dq_acc;
+    ck_tile::index_t batch_stride_dq;
+    ck_tile::index_t batch_stride_dk;
+    ck_tile::index_t batch_stride_dv;
+    ck_tile::index_t batch_stride_dbias;
+    ck_tile::index_t split_stride_dq_acc;
+    ck_tile::index_t window_size_left;
+    ck_tile::index_t window_size_right;
+    ck_tile::index_t mask_type;
+    float p_drop;
+    float p_undrop;
+    std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+        drop_seed_offset;
+};
+
+template <typename FmhaBwdDQDKDVKernel>
+auto fmha_bwd_dq_dk_dv_create_kargs_and_grids(fmha_bwd_args args)
+{
+    assert(args.nhead_q % args.nhead_k == 0);
+    auto kargs = [&] {
+        // create group mode kernel arguments
+        if constexpr(FmhaBwdDQDKDVKernel::kIsGroupMode)
+        {
+            return FmhaBwdDQDKDVKernel::MakeKargsImpl(args.q_ptr,
+                                                      args.k_ptr,
+                                                      args.v_ptr,
+                                                      args.bias_ptr,
+                                                      args.lse_ptr,
+                                                      args.do_ptr,
+                                                      args.d_ptr,
+                                                      args.rand_val_ptr,
+                                                      args.dk_ptr,
+                                                      args.dv_ptr,
+                                                      args.dbias_ptr,
+                                                      args.dq_acc_ptr,
+                                                      args.seqstart_q_ptr,
+                                                      args.seqstart_k_ptr,
+                                                      args.seqlen_k_ptr,
+                                                      args.hdim_q,
+                                                      args.hdim_v,
+                                                      args.nhead_q,
+                                                      args.nhead_q / args.nhead_k,
+                                                      args.scale,
+                                                      args.stride_q,
+                                                      args.stride_k,
+                                                      args.stride_v,
+                                                      args.stride_bias,
+                                                      args.stride_randval,
+                                                      args.stride_do,
+                                                      args.stride_dq_acc,
+                                                      args.stride_dk,
+                                                      args.stride_dv,
+                                                      args.stride_dbias,
+                                                      args.nhead_stride_q,
+                                                      args.nhead_stride_k,
+                                                      args.nhead_stride_v,
+                                                      args.nhead_stride_bias,
+                                                      args.nhead_stride_randval,
+                                                      args.nhead_stride_do,
+                                                      args.nhead_stride_lsed,
+                                                      args.nhead_stride_dq_acc,
+                                                      args.nhead_stride_dk,
+                                                      args.nhead_stride_dv,
+                                                      args.nhead_stride_dbias,
+                                                      args.split_stride_dq_acc,
+                                                      args.window_size_left,
+                                                      args.window_size_right,
+                                                      args.mask_type,
+                                                      args.p_drop,
+                                                      args.drop_seed_offset);
+        }
+        else
+        { // create batch mode kernel arguments
+            return FmhaBwdDQDKDVKernel::MakeKargsImpl(args.q_ptr,
+                                                      args.k_ptr,
+                                                      args.v_ptr,
+                                                      args.bias_ptr,
+                                                      args.lse_ptr,
+                                                      args.do_ptr,
+                                                      args.d_ptr,
+                                                      args.rand_val_ptr,
+                                                      args.dk_ptr,
+                                                      args.dv_ptr,
+                                                      args.dbias_ptr,
+                                                      args.dq_acc_ptr,
+                                                      args.seqlen_q,
+                                                      args.seqlen_k,
+                                                      args.hdim_q,
+                                                      args.hdim_v,
+                                                      args.nhead_q,
+                                                      args.nhead_q / args.nhead_k,
+                                                      args.scale,
+                                                      args.stride_q,
+                                                      args.stride_k,
+                                                      args.stride_v,
+                                                      args.stride_bias,
+                                                      args.stride_randval,
+                                                      args.stride_do,
+                                                      args.stride_dq_acc,
+                                                      args.stride_dk,
+                                                      args.stride_dv,
+                                                      args.stride_dbias,
+                                                      args.nhead_stride_q,
+                                                      args.nhead_stride_k,
+                                                      args.nhead_stride_v,
+                                                      args.nhead_stride_bias,
+                                                      args.nhead_stride_randval,
+                                                      args.nhead_stride_do,
+                                                      args.nhead_stride_lsed,
+                                                      args.nhead_stride_dq_acc,
+                                                      args.nhead_stride_dk,
+                                                      args.nhead_stride_dv,
+                                                      args.nhead_stride_dbias,
+                                                      args.batch_stride_q,
+                                                      args.batch_stride_k,
+                                                      args.batch_stride_v,
+                                                      args.batch_stride_bias,
+                                                      args.batch_stride_randval,
+                                                      args.batch_stride_do,
+                                                      args.batch_stride_lsed,
+                                                      args.batch_stride_dq_acc,
+                                                      args.batch_stride_dk,
+                                                      args.batch_stride_dv,
+                                                      args.batch_stride_dbias,
+                                                      args.split_stride_dq_acc,
+                                                      args.window_size_left,
+                                                      args.window_size_right,
+                                                      args.mask_type,
+                                                      args.p_drop,
+                                                      args.drop_seed_offset);
+        }
+    }();
+
+    dim3 grids = FmhaBwdDQDKDVKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_k);
+    return ck_tile::make_tuple(kargs, grids);
+}
+
+template <typename FmhaBwdOGradDotOKernel>
+auto fmha_bwd_dot_do_o_create_kargs_and_grids(fmha_bwd_args args)
+{
+    auto kargs = [&] {
+        // create group mode kernel arguments
+        if constexpr(FmhaBwdOGradDotOKernel::kIsGroupMode)
+        {
+            return FmhaBwdOGradDotOKernel::MakeKargs(args.o_ptr,
+                                                     args.do_ptr,
+                                                     args.d_ptr,
+                                                     args.p_undrop,
+                                                     args.seqstart_q_ptr,
+                                                     args.hdim_v,
+                                                     args.stride_do,
+                                                     args.stride_o,
+                                                     args.nhead_stride_do,
+                                                     args.nhead_stride_o,
+                                                     args.nhead_stride_lsed);
+        }
+        else
+        { // create batch mode kernel arguments
+            return FmhaBwdOGradDotOKernel::MakeKargs(args.o_ptr,
+                                                     args.do_ptr,
+                                                     args.d_ptr,
+                                                     args.p_undrop,
+                                                     args.seqlen_q,
+                                                     args.hdim_v,
+                                                     args.stride_do,
+                                                     args.stride_o,
+                                                     args.nhead_stride_do,
+                                                     args.nhead_stride_o,
+                                                     args.nhead_stride_lsed,
+                                                     args.batch_stride_do,
+                                                     args.batch_stride_o,
+                                                     args.batch_stride_lsed);
+        }
+    }();
+
+    dim3 grids = FmhaBwdOGradDotOKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q);
+    return ck_tile::make_tuple(kargs, grids);
+}
+
+template <typename FmhaBwdConvertQGradKernel>
+auto fmha_bwd_convert_dq_create_kargs_and_grids(fmha_bwd_args args)
+{
+    auto kargs = [&] {
+        // create group mode kernel arguments
+        if constexpr(FmhaBwdConvertQGradKernel::kIsGroupMode)
+        {
+            return FmhaBwdConvertQGradKernel::MakeKargs(args.dq_acc_ptr,
+                                                        args.dq_ptr,
+                                                        args.seqstart_q_ptr,
+                                                        args.seqstart_k_ptr,
+                                                        args.hdim_q,
+                                                        args.stride_dq,
+                                                        args.stride_dq_acc,
+                                                        args.nhead_stride_dq,
+                                                        args.nhead_stride_dq_acc,
+                                                        args.split_stride_dq_acc);
+        }
+        else
+        { // create batch mode kernel arguments
+            return FmhaBwdConvertQGradKernel::MakeKargs(args.dq_acc_ptr,
+                                                        args.dq_ptr,
+                                                        args.seqlen_q,
+                                                        args.seqlen_k,
+                                                        args.hdim_q,
+                                                        args.stride_dq,
+                                                        args.stride_dq_acc,
+                                                        args.nhead_stride_dq,
+                                                        args.nhead_stride_dq_acc,
+                                                        args.batch_stride_dq,
+                                                        args.batch_stride_dq_acc,
+                                                        args.split_stride_dq_acc);
+        }
+    }();
+
+    dim3 grids = FmhaBwdConvertQGradKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q);
+    return ck_tile::make_tuple(kargs, grids);
+}
+
+// this is used to pattern-match internl kernel implementation, not to instantiate kernel
+template <ck_tile::index_t HDim_,
+          typename DataType_,
+          bool kIsGroupMode_,
+          ck_tile::BlockFmhaBwdPipelineEnum FmhaBwdPipelineEnum_,
+          typename FmhaMask_,
+          typename FmhaDropout_,
+          ck_tile::BlockAttentionBiasEnum BiasEnum_,
+          bool kHasBiasGrad_,
+          bool kPadS_,
+          bool kPadSK_,
+          bool kPadD_,
+          bool kPadDv_,
+          bool kIsDeterministic_>
+struct fmha_bwd_dq_dk_dv_traits_
+{
+    static constexpr ck_tile::index_t HDim    = HDim_;
+    using DataType                            = ck_tile::remove_cvref_t<DataType_>;
+    static constexpr bool kIsGroupMode        = kIsGroupMode_;
+    static constexpr auto FmhaBwdPipelineEnum = FmhaBwdPipelineEnum_;
+    using FmhaMask                            = ck_tile::remove_cvref_t<FmhaMask_>;
+    using FmhaDropout                         = ck_tile::remove_cvref_t<FmhaDropout_>;
+    static constexpr auto BiasEnum            = BiasEnum_;
+    static constexpr bool kHasBiasGrad        = kHasBiasGrad_;
+    static constexpr bool kPadS               = kPadS_;
+    static constexpr bool kPadSK              = kPadSK_;
+    static constexpr bool kPadD               = kPadD_;
+    static constexpr bool kPadDv              = kPadDv_;
+    static constexpr bool kIsDeterministic    = kIsDeterministic_;
+};
+
+template <typename Traits_>
+float fmha_bwd_dq_dk_dv_(const ck_tile::stream_config&, fmha_bwd_args);
+
+template <typename Traits_>
+void fmha_bwd_dq_dk_dv_oneshot_(const ck_tile::stream_config&, fmha_bwd_args);
+
+template <typename Traits_>
+std::string fmha_bwd_dq_dk_dv_get_name_();
+
+template <ck_tile::index_t HDim_, typename DataType_, bool kIsGroupMode_, bool kPadS_, bool kPadDv_>
+struct fmha_bwd_dot_do_o_traits_
+{
+    static constexpr ck_tile::index_t HDim = HDim_;
+    using DataType                         = ck_tile::remove_cvref_t<DataType_>;
+    static constexpr bool kIsGroupMode     = kIsGroupMode_;
+    static constexpr bool kPadS            = kPadS_;
+    static constexpr bool kPadDv           = kPadDv_;
+};
+
+template <typename Traits_>
+float fmha_bwd_dot_do_o_(const ck_tile::stream_config&, fmha_bwd_args);
+
+template <typename Traits_>
+void fmha_bwd_dot_do_o_oneshot_(const ck_tile::stream_config&, fmha_bwd_args);
+
+template <typename Traits_>
+std::string fmha_bwd_dot_do_o_get_name_();
+
+template <ck_tile::index_t HDim_,
+          typename DataType_,
+          bool kIsGroupMode_,
+          bool kPadS_,
+          bool kPadD_,
+          bool kIsDeterministic_>
+struct fmha_bwd_convert_dq_traits_
+{
+    static constexpr ck_tile::index_t HDim = HDim_;
+    using DataType                         = ck_tile::remove_cvref_t<DataType_>;
+    static constexpr bool kIsGroupMode     = kIsGroupMode_;
+    static constexpr bool kPadS            = kPadS_;
+    static constexpr bool kPadD            = kPadD_;
+    static constexpr bool kIsDeterministic = kIsDeterministic_;
+};
+
+template <typename Traits_>
+float fmha_bwd_convert_dq_(const ck_tile::stream_config&, fmha_bwd_args);
+
+template <typename Traits_>
+void fmha_bwd_convert_dq_oneshot_(const ck_tile::stream_config&, fmha_bwd_args);
+
+template <typename Traits_>
+std::string fmha_bwd_convert_dq_get_name_();
+
+// This is the public API, will be generated by script
+struct fmha_bwd_traits
+{
+    int hdim_q;
+    int hdim_v;
+    std::string data_type;
+    bool is_group_mode;
+    mask_enum mask_type;
+    bias_enum bias_type; // 0:no bias, 1:elementwise bias, 2:alibi. sync with BlockAttentionBiasEnum
+    bool has_dbias;
+    bool has_dropout;
+    bool is_store_randval;
+    bool is_deterministic;
+    // TODO: padding check is inside this api
+};
+template <int Version = 2>
+float fmha_bwd(fmha_bwd_traits, fmha_bwd_args, const ck_tile::stream_config&);

example/ck_tile/01_fmha/fmha_fwd.hpp

@@ -0,0 +1,823 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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/epilogue.hpp"
+#include "ck_tile/ops/fmha.hpp"
+
+#include "bias.hpp"
+#include "mask.hpp"
+#include "rotary.hpp"
+
+#include <type_traits>
+#include <utility>
+#include <variant>
+
+struct FmhaFwdFp16
+{
+};
+
+struct FmhaFwdBf16
+{
+};
+
+struct FmhaFwdFp8
+{
+};
+
+struct FmhaFwdBf8
+{
+};
+
+struct FmhaFwdFp8Fp16
+{
+};
+
+struct FmhaFwdFp8Bf16
+{
+};
+
+template <typename DataType>
+struct FmhaFwdTypeConfig;
+
+template <>
+struct FmhaFwdTypeConfig<FmhaFwdFp16>
+{
+    using QDataType             = ck_tile::half_t;
+    using KDataType             = ck_tile::half_t;
+    using VDataType             = ck_tile::half_t;
+    using BiasDataType          = ck_tile::half_t;
+    using RandValOutputDataType = uint8_t;
+    using LSEDataType           = float; // data type for lse(logsumexp L_j = max_j + log(l_j))
+    using SaccDataType          = float; // data type for first gemm accumulation
+    using SMPLComputeDataType   = float; // data type for reduction, softmax
+    using PDataType             = ck_tile::half_t; // data type for A matrix of second gemm
+    using OaccDataType          = float;           // data type for second gemm accumulation
+    using ODataType             = ck_tile::half_t;
+};
+
+template <>
+struct FmhaFwdTypeConfig<FmhaFwdBf16>
+{
+    using QDataType             = ck_tile::bf16_t;
+    using KDataType             = ck_tile::bf16_t;
+    using VDataType             = ck_tile::bf16_t;
+    using BiasDataType          = ck_tile::bf16_t;
+    using RandValOutputDataType = uint8_t;
+    using LSEDataType           = float; // data type for lse(logsumexp L_j = max_j + log(l_j))
+    using SaccDataType          = float; // data type for first gemm accumulation
+    using SMPLComputeDataType   = float; // data type for reduction, softmax
+    using PDataType             = ck_tile::bf16_t; // data type for A matrix of second gemm
+    using OaccDataType          = float;           // data type for second gemm accumulation
+    using ODataType             = ck_tile::bf16_t;
+};
+
+template <>
+struct FmhaFwdTypeConfig<FmhaFwdFp8>
+{
+    using QDataType             = ck_tile::fp8_t;
+    using KDataType             = ck_tile::fp8_t;
+    using VDataType             = ck_tile::fp8_t;
+    using BiasDataType          = float;
+    using RandValOutputDataType = uint8_t;
+    using LSEDataType           = float; // data type for lse(logsumexp L_j = max_j + log(l_j))
+    using SaccDataType          = float; // data type for first gemm accumulation
+    using SMPLComputeDataType   = float; // data type for reduction, softmax
+    using PDataType             = ck_tile::fp8_t; // data type for A matrix of second gemm
+    using OaccDataType          = float;          // data type for second gemm accumulation
+    using ODataType             = ck_tile::fp8_t;
+};
+
+template <>
+struct FmhaFwdTypeConfig<FmhaFwdBf8>
+{
+    using QDataType             = ck_tile::bf8_t;
+    using KDataType             = ck_tile::bf8_t;
+    using VDataType             = ck_tile::bf8_t;
+    using BiasDataType          = ck_tile::bf8_t;
+    using RandValOutputDataType = uint8_t;
+    using LSEDataType           = float; // data type for lse(logsumexp L_j = max_j + log(l_j))
+    using SaccDataType          = float; // data type for first gemm accumulation
+    using SMPLComputeDataType   = float; // data type for reduction, softmax
+    using PDataType             = ck_tile::bf8_t; // data type for A matrix of second gemm
+    using OaccDataType          = float;          // data type for second gemm accumulation
+    using ODataType             = ck_tile::bf8_t;
+};
+
+struct FmhaMasks
+{
+    using NoMask      = ck_tile::GenericAttentionMask<false>;
+    using GenericMask = ck_tile::GenericAttentionMask<true, true>;
+    using CausalMask  = ck_tile::GenericAttentionMask<true, false>;
+};
+
+// runtime args, some will passed to karg, some will used to compute grids/blocks
+struct fmha_fwd_args
+{
+    const void* q_ptr;
+    const void* k_ptr;
+    const void* v_ptr;
+    const void* bias_ptr; // bias or alibi_slope pointer
+    void* rand_val_ptr;
+    void* lse_ptr;
+    void* o_ptr;
+
+    const void* seqstart_q_ptr;
+    const void* seqstart_k_ptr;
+    const void*
+        seqlen_k_ptr; // only used if both 'seqstart_q_ptr' & 'seqstart_k_ptr' are not nullptr
+
+    ck_tile::index_t seqlen_q;
+    ck_tile::index_t seqlen_k;
+    ck_tile::index_t batch;
+    ck_tile::index_t max_seqlen_q;
+    ck_tile::index_t hdim_q;
+    ck_tile::index_t hdim_v;
+    ck_tile::index_t nhead_q;
+    ck_tile::index_t nhead_k;
+
+    float scale_s;
+    float scale_p;
+    float scale_o;
+
+    ck_tile::index_t stride_q;
+    ck_tile::index_t stride_k;
+    ck_tile::index_t stride_v;
+    ck_tile::index_t stride_bias; // if alibi, b*h need set this to h, 1*h need set this to 0
+    ck_tile::index_t stride_randval;
+    ck_tile::index_t stride_o;
+    ck_tile::index_t nhead_stride_q;
+    ck_tile::index_t nhead_stride_k;
+    ck_tile::index_t nhead_stride_v;
+    ck_tile::index_t nhead_stride_bias;
+    ck_tile::index_t nhead_stride_randval;
+    ck_tile::index_t nhead_stride_lse;
+    ck_tile::index_t nhead_stride_o;
+    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_bias;
+    ck_tile::index_t batch_stride_randval;
+    ck_tile::index_t batch_stride_lse;
+    ck_tile::index_t batch_stride_o;
+
+    ck_tile::index_t window_size_left;
+    ck_tile::index_t window_size_right;
+    ck_tile::index_t mask_type;
+
+    float p_drop;
+    bool s_randval;
+
+    std::variant<std::pair<uint64_t, uint64_t>, std::pair<const void*, const void*>>
+        drop_seed_offset;
+};
+
+struct fmha_fwd_splitkv_args
+{
+    const void* q_ptr;
+    const void* k_ptr;
+    const void* v_ptr;
+    const void* bias_ptr; // bias or alibi_slope pointer
+    void* lse_acc_ptr;
+    void* o_acc_ptr;
+    void* lse_ptr;
+    void* o_ptr;
+
+    void* block_table_ptr;
+    ck_tile::index_t batch_stride_block_table; // only used if 'block_table_ptr' is not nullptr
+    ck_tile::index_t page_block_size;          // only used if 'block_table_ptr' is not nullptr
+    bool is_gappy; // differentiate seqstart_k_ptr usage. only used if 'block_table_ptr' is not
+                   // nullptr.
+
+    const void* cache_batch_idx;
+
+    // the real seqlen_q & seqlen_k are decided by following:
+    // batch mode: seqlen_q = kargs.seqlen_q
+    //             seqlen_k = kargs.seqlen_k
+    // group mode: seqlen_q = kargs.seqstart_q_ptr[b + 1] - kargs.seqstart_q_ptr[b]
+    //             seqlen_k = kargs.seqstart_k_ptr[b + 1] - kargs.seqstart_k_ptr[b]
+    //                      or kargs.seqlen_k_ptr[b]
+    //
+    // batch mode (kvcache):
+    //             seqlen_q = kargs.seqlen_q
+    //             seqlen_k = kargs.seqlen_k_ptr[b]
+    // group mode (kvcache):
+    //             seqlen_q = kargs.seqstart_q_ptr[b + 1] - kargs.seqstart_q_ptr[b]
+    //
+    //     when is_gappy=true:
+    //             seqlen_k = kargs.seqlen_k_ptr[b]
+    //             seqstart_k_ptr[b] now store local offset of each batch
+    //
+    //     when is_gappy=false:
+    //             seqlen_k = kargs.seqstart_k_ptr[b + 1] - kargs.seqstart_k_ptr[b]
+    //                      or kargs.seqlen_k_ptr[b]
+    const void* seqstart_q_ptr;
+    const void* seqstart_k_ptr;
+    const void* seqlen_k_ptr;
+
+    ck_tile::index_t seqlen_q;
+    ck_tile::index_t seqlen_k;
+    ck_tile::index_t batch;
+    ck_tile::index_t max_seqlen_q;
+    ck_tile::index_t hdim_q;
+    ck_tile::index_t hdim_v;
+    ck_tile::index_t nhead_q;
+    ck_tile::index_t nhead_k;
+    ck_tile::index_t num_splits;
+
+    float scale_s;
+    float scale_p;
+    float scale_o;
+
+    ck_tile::index_t stride_q;
+    ck_tile::index_t stride_k;
+    ck_tile::index_t stride_v;
+    ck_tile::index_t stride_bias; // if alibi, b*h need set this to h, 1*h need set this to 0
+    ck_tile::index_t stride_o_acc;
+    ck_tile::index_t stride_o;
+    ck_tile::index_t nhead_stride_q;
+    ck_tile::index_t nhead_stride_k;
+    ck_tile::index_t nhead_stride_v;
+    ck_tile::index_t nhead_stride_bias;
+    ck_tile::index_t nhead_stride_lse;
+    ck_tile::index_t nhead_stride_lse_acc;
+    ck_tile::index_t nhead_stride_o_acc;
+    ck_tile::index_t nhead_stride_o;
+    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_bias;
+    ck_tile::index_t batch_stride_lse;
+    ck_tile::index_t batch_stride_lse_acc;
+    ck_tile::index_t batch_stride_o_acc;
+    ck_tile::index_t batch_stride_o;
+    ck_tile::index_t split_stride_lse_acc;
+    ck_tile::index_t split_stride_o_acc;
+
+    ck_tile::index_t window_size_left;
+    ck_tile::index_t window_size_right;
+    ck_tile::index_t mask_type;
+};
+
+struct fmha_fwd_appendkv_args
+{
+    void* q_ptr;
+    void* k_ptr;
+    const void* knew_ptr;
+    void* v_ptr;
+    const void* vnew_ptr;
+
+    const void* seqlen_k_ptr;
+
+    ck_tile::index_t seqlen_q;
+    ck_tile::index_t seqlen_knew;
+    ck_tile::index_t batch;
+    ck_tile::index_t hdim_q;
+    ck_tile::index_t hdim_v;
+    ck_tile::index_t nhead_q;
+    ck_tile::index_t nhead_k;
+
+    const void* rotary_cos_ptr; // only used if 'rotary_dim' > 0
+    const void* rotary_sin_ptr; // only used if 'rotary_dim' > 0
+    ck_tile::index_t rotary_dim;
+    bool has_mask;
+
+    void* block_table_ptr;
+    ck_tile::index_t batch_stride_block_table; // only used if 'block_table_ptr' is not nullptr
+    ck_tile::index_t page_block_size;          // only used if 'block_table_ptr' is not nullptr
+
+    const void* cache_batch_idx; // only used if block_table_ptr is nullptr -> batch mode (kvcache)
+
+    ck_tile::index_t stride_q;
+    ck_tile::index_t stride_k;
+    ck_tile::index_t stride_knew;
+    ck_tile::index_t stride_v;
+    ck_tile::index_t stride_vnew;
+    ck_tile::index_t nhead_stride_q;
+    ck_tile::index_t nhead_stride_k;
+    ck_tile::index_t nhead_stride_knew;
+    ck_tile::index_t nhead_stride_v;
+    ck_tile::index_t nhead_stride_vnew;
+    ck_tile::index_t batch_stride_q;
+    ck_tile::index_t batch_stride_k;
+    ck_tile::index_t batch_stride_knew;
+    ck_tile::index_t batch_stride_v;
+    ck_tile::index_t batch_stride_vnew;
+};
+
+template <typename FmhaKernel>
+auto fmha_fwd_create_kargs_and_grids(fmha_fwd_args args)
+{
+    assert(args.nhead_q % args.nhead_k == 0);
+    auto kargs = [&] {
+        // create group mode kernel arguments
+        if constexpr(FmhaKernel::kIsGroupMode)
+        {
+            return FmhaKernel::MakeKargsImpl(args.q_ptr,
+                                             args.k_ptr,
+                                             args.v_ptr,
+                                             args.bias_ptr,
+                                             args.rand_val_ptr,
+                                             args.lse_ptr,
+                                             args.o_ptr,
+                                             args.seqstart_q_ptr,
+                                             args.seqstart_k_ptr,
+                                             args.seqlen_k_ptr,
+                                             args.hdim_q,
+                                             args.hdim_v,
+                                             args.nhead_q,
+                                             args.nhead_q / args.nhead_k,
+                                             args.scale_s,
+                                             args.scale_p,
+                                             args.scale_o,
+                                             args.stride_q,
+                                             args.stride_k,
+                                             args.stride_v,
+                                             args.stride_bias,
+                                             args.stride_randval,
+                                             args.stride_o,
+                                             args.nhead_stride_q,
+                                             args.nhead_stride_k,
+                                             args.nhead_stride_v,
+                                             args.nhead_stride_bias,
+                                             args.nhead_stride_randval,
+                                             args.nhead_stride_lse,
+                                             args.nhead_stride_o,
+                                             args.window_size_left,
+                                             args.window_size_right,
+                                             args.mask_type,
+                                             args.p_drop,
+                                             args.s_randval,
+                                             args.drop_seed_offset);
+        }
+        else
+        { // create batch mode kernel arguments
+            return FmhaKernel::MakeKargsImpl(args.q_ptr,
+                                             args.k_ptr,
+                                             args.v_ptr,
+                                             args.bias_ptr,
+                                             args.rand_val_ptr,
+                                             args.lse_ptr,
+                                             args.o_ptr,
+                                             args.seqlen_q,
+                                             args.seqlen_k,
+                                             args.hdim_q,
+                                             args.hdim_v,
+                                             args.nhead_q,
+                                             args.nhead_q / args.nhead_k,
+                                             args.scale_s,
+                                             args.scale_p,
+                                             args.scale_o,
+                                             args.stride_q,
+                                             args.stride_k,
+                                             args.stride_v,
+                                             args.stride_bias,
+                                             args.stride_randval,
+                                             args.stride_o,
+                                             args.nhead_stride_q,
+                                             args.nhead_stride_k,
+                                             args.nhead_stride_v,
+                                             args.nhead_stride_bias,
+                                             args.nhead_stride_randval,
+                                             args.nhead_stride_lse,
+                                             args.nhead_stride_o,
+                                             args.batch_stride_q,
+                                             args.batch_stride_k,
+                                             args.batch_stride_v,
+                                             args.batch_stride_bias,
+                                             args.batch_stride_randval,
+                                             args.batch_stride_lse,
+                                             args.batch_stride_o,
+                                             args.window_size_left,
+                                             args.window_size_right,
+                                             args.mask_type,
+                                             args.p_drop,
+                                             args.s_randval,
+                                             args.drop_seed_offset);
+        }
+    }();
+
+    if constexpr(FmhaKernel::kIsGroupMode)
+    {
+        dim3 grids = FmhaKernel::GridSize(
+            args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v, args.seqlen_k_ptr != nullptr);
+        return ck_tile::make_tuple(kargs, grids);
+    }
+    else
+    {
+        dim3 grids =
+            FmhaKernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v, false);
+        return ck_tile::make_tuple(kargs, grids);
+    }
+}
+
+template <typename Kernel>
+auto fmha_fwd_splitkv_create_kargs_and_grids(fmha_fwd_splitkv_args args)
+{
+    assert(args.nhead_q % args.nhead_k == 0);
+    auto kargs = [&] {
+        // create group mode kernel arguments
+        if constexpr(Kernel::kIsGroupMode)
+        {
+            return Kernel::MakeKargs(args.q_ptr,
+                                     args.k_ptr,
+                                     args.v_ptr,
+                                     args.bias_ptr,
+                                     args.lse_acc_ptr,
+                                     args.o_acc_ptr,
+                                     args.batch,
+                                     args.seqstart_q_ptr,
+                                     args.seqstart_k_ptr,
+                                     args.seqlen_k_ptr,
+                                     args.hdim_q,
+                                     args.hdim_v,
+                                     args.nhead_q,
+                                     args.nhead_q / args.nhead_k,
+                                     args.num_splits,
+                                     args.block_table_ptr,
+                                     args.batch_stride_block_table,
+                                     args.page_block_size,
+                                     args.is_gappy,
+                                     args.scale_s,
+                                     args.scale_p,
+                                     args.stride_q,
+                                     args.stride_k,
+                                     args.stride_v,
+                                     args.stride_bias,
+                                     args.stride_o_acc,
+                                     args.nhead_stride_q,
+                                     args.nhead_stride_k,
+                                     args.nhead_stride_v,
+                                     args.nhead_stride_bias,
+                                     args.nhead_stride_lse_acc,
+                                     args.nhead_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,
+                                     args.window_size_right,
+                                     args.mask_type);
+        }
+        else
+        { // create batch mode kernel arguments
+            return Kernel::MakeKargs(args.q_ptr,
+                                     args.k_ptr,
+                                     args.v_ptr,
+                                     args.bias_ptr,
+                                     args.lse_acc_ptr,
+                                     args.o_acc_ptr,
+                                     args.batch,
+                                     args.seqlen_q,
+                                     args.seqlen_k,
+                                     args.seqlen_k_ptr,
+                                     args.hdim_q,
+                                     args.hdim_v,
+                                     args.nhead_q,
+                                     args.nhead_q / args.nhead_k,
+                                     args.num_splits,
+                                     args.block_table_ptr,
+                                     args.batch_stride_block_table,
+                                     args.page_block_size,
+                                     args.cache_batch_idx,
+                                     args.scale_s,
+                                     args.scale_p,
+                                     args.stride_q,
+                                     args.stride_k,
+                                     args.stride_v,
+                                     args.stride_bias,
+                                     args.stride_o_acc,
+                                     args.nhead_stride_q,
+                                     args.nhead_stride_k,
+                                     args.nhead_stride_v,
+                                     args.nhead_stride_bias,
+                                     args.nhead_stride_lse_acc,
+                                     args.nhead_stride_o_acc,
+                                     args.batch_stride_q,
+                                     args.batch_stride_k,
+                                     args.batch_stride_v,
+                                     args.batch_stride_bias,
+                                     args.batch_stride_lse_acc,
+                                     args.batch_stride_o_acc,
+                                     args.split_stride_lse_acc,
+                                     args.split_stride_o_acc,
+                                     args.window_size_left,
+                                     args.window_size_right,
+                                     args.mask_type);
+        }
+    }();
+
+    dim3 grids = Kernel::GridSize(
+        args.batch, args.nhead_q, args.nhead_k, args.max_seqlen_q, args.hdim_v, args.num_splits);
+
+    return ck_tile::make_tuple(kargs, grids);
+}
+
+template <typename Kernel>
+auto fmha_fwd_splitkv_combine_create_kargs_and_grids(fmha_fwd_splitkv_args args)
+{
+    assert(args.nhead_q % args.nhead_k == 0);
+    auto kargs = [&] {
+        // create group mode kernel argumentszs
+        if constexpr(Kernel::kIsGroupMode)
+        {
+            return Kernel::MakeKargs(args.lse_acc_ptr,
+                                     args.o_acc_ptr,
+                                     args.lse_ptr,
+                                     args.o_ptr,
+                                     args.batch,
+                                     args.seqstart_q_ptr,
+                                     args.hdim_v,
+                                     args.num_splits,
+                                     args.scale_o,
+                                     args.stride_o_acc,
+                                     args.stride_o,
+                                     args.nhead_stride_lse_acc,
+                                     args.nhead_stride_o_acc,
+                                     args.nhead_stride_lse,
+                                     args.nhead_stride_o,
+                                     args.split_stride_lse_acc,
+                                     args.split_stride_o_acc);
+        }
+        else
+        { // create batch mode kernel arguments
+            return Kernel::MakeKargs(args.lse_acc_ptr,
+                                     args.o_acc_ptr,
+                                     args.lse_ptr,
+                                     args.o_ptr,
+                                     args.batch,
+                                     args.seqlen_q,
+                                     args.hdim_v,
+                                     args.num_splits,
+                                     args.scale_o,
+                                     args.stride_o_acc,
+                                     args.stride_o,
+                                     args.nhead_stride_lse_acc,
+                                     args.nhead_stride_o_acc,
+                                     args.nhead_stride_lse,
+                                     args.nhead_stride_o,
+                                     args.batch_stride_lse_acc,
+                                     args.batch_stride_o_acc,
+                                     args.batch_stride_lse,
+                                     args.batch_stride_o,
+                                     args.split_stride_lse_acc,
+                                     args.split_stride_o_acc);
+        }
+    }();
+
+    dim3 grids = Kernel::GridSize(args.batch, args.nhead_q, args.max_seqlen_q, args.hdim_v);
+
+    return ck_tile::make_tuple(kargs, grids);
+}
+
+template <typename Kernel>
+auto fmha_fwd_appendkv_create_kargs_and_grids(fmha_fwd_appendkv_args args)
+{
+    assert(args.nhead_q % args.nhead_k == 0);
+    auto kargs = Kernel::MakeKargs(args.q_ptr,
+                                   args.k_ptr,
+                                   args.knew_ptr,
+                                   args.v_ptr,
+                                   args.vnew_ptr,
+                                   args.seqlen_q,
+                                   args.seqlen_k_ptr,
+                                   args.seqlen_knew,
+                                   args.hdim_q,
+                                   args.hdim_v,
+                                   args.nhead_q,
+                                   args.nhead_q / args.nhead_k,
+                                   args.rotary_cos_ptr,
+                                   args.rotary_sin_ptr,
+                                   args.rotary_dim,
+                                   args.has_mask,
+                                   args.block_table_ptr,
+                                   args.batch_stride_block_table,
+                                   args.page_block_size,
+                                   args.cache_batch_idx,
+                                   args.stride_q,
+                                   args.stride_k,
+                                   args.stride_knew,
+                                   args.stride_v,
+                                   args.stride_vnew,
+                                   args.nhead_stride_q,
+                                   args.nhead_stride_k,
+                                   args.nhead_stride_knew,
+                                   args.nhead_stride_v,
+                                   args.nhead_stride_vnew,
+                                   args.batch_stride_q,
+                                   args.batch_stride_k,
+                                   args.batch_stride_knew,
+                                   args.batch_stride_v,
+                                   args.batch_stride_vnew);
+
+    dim3 grids = Kernel::GridSize(args.batch, args.nhead_q, args.seqlen_q, args.seqlen_knew);
+
+    return ck_tile::make_tuple(kargs, grids);
+}
+
+// this is used to pattern-match internl kernel implementation, not to instantiate kernel
+template <ck_tile::index_t HDim_,
+          typename DataType_,
+          bool kIsGroupMode_,
+          ck_tile::index_t kM0_,
+          ck_tile::index_t kN0_,
+          ck_tile::index_t kK0_,
+          ck_tile::index_t kN1_,
+          ck_tile::index_t kK1_,
+          ck_tile::index_t kK0BlockLength_,
+          bool kIsVLayoutRowMajor_,
+          ck_tile::BlockFmhaPipelineEnum FmhaPipelineEnum_,
+          typename FmhaMask_,
+          ck_tile::BlockAttentionBiasEnum BiasEnum_,
+          bool kStoreLse_,
+          bool kHasDropout_,
+          bool kDoFp8StaticQuant_,
+          bool kPadS_,
+          bool kPadSK_,
+          bool kPadD_,
+          bool kPadDv_>
+struct fmha_fwd_traits_
+{
+    static constexpr ck_tile::index_t HDim           = HDim_;
+    using DataType                                   = ck_tile::remove_cvref_t<DataType_>;
+    static constexpr bool kIsGroupMode               = kIsGroupMode_;
+    static constexpr ck_tile::index_t kM0            = kM0_;
+    static constexpr ck_tile::index_t kN0            = kN0_;
+    static constexpr ck_tile::index_t kK0            = kK0_;
+    static constexpr ck_tile::index_t kN1            = kN1_;
+    static constexpr ck_tile::index_t kK1            = kK1_;
+    static constexpr ck_tile::index_t kK0BlockLength = kK0BlockLength_;
+    static constexpr bool kIsVLayoutRowMajor         = kIsVLayoutRowMajor_;
+    static constexpr auto FmhaPipelineEnum           = FmhaPipelineEnum_;
+    using FmhaMask                                   = ck_tile::remove_cvref_t<FmhaMask_>;
+    static constexpr auto BiasEnum                   = BiasEnum_;
+    static constexpr bool kStoreLse                  = kStoreLse_;
+    static constexpr bool kHasDropout                = kHasDropout_;
+    static constexpr bool kDoFp8StaticQuant          = kDoFp8StaticQuant_;
+    static constexpr bool kPadS                      = kPadS_;
+    static constexpr bool kPadSK                     = kPadSK_;
+    static constexpr bool kPadD                      = kPadD_;
+    static constexpr bool kPadDv                     = kPadDv_;
+};
+
+template <typename Traits_>
+float fmha_fwd_(const ck_tile::stream_config&, fmha_fwd_args);
+
+template <ck_tile::index_t HDim_,
+          typename DataType_,
+          bool kIsGroupMode_,
+          ck_tile::index_t kM0_,
+          ck_tile::index_t kN0_,
+          ck_tile::index_t kK0_,
+          ck_tile::index_t kN1_,
+          ck_tile::index_t kK1_,
+          ck_tile::index_t kK0BlockLength_,
+          bool kIsVLayoutRowMajor_,
+          ck_tile::BlockFmhaPipelineEnum FmhaPipelineEnum_,
+          typename FmhaMask_,
+          ck_tile::BlockAttentionBiasEnum BiasEnum_,
+          bool kStoreLse_,
+          bool kDoFp8StaticQuant_,
+          bool kIsPagedKV_,
+          bool kPadS_,
+          bool kPadSK_,
+          bool kPadD_,
+          bool kPadDv_>
+struct fmha_fwd_splitkv_traits_
+{
+    static constexpr ck_tile::index_t HDim           = HDim_;
+    using DataType                                   = ck_tile::remove_cvref_t<DataType_>;
+    static constexpr bool kIsGroupMode               = kIsGroupMode_;
+    static constexpr ck_tile::index_t kM0            = kM0_;
+    static constexpr ck_tile::index_t kN0            = kN0_;
+    static constexpr ck_tile::index_t kK0            = kK0_;
+    static constexpr ck_tile::index_t kN1            = kN1_;
+    static constexpr ck_tile::index_t kK1            = kK1_;
+    static constexpr ck_tile::index_t kK0BlockLength = kK0BlockLength_;
+    static constexpr bool kIsVLayoutRowMajor         = kIsVLayoutRowMajor_;
+    static constexpr auto FmhaPipelineEnum           = FmhaPipelineEnum_;
+    using FmhaMask                                   = ck_tile::remove_cvref_t<FmhaMask_>;
+    static constexpr auto BiasEnum                   = BiasEnum_;
+    static constexpr bool kStoreLse                  = kStoreLse_;
+    static constexpr bool kDoFp8StaticQuant          = kDoFp8StaticQuant_;
+    static constexpr bool kPadS                      = kPadS_;
+    static constexpr bool kPadSK                     = kPadSK_;
+    static constexpr bool kPadD                      = kPadD_;
+    static constexpr bool kPadDv                     = kPadDv_;
+    static constexpr bool kIsPagedKV                 = kIsPagedKV_;
+};
+
+template <typename Traits_>
+void fmha_fwd_splitkv_oneshot_(const ck_tile::stream_config&, fmha_fwd_splitkv_args);
+
+template <typename Traits_>
+std::string fmha_fwd_splitkv_get_name_();
+
+template <ck_tile::index_t HDim_,
+          typename DataType_,
+          bool kIsGroupMode_,
+          ck_tile::index_t kN1_,
+          bool kStoreLse_,
+          bool kDoFp8StaticQuant_,
+          bool kPadS_,
+          bool kPadDv_>
+struct fmha_fwd_splitkv_combine_traits_
+{
+    static constexpr ck_tile::index_t HDim  = HDim_;
+    using DataType                          = ck_tile::remove_cvref_t<DataType_>;
+    static constexpr bool kIsGroupMode      = kIsGroupMode_;
+    static constexpr ck_tile::index_t kN1   = kN1_;
+    static constexpr bool kStoreLse         = kStoreLse_;
+    static constexpr bool kDoFp8StaticQuant = kDoFp8StaticQuant_;
+    static constexpr bool kPadS             = kPadS_;
+    static constexpr bool kPadDv            = kPadDv_;
+};
+
+template <typename Traits_>
+void fmha_fwd_splitkv_combine_oneshot_(const ck_tile::stream_config&, fmha_fwd_splitkv_args);
+
+template <typename Traits_>
+std::string fmha_fwd_splitkv_combine_get_name_();
+
+// this is used to pattern-match internl kernel implementation, not to instantiate kernel
+template <ck_tile::index_t HDim_,
+          typename DataType_,
+          ck_tile::index_t kTileSizeS_,
+          ck_tile::index_t kTileSizeSk_,
+          ck_tile::index_t kTileSizeD_,
+          ck_tile::index_t kTileSizeDv_,
+          bool kIsVLayoutRowMajor_,
+          bool kPadS_,
+          bool kPadSk_,
+          bool kPadD_,
+          bool kPadDv_,
+          ck_tile::RotaryEmbeddingEnum RotaryEnum_,
+          bool kIsPagedKV_>
+struct fmha_fwd_appendkv_traits_
+{
+    static constexpr ck_tile::index_t HDim        = HDim_;
+    using DataType                                = ck_tile::remove_cvref_t<DataType_>;
+    static constexpr ck_tile::index_t kTileSizeS  = kTileSizeS_;
+    static constexpr ck_tile::index_t kTileSizeSk = kTileSizeSk_;
+    static constexpr ck_tile::index_t kTileSizeD  = kTileSizeD_;
+    static constexpr ck_tile::index_t kTileSizeDv = kTileSizeDv_;
+    static constexpr bool kIsVLayoutRowMajor      = kIsVLayoutRowMajor_;
+    static constexpr bool kPadS                   = kPadS_;
+    static constexpr bool kPadSk                  = kPadSk_;
+    static constexpr bool kPadD                   = kPadD_;
+    static constexpr bool kPadDv                  = kPadDv_;
+    static constexpr auto RotaryEnum              = RotaryEnum_;
+    static constexpr bool kIsPagedKV              = kIsPagedKV_;
+};
+
+template <typename Traits_>
+float fmha_fwd_appendkv_(const ck_tile::stream_config&, fmha_fwd_appendkv_args);
+
+// This is the public API, will be generated by script
+struct fmha_fwd_traits
+{
+    int hdim_q;
+    int hdim_v;
+    std::string data_type;
+    bool is_group_mode;
+    bool is_v_rowmajor;
+    mask_enum mask_type;
+    bias_enum bias_type; // 0:no bias, 1:elementwise bias, 2:alibi. sync with BlockAttentionBiasEnum
+    bool has_lse;
+    bool has_dropout;
+    bool do_fp8_static_quant;
+    // TODO: padding check is inside this api
+};
+float fmha_fwd(fmha_fwd_traits, fmha_fwd_args, const ck_tile::stream_config&);
+
+struct fmha_fwd_splitkv_traits
+{
+    int hdim_q;
+    int hdim_v;
+    std::string data_type;
+    bool is_group_mode;
+    bool is_v_rowmajor;
+    mask_enum mask_type;
+    bias_enum bias_type; // 0:no bias, 1:elementwise bias, 2:alibi. sync with BlockAttentionBiasEnum
+    bool has_lse;
+    bool do_fp8_static_quant;
+    // TODO: padding check is inside this api
+};
+float fmha_fwd_splitkv(fmha_fwd_splitkv_traits,
+                       fmha_fwd_splitkv_args,
+                       const ck_tile::stream_config&);
+
+struct fmha_fwd_appendkv_traits
+{
+    int hdim_q;
+    int hdim_v;
+    std::string data_type;
+    bool is_v_rowmajor;
+    rope_enum rope_type;
+};
+float fmha_fwd_appendkv(fmha_fwd_appendkv_traits,
+                        fmha_fwd_appendkv_args,
+                        const ck_tile::stream_config&);

example/ck_tile/01_fmha/generate.py

@@ -0,0 +1,136 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import argparse
+from enum import IntEnum
+from pathlib import Path
+import pkgutil
+import sys
+from typing import List, Optional
+
+import codegen.ops
+from codegen.cmake_config import *
+
+
+class HandlerId(IntEnum):
+    LIST_BLOBS = 0
+    WRITE_BLOBS = 1
+
+# inspect all modules under 'codegen.ops' and register API handlers
+ops = []
+for importer, module_name, _ in pkgutil.iter_modules(codegen.ops.__path__):
+    full_module_name = '%s.%s' % (codegen.ops.__name__, module_name)
+    if full_module_name not in sys.modules:
+        ops.append(importer.find_spec(module_name).loader.load_module(module_name))
+unwanted_prefix = 'fmha_'
+handlers = dict(
+    [(op.__name__[len(unwanted_prefix):] if op.__name__.startswith(unwanted_prefix) else op.__name__,
+        (op.list_blobs, op.write_blobs)) for op in ops]
+)
+assert 0 < len(handlers)
+
+def write_blobs(output_dir: Optional[str], api_list : List[str], filters_list : List[str], optdim_list : List[int], receipt, mask_impl) -> None:
+    if output_dir is None:
+        output_dir = Path(__file__).parent
+    else:
+        output_dir = Path(output_dir) / GEN_DIR
+
+    output_dir.mkdir(parents=True, exist_ok=True)
+
+    for api, kernel_filter in zip(api_list, filters_list):
+        handler = handlers[api][HandlerId.WRITE_BLOBS]
+        handler(output_dir, kernel_filter, receipt, optdim_list, mask_impl)
+
+# list all the files that will be generated
+def list_blobs(output_file : Optional[str], api_list : List[str], filters_list : List[str], optdim_list : List[int], receipt, mask_impl) -> None:
+    assert output_file is not None
+    file_path = Path(output_file)
+
+    # create an empty file / drop its contents if it exists
+    open(file_path, "w").close()
+
+    for api, kernel_filter in zip(api_list, filters_list):
+        handler = handlers[api][HandlerId.LIST_BLOBS]
+        handler(file_path, kernel_filter, receipt, optdim_list, mask_impl)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        prog="generate",
+        description="gen API for CK fmha kernel",
+    )
+    parser.add_argument(
+        "-d",
+        "--direction", # we keep 'direction' option for backward compatibility
+        "-a",
+        "--api",
+        default='fwd',
+        required=False,
+        help="supply API(s) to generate (default: fwd). separated by comma."
+    )
+    parser.add_argument(
+        "-o",
+        "--output_dir",
+        required=False,
+        help="write all the blobs into a directory"
+    )
+    parser.add_argument(
+        "-l",
+        "--list_blobs",
+        required=False,
+        help="list all the kernels to a file"
+    )
+    # TODO: if using filter, must apply same value to output_dir and list_blobs
+    parser.add_argument(
+        "-f",
+        "--filter",
+        default='',
+        required=False,
+        help="filter out kernels that need to generate, using fnmatch module"
+    )
+
+    parser.add_argument(
+        "-m",
+        "--mask",
+        default="simplified",
+        required=False,
+        help="mask implementation, simplified/generic"
+    )
+
+    parser.add_argument(
+        "-r",
+        "--receipt",
+        default=0,
+        required=False,
+        help="codegen receipt. 0: generate only 8xhdim coverage\n"  + \
+             "  1: generate more instance to cover all hdim\n"  + \
+             "  2: Only generate instance for Flash attention integration\n"  + \
+             "  4: Only generate instance for PyTorch integration\n" + \
+             "  100-199: Only generate instance for Aiter(mha_fwd) integration\n" + \
+             "  200-299: Only generate instance for Aiter(mha_varlen_fwd) integration\n" + \
+             "  300-399: Only generate instance for Aiter(mha_bwd) integration\n" + \
+             "  400-499: Only generate instance for Aiter(mha_varlen_bwd) integration\n" + \
+             "  600-699: Only generate instance for aiter::mha_fwd && aiter::mha_fwd_splitkv && aiter::mha_bwd C++ api integration"
+    )
+
+    parser.add_argument(
+        "--optdim",
+        default='-1',
+        required=False,
+        help="only optimize the hdim in the list. separated by comma. -1 is the default choice" + \
+              "eg. --optdim=32,64,128,256"
+    )
+
+    args = parser.parse_args()
+    api_list = args.direction.split(',')
+    filter_list = args.filter.split(',')
+    filter_list.extend([''] * (len(api_list) - len(filter_list)))
+    optdim_list = [int(hdim) for hdim in args.optdim.split(',')]
+
+    if len(api_list) > 1:
+        assert optdim_list == [-1]
+
+    if args.list_blobs is not None:
+        list_blobs(args.list_blobs, api_list, filter_list, optdim_list, int(args.receipt), mask_impl=args.mask)
+    else:
+        write_blobs(args.output_dir, api_list, filter_list, optdim_list, int(args.receipt), mask_impl=args.mask)

example/ck_tile/01_fmha/mask.hpp

@@ -0,0 +1,157 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <ostream>
+#include <string>
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/fmha.hpp"
+
+// keep this in sync with ck_tile::GenericAttentionMaskEnum
+enum class mask_enum
+{
+    no_mask = 0,
+    mask_top_left,
+    mask_bottom_right,
+    window_generic,
+};
+
+struct mask_info
+{
+    mask_enum type;
+    ck_tile::index_t y, x;
+    ck_tile::index_t left, right; // FA style SWA left/right
+
+    void serialize(std::ostream& os) const
+    {
+        if(type == mask_enum::no_mask)
+            os << "n";
+        else if(type == mask_enum::mask_top_left)
+            os << "t(" << left << ":" << right << ")";
+        else if(type == mask_enum::mask_bottom_right)
+            os << "b(" << left << ":" << right << ")";
+        else
+        {
+            os << "g(" << y << ":" << x << ")";
+        }
+    }
+    static mask_info decode(std::string str, ck_tile::index_t seqlen_q, ck_tile::index_t seqlen_k)
+    {
+        ck_tile::index_t x_total = seqlen_k;
+        ck_tile::index_t y_total = seqlen_q;
+        mask_info tmp;
+        auto found_0 = str.find(':');
+        if(found_0 != std::string::npos)
+        {
+            std::string t = str.substr(0, found_0);
+            std::string v = str.substr(found_0 + 1);
+            if(t == "xt" || t == "xb")
+            {
+                // xformer style sliding window attn from top-left
+                ck_tile::index_t window_size = atoi(v.c_str());
+                ck_tile::index_t left_size   = -1;
+                ck_tile::index_t right_size  = 0;
+                if(window_size > 0)
+                {
+                    left_size  = window_size / 2;
+                    right_size = window_size - 1 - left_size;
+                }
+                auto r = ck_tile::make_generic_attention_mask_coordinates_from_lr_window(
+                    left_size, right_size, y_total, x_total, t == "xt");
+
+                tmp.type  = t == "xt" ? mask_enum::mask_top_left : mask_enum::mask_bottom_right;
+                tmp.y     = r.at(ck_tile::number<0>{});
+                tmp.x     = r.at(ck_tile::number<1>{});
+                tmp.left  = left_size;
+                tmp.right = right_size;
+            }
+            else
+            {
+                auto found_1 = v.find(",");
+                if(found_1 == std::string::npos)
+                {
+                    printf("not supported value %s, %s\n", v.c_str(), str.c_str());
+                    assert(0);
+                }
+                tmp.type            = mask_enum::window_generic;
+                ck_tile::index_t v0 = atoi(v.substr(0, found_1).c_str());
+                ck_tile::index_t v1 = atoi(v.substr(found_1 + 1).c_str());
+                // TODO: some validation
+                if(t == "t")
+                {
+                    tmp.type = mask_enum::mask_top_left;
+                    auto r   = ck_tile::make_generic_attention_mask_coordinates_from_lr_window(
+                        v0, v1, y_total, x_total, true);
+                    tmp.y     = r.at(ck_tile::number<0>{});
+                    tmp.x     = r.at(ck_tile::number<1>{});
+                    tmp.left  = v0;
+                    tmp.right = v1;
+                }
+                else if(t == "b")
+                {
+                    tmp.type = mask_enum::mask_bottom_right;
+                    auto r   = ck_tile::make_generic_attention_mask_coordinates_from_lr_window(
+                        v0, v1, y_total, x_total, false);
+                    tmp.y     = r.at(ck_tile::number<0>{});
+                    tmp.x     = r.at(ck_tile::number<1>{});
+                    tmp.left  = v0;
+                    tmp.right = v1;
+                }
+                else if(t == "g")
+                {
+                    tmp.y     = v0;
+                    tmp.x     = v1;
+                    tmp.left  = v0; // TODO: don't use this?
+                    tmp.right = v1;
+                }
+                else
+                {
+                    printf("not supported type %s, %s\n", t.c_str(), str.c_str());
+                    assert(0);
+                }
+            }
+        }
+        else
+        {
+            auto set_causal_top_left = [&]() {
+                tmp.type  = mask_enum::mask_top_left;
+                tmp.y     = seqlen_q;
+                tmp.x     = 1;
+                tmp.left  = -1;
+                tmp.right = 0;
+            };
+            auto set_causal_bottom_right = [&]() {
+                tmp.type  = mask_enum::mask_bottom_right;
+                tmp.y     = seqlen_q;
+                tmp.x     = seqlen_k - seqlen_q + 1;
+                tmp.left  = -1;
+                tmp.right = 0;
+            };
+            if(str == "t")
+                set_causal_top_left();
+            else if(str == "b")
+                set_causal_bottom_right();
+            else
+            {
+                tmp.type = static_cast<mask_enum>(atoi(str.c_str()));
+                if(tmp.type == mask_enum::mask_top_left)
+                {
+                    set_causal_top_left();
+                }
+                else if(tmp.type == mask_enum::mask_bottom_right)
+                {
+                    set_causal_bottom_right();
+                }
+            }
+        }
+        return tmp;
+    }
+
+    friend std::ostream& operator<<(std::ostream& os, const mask_info& mi)
+    {
+        mi.serialize(os);
+        return os;
+    }
+};

example/ck_tile/01_fmha/rotary.hpp

@@ -0,0 +1,84 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/host_tensor.hpp"
+
+#include <cassert>
+#include <cmath>
+#include <functional>
+#include <iterator>
+#include <optional>
+#include <random>
+#include <tuple>
+
+// keep sync with RotaryEmbeddingEnum
+enum class rope_enum
+{
+    none         = 0,
+    interleaved  = 1,
+    half_rotated = 2,
+};
+
+template <typename DataType>
+std::tuple<ck_tile::HostTensor<DataType>, ck_tile::HostTensor<DataType>>
+generate_rotary_cos_sin(ck_tile::index_t seqlen,
+                        ck_tile::index_t rotary_dim,
+                        std::optional<unsigned> seed = std::nullopt)
+{
+    // return dummy tensors if we won't apply RoPE at all
+    if(rotary_dim <= 0)
+    {
+        ck_tile::HostTensor<DataType> dummy({1, 1});
+        return std::make_tuple(dummy, dummy);
+    }
+
+    std::mt19937 random_engine(seed.has_value() ? *seed : std::random_device{}());
+    std::uniform_real_distribution<float> generator(0.0f, 1.0f);
+
+    const ck_tile::index_t num_rows = seqlen * 2;
+    const ck_tile::index_t num_cols = rotary_dim / 2;
+
+    using std::begin, std::end;
+
+    ck_tile::HostTensor<float> angle({num_rows, num_cols});
+    std::generate(begin(angle), end(angle), [&] { return generator(random_engine) * 2 * M_PI; });
+
+    ck_tile::HostTensor<DataType> cos({num_rows, num_cols});
+    std::transform(begin(angle), end(angle), begin(cos), [](float origin_value) {
+        return ck_tile::type_convert<DataType>(std::cos(origin_value));
+    });
+
+    ck_tile::HostTensor<DataType> sin({num_rows, num_cols});
+    std::transform(begin(angle), end(angle), begin(sin), [](float origin_value) {
+        return ck_tile::type_convert<DataType>(std::sin(origin_value));
+    });
+
+    return std::make_tuple(cos, sin);
+}
+
+template <typename DataType>
+std::tuple<ck_tile::HostTensor<DataType>, ck_tile::HostTensor<DataType>>
+slice_rotary_cos_sin(const ck_tile::HostTensor<DataType>& cos,
+                     const ck_tile::HostTensor<DataType>& sin,
+                     ck_tile::index_t seqlen_offset,
+                     ck_tile::index_t seqlen)
+{
+    assert(cos.get_num_of_dimension() == 2 && sin.get_num_of_dimension() == 2);
+    assert(cos.get_length(0) == sin.get_length(0) && cos.get_length(1) == sin.get_length(1));
+
+    assert(static_cast<std::size_t>(seqlen_offset + seqlen) <= cos.get_length(0));
+
+    const ck_tile::index_t num_rows = seqlen;
+    const ck_tile::index_t num_cols = cos.get_length(1);
+
+    ck_tile::HostTensor<DataType> cos_pt({num_rows, num_cols});
+    cos_pt.ForEach([&](auto& self, auto i) { self(i) = cos(i[0] + seqlen_offset, i[1]); });
+
+    ck_tile::HostTensor<DataType> sin_pt({num_rows, num_cols});
+    sin_pt.ForEach([&](auto& self, auto i) { self(i) = sin(i[0] + seqlen_offset, i[1]); });
+
+    return std::make_tuple(cos_pt, sin_pt);
+}

example/ck_tile/01_fmha/script/benchmark_bwd.sh

@@ -0,0 +1,20 @@
+#!/bin/sh
+# TODO: run this script from CK root or build directory
+EXE="$(find . -name tile_example_fmha_bwd -type f | head -n 1)"
+VALID=0
+
+for prec in "fp16" "bf16" ; do
+for perm in 0 1 ; do
+for hdim in 32 64 128 ; do
+
+nhead=$((2048 / $hdim))     # follow fav2 setup
+$EXE -prec=$prec -b=32 -h=$nhead -d=$hdim -s=512   -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=16 -h=$nhead -d=$hdim -s=1024  -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=8  -h=$nhead -d=$hdim -s=2048  -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=4  -h=$nhead -d=$hdim -s=4096  -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=2  -h=$nhead -d=$hdim -s=8192  -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=1  -h=$nhead -d=$hdim -s=16384 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+
+done
+done
+done

example/ck_tile/01_fmha/script/benchmark_fwd.sh

@@ -0,0 +1,31 @@
+#!/bin/sh
+# TODO: run this script from CK root or build directory
+EXE="$(find . -name tile_example_fmha_fwd -type f | head -n 1)"
+VALID=0
+
+for prec in "fp16" "bf16" ; do
+for perm in 0 1 ; do
+for hdim in 64 128 256 ; do
+
+nhead=$((2048 / $hdim))     # follow fav2 setup
+$EXE -prec=$prec -b=32 -h=$nhead -d=$hdim -s=512   -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=16 -h=$nhead -d=$hdim -s=1024  -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=8  -h=$nhead -d=$hdim -s=2048  -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=4  -h=$nhead -d=$hdim -s=4096  -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=2  -h=$nhead -d=$hdim -s=8192  -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=$prec -b=1  -h=$nhead -d=$hdim -s=16384 -iperm=$perm -operm=$perm -kname=1 -v=$VALID ; sleep 3
+
+done
+done
+done
+
+for perm in 0 1 ; do
+
+$EXE -prec=fp8 -squant=1 -b=32 -h=16 -d=128 -s=512   -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=fp8 -squant=1 -b=16 -h=16 -d=128 -s=1024  -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=fp8 -squant=1 -b=8  -h=16 -d=128 -s=2048  -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=fp8 -squant=1 -b=4  -h=16 -d=128 -s=4096  -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=fp8 -squant=1 -b=2  -h=16 -d=128 -s=8192  -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
+$EXE -prec=fp8 -squant=1 -b=1  -h=16 -d=128 -s=16384 -iperm=$perm -operm=$perm -vlayout=c -range_q=240 -range_k=240 -range_v=240 -range_p=240 -range_o=240 -kname=1 -v=$VALID ; sleep 3
+
+done

example/ck_tile/01_fmha/script/run_full_test.sh

@@ -0,0 +1,46 @@
+#!/bin/bash 
+#
+# in order to run this script you'd first need to build the tile_example_fmha_fwd and tile_eaxmple_fmha_bwd executables in ../build/bin/
+#
+# run the script as "./run_full_test.sh <tag for your test environment> <branch name> <host name> <gpu_arch>
+# input arguments: 
+# environment tag  : a string describing the specifics of your test environment
+# branch name      : name of the branch in git repo (git status | grep -e 'On branch')
+# host name        : $hostname
+# gpu architecture: e.g., gfx90a, or gfx942, etc.
+
+#get the command line arguments:
+export env_type=$1
+echo 'Environment type: ' $env_type
+export branch=$2
+echo 'Branch name: ' $branch
+export host_name=$3
+echo 'Host name: ' $host_name
+export GPU_arch=$4
+echo 'GPU_arch: ' $GPU_arch
+
+function print_log_header(){
+	rm -f $1;
+	echo 'On branch ' $3 &> $1;
+	echo 'Node name: ' $4 >> $1;
+	#get GPU_arch and number of compute units from rocminfo
+	echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
+	rocminfo | grep "Compute Unit:" >> $1;
+	hipcc --version | grep -e 'HIP version'  >> $1;
+	echo 'Environment type: ' $2 >> $1;
+	/opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
+}
+
+#run verification tests
+example/ck_tile/01_fmha/script/smoke_test_fwd.sh
+example/ck_tile/01_fmha/script/smoke_test_bwd.sh
+
+#run performance benchmarks
+export fmha_fwd_log="perf_fmha_fwd_$GPU_arch.log"
+print_log_header $fmha_fwd_log $env_type $branch $host_name
+example/ck_tile/01_fmha/script/benchmark_fwd.sh 2>&1 | tee -a $fmha_fwd_log
+
+export fmha_bwd_log="perf_fmha_bwd_$GPU_arch.log"
+print_log_header $fmha_bwd_log $env_type $branch $host_name
+example/ck_tile/01_fmha/script/benchmark_bwd.sh 2>&1 | tee -a $fmha_bwd_log
+

example/ck_tile/01_fmha/script/smoke_test_bwd.sh

@@ -0,0 +1,35 @@
+#!/bin/sh
+# TODO: run this script from CK root or build directory
+EXE="$(find . -name tile_example_fmha_bwd -type f | head -n 1)"
+KNAME=1
+
+export CK_WARMUP=0
+export CK_REPEAT=1
+
+COMMON_ARGS='-v=1'
+set -x
+for prec in "fp16" "bf16" ; do
+for perm in 0 1 ; do
+for hdim in 32 64 128 256 ; do
+for mode in 0 1 ; do
+for bias in "n" "a" ; do
+for dbias in 0 ; do
+for p_drop in 0.0 0.2 ; do
+for deterministic in 0 ; do
+
+$EXE -prec=$prec -b=1 -h=4 -h_k=2 -d=$hdim -s=259 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
+$EXE -prec=$prec -b=2 -h=2 -d=$hdim -s=516 -s_k=253 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
+$EXE -prec=$prec -b=1 -h=4 -h_k=1 -d=$hdim -s=500 -s_k=251 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=1 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
+$EXE -prec=$prec -b=1 -h=2 -d=$hdim -s=900 -s_k=258 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=2 -v=1 -deterministic=$deterministic -mode=$mode -kname=$KNAME $COMMON_ARGS
+$EXE -prec=$prec -b=2 -h=1 -d=$hdim -s=987 -s_k=219 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=t:128,30 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
+$EXE -prec=$prec -b=2 -h=3 -h_k=1 -d=$hdim -s=244 -s_k=499 -bias=$bias -dbias=$dbias -p_drop=$p_drop -iperm=$perm -operm=$perm -mask=b:4,35 -deterministic=$deterministic -v=1 -mode=$mode -kname=$KNAME $COMMON_ARGS
+
+done
+done
+done
+done
+done
+done
+done
+done
+set +x

example/ck_tile/01_fmha/script/smoke_test_fwd.sh

@@ -0,0 +1,106 @@
+#!/bin/bash
+# TODO: run this script from CK root or build directory
+EXE="$(find . -name tile_example_fmha_fwd -type f | head -n 1)"
+KNAME=1
+
+export CK_WARMUP=0
+export CK_REPEAT=1
+
+COMMON_ARGS='-v=1 -warmup=0 -repeat=1'
+# mode=0
+# export HIP_VISIBLE_DEVICES=4
+
+TEST_SPLITKV=0
+TEST_APPENDKV=0
+# options:
+#    -s: run splitkv tests
+#    -a: run appendkv tests
+while getopts ":sa" opt; do
+    case "${opt}" in
+        s)
+            TEST_SPLITKV=1
+            ;;
+        a)
+            TEST_APPENDKV=1
+            ;;
+        *)
+            ;;
+    esac
+done
+
+run_fp16_bf16_tests() {
+    local NUM_SPLITS="1"
+    local PAGE_BLOCK_SIZE="0"
+    local CACHE_BATCH_IDX="0"
+
+    if [ $TEST_SPLITKV -eq 1 ] ; then
+        NUM_SPLITS="$NUM_SPLITS 2 3"
+        PAGE_BLOCK_SIZE="$PAGE_BLOCK_SIZE 128"
+        CACHE_BATCH_IDX="$CACHE_BATCH_IDX 1"
+    fi
+
+    for prec in "fp16" "bf16" ; do
+    for mode in 1 0 ; do
+    for perm in 0 1 ; do
+    for vlayout in "r" "c" ; do
+    for hdim in 32 64 128 256 ; do
+    for lse in 0 1 ; do
+    for bias in "n" "e" "a" ; do
+    for p_drop in 0.0 0.2 ; do
+    for num_splits in $NUM_SPLITS ; do
+    for page_block_size in $PAGE_BLOCK_SIZE ; do
+    for cache_batch_idx in $CACHE_BATCH_IDX ; do
+
+    # $EXE -prec=$prec -mode=$mode -b=1 -h=1 -d=$hdim -s=1024 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -kname=$KNAME $COMMON_ARGS  
+    $EXE -prec=$prec -mode=$mode -b=2 -h=2 -h_k=1 -d=16, -d_v=$hdim -s=55 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS  
+    $EXE -prec=$prec -mode=$mode -b=1 -h=3 -d=$hdim -s=100 -s_k=51 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS  
+    $EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=16 -d_v=$hdim -s=99 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=1 -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS  
+    $EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1024 -s_k=256 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS  
+    $EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -d_v=24 -s=3 -s_k=99 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS  
+    $EXE -prec=$prec -mode=$mode -b=3 -h=2 -h_k=1 -d=$hdim -s=200 -s_k=520 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=t:128,30 -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS  
+    $EXE -prec=$prec -mode=$mode -b=2 -h=1 -d=$hdim -s=99 -s_k=32 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=b:4,35 -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS  
+    $EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=33 -s_k=0 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS  
+    $EXE -prec=$prec -mode=$mode -b=1 -h=2 -h_k=1 -d=$hdim -s=1 -s_k=10 -s_kpad=32 -bias=$bias -p_drop=$p_drop -lse=$lse -iperm=$perm -operm=$perm -mask=2 -vlayout=$vlayout -num_splits=$num_splits -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -kname=$KNAME $COMMON_ARGS  
+
+    done ; done ; done ; done ; done
+    done ; done ; done ; done ; done
+    done ;
+}
+
+run_fp8_tests() {
+    for perm in 0 1 ; do
+    for bias in "n" "e" "a" ; do
+    for b in 1 2 ; do
+    for hdim in 64 128 256 ; do
+    
+    $EXE -prec=fp8 -init=3 -b=$b -h=1 -d=128 -s=128 -bias=$bias -iperm=$perm -operm=$perm -vlayout=c -squant=1 -kname=$KNAME $COMMON_ARGS  
+
+    done ; done ; done ; done
+}
+
+run_fp16_appendkv_tests() {
+    for s in $(seq 63 1 65) ; do
+    for s_k in 65 129 ; do
+    for s_knew in 0 64 $s_k ; do
+    for hdim in 32 64 128 256 ; do
+    for ri in 0 1 ; do
+    for rdim in 0 16 32 $hdim ; do
+    for page_block_size in 0 128 ; do
+    for cache_batch_idx in 0 1 ; do
+
+    $EXE -prec=fp16 -b=3 -h=3 -d=$hdim -s=$s -s_k=$s_k -s_knew=$s_knew -rotary_dim=$rdim -rotary_interleaved=$ri -page_block_size=$page_block_size -cache_batch_idx=$cache_batch_idx -iperm=1 -operm=1 -kname=1 $COMMON_ARGS
+
+    done ; done ; done ; done ; done 
+    done ; done ; done
+}
+
+set -x
+
+run_fp16_bf16_tests
+run_fp8_tests
+
+if [ $TEST_APPENDKV -eq 1 ] ; then
+    run_fp16_appendkv_tests
+fi
+
+set +x

example/ck_tile/01_fmha/utils.hpp

@@ -0,0 +1,266 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <algorithm>
+#include <cstdint>
+#include <cstdlib>
+#include <functional>
+#include <optional>
+#include <ostream>
+#include <sstream>
+#include <string>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+#include "ck_tile/core/container/span.hpp"
+
+enum class mode_enum
+{
+    batch = 0,
+    group
+};
+
+std::ostream& operator<<(std::ostream& stream, mode_enum mode)
+{
+    return stream << (mode == mode_enum::batch ? "batch" : "group");
+}
+
+std::vector<int32_t> to_seqstarts(ck_tile::span<const int32_t> seqlens)
+{
+    std::vector<int32_t> seqstarts = {0};
+    for(int32_t seqlen : seqlens)
+    {
+        seqstarts.push_back(seqstarts.back() + seqlen);
+    }
+    assert(seqstarts.size() == seqlens.size() + 1);
+    return seqstarts;
+}
+
+std::vector<int32_t> generate_seqlens(mode_enum mode,
+                                      unsigned count,
+                                      int32_t seqlen_avg,
+                                      int32_t seqlen_min = -1, // if not negative, clamp min
+                                      int32_t seqlen_max = -1, // if not negative, clamp max
+                                      std::optional<unsigned> seed = std::nullopt)
+{
+    assert(0 < count);
+
+    seqlen_min = (0 < seqlen_min ? seqlen_min : 1);
+    seqlen_max = (0 < seqlen_max ? seqlen_max : std::numeric_limits<int32_t>::max());
+    assert(seqlen_min <= seqlen_max);
+
+    std::vector<int32_t> seqlens(count, std::clamp(seqlen_avg, seqlen_min, seqlen_max));
+
+    if(mode == mode_enum::group && 1 < count)
+    {
+        using size_type = std::vector<int32_t>::size_type;
+
+        std::mt19937 random_engine(seed.has_value() ? *seed : std::random_device{}());
+        std::uniform_int_distribution<size_type> idx_dist(0, count - 1);
+        auto next_idx = std::bind(idx_dist, std::ref(random_engine));
+
+        std::uniform_int_distribution<size_type> step_dist(1, count - 1);
+        auto next_step = std::bind(step_dist, std::ref(random_engine));
+
+        for(unsigned repeat = seqlen_avg * (count / 2); 0 < repeat; --repeat)
+        {
+            const size_type to_decrease = next_idx();
+            // make sure each elements of seqlens is in range [seqlen_min, seqlen_max]
+            if(seqlens[to_decrease] == seqlen_min)
+            {
+                continue;
+            }
+
+            const size_type to_increase = (to_decrease + next_step()) % count;
+
+            if(seqlens[to_increase] >= seqlen_max)
+            {
+                continue;
+            }
+
+            --seqlens[to_decrease];
+            ++seqlens[to_increase];
+        }
+    }
+
+    return seqlens;
+}
+
+std::vector<int32_t> generate_seqstarts(mode_enum mode,
+                                        unsigned count,
+                                        int32_t seqlen_avg,
+                                        int32_t seqlen_min           = -1,
+                                        int32_t seqlen_max           = -1,
+                                        std::optional<unsigned> seed = std::nullopt)
+{
+    return to_seqstarts(generate_seqlens(mode, count, seqlen_avg, seqlen_min, seqlen_max, seed));
+}
+
+// return random integer generated uniformly in range [low, high]
+template <typename Int = int>
+auto randint(Int low, Int high, std::optional<unsigned> seed = std::nullopt)
+    -> std::enable_if_t<std::is_integral_v<Int>, Int>
+{
+    std::mt19937 engine(seed.has_value() ? *seed : std::random_device{}());
+    std::uniform_int_distribution<Int> dist(low, high);
+    return dist(engine);
+}
+
+// return random integers generated uniformly in range [low, high]
+template <typename Int, typename ForwardIterator>
+auto randints(ForwardIterator first,
+              ForwardIterator last,
+              Int low,
+              Int high,
+              std::optional<unsigned> seed = std::nullopt)
+    -> std::enable_if_t<std::is_integral_v<Int>>
+{
+    std::mt19937 engine(seed.has_value() ? *seed : std::random_device{}());
+    std::uniform_int_distribution<Int> dist(low, high);
+
+    std::generate(first, last, [&] { return dist(engine); });
+}
+
+/*
+ * decode the seqlen string from cmdline
+ * example (assume batch=3)
+ *   q_val=1,2,3 k_val=4,5,6 -> OK
+ *   q_val=1,2,3             -> OK, k same as q
+ *   q_val=1,2               -> OK, q will rand remaining 1 element, k same as q
+ *   q_val=1,2   k_val=4,5   -> OK, q/k will rand remaining 1 element
+ *   q_val=1,2,3,4           -> OK, but ignore exceed one
+ *
+ *   q_val=1,2   k_val=4,5,6 -> not OK, k must have same splits with q
+ *   q_val=1,2   k_val=4     -> not OK, k must have same splits with q
+ */
+std::tuple<std::vector<ck_tile::index_t>,
+           std::vector<ck_tile::index_t>,
+           std::vector<ck_tile::index_t>>
+decode_seqlen(mode_enum mode,
+              ck_tile::index_t batch,
+              std::string q_val,
+              std::string k_val,
+              std::string k_pad_val,
+              ck_tile::index_t seqlen_k_min = 0,
+              bool need_append_kvcache      = false,
+              std::optional<unsigned> seed  = std::nullopt)
+{
+#define _S2I_(str_) static_cast<ck_tile::index_t>(std::atoi((str_).c_str()))
+    if(mode == mode_enum::batch)
+    {
+        ck_tile::index_t q = _S2I_(q_val);
+        ck_tile::index_t k = _S2I_(k_val);
+
+        auto s_q = std::vector<ck_tile::index_t>(batch, q);
+        auto s_k = [&] {
+            const ck_tile::index_t seqlen_k_max = (k < 0 ? q : k);
+            std::vector<ck_tile::index_t> seqlen_ks(batch, seqlen_k_max);
+
+            if(1 < batch && need_append_kvcache)
+            {
+                // to keep the original s_k value, we always use seqlen_k_max in first batch
+                randints(std::next(seqlen_ks.begin()),
+                         seqlen_ks.end(),
+                         seqlen_k_min,
+                         seqlen_k_max,
+                         seed);
+                return seqlen_ks;
+            }
+
+            return seqlen_ks;
+        }();
+        auto s_kpad = std::vector<ck_tile::index_t>(batch, -1); // TODO: batch not support k_padding
+
+        // s_k should be greater than or equal to seqlen_k_min if provided
+        if(s_k.back() < seqlen_k_min)
+        {
+            std::ostringstream msg;
+            msg << __FILE__ << ":" << __LINE__ << ": seqlen_k (=" << s_k.back()
+                << ") is less than minimum seqlen_k (=" << seqlen_k_min << ")";
+            throw std::runtime_error(msg.str());
+        }
+
+        return std::make_tuple(s_q, s_k, s_kpad);
+    }
+    else
+    {
+        ck_tile::index_t idx          = 0;
+        std::string::size_type pos_q  = 0;
+        std::string::size_type pos_k  = 0;
+        std::string::size_type pos_kp = 0;
+        std::vector<ck_tile::index_t> s_q;
+        std::vector<ck_tile::index_t> s_k;
+        std::vector<ck_tile::index_t> s_kpad;
+        while(true)
+        {
+            auto found_q  = q_val.find(',', pos_q);
+            auto found_k  = k_val.find(',', pos_k);
+            auto found_kp = k_pad_val.find(',', pos_kp);
+
+            ck_tile::index_t q = _S2I_(
+                q_val.substr(pos_q, found_q == std::string::npos ? found_q : found_q - pos_q));
+            ck_tile::index_t k = _S2I_(
+                k_val.substr(pos_k, found_k == std::string::npos ? found_k : found_k - pos_k));
+            ck_tile::index_t kp = _S2I_(k_pad_val.substr(
+                pos_kp, found_kp == std::string::npos ? found_kp : found_kp - pos_kp));
+
+            s_q.push_back(q);
+            s_k.push_back(k < 0 ? q : k);
+            s_kpad.push_back(kp);
+
+            // s_k should be greater than or equal to seqlen_k_min
+            if(s_k.back() < seqlen_k_min)
+            {
+                std::ostringstream msg;
+                msg << __FILE__ << ":" << __LINE__ << ": seqlen_k (=" << s_k.back()
+                    << ") is less than minimum seqlen_k (=" << seqlen_k_min << ")";
+                throw std::runtime_error(msg.str());
+            }
+
+            idx++;
+            if(found_q == std::string::npos || idx >= batch)
+            {
+                break;
+            }
+            pos_q  = found_q + 1;
+            pos_k  = found_k == std::string::npos ? pos_k : found_k + 1;
+            pos_kp = found_kp == std::string::npos ? pos_kp : found_kp + 1;
+        }
+        if(idx < batch)
+        {
+            auto rem_q = generate_seqlens(mode, batch - idx, s_q.back(), 1, s_kpad.back(), seed);
+            auto rem_k =
+                generate_seqlens(mode, batch - idx, s_k.back(), seqlen_k_min, s_kpad.back(), seed);
+
+            s_q.insert(s_q.end(), rem_q.begin(), rem_q.end());
+            s_k.insert(s_k.end(), rem_k.begin(), rem_k.end());
+            s_kpad.insert(s_kpad.end(), batch - idx, s_kpad.back());
+        }
+        return std::make_tuple(s_q, s_k, s_kpad);
+    }
+#undef _S2I_
+}
+
+int env_get_int(const char* var_name, int default_int)
+{
+    char* v = getenv(var_name);
+    int r   = default_int;
+    if(v)
+        r = std::atoi(v);
+    return r;
+}
+
+template <typename RandomAccessIterator, typename Int>
+std::enable_if_t<std::is_integral_v<Int>> iota_shuffle(RandomAccessIterator first,
+                                                       RandomAccessIterator last,
+                                                       Int value,
+                                                       std::optional<unsigned> seed = std::nullopt)
+{
+    std::iota(first, last, value);
+
+    std::mt19937 engine(seed.has_value() ? *seed : std::random_device{}());
+    std::shuffle(first, last, engine);
+}

example/ck_tile/02_layernorm2d/CMakeLists.txt

@@ -0,0 +1,45 @@
+set(LAYERNORM2D_FWD_KNOWN_APIS "fwd;bwd")
+set(LAYERNORM2D_FWD_ENABLE_APIS  "fwd" CACHE STRING
+    "semicolon-separated list of APIs to generate (${LAYERNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
+if(LAYERNORM2D_FWD_ENABLE_APIS  STREQUAL "all")
+  set(LAYERNORM2D_FWD_ENABLE_APIS  ${LAYERNORM2D_FWD_KNOWN_APIS})
+endif()
+
+# generate a list of kernels, but not actually emit files at config sta
+execute_process(
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${LAYERNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --list_blobs
+  RESULT_VARIABLE ret
+)
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "Fail to generate kernels via Python. ${ret}")
+endif()
+
+file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/layernorm2d_fwd_blobs.txt LAYERNORM2D_FWD_GEN_BLOBS)
+
+add_custom_command(
+  OUTPUT ${LAYERNORM2D_FWD_GEN_BLOBS}
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${LAYERNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --gen_blobs
+)
+
+set(EXAMPLE_LAYERNORM2D_FWD "tile_example_layernorm2d_fwd")
+
+message("adding example ${EXAMPLE_LAYERNORM2D_FWD}")
+add_executable(${EXAMPLE_LAYERNORM2D_FWD} layernorm2d_fwd.cpp)
+rocm_install(TARGETS ${EXAMPLE_LAYERNORM2D_FWD} COMPONENT examples)
+target_include_directories(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+target_sources(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${LAYERNORM2D_FWD_GEN_BLOBS})
+
+set(EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS)
+
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+list(APPEND EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal --offload-compress)
+
+target_compile_options(${EXAMPLE_LAYERNORM2D_FWD} PRIVATE ${EXAMPLE_LAYERNORM2D_FWD_COMPILE_OPTIONS})
+
+# TODO: we have to turn off this global prop, otherwise the progress bar generated
+# by cmake will print too many files, execvp: /bin/sh: Argument list too long
+# however, this property may affect global
+# TODO: consider codegen a makefile by us
+set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)

example/ck_tile/02_layernorm2d/README.md

@@ -0,0 +1,85 @@
+# Layernorm2D forward
+
+This folder contains example for Layernorm2D forward using `ck_tile` tile-programming implementation.
+
+# Implementation and feature support
+
+## welford online algorithm
+We use welfold algorithm to update `mean`/`variance` block by block. For `N <=4096` case we can compute `mean`/`var`/`normalization` within one loop, we call it `one-pass`. For large N case, it is hard to keep `mean`/`var` inside register/LDS and then computation `normalization`, so we need to load input twice, first time to compute `mean`/`var` block-by-block, then load input another time to compute the `normalization`. We call it `two-pass`.
+
+## mean/variance save
+In training case the mean/variance need to store out (TBD, not supported yet)
+
+## prenorm/postnorm
+
+![](misc/pnorm.png)
+
+since [prenorm/postnorm](https://arxiv.org/pdf/1906.01787) is quite common in LLM blocks, this example boosts this feature by kernel fusion. Note that `prenorm`/`postnorm` always need to do elementwise-add a `shortcut` before the actual layernorm computation, and optionally store out the result to global. You can use `-fadd=1` to test `pre-add+store`, or `-fadd=2` to test `pre-add` without store out (not codegen by default).
+
+## smooth-quant/dynamic-quant
+we support smooth/dynamic quantization for `int8` output, by setting `-fquant=1` and `-prec_o=int8`. In this case the output will doing a rowwise dynamic quantization like below. Note that smooth-quant require input a `(1*N)` size per-channel scale(in fp32 in our example, though this is customizable), then elememt-wise multiply the tensor for each row, then compute the rowwise dynamic quant. if set `-fquant=2` will have the input per-channel scale stage, only the dynamic quant. This case is supported in our kernel but by default not generated (TBD: add some filter in generate.py support on-demand codegen)
+![](misc/dquant.png)
+
+```
+# assume output int8, hidden_states is [m, n] shape and in fp16/bf16
+# [m, 1]
+per_token_amax, _ = torch.max(
+     input=torch.abs(hidden_states), 
+     dim=-1, 
+     keepdim=True
+)
+per_token_scale = per_token_amax.to(dtype=torch.float32) / 127.0
+
+# quant hidden_states
+hidden_states = (hidden_states / per_token_scale).to(dtype=torch.int8)
+
+return hidden_states, per_token_scale
+# hidden_states now is int8 will feed to next layer as intput
+# per_token_scale will be used as dequant factor later layer
+```
+
+## build
+```
+# 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...
+make tile_example_layernorm2d_fwd -j
+```
+This will result in an executable `build/bin/tile_example_layernorm2d_fwd`
+
+## example
+```
+args:
+          -m    m dimension (default:3328)
+          -n    n dimension (default:4096)
+     -stride    stride per row, if -1 then equal to n (default:-1)
+          -e    epsilon (default:1e-5)
+    -save_mv    save mean/variance(invstd) or not. set to 1 in training case (default:0)
+          -v    cpu validation or not (default:1)
+      -kname    print kernel name or not (default:1)
+     -prec_i    input precision (default:fp16)
+     -prec_o    output precision, set auto will be the same as input (default:auto)
+    -prec_sm    output quant scale type, set auto will be the same as input. used when fquant=1 (default:auto)
+    -prec_sy    output quant scale type, set auto will be the same as input. used when fquant=1 or 2 (default:auto)
+       -fadd    fused-add, 0:no fused add, 1:preadd+store, 2:preadd only (default:0)
+     -fquant    fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant (default:0)
+     -warmup    cold iter (default:5)
+     -repeat    hot iter (default:20)
+
+```
+
+## limitations
+Note that `fquant=2`, `fadd=2`, `prec_sm/prec_sy` other than `fp32` are not by default generated. Though our kernel template suppor this. (TBD: add some flag in generate.py) to generate those instance on demand. Beside, `N>8192` case will by default using two-pass pipeline, and `-fquant=1/2` are not supported yet. If need suport `N>8192` and `fused+residual+store`, you can use this example together with `12_smoothquant`, to construct layernorm+residual, and smoothquant, 2 kernels for this purpose.
+
+```
+# some case
+# standard fp16 layernorm 2d, m=10. n=1024
+./build/bin/tile_example_layernorm2d_fwd  -m=10 -n=1024
+
+# standard fp16 layernorm 2d, m=10. n=1024, fused-smooth-quant, output in int8
+./build/bin/tile_example_layernorm2d_fwd  -m=10 -n=1024 -prec_o=int8 -fquant=1
+
+# standard fp16 layernorm 2d, m=10. n=1024, fused-smooth-quant+fused-add-store, output in int8
+./build/bin/tile_example_layernorm2d_fwd  -m=10 -n=1024 -prec_o=int8 -fquant=1 -fadd=1
+
+```

example/ck_tile/02_layernorm2d/generate.py

@@ -0,0 +1,730 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import argparse
+from enum import IntEnum
+from pathlib import Path
+import sys
+from typing import List, Optional, Any
+import functools
+import itertools
+import copy
+from dataclasses import dataclass
+
+def get_if_str(idx, total, lase_else = True):
+    if idx == 0:
+        return 'if'
+    elif idx < total - 1:
+        return 'else if'
+    else:
+        if lase_else:
+            return 'else'
+        else:
+            return 'else if'
+
+XBIAS_ENUM_STR_MAP = [
+    'no',
+    'xbias']      # pre-norm add bias
+
+FUSED_ADD_ENUM_STR_MAP = [
+    'no',
+    'pras',      # pre-norm
+    'pra' ]      # post-norm
+
+FUSED_FUSED_SWEEP_STR_MAP = [
+    'no',
+    'dquant' ]
+
+DATA_TYPE_MAP = {'fp32' : 'float',
+                 'fp16' : 'ck_tile::fp16_t',
+                 'bf16' : 'ck_tile::bf16_t',
+                 'int8' : 'ck_tile::int8_t',
+                 'fp8'  : 'ck_tile::fp8_t'}
+
+def BOOL_MAP(b_) -> str:
+    if b_:
+        return 'true'
+    else:
+        return 'false'
+
+class layernorm_fwd_codegen:
+    API_TRAITS_DEFINE = """
+// this is used to pattern-match internl kernel implementation, not to instantiate kernel
+template <typename XDataType_,
+          typename YDataType_,
+          typename SmoothScaleDataType_,
+          typename YScaleDataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kSaveMeanInvStd_,
+          bool kFastFDiv_,
+          bool kWelford_,
+          bool kTwoPass_,
+          ck_tile::index_t kXbias_ = 0,
+          ck_tile::index_t kFusedAdd_ = 0,
+          ck_tile::index_t kFusedQuant_ = 0>
+struct layernorm2d_fwd_traits_
+{
+    using XDataType = ck_tile::remove_cvref_t<XDataType_>;
+    using YDataType = ck_tile::remove_cvref_t<YDataType_>;
+    using SmoothScaleDataType = ck_tile::remove_cvref_t<SmoothScaleDataType_>;
+    using YScaleDataType = ck_tile::remove_cvref_t<YScaleDataType_>;
+
+    static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize;
+    static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % warpSize == 0);
+    static constexpr ck_tile::index_t total_warps =
+        (ThreadPerBlock_M_ * ThreadPerBlock_N_) / warpSize;
+
+    // num of warps along m
+    static constexpr ck_tile::index_t BlockWarps_M = []() {
+        if constexpr(is_warp_per_row)
+        {
+            static_assert(warpSize % ThreadPerBlock_N_ == 0);
+            return total_warps * (warpSize / ThreadPerBlock_N_);
+        }
+        else
+        {
+            // static_assert(warpSize % ThreadPerBlock_M_ == 0);
+            return total_warps / (ThreadPerBlock_N_ / warpSize);
+        }
+    }();
+
+    // num of warps along n
+    static constexpr ck_tile::index_t BlockWarps_N = []() {
+        if constexpr(is_warp_per_row)
+        {
+            static_assert(warpSize % ThreadPerBlock_N_ == 0);
+            return 1;
+        }
+        else
+        {
+            static_assert(ThreadPerBlock_N_ % warpSize == 0);
+            return ThreadPerBlock_N_ / warpSize;
+        }
+    }();
+
+    static constexpr ck_tile::index_t Repeat_M = Repeat_M_;
+    static constexpr ck_tile::index_t Repeat_N = Repeat_N_;
+
+    static constexpr ck_tile::index_t Block_M = Repeat_M_ * ThreadPerBlock_M_;
+    static constexpr ck_tile::index_t Block_N = Repeat_N_ * ThreadPerBlock_N_ * Vector_N_;
+
+    static constexpr ck_tile::index_t Warp_M = ThreadPerBlock_M_ / BlockWarps_M;
+    static constexpr ck_tile::index_t Warp_N = ThreadPerBlock_N_ / BlockWarps_N * Vector_N_;
+
+    using BlockTile  = ck_tile::sequence<Block_M, Block_N>;
+    using BlockWarps = ck_tile::sequence<BlockWarps_M, BlockWarps_N>;
+    using WarpTile   = ck_tile::sequence<Warp_M, Warp_N>;
+    using Vector     = ck_tile::sequence<1, Vector_N_>;
+
+    using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
+
+    static constexpr bool kPadN           = kPadN_;
+    static constexpr bool kSaveMeanInvStd = kSaveMeanInvStd_;
+    static constexpr bool kFastFDiv       = kFastFDiv_;
+    static constexpr bool kWelford        = kWelford_;
+    static constexpr bool kTwoPass        = kTwoPass_;
+    static constexpr ck_tile::index_t kXbias = kXbias_;
+    static constexpr ck_tile::index_t kFusedAdd = kFusedAdd_;
+    static constexpr ck_tile::index_t kFusedQuant = kFusedQuant_;
+};
+
+template <typename XDataType_,
+          typename YDataType_,
+          typename SmoothScaleDataType_,
+          typename YScaleDataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kSaveMeanInvStd_,
+          bool kFastFDiv_,
+          bool kWelford_,
+          bool kTwoPass_,
+          int  kXbias_,
+          int  kFusedAdd_,
+          int  kFusedQuant_>
+using traits_ = layernorm2d_fwd_traits_<XDataType_,
+                                       YDataType_,
+                                       SmoothScaleDataType_,
+                                       YScaleDataType_,
+                                       Repeat_M_,
+                                       Repeat_N_,
+                                       ThreadPerBlock_M_,
+                                       ThreadPerBlock_N_,
+                                       Vector_N_,
+                                       kPadN_,
+                                       kSaveMeanInvStd_,
+                                       kFastFDiv_,
+                                       kWelford_,
+                                       kTwoPass_,
+                                       kXbias_,
+                                       kFusedAdd_,
+                                       kFusedQuant_>;
+"""
+    API_COMMON_HEADER = """
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "layernorm2d_fwd.hpp"
+#include <ck_tile/ops/epilogue.hpp>
+#include <iostream>
+
+#pragma once
+
+using S = ck_tile::stream_config;
+using A = layernorm2d_fwd_args;
+
+{F_traits_define}
+
+template <typename Traits_>
+float layernorm2d_fwd_(const S& s, A a)
+{{
+    using XDataType = typename Traits_::XDataType;
+    using YDataType = typename Traits_::YDataType;
+    using SmoothScaleDataType = typename Traits_::SmoothScaleDataType;
+    using YScaleDataType = typename Traits_::YScaleDataType;
+    using ComputeDataType = typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::ComputeDataType;
+
+    using PipelineTraits = ck_tile::Layernorm2dFwdTraits<Traits_::kPadN,
+        Traits_::kSaveMeanInvStd,
+        Traits_::kFastFDiv,
+        Traits_::kWelford,
+        Traits_::kTwoPass,
+        static_cast<ck_tile::Layernorm2dXBiasEnum>(Traits_::kXbias),
+        static_cast<ck_tile::Layernorm2dFusedAddEnum>(Traits_::kFusedAdd),
+        static_cast<ck_tile::Layernorm2dFusedQuantEnum>(Traits_::kFusedQuant)>;
+    using PipelineProblem = ck_tile::Layernorm2dFwdPipelineProblem<
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::XDataType,
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::XBiasDataType,
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::GammaDataType,
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::BetaDataType,
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::ComputeDataType,
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::YDataType,
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::MeanDataType,
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::InvStdDataType,
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::SmoothScaleDataType,
+        typename LayerNormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::YScaleDataType,
+        typename Traits_::Shape,
+        PipelineTraits>;
+
+    using OnePassPipeline = ck_tile::Layernorm2dFwdPipelineOnePass<PipelineProblem>;
+    using TwoPassPipeline = ck_tile::Layernorm2dFwdPipelineTwoPass<PipelineProblem>;
+    using Pipeline        = std::conditional_t<Traits_::kTwoPass, TwoPassPipeline, OnePassPipeline>;
+
+    using Default2DEpilogueProblem = ck_tile::Default2DEpilogueProblem<ComputeDataType, YDataType, false, Traits_::kPadN, true>;
+    using Default2DEpilogue = ck_tile::Default2DEpilogue<Default2DEpilogueProblem>;
+
+    static constexpr bool UseSmoothInputScale = Traits_::kFusedQuant == 1;
+    static constexpr bool UseRawStore = sizeof(YDataType) == 4;
+    using DynamicQuantEpilogueProblem = ck_tile::DynamicQuantEpilogueProblem<ComputeDataType, SmoothScaleDataType, YScaleDataType, YDataType, typename Traits_::Shape,
+            ck_tile::DynamicQuantEpilogueTraits<false, Traits_::kPadN, UseSmoothInputScale, UseRawStore,  true/*max3*/>>;
+
+    using DynamicQuantEpilogue = ck_tile::DynamicQuantEpilogue<DynamicQuantEpilogueProblem>;
+
+    using Epilogue = std::conditional_t<Traits_::kFusedQuant == 1, DynamicQuantEpilogue,  Default2DEpilogue>;
+
+    using Kernel = ck_tile::Layernorm2dFwd<Pipeline, Epilogue>;
+
+    const dim3 grids                       = Kernel::GridSize(a);
+    constexpr dim3 blocks                  = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+
+    auto kargs = Kernel::MakeKargs(a);
+    if(s.log_level_ > 0)
+        std::cout << ", " << Kernel::GetName() << std::flush;
+
+    return ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{{}}, grids, blocks, 0, kargs));
+}}
+
+"""
+
+    API_BASE = """
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "layernorm2d_fwd.hpp"
+
+{F_traits_define}
+
+// Note: this internal API only declare, not define here, otherwise will block `make -j`
+template <typename Traits_>
+float layernorm2d_fwd_(const ck_tile::stream_config& s, layernorm2d_fwd_args a);
+
+float layernorm2d_fwd(layernorm2d_fwd_traits t,
+                      layernorm2d_fwd_args a,
+                      const ck_tile::stream_config& s)
+{{
+    float r = -1;
+{F_dispatch}
+    return r;
+}}
+
+"""
+
+    API_PER_DTYPE="""    {F_if}(t.prec_i == \"{F_i_type}\" && t.prec_o == \"{F_o_type}\"){{
+{F_per_n_case}
+    }}
+"""
+    API_PER_N_CASE="""        {F_if} {F_N_COND} {{
+{F_inner_dispatch}
+        }}
+"""
+    API_INNER_CASE="""            {F_if} {F_VEC_COND}
+                r={F_instance_func}(s, a);
+"""
+
+    INSTANCE_BASE = """
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "layernorm2d_fwd_api_common.hpp"
+
+// clang-format off
+//                                      prec_i           prec_o           prec_sy           rm  rn  tm    tn  vn  pd     mv    rpcf   welford   2p   xbias   add  sweep
+{F_instance_def}
+// clang-format on
+
+"""
+
+    def __init__(self, working_path, kernel_filter):
+        self.working_path = working_path
+        self.kernel_filter = kernel_filter
+
+    class k_xbias_enum(IntEnum):
+        F_NO_XBIAS = 0
+        F_ADD_XBIAS = 1
+
+    class k_fuesd_add_enum(IntEnum):
+        F_NO_ADD = 0
+        F_PRE_ADD = 1
+        F_PRE_ADD_STORE_RESIDUAL = 2
+
+    class k_fused_sweep_enum(IntEnum):
+        F_NO_SWEEP = 0
+        F_RENORM = 1
+        F_DYNAMIC_QUANT = 2
+
+    @dataclass
+    class k_traits:
+        F_kPadN : bool
+        F_kSaveMeanInvStd : bool
+        F_kTwoPass : bool
+        F_kXbias : Any #: layernorm_fwd_codegen.k_bias_enum
+        F_kFusedAdd : Any #: layernorm_fwd_codegen.k_fuesd_add_enum
+        F_kFusedQuant : Any  #: layernorm_fwd_codegen.k_fused_sweep_enum
+
+    @dataclass
+    class k_shape:
+        F_BlockTile    : List[int]
+        F_WarpPerBlock : List[int]
+        F_WarpTile     : List[int]
+        F_Vector_      : List[int]
+        @property
+        def F_BlockSize(self) -> int:
+            return functools.reduce(lambda a, b: a*b, self.F_WarpTile)
+
+    @dataclass
+    class k_problem:
+        F_XDataType       : str
+        F_XBiasDataType   : str
+        F_GammaDataType   : str
+        F_BetaDataType    : str
+        F_ComputeDataType : str
+        F_YDataType       : str
+        F_MeanDataType    : str
+        F_InvStdDataType  : str
+        F_BlockShape      : str
+        F_Traits          : Any #k_traits
+
+    @dataclass
+    class k_pipeline_one_pass:
+        F_Problem         : Any #k_problem
+    
+    @dataclass
+    class k_pipeline_two_pass:
+        F_Problem         : Any #k_problem
+
+    @dataclass
+    class default_2d_epilogue_problem:
+        F_AccDataType : str
+        F_ODataType : str
+        F_kPadM : bool
+        F_kPadN : bool
+
+    @dataclass
+    class default_2d_epilogue:
+        F_problem : Any
+
+    @dataclass
+    class k_kernel:
+        F_pipeline : Any
+        F_epilogue : Any
+
+    @dataclass
+    class h_traits:
+        F_XDataType : str
+        F_YDataType : str
+        F_SmoothScaleDataType : str
+        F_YScaleDataType : str
+        F_Repeat_M : int
+        F_Repeat_N : int
+        F_ThreadPerBlock_M : int
+        F_ThreadPerBlock_N : int
+        F_Vector_N : int
+        F_kPadN : bool
+        F_kSaveMeanInvStd_ : bool
+        F_kFastFDiv_ : bool
+        F_kWelford_ : bool
+        F_kTwoPass_ : bool
+        F_kXbias_ : int
+        F_kFusedAdd : int
+        F_kFusedQuant : int
+
+        @property
+        def trait_name(self) ->str:
+            t_ = f'{DATA_TYPE_MAP[self.F_XDataType]}, {DATA_TYPE_MAP[self.F_YDataType]}, {DATA_TYPE_MAP[self.F_SmoothScaleDataType]}, {DATA_TYPE_MAP[self.F_YScaleDataType]}, {self.F_Repeat_M:2}, {self.F_Repeat_N:2}, {self.F_ThreadPerBlock_M:2}, {self.F_ThreadPerBlock_N:4}'
+            t_ += f', {self.F_Vector_N:2}, {BOOL_MAP(self.F_kPadN):5}, {BOOL_MAP(self.F_kSaveMeanInvStd_):5}, {BOOL_MAP(self.F_kFastFDiv_):5}, {BOOL_MAP(self.F_kWelford_):5}'
+            t_ += f', {BOOL_MAP(self.F_kTwoPass_):5}, {self.F_kXbias:4}, {self.F_kFusedAdd:4}, {self.F_kFusedQuant:4}'
+            return t_
+
+        # string when calling this kernel
+        @property
+        def call_name(self) -> str:
+            return f'layernorm2d_fwd_<traits_<{self.trait_name}>>'
+
+        # string when define this kernel
+        @property
+        def def_name(self) -> str:
+            return f'template float layernorm2d_fwd_<traits_<{self.trait_name}>>(const S&, A);'
+
+    # this class hold kernel under same source file
+    @dataclass
+    class h_instance:
+        F_DataTypePair : str
+        F_N : str
+        F_xbias : int
+        F_add : int
+        F_sweep : int
+        instance_list : List[Any] # List[h_traits]
+
+        @property
+        def name(self) -> str:
+            prec_i, prec_o = self.F_DataTypePair.split(',')
+            dtype_str = f'{prec_i}' if prec_i == prec_o else f'{prec_i}_{prec_o}'
+            nnn = f'layernorm2d_fwd_{dtype_str}_n{self.F_N}'
+            if self.F_xbias != 0:
+                nnn = nnn + '_' + XBIAS_ENUM_STR_MAP[self.F_xbias] 
+            if self.F_add != 0:
+                nnn = nnn + '_' + FUSED_ADD_ENUM_STR_MAP[self.F_add]
+            if self.F_sweep != 0:
+                nnn = nnn + '_' + FUSED_FUSED_SWEEP_STR_MAP[self.F_sweep]
+            return nnn
+
+        @property
+        def instance_name(self) ->str:
+            return self.name
+
+        @property
+        def content(self) ->str:
+            instance_defs = ''
+            for ins in self.instance_list:
+                instance_defs += ins.def_name + '\n'
+            return layernorm_fwd_codegen.INSTANCE_BASE.format(F_instance_def=instance_defs)
+
+    @property
+    def name_api(self) -> str:
+        return 'layernorm2d_fwd_api'
+
+    @property
+    def name_common_header(self) -> str:
+        return 'layernorm2d_fwd_api_common'
+
+    def content_api(self, args) -> str:
+        # 1 sort based on dtype
+        t_dtype_dict = dict()
+        blobs = self.get_blobs(args)
+        for blob in blobs:
+            if blob.F_DataTypePair not in t_dtype_dict:
+                t_dtype_dict[blob.F_DataTypePair] = {}
+            if blob.F_N not in t_dtype_dict[blob.F_DataTypePair]:
+                t_dtype_dict[blob.F_DataTypePair][blob.F_N] = []
+            t_dtype_dict[blob.F_DataTypePair][blob.F_N].append(blob)
+
+        d_str = ''
+        for i_d, dtype_ in enumerate(t_dtype_dict):
+            blob_per_t = t_dtype_dict[dtype_]
+            n_str = ''
+            for i_n, n_ in enumerate(blob_per_t):
+                blob_per_n = blob_per_t[n_]
+                inner_str = ""
+                for i_b, b_ in enumerate(blob_per_n):
+                    # generate single kernel instance file
+                    #vec_str = ""
+                    for i_ins, ins in enumerate(b_.instance_list):
+                        idx_in_n = i_b * len(b_.instance_list) + i_ins
+                        len_in_n = len(blob_per_n) * len(b_.instance_list)
+                        # _if = 'if' if i_ins == 0 else 'else if'
+                        if ins.F_kFusedQuant == 0:
+                            _sweep_cond = 't.fused_quant == {f_fused_sweep}'.format(f_fused_sweep = ins.F_kFusedQuant)
+                        elif ins.F_kFusedQuant == 1:
+                            _sweep_cond = 't.fused_quant == {f_fused_sweep} && (t.prec_sm == \"{f_sx_type}\" && t.prec_sy == \"{f_sy_type}\")'.format(
+                                f_fused_sweep = ins.F_kFusedQuant, f_sx_type=ins.F_SmoothScaleDataType, f_sy_type=ins.F_YScaleDataType)
+                        elif ins.F_kFusedQuant == 2:
+                            _sweep_cond = 't.fused_quant == {f_fused_sweep} && (t.prec_sy == \"{f_sy_type}\")'.format(
+                                f_fused_sweep = ins.F_kFusedQuant, f_sy_type=ins.F_YScaleDataType)
+                        _cond = '((a.n % {f_vec_n} == 0) && (t.xbias == {f_xbias}) && (t.fused_add == {f_fused_add}) && ({f_sweep_cond}))'.format(
+                                        f_vec_n = ins.F_Vector_N, f_xbias = ins.F_kXbias, f_fused_add = ins.F_kFusedAdd,
+                                        f_sweep_cond = _sweep_cond)
+                        inner_str += self.API_INNER_CASE.format(F_if = get_if_str(idx_in_n, len_in_n, False),
+                                            F_VEC_COND = _cond, F_instance_func=ins.call_name)
+                    #inner_str = inner_str + vec_str
+                n_cnd = f'(a.n <= {n_})' if isinstance(n_, int) else ''
+                n_str += self.API_PER_N_CASE.format(F_if = get_if_str(i_n, len(blob_per_t), not isinstance(n_, int)), F_N_COND=n_cnd, F_inner_dispatch=inner_str)
+            prec_i, prec_o = dtype_.split(',')
+            d_str += self.API_PER_DTYPE.format(F_if = get_if_str(i_d, len(t_dtype_dict), False), F_i_type=prec_i, F_o_type=prec_o, F_per_n_case=n_str)
+
+        api_base = self.API_BASE.format(F_traits_define=self.API_TRAITS_DEFINE, F_dispatch=d_str)
+        return api_base
+
+    @property
+    def content_common_header(self) -> str:
+        return self.API_COMMON_HEADER.format(F_traits_define=self.API_TRAITS_DEFINE)
+
+    def get_blobs(self, args):
+        h_traits = layernorm_fwd_codegen.h_traits
+        h_instance = layernorm_fwd_codegen.h_instance
+
+        dynamic_quant_out_dtype = ['int8', 'fp8']
+        # some predefined support range
+        # (prec_i,prec_o) for simplicity this string will be used as key for dict
+        scale_list = [('fp32,fp32')]
+        dtype_list = [('fp16,fp16'), ('bf16,bf16'),
+                        ('fp16,int8'), ('bf16,int8'),
+                        ('fp16,fp8'), ('bf16,fp8')] # NOTE: only fused-dynamic-quant use int8 or fp8 out
+        types_8bit = ('int8', 'fp8')
+        types_16bit = ('int16', 'fp16', 'bf16')
+        #fused_add_list = [0, 1, 2]
+        #fused_sweep_list = [0, 1, 2] # NOTE: only single pass can use fused dynamic quant
+        xbias_list = [0, 1]
+        fused_add_list = [0, 1]
+        fused_sweep_list = [0, 1] # NOTE: only single pass can use fused dynamic quant
+        #                                                       rm  rn  tm   tn  vn  pd     mv     fdiv  welford   2p     xbias    add   sweep
+        h_trait_dict = {'64'  : [ h_traits('x', 'y', 'xs', 'ys', 1,  1,  8,  8,  8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  1,  4,  16, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  1,  4,  64, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '128' : [ h_traits('x', 'y', 'xs', 'ys', 1,  1,  4,  16, 8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  1,  4,  64, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  2,  4,  64, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '256' : [ h_traits('x', 'y', 'xs', 'ys', 1,  1,  4,  64, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  2,  4,  64, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  4,  64, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '512' : [ h_traits('x', 'y', 'xs', 'ys', 1,  1,  4,  64, 8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  2,  4,  64, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  4,  64, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  8,  4,  64, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '768' : [ h_traits('x', 'y', 'xs', 'ys', 1,  3,  4,  64, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  6,  4,  64, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1, 12,  4,  64, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '1024' :[ h_traits('x', 'y', 'xs', 'ys', 1,  1,  2, 128, 8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  2,  2, 128, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  2, 128, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  1, 256, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '1536' :[ h_traits('x', 'y', 'xs', 'ys', 1,  3,  4,  64, 8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  3,  2, 128, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  3,  1, 256, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  6,  1, 256, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '2048' :[ h_traits('x', 'y', 'xs', 'ys', 1,  1,  1, 256, 8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  2,  1, 256, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  1, 256, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  8,  1, 256, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '3072' :[ h_traits('x', 'y', 'xs', 'ys', 1,  3,  1, 128, 8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  3,  1, 256, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  6,  1, 256, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  3,  1,1024, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '4096' :[ h_traits('x', 'y', 'xs', 'ys', 1,  2,  1, 256, 8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  1, 256, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  2,  1,1024, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  1,1024, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '6144' :[ h_traits('x', 'y', 'xs', 'ys', 1,  3,  1, 256, 8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  3,  1, 512, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  3,  1,1024, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  6,  1,1024, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        '8192' :[ h_traits('x', 'y', 'xs', 'ys', 1,  4,  1, 256, 8,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  1, 512, 4,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  1,1024, 2,  True,  False, True, True,   False,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  8,  1,1024, 1,  True,  False, True, True,   False,   0,    0,    0)],
+                        'big'  :[ h_traits('x', 'y', 'xs', 'ys', 1,  1,  1,1024, 8,  True,  False, True, True,    True,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  1, 256, 4,  True,  False, True, True,    True,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1, 12,  1, 256, 2,  True,  False, True, True,    True,   0,    0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 1,  4,  1,1024, 1,  True,  False, True, True,    True,   0,    0,    0)]}
+        total_blob = list()
+        for hs_key in h_trait_dict:
+            hs = h_trait_dict[hs_key]
+            current_n = hs[0].F_Repeat_N * hs[0].F_ThreadPerBlock_N * hs[0].F_Vector_N
+            for dtype, scale_type, xbias, fused_add, fused_quant in itertools.product(dtype_list, scale_list, xbias_list, fused_add_list, fused_sweep_list):
+                prec_i, prec_o = dtype.split(',')
+                scale_sm, scale_y = scale_type.split(',')
+                if prec_o in dynamic_quant_out_dtype and fused_quant != 1:
+                    continue # skip non dynamic quant case
+                if fused_quant == 1 and hs_key == 'big':
+                    continue
+                current_hs = list()
+                for chs_ in hs:
+                    h_ = copy.copy(chs_) # copy the base instance out
+                    h_.F_XDataType = prec_i
+                    h_.F_YDataType = prec_o
+                    h_.F_SmoothScaleDataType = scale_sm
+                    h_.F_YScaleDataType = scale_y
+                    h_.F_kXbias = xbias
+                    h_.F_kFusedAdd = fused_add
+                    h_.F_kFusedQuant = fused_quant
+                    # disable welford update for 8bit and 16 bit smallN
+                    if not h_.F_kTwoPass_:
+                        #disable 16 bit when set args disable_16b_welford
+                        if args.disable_16b_welford and prec_i in types_16bit:
+                            h_.F_kWelford_ = False
+                        #disable 8bit by default
+                        elif prec_i in types_8bit or prec_o in types_8bit:
+                            h_.F_kWelford_ = False
+                        #disable 16bit small N
+                        elif prec_i in types_16bit and hs_key == '64':
+                            h_.F_kWelford_ = False
+                    current_hs.append(h_) # + "\n"
+                #f.write(str(f.parent / GEN_DIR / (blobs.api_common_header_
+                current_n_str = 'big' if hs_key == 'big' else current_n
+                total_blob.append(h_instance(dtype, current_n_str, xbias, fused_add, fused_quant, current_hs))
+        return total_blob
+
+    def list_blobs(self, args) -> None:
+        w_p = Path(self.working_path)
+        list_p = w_p / 'layernorm2d_fwd_blobs.txt'
+        blobs = self.get_blobs(args)
+        with list_p.open('w') as list_f:
+            # api related file
+            list_f.write(str(w_p / (self.name_api + ".cpp"))  + "\n")
+            list_f.write(str(w_p / (self.name_common_header + ".hpp"))  + "\n")
+            # kernel instance file
+            for b in blobs:
+                list_f.write(str(w_p / (b.name + ".cpp")) + "\n")
+
+    def gen_blobs(self, args) -> None:
+        w_p = Path(self.working_path)
+        w_str = self.content_api(args)
+        (w_p / (self.name_api + ".cpp")).write_text(w_str)
+        (w_p / (self.name_common_header + ".hpp")).write_text(self.content_common_header)
+        blobs = self.get_blobs(args)
+        for b in blobs:
+            (w_p / (b.name + ".cpp")).write_text(b.content)
+
+def list_blobs(args):
+    api_list = args.api.split(',')
+    for api in api_list:
+        if api == 'fwd':
+            layernorm_fwd_codegen(args.working_path, args.filter).list_blobs(args)
+
+
+def gen_blobs(args):
+    api_list = args.api.split(',')
+    for api in api_list:
+        if api == 'fwd':
+            layernorm_fwd_codegen(args.working_path, args.filter).gen_blobs(args)
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        prog="generate",
+        description="gen API for CK layernorm kernel",
+    )
+    parser.add_argument(
+        "-a",
+        "--api",
+        default='fwd[all]',
+        required=False,
+        help="supply API(s) to generate (default: fwd). separated by comma."
+    )
+
+    # the directory for list_blobs/gen_blobs to write files into
+    parser.add_argument(
+        "-w",
+        "--working_path",
+        default="./",
+        required=False,
+        help="the path where all the blobs are going to be generated"
+    )
+
+    # this script have 2 modes
+    # 1) list_blobs mode, will generate a txt file with all the files going to be generated.
+    #    this is useful in build system like cmake to construct source code dependency, by
+    #    reading the content out of this file
+    # 2) gen_blobs mode, will generate the actuall kernel instance and api. If in framework
+    #    like FA, only need to use this mode
+    parser.add_argument(
+        "-l",
+        "--list_blobs",
+        action='store_true',
+        help="list all the kernels to a file, "
+    )
+
+    parser.add_argument(
+        "-g",
+        "--gen_blobs",
+        action='store_true',
+        help="generate all kernels into different tile"
+    )
+
+    # TODO: if using filter, must apply same value to output_dir and list_blobs
+    parser.add_argument(
+        "-f",
+        "--filter",
+        required=False,
+        help="filter out kernels that need to generate, using fnmatch module"
+    )
+
+    parser.add_argument(
+        "-t",
+        "--traits",
+        default="all",
+        required=False,
+        help="enable/disable some feature. default generate all"
+    )
+
+    parser.add_argument(
+        "-r",
+        "--receipt",
+        default=0,
+        required=False,
+        help="codegen receipt."
+    )
+
+    parser.add_argument(
+        "--disable_16b_welford",
+        default=False,
+        required=False,
+        help="enable/disable welford for 16bit datatype n > 64"
+    )
+
+    args = parser.parse_args()
+
+    # print(f'{args.list_blobs}-{args.gen_blobs}')
+    if (args.gen_blobs and args.list_blobs) or ((not args.gen_blobs) and (not args.list_blobs)):
+        print('gen_blobs/list_blobs must specify only one option')
+        sys.exit()
+
+    p = Path(args.working_path)
+    if not p.exists():
+        p.mkdir()
+
+    if args.list_blobs:
+        list_blobs(args)
+    else:
+        gen_blobs(args)

example/ck_tile/02_layernorm2d/layernorm2d_fwd.cpp

@@ -0,0 +1,481 @@
+#include "ck_tile/host.hpp"
+#include "layernorm2d_fwd.hpp"
+#include <algorithm>
+#include <cstring>
+
+// different threshold for different dtype
+template <typename DataType>
+auto get_elimit()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::bf16_t>()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::int8_t>()
+{
+    double rtol = 1e-2;
+    double atol = 1.0;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("m", "3328", "m dimension")
+        .insert("n", "4096", "n dimension")
+        .insert("x_stride", "-1", "x row_stride, if -1 then equal to n")
+        .insert("xr_stride", "-1", "x residule row_stride, if -1 then equal to n")
+        .insert("y_stride", "-1", "y row_stride, if -1 then equal to n")
+        .insert("yr_stride", "-1", "y residule row_stride, if -1 then equal to n")
+        .insert("e", "1e-5", "epsilon")
+        .insert("save_mv", "0", "save mean/variance(invstd) or not. set to 1 in training case")
+        .insert("v", "1", "cpu validation or not")
+        .insert("kname", "1", "print kernel name or not")
+        .insert("prec_i", "fp16", "input precision")
+        .insert("prec_o", "auto", "output precision, set auto will be the same as input")
+        .insert("prec_sm",
+                "auto",
+                "output quant scale type, set auto will use fp32. used when fquant=1")
+        .insert("prec_sy",
+                "auto",
+                "output quant scale type, set auto will use fp32. used when fquant=1 or 2")
+        .insert("xbias", "0", "add bias, 0:no add, 1:add bias before fadd")
+        .insert("fadd", "0", "fused-add, 0:no fused add, 1:preadd+store, 2:preadd only")
+        .insert("fquant", "0", "fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant")
+        .insert("warmup", "5", "cold iter")
+        .insert("repeat", "20", "hot iter");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename InDataType,
+          typename OutDataType,
+          typename SmoothScaleDataType,
+          typename YScaleDataType,
+          bool SaveMeanVar>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    ck_tile::index_t m        = arg_parser.get_int("m");
+    ck_tile::index_t n        = arg_parser.get_int("n");
+    ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
+    if(x_stride < 0)
+        x_stride = n;
+    ck_tile::index_t xr_stride = arg_parser.get_int("xr_stride");
+    if(xr_stride < 0)
+        xr_stride = n;
+    ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
+    if(y_stride < 0)
+        y_stride = n;
+    ck_tile::index_t yr_stride = arg_parser.get_int("yr_stride");
+    if(yr_stride < 0)
+        yr_stride = n;
+    float epsilon       = arg_parser.get_float("e");
+    std::string prec_i  = arg_parser.get_str("prec_i");
+    std::string prec_o  = arg_parser.get_str("prec_o");
+    std::string prec_sm = arg_parser.get_str("prec_sm");
+    std::string prec_sy = arg_parser.get_str("prec_sy");
+    if(prec_o == "auto")
+    {
+        prec_o = prec_i;
+    }
+    if(prec_sm == "auto")
+    {
+        prec_sm = "fp32";
+    }
+    if(prec_sy == "auto")
+    {
+        prec_sy = "fp32";
+    }
+
+    int kname         = arg_parser.get_int("kname");
+    int do_validation = arg_parser.get_int("v");
+    int warmup        = arg_parser.get_int("warmup");
+    int repeat        = arg_parser.get_int("repeat");
+    int xbias         = arg_parser.get_int("xbias");
+    int fused_add     = arg_parser.get_int("fadd");
+    int fused_quant   = arg_parser.get_int("fquant");
+    if(fused_quant == 1 && prec_o != "int8" && prec_o != "fp8")
+    {
+        std::cout
+            << "if fused_quant is 1 or 2, only support \"-prec_o=int8\" or \"-prec_o=fp8\" cases."
+            << std::endl;
+        return false;
+    }
+
+    assert(x_stride >= n);
+
+    using TypeConfig =
+        LayerNormTypeConfig<InDataType, OutDataType, SmoothScaleDataType, YScaleDataType>;
+
+    using XDataType         = typename TypeConfig::XDataType;
+    using YDataType         = typename TypeConfig::YDataType;
+    using XBiasDataType     = typename TypeConfig::XBiasDataType;
+    using GammaDataType     = typename TypeConfig::GammaDataType;
+    using BetaDataType      = typename TypeConfig::BetaDataType;
+    using XResidualDataType = XDataType;
+    using YResidualDataType = XDataType;
+
+    using MeanDataType =
+        std::conditional_t<SaveMeanVar, typename TypeConfig::MeanDataType, ck_tile::null_type>;
+    using InvStdDataType =
+        std::conditional_t<SaveMeanVar, typename TypeConfig::InvStdDataType, ck_tile::null_type>;
+
+    using ComputeDataType = typename TypeConfig::ComputeDataType;
+
+    // host verify
+    ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
+    ck_tile::HostTensor<XBiasDataType> x_bias_host({n});
+    ck_tile::HostTensor<GammaDataType> gamma_host({n});
+    ck_tile::HostTensor<BetaDataType> beta_host({n});
+
+    ck_tile::HostTensor<XResidualDataType> x_residual_host({m, n}, {xr_stride, 1});
+    ck_tile::HostTensor<YResidualDataType> y_residual_host({m, n}, {yr_stride, 1});
+
+    ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {y_stride, 1});
+
+    ck_tile::HostTensor<MeanDataType> mean_host_ref({m});
+    ck_tile::HostTensor<InvStdDataType> invStd_host_ref({m});
+    ck_tile::HostTensor<YScaleDataType> y_scale_host_ref({m});
+    ck_tile::HostTensor<YScaleDataType> y_scale_host_dev({m});
+
+    ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host({n});
+    ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host_dev({n});
+
+    ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
+    ck_tile::FillUniformDistribution<XResidualDataType>{-.5f, .5f}(x_residual_host);
+    ck_tile::FillUniformDistribution<SmoothScaleDataType>{-1.f, 1.f}(sm_scale_host);
+    ck_tile::FillUniformDistribution<XBiasDataType>{-.5f, .5f}(x_bias_host);
+    ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host);
+    ck_tile::FillUniformDistribution<BetaDataType>{-.5f, .5f}(beta_host);
+
+    ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem x_bias_buf(x_bias_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem gamma_buf(gamma_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem beta_buf(beta_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_buf(y_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_scale_buf(y_scale_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem sm_scale_buf(sm_scale_host_dev.get_element_space_size_in_bytes());
+
+    ck_tile::DeviceMem x_residual_buf(x_residual_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_residual_buf(y_residual_host.get_element_space_size_in_bytes());
+
+    x_buf.ToDevice(x_host.data());
+    x_bias_buf.ToDevice(x_bias_host.data());
+    gamma_buf.ToDevice(gamma_host.data());
+    beta_buf.ToDevice(beta_host.data());
+    x_residual_buf.ToDevice(x_residual_host.data());
+    sm_scale_buf.ToDevice(sm_scale_host.data());
+
+    auto prec_str = [&]() {
+        auto base_str = prec_i;
+        if(prec_i != prec_o)
+        {
+            base_str += "|" + prec_o;
+        }
+        if(fused_quant == 1)
+        {
+            base_str += std::string("(") + prec_sy + ")";
+        }
+        return base_str;
+    }();
+
+    std::cout << "[" << prec_str << "]"
+              << " m:" << m << ", n:" << n << ", x_stride:" << x_stride
+              << ", xr_stride:" << xr_stride << ", y_stride:" << y_stride
+              << ", yr_stride:" << yr_stride << std::flush;
+
+    layernorm2d_fwd_traits traits{
+        prec_i, prec_o, prec_sm, prec_sy, SaveMeanVar, xbias, fused_add, fused_quant};
+
+    layernorm2d_fwd_args args{x_buf.GetDeviceBuffer(),
+                              fused_add != 0 ? x_residual_buf.GetDeviceBuffer() : nullptr,
+                              fused_quant == 1 ? sm_scale_buf.GetDeviceBuffer() : nullptr,
+                              x_bias_buf.GetDeviceBuffer(),
+                              gamma_buf.GetDeviceBuffer(),
+                              beta_buf.GetDeviceBuffer(),
+
+                              y_buf.GetDeviceBuffer(),
+                              fused_add == 1 ? y_residual_buf.GetDeviceBuffer() : nullptr,
+                              fused_quant != 0 ? y_scale_buf.GetDeviceBuffer() : nullptr,
+                              nullptr, // p_mean, unsupported yet
+                              nullptr, // p_invStd, unsupported yet
+
+                              epsilon,
+                              m,
+                              n,
+                              x_stride,   // x row_stride
+                              xr_stride,  // x residule row stride
+                              y_stride,   // y row stride
+                              yr_stride}; // y residule row stride
+
+    float ave_time = layernorm2d_fwd(
+        traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
+
+    if(ave_time < 0)
+    {
+        std::cout << " not supported!" << std::endl << std::flush;
+        return false;
+    }
+
+    std::size_t num_byte = sizeof(XDataType) * m * n + sizeof(XBiasDataType) * n +
+                           sizeof(GammaDataType) * n + sizeof(BetaDataType) * n +
+                           sizeof(YDataType) * m * n;
+
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+    std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush;
+
+    bool pass = true;
+
+    if(do_validation)
+    {
+        // reference
+        if(xbias != 0)
+        {
+            // add bias before fadd
+            int M = x_host.mDesc.get_lengths()[0];
+            int N = x_host.mDesc.get_lengths()[1];
+            for(int idx_m = 0; idx_m < M; ++idx_m)
+            {
+                for(int idx_n = 0; idx_n < N; ++idx_n)
+                {
+                    x_host(idx_m, idx_n) = ck_tile::type_convert<XDataType>(
+                        ck_tile::type_convert<ComputeDataType>(x_host(idx_m, idx_n)) +
+                        ck_tile::type_convert<ComputeDataType>(x_bias_host(idx_n)));
+                }
+            }
+        }
+
+        if(fused_add != 0)
+        {
+            // fused pre_add/pre_add_store
+            // TODO we accumulate directly to x_host for simplcity here...
+
+            std::transform(x_host.mData.cbegin(),
+                           x_host.mData.cend(),
+                           x_residual_host.mData.cbegin(),
+                           x_host.mData.begin(),
+                           [](auto x_, auto r_) {
+                               auto o_ = ck_tile::type_convert<ComputeDataType>(x_) +
+                                         ck_tile::type_convert<ComputeDataType>(r_);
+                               return ck_tile::type_convert<XDataType>(o_);
+                           });
+        }
+        ck_tile::reference_layernorm2d_fwd<XDataType,
+                                           GammaDataType,
+                                           BetaDataType,
+                                           ComputeDataType,
+                                           YDataType,
+                                           MeanDataType,
+                                           InvStdDataType>(
+            x_host, gamma_host, beta_host, y_host_ref, mean_host_ref, invStd_host_ref, epsilon);
+
+        if(fused_quant != 0)
+        {
+            auto dquant_functor = [&](int m_, auto& o_, auto& acc_) {
+                int N_ = acc_.mDesc.get_lengths()[1];
+                if(fused_quant == 1)
+                {
+                    for(int n_ = 0; n_ < N_; n_++)
+                    {
+                        // input smooth outlier
+                        acc_(m_, n_) = acc_(m_, n_) *
+                                       ck_tile::type_convert<ComputeDataType>(sm_scale_host(n_));
+                    }
+                }
+                ComputeDataType absmax = static_cast<ComputeDataType>(0);
+                for(int n_ = 0; n_ < N_; n_++)
+                {
+                    const auto a = ck_tile::abs(acc_(m_, n_));
+                    absmax       = a > absmax ? a : absmax;
+                }
+                // printf("cpu:absmax:%f\n", absmax);
+                constexpr ComputeDataType kMaxY =
+                    std::is_same<YDataType, ck_tile::fp8_t>::value    ? 240.0
+                    : std::is_same<YDataType, ck_tile::int8_t>::value ? 127.0
+                                                                      : 0.0;
+                ComputeDataType y_scale = absmax / kMaxY;
+                y_scale_host_ref(m_)    = ck_tile::type_convert<YScaleDataType>(y_scale);
+                for(int n_ = 0; n_ < N_; n_++)
+                {
+                    o_(m_, n_) = ck_tile::type_convert<YDataType>(acc_(m_, n_) / y_scale);
+                }
+            };
+
+            ck_tile::reference_layernorm2d_fwd<XDataType,
+                                               GammaDataType,
+                                               BetaDataType,
+                                               ComputeDataType,
+                                               YDataType,
+                                               MeanDataType,
+                                               InvStdDataType>(x_host,
+                                                               gamma_host,
+                                                               beta_host,
+                                                               y_host_ref,
+                                                               mean_host_ref,
+                                                               invStd_host_ref,
+                                                               epsilon,
+                                                               dquant_functor);
+        }
+        else
+        {
+            ck_tile::reference_layernorm2d_fwd<XDataType,
+                                               GammaDataType,
+                                               BetaDataType,
+                                               ComputeDataType,
+                                               YDataType,
+                                               MeanDataType,
+                                               InvStdDataType>(
+                x_host, gamma_host, beta_host, y_host_ref, mean_host_ref, invStd_host_ref, epsilon);
+        }
+
+        y_buf.FromDevice(y_host_dev.data());
+
+        ck_tile::HostTensor<YResidualDataType> y_residual_host_dev({m, n}, {yr_stride, 1});
+        if(fused_add == 1)
+        {
+            y_residual_buf.FromDevice(y_residual_host_dev.data());
+        }
+
+        auto [rtol, atol] = get_elimit<OutDataType>();
+
+        if(x_stride == n)
+        {
+            pass = ck_tile::check_err(
+                y_host_dev, y_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol);
+            if(fused_add == 1)
+            {
+                pass &= ck_tile::check_err(y_residual_host_dev,
+                                           x_host,
+                                           std::string("ADD Error: Incorrect results!"),
+                                           rtol,
+                                           atol);
+            }
+        }
+        else
+        {
+            for(int i_r = 0; i_r < m; i_r++)
+            {
+                std::vector<YDataType> y_host_dev_row(y_host_dev.begin() + i_r * y_stride,
+                                                      y_host_dev.begin() + i_r * y_stride + n);
+                std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * y_stride,
+                                                      y_host_ref.begin() + i_r * y_stride + n);
+                pass &= ck_tile::check_err(y_host_dev_row,
+                                           y_host_ref_row,
+                                           std::string("OUT[") + std::to_string(i_r) +
+                                               std::string("] Error: Incorrect results!"),
+                                           rtol,
+                                           atol);
+                if(fused_add == 1)
+                {
+                    std::vector<YResidualDataType> y_residual_host_dev_row(
+                        y_residual_host_dev.begin() + i_r * yr_stride,
+                        y_residual_host_dev.begin() + i_r * yr_stride + n);
+                    std::vector<YResidualDataType> y_residual_host_ref_row(
+                        x_host.begin() + i_r * yr_stride, x_host.begin() + i_r * yr_stride + n);
+                    pass &= ck_tile::check_err(y_residual_host_dev_row,
+                                               y_residual_host_ref_row,
+                                               std::string("ADD[") + std::to_string(i_r) +
+                                                   std::string("] Error: Incorrect results!"),
+                                               rtol,
+                                               atol);
+                }
+            }
+        }
+        if(fused_quant == 1)
+        {
+            y_scale_buf.FromDevice(y_scale_host_dev.data());
+            pass &= ck_tile::check_err(y_scale_host_dev,
+                                       y_scale_host_ref,
+                                       std::string("SCALE Error: Incorrect results!"),
+                                       rtol,
+                                       atol);
+        }
+
+        std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    std::string prec_i  = arg_parser.get_str("prec_i");
+    std::string prec_o  = arg_parser.get_str("prec_o");
+    std::string prec_sm = arg_parser.get_str("prec_sm");
+    std::string prec_sy = arg_parser.get_str("prec_sy");
+
+    if(prec_o == "auto")
+    {
+        prec_o = prec_i;
+    }
+    if(prec_sm == "auto")
+    {
+        prec_sm = "fp32";
+    }
+    if(prec_sy == "auto")
+    {
+        prec_sy = "fp32";
+    }
+    int save_mv = arg_parser.get_int("save_mv");
+
+    // no dynamic quant case
+    if(prec_i == "fp16" && prec_o == "fp16" && prec_sm == "fp32" && prec_sy == "fp32" && save_mv)
+    {
+        return run<ck_tile::half_t, ck_tile::half_t, float, float, true>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "fp16" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_mv)
+    {
+        return run<ck_tile::half_t, ck_tile::half_t, float, float, false>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "bf16" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            save_mv)
+    {
+        return run<ck_tile::bf16_t, ck_tile::bf16_t, float, float, true>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "bf16" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_mv)
+    {
+        return run<ck_tile::bf16_t, ck_tile::bf16_t, float, float, true>(arg_parser) ? 0 : -2;
+    }
+
+    // dynamic quant case, only in inference
+    else if(prec_i == "fp16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_mv)
+    {
+        return run<ck_tile::half_t, ck_tile::int8_t, float, float, false>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_mv)
+    {
+        return run<ck_tile::bf16_t, ck_tile::int8_t, float, float, false>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_mv)
+    {
+        return run<ck_tile::half_t, ck_tile::fp8_t, float, float, false>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_mv)
+    {
+        return run<ck_tile::bf16_t, ck_tile::fp8_t, float, float, false>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/02_layernorm2d/layernorm2d_fwd.hpp

@@ -0,0 +1,70 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, 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/layernorm2d.hpp"
+#include <string>
+
+template <typename InType,
+          typename OutType,
+          typename SmoothSScaleDataType_,
+          typename YScaleDataType_>
+struct LayerNormTypeConfig;
+
+template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
+struct LayerNormTypeConfig<ck_tile::half_t, OutType, SmoothScaleDataType_, YScaleDataType_>
+{
+    using XDataType           = ck_tile::half_t;
+    using YDataType           = OutType;
+    using XBiasDataType       = ck_tile::half_t;
+    using GammaDataType       = ck_tile::half_t;
+    using BetaDataType        = ck_tile::half_t;
+    using MeanDataType        = ck_tile::half_t;
+    using InvStdDataType      = ck_tile::half_t;
+    using ComputeDataType     = float;
+    using SmoothScaleDataType = SmoothScaleDataType_;
+    using YScaleDataType      = YScaleDataType_;
+};
+
+template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
+struct LayerNormTypeConfig<ck_tile::bf16_t, OutType, SmoothScaleDataType_, YScaleDataType_>
+{
+    using XDataType           = ck_tile::bf16_t;
+    using YDataType           = OutType;
+    using XBiasDataType       = ck_tile::bf16_t;
+    using GammaDataType       = ck_tile::bf16_t;
+    using BetaDataType        = ck_tile::bf16_t;
+    using MeanDataType        = ck_tile::bf16_t;
+    using InvStdDataType      = ck_tile::bf16_t;
+    using ComputeDataType     = float;
+    using SmoothScaleDataType = SmoothScaleDataType_;
+    using YScaleDataType      = YScaleDataType_;
+};
+
+// runtime args
+struct layernorm2d_fwd_args : public ck_tile::Layernorm2dFwdHostArgs
+{
+};
+
+// This is the public API, will be generated by script
+struct layernorm2d_fwd_traits
+{
+    std::string prec_i; // input precision
+    std::string prec_o; // output precision
+
+    // if fused_quant == 1, need set prec_sm/prec_sy to proper string, otherwise can set
+    // arbitrary(will skip check) if fused_quant == 2, need set prec_sy to proper string, otherwise
+    // can set arbitrary(will skip check)
+    std::string prec_sm; // x-scale, used for [1*N] input smooth quant
+    std::string prec_sy; // y-scale, used for [M*1] output for next layer
+
+    bool save_mean_var; //
+    int xbias;          // 0:no-bias, 1:add bias
+    int fused_add;      // 0:no-add, 1:pre-add-store, 2:pre-add
+    int fused_quant;    // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
+};
+
+float layernorm2d_fwd(layernorm2d_fwd_traits, layernorm2d_fwd_args, const ck_tile::stream_config&);

example/ck_tile/02_layernorm2d/script/perf_test.sh

@@ -0,0 +1,37 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_layernorm2d_fwd -type f | head -n 1)"
+
+$EXE -m=1 -n=1 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec_i=bf16 -repeat=1000
+
+$EXE -m=700 -n=80 -e=1e-12 -v=1  -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec_i=fp16 -repeat=1000

example/ck_tile/02_layernorm2d/script/smoke_test.sh

@@ -0,0 +1,35 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_layernorm2d_fwd -type f | head -n 1)"
+
+for fquant in "" "-fquant=1 -prec_o=int8" "-fquant=1 -prec_o=fp8"; do
+for pr_i in "fp16" "bf16" ; do
+for fadd in "0" "1"; do
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=99  -n=13
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=17  -n=16
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=100
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=4   -n=128
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=80  -n=127
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=22  -n=255 -stride=256
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7   -n=599
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=19  -n=512
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=33  -n=313 -stride=1000
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=11  -n=510
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=171 -n=676 -stride=818
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=91  -n=636
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=12  -n=768 -stride=800
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=100 -n=766 -stride=812
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=31  -n=1024
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=64  -n=1000 -stride=1004
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=8   -n=1501
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=1826
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=5   -n=2040
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7   -n=2734
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=3182
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=9   -n=4096
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=8192
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=9120
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=10547
+#$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=17134
+done
+done
+done

example/ck_tile/03_gemm/CMakeLists.txt

@@ -0,0 +1,12 @@
+add_executable(tile_example_gemm_basic gemm_basic.cpp)
+rocm_install(TARGETS tile_example_gemm_basic COMPONENT examples)
+add_executable(tile_example_gemm_universal universal_gemm.cpp)
+rocm_install(TARGETS tile_example_gemm_universal COMPONENT examples)
+
+set(EXAMPLE_GEMM_COMPILE_OPTIONS)
+if(CK_USE_OCP_FP8)
+  list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -DCK_TILE_USE_OCP_FP8)
+endif()
+list(APPEND EXAMPLE_GEMM_COMPILE_OPTIONS -mllvm -enable-noalias-to-md-conversion=0)
+target_compile_options(tile_example_gemm_basic PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})
+target_compile_options(tile_example_gemm_universal PRIVATE ${EXAMPLE_GEMM_COMPILE_OPTIONS})

example/ck_tile/03_gemm/README.md

@@ -0,0 +1,37 @@
+# GEMM Matrix Multiplication
+
+This folder contains example for GEMM using ck_tile tile-programming implementation. Currently, it only supports the basic feature of the CK Tile GEMM, but creates the placeholders for the future support on different GEMM pipeline and different GEMM modules. In the near future, we will gradually migrate all the GEMM features from old CK to CK Tile.
+
+## 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>
+# The basic pipeline method on the gemm calculation
+make tile_example_gemm_basic -j
+# The memory bound pipeline on the gemm calculation
+make tile_example_gemm_universal -j
+```
+This will result in an executable `build/bin/tile_example_gemm_basic` & `build/bin/tile_example_gemm_universal`
+
+## example
+```
+args:
+          -b    batch size (default:1)
+          -m    m dimension (default:1024)
+          -n    n dimension (default:2048)
+          -k    k dimension (default:64)
+   -a_layout    Tensor A data layout (default: R)
+   -b_layout    Tensor B data layout (default: R)
+   -c_layout    Tensor C data layout (default: R)
+   -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

@@ -0,0 +1,200 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <hip/hip_runtime.h>
+
+#include <cstring>
+#include <iostream>
+#include <ostream>
+#include <string>
+#include <tuple>
+
+#include "ck_tile/host.hpp"
+#include "gemm_utils.hpp"
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
+{
+    // The kPadM, kPadN, kPadK & kBlockPerCu should also come from the Codegen part.
+    constexpr bool kPadM = false;
+    constexpr bool kPadN = false;
+    constexpr bool kPadK = false;
+
+    constexpr int kBlockPerCu = 1;
+
+    // This part comes from the Codegen
+    constexpr ck_tile::index_t M_Tile = 256;
+    constexpr ck_tile::index_t N_Tile = 256;
+    constexpr ck_tile::index_t K_Tile = 64;
+
+    constexpr ck_tile::index_t M_Warp = 2;
+    constexpr ck_tile::index_t N_Warp = 2;
+    constexpr ck_tile::index_t K_Warp = 1;
+
+    constexpr ck_tile::index_t M_Warp_Tile = 32;
+    constexpr ck_tile::index_t N_Warp_Tile = 32;
+    constexpr ck_tile::index_t K_Warp_Tile = 16;
+
+    using CodegenGemmShape =
+        ck_tile::TileGemmShape<ck_tile::sequence<M_Tile, N_Tile, K_Tile>,
+                               ck_tile::sequence<M_Warp, N_Warp, K_Warp>,
+                               ck_tile::sequence<M_Warp_Tile, N_Warp_Tile, K_Warp_Tile>>;
+
+    using TilePartitioner = ck_tile::GemmTile1DPartitioner<CodegenGemmShape>;
+
+    using CodegenGemmTraits =
+        ck_tile::TileGemmTraits<kPadM, kPadN, kPadK, ALayout, BLayout, CLayout>;
+    using CodegenPipelineProblem = ck_tile::
+        GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
+    using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
+    using GemmEpilogue        = ck_tile::CShuffleEpilogue<
+        ck_tile::CShuffleEpilogueProblem<ADataType,
+                                         BDataType,
+                                         AccDataType,
+                                         CDataType,
+                                         CLayout,
+                                         CodegenPipelineProblem::kBlockSize,
+                                         TilePartitioner::MPerBlock,
+                                         TilePartitioner::NPerBlock,
+                                         M_Warp,
+                                         N_Warp,
+                                         M_Warp_Tile,
+                                         N_Warp_Tile,
+                                         K_Warp_Tile,
+                                         CodegenPipelineProblem::TransposeC>>;
+    // 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, CodegenGemmPipeline, GemmEpilogue>;
+
+    auto kargs = Kernel::MakeKernelArgs(args);
+
+    const dim3 grids      = Kernel::GridSize(args.M, args.N, args.k_batch);
+    constexpr dim3 blocks = Kernel::BlockSize();
+
+    if(!Kernel::IsSupportedArgument(kargs))
+    {
+        throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
+    }
+
+    if(s.log_level_ > 0)
+    {
+        std::cout << "Launching kernel with args: " << Kernel::GetName() << '\n'
+                  << "shape: " << CodegenGemmShape::GetName() << '\n'
+                  << "problem: " << CodegenPipelineProblem::GetName() << '\n'
+                  << "pipeline: " << CodegenGemmPipeline::GetName() << '\n'
+                  << "grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
+                  << ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
+                  << std::endl;
+    }
+
+    float ave_time = ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+
+    return ave_time;
+}
+
+#include "run_gemm_example.inc"
+
+template <typename APrecType, typename BPrecType = APrecType, typename CPrecType = APrecType>
+int run_gemm_example_prec_type(std::string a_layout, std::string b_layout, int argc, char* argv[])
+{
+    using Row = ck_tile::tensor_layout::gemm::RowMajor;
+    using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
+
+    if constexpr(std::is_same_v<BPrecType, ck_tile::pk_int4_t>)
+    {
+        if(a_layout == "R" && b_layout == "C")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Row{}, Col{}, Row{});
+        }
+        else if(a_layout == "C" && b_layout == "C")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Col{}, Col{}, Row{});
+        }
+        else
+        {
+            throw std::runtime_error("Unsupported memory layout for the input matrices when "
+                                     "BPrecType is ck_tile::pk_int4_t!");
+        }
+    }
+    else
+    {
+        if(a_layout == "R" && b_layout == "R")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Row{}, Row{}, Row{});
+        }
+        else if(a_layout == "R" && b_layout == "C")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Row{}, Col{}, Row{});
+        }
+        else if(a_layout == "C" && b_layout == "R")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Col{}, Row{}, Row{});
+        }
+        else if(a_layout == "C" && b_layout == "C")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Col{}, Col{}, Row{});
+        }
+        else
+        {
+            throw std::runtime_error("Unsupported memory layout for the input matrices!");
+        }
+    }
+}
+
+int run_gemm_example(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    std::string data_type = arg_parser.get_str("prec");
+    std::string a_layout  = arg_parser.get_str("a_layout");
+    std::string b_layout  = arg_parser.get_str("b_layout");
+
+    if(data_type == "fp16")
+    {
+        return run_gemm_example_prec_type<ck_tile::half_t>(a_layout, b_layout, argc, argv);
+    }
+    else if(data_type == "bf16")
+    {
+        return run_gemm_example_prec_type<ck_tile::bf16_t>(a_layout, b_layout, argc, argv);
+    }
+    else if(data_type == "fp8")
+    {
+        return run_gemm_example_prec_type<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t>(
+            a_layout, b_layout, argc, argv);
+    }
+    else if(data_type == "bf8")
+    {
+        return run_gemm_example_prec_type<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t>(
+            a_layout, b_layout, argc, argv);
+    }
+
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
+    else if(data_type == "pk_int4_t")
+    {
+        // TODO: Add support for bhalf_t ADataType
+        return run_gemm_example_prec_type<ck_tile::half_t, ck_tile::pk_int4_t, ck_tile::half_t>(
+            a_layout, b_layout, argc, argv);
+    }
+#endif
+    else
+    {
+        throw std::runtime_error("Unsupported data type for this operation !!!");
+    }
+}
+
+int main(int argc, char* argv[]) { return !run_gemm_example(argc, argv); }

example/ck_tile/03_gemm/gemm_utils.hpp

@@ -0,0 +1,223 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <string>
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/kernel_launch.hpp"
+#include "ck_tile/ops/epilogue.hpp"
+#include "ck_tile/ops/gemm.hpp"
+
+#define CK_TILE_PIPELINE_COMPUTE_V3 1
+#define CK_TILE_PIPELINE_MEMORY 2
+#define CK_TILE_PIPELINE_COMPUTE_V4 3
+
+#ifndef CK_TILE_PIPELINE_DEFAULT
+#define CK_TILE_PIPELINE_DEFAULT CK_TILE_PIPELINE_COMPUTE_V3
+#endif
+
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
+#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrMem
+#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrMem
+#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Interwave
+#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
+#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrCompV3
+#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrCompV3
+#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Intrawave
+#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V4)
+#define GEMM_PIPELINE ck_tile::GemmPipelineAgBgCrCompV4
+#define UNIVERSAL_GEMM_PIPELINE ck_tile::BaseGemmPipelineAgBgCrCompV4
+#define GEMM_PIPELINE_SCHEDULER ck_tile::GemmPipelineScheduler::Intrawave
+#else
+#error "unsupported CK_TILE_PIPELINE_DEFAULT value"
+#endif
+
+struct GemmConfig
+{
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
+    // Memory friendly for Interwave scheduler
+    static constexpr ck_tile::index_t M_Tile = 128;
+    static constexpr ck_tile::index_t N_Tile = 32;
+    static constexpr ck_tile::index_t K_Tile = 64;
+
+    static constexpr ck_tile::index_t M_Warp = 4;
+    static constexpr ck_tile::index_t N_Warp = 1;
+    static constexpr ck_tile::index_t K_Warp = 1;
+
+    static constexpr ck_tile::index_t M_Warp_Tile = 32;
+    static constexpr ck_tile::index_t N_Warp_Tile = 32;
+    static constexpr ck_tile::index_t K_Warp_Tile = 8;
+
+    static constexpr bool DoubleSmemBuffer = false;
+#endif
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
+    // Compute friendly for Intrawave scheduler
+    static constexpr ck_tile::index_t M_Tile = 128;
+    static constexpr ck_tile::index_t N_Tile = 128;
+    static constexpr ck_tile::index_t K_Tile = 128;
+
+    static constexpr ck_tile::index_t M_Warp = 2;
+    static constexpr ck_tile::index_t N_Warp = 2;
+    static constexpr ck_tile::index_t K_Warp = 1;
+
+    static constexpr ck_tile::index_t M_Warp_Tile = 16;
+    static constexpr ck_tile::index_t N_Warp_Tile = 16;
+    static constexpr ck_tile::index_t K_Warp_Tile = 32;
+
+    static constexpr bool DoubleSmemBuffer = false;
+#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V4)
+    // Compute friendly for Intrawave scheduler
+    // Using the ping pong reader in the lds level
+    static constexpr ck_tile::index_t M_Tile = 256;
+    static constexpr ck_tile::index_t N_Tile = 256;
+    static constexpr ck_tile::index_t K_Tile = 32;
+
+    static constexpr ck_tile::index_t M_Warp = 2;
+    static constexpr ck_tile::index_t N_Warp = 2;
+    static constexpr ck_tile::index_t K_Warp = 1;
+
+    static constexpr ck_tile::index_t M_Warp_Tile = 32;
+    static constexpr ck_tile::index_t N_Warp_Tile = 32;
+    static constexpr ck_tile::index_t K_Warp_Tile = 16;
+
+    static constexpr bool DoubleSmemBuffer = true;
+#endif
+
+    static constexpr bool kPadM = false;
+    static constexpr bool kPadN = false;
+    static constexpr bool kPadK = false;
+
+    static constexpr bool PermuteA = false;
+    static constexpr bool PermuteB = false;
+
+    static constexpr bool TransposeC            = false;
+    static constexpr bool UseStructuredSparsity = false;
+
+    static constexpr int kBlockPerCu                         = 1;
+    static constexpr ck_tile::index_t TileParitionerGroupNum = 8;
+    static constexpr ck_tile::index_t TileParitionerM01      = 4;
+};
+
+template <typename ADataType, typename BDataType = ADataType, typename CDataType = ADataType>
+struct GemmTypeConfig;
+
+template <>
+struct GemmTypeConfig<ck_tile::half_t>
+{
+    using ADataType   = ck_tile::half_t;
+    using BDataType   = ck_tile::half_t;
+    using AccDataType = float;
+    using CDataType   = ck_tile::half_t;
+    // ToDo: Add more bias config to support different categories of GEMM.
+};
+
+template <>
+struct GemmTypeConfig<ck_tile::bf16_t, ck_tile::bf16_t, ck_tile::bf16_t>
+{
+    using ADataType   = ck_tile::bf16_t;
+    using BDataType   = ck_tile::bf16_t;
+    using AccDataType = float;
+    using CDataType   = ck_tile::bf16_t;
+};
+
+template <>
+struct GemmTypeConfig<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t>
+{
+    using ADataType   = ck_tile::fp8_t;
+    using BDataType   = ck_tile::fp8_t;
+    using AccDataType = float;
+    using CDataType   = ck_tile::half_t;
+};
+
+template <>
+struct GemmTypeConfig<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t>
+{
+    using ADataType   = ck_tile::bf8_t;
+    using BDataType   = ck_tile::bf8_t;
+    using AccDataType = float;
+    using CDataType   = ck_tile::half_t;
+};
+
+template <>
+struct GemmTypeConfig<ck_tile::half_t, ck_tile::pk_int4_t, ck_tile::half_t>
+{
+    using ADataType   = ck_tile::half_t;
+    using BDataType   = ck_tile::pk_int4_t;
+    using AccDataType = float;
+    using CDataType   = ck_tile::half_t;
+};
+
+template <typename T>
+struct DataTypeTraits;
+
+template <>
+struct DataTypeTraits<float>
+{
+    static constexpr const char* name = "fp32";
+};
+
+template <>
+struct DataTypeTraits<double>
+{
+    static constexpr const char* name = "fp64";
+};
+
+template <>
+struct DataTypeTraits<ck_tile::half_t>
+{
+    static constexpr const char* name = "fp16";
+};
+
+template <>
+struct DataTypeTraits<ck_tile::bf16_t>
+{
+    static constexpr const char* name = "bf16";
+};
+
+template <>
+struct DataTypeTraits<ck_tile::fp8_t>
+{
+    static constexpr const char* name = "fp8";
+};
+
+template <>
+struct DataTypeTraits<ck_tile::bf8_t>
+{
+    static constexpr const char* name = "bf8";
+};
+
+template <>
+struct DataTypeTraits<ck_tile::pk_int4_t>
+{
+    static constexpr const char* name = "pk_int4_t";
+};
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("m", "3840", "m dimension")
+        .insert("n", "4096", "n dimension")
+        .insert("k", "2048", "k dimension")
+        .insert("a_layout", "R", "A tensor data layout - Row by default")
+        .insert("b_layout", "C", "B tensor data layout - Column by default")
+        .insert("c_layout", "R", "C tensor data layout - Row by default")
+        .insert("stride_a", "0", "Tensor A stride")
+        .insert("stride_b", "0", "Tensor B stride")
+        .insert("stride_c", "0", "Tensor C stride")
+        .insert("v", "2", "0. No validation, 1. Validation on CPU, 2. Validation on GPU")
+        .insert("prec", "fp16", "data type. fp16/bf16/fp8/bf8")
+        .insert("warmup", "50", "number of iterations before benchmark the kernel")
+        .insert("repeat", "100", "number of iterations to benchmark the kernel")
+        .insert("timer", "gpu", "gpu:gpu timer, cpu:cpu timer")
+        .insert("split_k", "1", "splitK value")
+        .insert("init", "0", "0:random, 1:linear, 2:constant(1)");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+// host API
+float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s);

example/ck_tile/03_gemm/run_gemm_example.inc

@@ -0,0 +1,412 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+#pragma once
+
+template <typename Layout>
+static constexpr inline auto is_row_major(Layout layout_)
+{
+    return ck_tile::bool_constant<std::is_same_v<ck_tile::remove_cvref_t<decltype(layout_)>,
+                                                 ck_tile::tensor_layout::gemm::RowMajor>>{};
+}
+
+template <typename ADataType, typename BDataType, typename AccDataType, typename CDataType>
+auto calculate_rtol_atol(const ck_tile::index_t K,
+                         const ck_tile::index_t kbatch,
+                         const float max_accumulated_value)
+{
+    using ComputeType =
+        std::conditional_t<sizeof(ADataType) < sizeof(BDataType), ADataType, BDataType>;
+    // Calculate thresholds
+    const auto rtol = ck_tile::get_relative_threshold<ComputeType, CDataType, AccDataType>(
+        ck_tile::integer_divide_ceil(K, kbatch));
+    const auto atol = ck_tile::get_absolute_threshold<ComputeType, CDataType, AccDataType>(
+        max_accumulated_value / kbatch, ck_tile::integer_divide_ceil(K, kbatch));
+    // Calculate error due to split_k accumulation
+    const auto rtol_split_k =
+        ck_tile::get_relative_threshold<CDataType, CDataType, CDataType>(kbatch);
+    const auto atol_split_k = ck_tile::get_absolute_threshold<CDataType, CDataType, CDataType>(
+        max_accumulated_value, kbatch);
+    // Use higher threshold
+    return ck_tile::make_tuple(std::max(rtol, rtol_split_k), std::max(atol, atol_split_k));
+}
+
+template <typename Tensor,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+void permute_tensor_b(Tensor& tensor)
+{
+    using GemmShape = ck_tile::TileGemmShape<
+        ck_tile::sequence<GemmConfig::M_Tile, GemmConfig::N_Tile, GemmConfig::K_Tile>,
+        ck_tile::sequence<GemmConfig::M_Warp, GemmConfig::N_Warp, GemmConfig::K_Warp>,
+        ck_tile::
+            sequence<GemmConfig::M_Warp_Tile, GemmConfig::N_Warp_Tile, GemmConfig::K_Warp_Tile>,
+        GemmConfig::PermuteA,
+        GemmConfig::PermuteB>;
+
+    using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<GemmConfig::kPadM,
+                                                                 GemmConfig::kPadN,
+                                                                 GemmConfig::kPadK,
+                                                                 GemmConfig::DoubleSmemBuffer,
+                                                                 ALayout,
+                                                                 BLayout,
+                                                                 CLayout,
+                                                                 GemmConfig::TransposeC,
+                                                                 GemmConfig::UseStructuredSparsity>;
+
+    using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
+                                                                       BDataType,
+                                                                       AccDataType,
+                                                                       GemmShape,
+                                                                       GemmUniversalTraits,
+                                                                       GEMM_PIPELINE_SCHEDULER,
+                                                                       true,
+                                                                       ck_tile::TailNumber::Full>;
+
+    using GemmPipeline = GEMM_PIPELINE<UniversalGemmProblem>;
+
+    const ck_tile::index_t K  = tensor.get_length(0);
+    const ck_tile::index_t N  = tensor.get_length(1);
+    const ck_tile::index_t K1 = GemmPipeline::GetSmemPackB();
+    const ck_tile::index_t K0 = K / K1;
+
+    Tensor tensor_copy = tensor;
+
+    // int K0, N, K1
+    for(int j = 0; j < K0; j++)
+    {
+        for(int i = 0; i < N; i++)
+        {
+            for(int jj = 0; jj < K1; jj++)
+            {
+                tensor(j * N * K1 + i * K1 + jj) = tensor_copy(i * K + (j * K1 + jj));
+            }
+        }
+    }
+}
+
+template <typename Tensor>
+void permute_vectors_i4x4_b(Tensor& tensor)
+{
+    const ck_tile::index_t K = tensor.get_length(0);
+    const ck_tile::index_t N = tensor.get_length(1);
+    // vector pk_i4x4 permute
+    for(int i = 0; i < N; i++)
+    {
+        for(int j = 0; j < K; j += 8)
+        {
+            int8_t input[8];
+
+            for(int k = 0; k < 4; k++)
+            {
+                int8_t i4x2      = tensor(j + k * 2, i).data;
+                input[k * 2 + 0] = (i4x2 >> 4) & 0xf;
+                input[k * 2 + 1] = (i4x2 >> 0) & 0xf;
+            }
+
+            // permute 01234567->20643175
+            {
+                int8_t hi   = input[2];
+                int8_t lo   = input[0];
+                int8_t i4x2 = (hi << 4) | lo;
+
+                tensor(j + 0, i) = i4x2;
+            }
+
+            {
+                int8_t hi   = input[6];
+                int8_t lo   = input[4];
+                int8_t i4x2 = (hi << 4) | lo;
+
+                tensor(j + 2, i) = i4x2;
+            }
+
+            {
+                int8_t hi   = input[3];
+                int8_t lo   = input[1];
+                int8_t i4x2 = (hi << 4) | lo;
+
+                tensor(j + 4, i) = i4x2;
+            }
+
+            {
+                int8_t hi   = input[7];
+                int8_t lo   = input[5];
+                int8_t i4x2 = (hi << 4) | lo;
+
+                tensor(j + 6, i) = i4x2;
+            }
+        }
+    }
+}
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+float invoke_gemm(ck_tile::DeviceMem& a_m_k_dev_buf,
+                  ck_tile::DeviceMem& b_k_n_dev_buf,
+                  ck_tile::DeviceMem& c_m_n_dev_buf,
+                  ck_tile::index_t M,
+                  ck_tile::index_t N,
+                  ck_tile::index_t K,
+                  ck_tile::index_t stride_A,
+                  ck_tile::index_t stride_B,
+                  ck_tile::index_t stride_C,
+                  ck_tile::index_t kbatch,
+                  int n_warmup,
+                  int n_repeat)
+{
+    ck_tile::GemmHostArgs args;
+    args.a_ptr    = a_m_k_dev_buf.GetDeviceBuffer();
+    args.b_ptr    = b_k_n_dev_buf.GetDeviceBuffer();
+    args.c_ptr    = c_m_n_dev_buf.GetDeviceBuffer();
+    args.k_batch  = kbatch;
+    args.M        = M;
+    args.N        = N;
+    args.K        = K;
+    args.stride_A = stride_A;
+    args.stride_B = stride_B;
+    args.stride_C = stride_C;
+
+    float ave_time =
+        gemm_calc<ADataType, BDataType, AccDataType, CDataType, ALayout, BLayout, CLayout>(
+            args, ck_tile::stream_config{nullptr, true, 1, n_warmup, n_repeat});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_byte =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * N * K + sizeof(CDataType) * M * N;
+    float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+
+    std::cout << "Run Gemm kernel with M=" << M << " N=" << N << " K=" << K
+              << " StrideA=" << stride_A << " StrideB=" << stride_B << " StrideC=" << stride_C
+              << " A_Layout=" << ALayout::name << " B_Layout =" << BLayout::name
+              << " C_Layout=" << CLayout::name << " A_Type=" << DataTypeTraits<ADataType>::name
+              << " B_Type=" << DataTypeTraits<BDataType>::name
+              << " C_Type=" << DataTypeTraits<CDataType>::name
+              << " StructuredSparsity=" << (GemmConfig::UseStructuredSparsity ? "on" : "off")
+              << " : " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
+              << std::endl;
+
+    return ave_time;
+}
+
+template <typename ADataType,
+          typename BDataType = ADataType,
+          typename CDataType = ADataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+int run_gemm_example_with_layouts(int argc,
+                                  char* argv[],
+                                  const ALayout a_layout                  = ALayout{},
+                                  const BLayout b_layout                  = BLayout{},
+                                  [[maybe_unused]] const CLayout c_layout = CLayout{})
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    using AccDataType = typename GemmTypeConfig<ADataType, BDataType, CDataType>::AccDataType;
+
+    ck_tile::index_t M = arg_parser.get_int("m");
+    ck_tile::index_t N = arg_parser.get_int("n");
+    ck_tile::index_t K = arg_parser.get_int("k");
+
+    ck_tile::index_t stride_A = arg_parser.get_int("stride_a");
+    ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
+    ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
+
+    ck_tile::index_t kbatch      = arg_parser.get_int("split_k");
+    int n_warmup                 = arg_parser.get_int("warmup");
+    int n_repeat                 = arg_parser.get_int("repeat");
+    ck_tile::index_t init_method = arg_parser.get_int("init");
+
+    stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
+    stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
+    stride_C = ck_tile::get_default_stride(M, N, stride_C, is_row_major(CLayout{}));
+
+    ck_tile::HostTensor<ADataType> a_m_k(
+        ck_tile::host_tensor_descriptor(M, K, stride_A, is_row_major(a_layout)));
+    ck_tile::HostTensor<BDataType> b_k_n(
+        ck_tile::host_tensor_descriptor(K, N, stride_B, is_row_major(b_layout)));
+    ck_tile::HostTensor<CDataType> c_m_n_dev_result(
+        ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
+
+    if(init_method == 0)
+    {
+        ck_tile::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
+        ck_tile::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
+    }
+    else if(init_method == 1)
+    {
+        ck_tile::FillMonotonicSeq<ADataType>{}(a_m_k);
+        ck_tile::FillMonotonicSeq<BDataType>{}(b_k_n);
+    }
+    else if(init_method == 2)
+    {
+        ck_tile::FillUniformDistribution<ADataType>{1.f, 1.f}(a_m_k);
+        ck_tile::FillUniformDistribution<BDataType>{1.f, 1.f}(b_k_n);
+    }
+    else
+    {
+        a_m_k.SetZero();
+        b_k_n.SetZero();
+    }
+
+    if(GemmConfig::UseStructuredSparsity)
+    {
+        ck_tile::AdjustToStructuredSparsity<ADataType>{}(a_m_k);
+    }
+
+    ck_tile::DeviceMem a_m_k_dev_buf(a_m_k.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem b_k_n_dev_buf(b_k_n.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem c_m_n_dev_buf(c_m_n_dev_result.get_element_space_size_in_bytes());
+
+    static_assert(!GemmConfig::PermuteA, "Not implemented");
+    if constexpr(std::is_same_v<BDataType, ck_tile::pk_int4_t>)
+    {
+        // Permute vector pk_i4x4 data for device implementation
+        ck_tile::HostTensor<BDataType> b_k_n_dev = b_k_n;
+        if constexpr(GemmConfig::PermuteB)
+        {
+            permute_tensor_b<decltype(b_k_n_dev),
+                             ADataType,
+                             BDataType,
+                             AccDataType,
+                             CDataType,
+                             ALayout,
+                             BLayout,
+                             CLayout>(b_k_n_dev);
+        }
+        permute_vectors_i4x4_b(b_k_n_dev);
+        b_k_n_dev_buf.ToDevice(b_k_n_dev.data());
+    }
+    else
+    {
+        if constexpr(GemmConfig::PermuteB)
+        {
+            std::cout << "Permute for this DataType is not implemented." << std::endl;
+            return false;
+        }
+        b_k_n_dev_buf.ToDevice(b_k_n.data());
+    }
+
+    a_m_k_dev_buf.ToDevice(a_m_k.data());
+    c_m_n_dev_buf.SetZero();
+    c_m_n_dev_result.SetZero();
+
+    invoke_gemm<ADataType, BDataType, AccDataType, CDataType, ALayout, BLayout, CLayout>(
+        a_m_k_dev_buf,
+        b_k_n_dev_buf,
+        c_m_n_dev_buf,
+        M,
+        N,
+        K,
+        stride_A,
+        stride_B,
+        stride_C,
+        kbatch,
+        n_warmup,
+        n_repeat);
+
+    c_m_n_dev_buf.FromDevice(c_m_n_dev_result.data());
+    bool pass = true;
+
+    if(arg_parser.get_int("v") == 1)
+    {
+        ck_tile::HostTensor<CDataType> c_m_n_host_ref(
+            ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
+        c_m_n_host_ref.SetZero();
+
+        ck_tile::reference_gemm<ADataType, BDataType, AccDataType, CDataType>(
+            a_m_k, b_k_n, c_m_n_host_ref);
+        const float max_accumulated_value =
+            *std::max_element(c_m_n_host_ref.mData.begin(), c_m_n_host_ref.mData.end());
+        const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
+            K, kbatch, max_accumulated_value);
+        pass = ck_tile::check_err(c_m_n_dev_result,
+                                  c_m_n_host_ref,
+                                  "Error: Incorrect results!",
+                                  rtol_atol.at(ck_tile::number<0>{}),
+                                  rtol_atol.at(ck_tile::number<1>{}));
+
+        std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
+                  << " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
+                  << std::endl;
+        std::cout << "The CPU verification result is:" << (pass ? "correct" : "fail") << std::endl;
+    }
+    else if(arg_parser.get_int("v") == 2)
+    {
+        if constexpr(std::is_same_v<BDataType, ck_tile::pk_int4_t>)
+        {
+            // Restore input for B for gpu reference
+            b_k_n_dev_buf.ToDevice(b_k_n.data());
+        }
+        ck_tile::HostTensor<CDataType> c_m_n_gpu_ref(
+            ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
+        ck_tile::DeviceMem c_m_n_gpu_buf_ref(c_m_n_gpu_ref.get_element_space_size_in_bytes());
+        c_m_n_gpu_ref.SetZero();
+        c_m_n_gpu_buf_ref.SetZero();
+
+        ADataType* d_A;
+        BDataType* d_B;
+        CDataType* d_C;
+
+        ck_tile::hip_check_error(hipMalloc(&d_A, a_m_k.get_element_space_size_in_bytes()));
+        ck_tile::hip_check_error(hipMalloc(&d_B, b_k_n.get_element_space_size_in_bytes()));
+        ck_tile::hip_check_error(
+            hipMalloc(&d_C, c_m_n_dev_result.get_element_space_size_in_bytes()));
+
+        ck_tile::hip_check_error(hipMemcpy(d_A,
+                                           a_m_k_dev_buf.GetDeviceBuffer(),
+                                           a_m_k.get_element_space_size_in_bytes(),
+                                           hipMemcpyHostToDevice));
+        ck_tile::hip_check_error(hipMemcpy(d_B,
+                                           b_k_n_dev_buf.GetDeviceBuffer(),
+                                           b_k_n.get_element_space_size_in_bytes(),
+                                           hipMemcpyHostToDevice));
+
+        ck_tile::reference_gemm_gpu<ADataType,
+                                    BDataType,
+                                    AccDataType,
+                                    CDataType,
+                                    ALayout,
+                                    BLayout,
+                                    CLayout>(d_A, d_B, d_C, M, N, K, stride_A, stride_B, stride_C);
+
+        ck_tile::hip_check_error(hipMemcpy(c_m_n_gpu_buf_ref.GetDeviceBuffer(),
+                                           d_C,
+                                           c_m_n_dev_result.get_element_space_size_in_bytes(),
+                                           hipMemcpyDeviceToHost));
+
+        ck_tile::hip_check_error(hipFree(d_A));
+        ck_tile::hip_check_error(hipFree(d_B));
+        ck_tile::hip_check_error(hipFree(d_C));
+
+        c_m_n_gpu_buf_ref.FromDevice(c_m_n_gpu_ref.data());
+        const float max_accumulated_value =
+            *std::max_element(c_m_n_gpu_ref.mData.begin(), c_m_n_gpu_ref.mData.end());
+        const auto rtol_atol = calculate_rtol_atol<ADataType, BDataType, AccDataType, CDataType>(
+            K, kbatch, max_accumulated_value);
+        pass = ck_tile::check_err(c_m_n_dev_result,
+                                  c_m_n_gpu_ref,
+                                  "Error: Incorrect results!",
+                                  rtol_atol.at(ck_tile::number<0>{}),
+                                  rtol_atol.at(ck_tile::number<1>{}));
+        std::cout << "Relative error threshold: " << rtol_atol.at(ck_tile::number<0>{})
+                  << " Absolute error threshold: " << rtol_atol.at(ck_tile::number<1>{})
+                  << std::endl;
+        std::cout << "The GPU verification result is: " << (pass ? "correct" : "fail") << std::endl;
+    }
+
+    return pass;
+}

example/ck_tile/03_gemm/script/benchmark_basic_bf16.sh

@@ -0,0 +1,14 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_gemm_basic -type f | head -n 1)"
+VALID=1
+
+
+for b_matrix_layout in "C"; do
+    for m in "64" "512" "1024" "2048"; do
+        for n in "512" "1024" "2048"; do
+            for k in "64" "512" "1024" "2048"; do
+                $EXE -prec=bf16 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
+            done
+        done
+    done
+done

example/ck_tile/03_gemm/script/benchmark_basic_bf8.sh

@@ -0,0 +1,14 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_gemm_basic -type f | head -n 1)"
+VALID=1
+
+
+for b_matrix_layout in "C"; do
+    for m in "64" "512" "1024" "2048"; do
+        for n in "512" "1024" "2048"; do
+            for k in "64" "512" "1024" "2048"; do
+                $EXE -prec=bf8 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
+            done
+        done
+    done
+done

example/ck_tile/03_gemm/script/benchmark_basic_fp16.sh

@@ -0,0 +1,13 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_gemm_basic -type f | head -n 1)"
+VALID=1
+
+for b_matrix_layout in "C"; do
+    for m in "64" "512" "1024" "2048"; do
+        for n in "512" "1024" "2048"; do
+            for k in "64" "512" "1024" "2048"; do
+                $EXE -prec=fp16 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
+            done
+        done
+    done
+done

example/ck_tile/03_gemm/script/benchmark_basic_fp8.sh

@@ -0,0 +1,14 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_gemm_basic -type f | head -n 1)"
+VALID=1
+
+
+for b_matrix_layout in "C"; do
+    for m in "64" "512" "1024" "2048"; do
+        for n in "512" "1024" "2048"; do
+            for k in "64" "512" "1024" "2048"; do
+                $EXE -prec=fp8 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
+            done
+        done
+    done
+done

example/ck_tile/03_gemm/script/benchmark_mem_pipeline_bf16.sh

@@ -0,0 +1,13 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_gemm_universal -type f | head -n 1)"
+VALID=1
+
+for b_matrix_layout in "C"; do
+    for m in "512" "1024" "2048" "4096"; do
+        for n in "512" "1024" "2048"; do
+            for k in "512" "1024" "2048"; do
+                $EXE -prec=bf16 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
+            done
+        done
+    done
+done

example/ck_tile/03_gemm/script/benchmark_mem_pipeline_bf8.sh

@@ -0,0 +1,13 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_gemm_universal -type f | head -n 1)"
+VALID=1
+
+for b_matrix_layout in "C"; do
+    for m in "512" "1024" "2048" "4096"; do
+        for n in "512" "1024" "2048"; do
+            for k in "512" "1024" "2048"; do
+                $EXE -prec=bf8 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
+            done
+        done
+    done
+done

example/ck_tile/03_gemm/script/benchmark_mem_pipeline_fp16.sh

@@ -0,0 +1,13 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_gemm_universal -type f | head -n 1)"
+VALID=1
+
+for b_matrix_layout in "C"; do
+    for m in "512" "1024" "2048" "4096"; do
+        for n in "512" "1024" "2048"; do
+            for k in "512" "1024" "2048"; do
+                $EXE -prec=fp16 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
+            done
+        done
+    done
+done

example/ck_tile/03_gemm/script/benchmark_mem_pipeline_fp8.sh

@@ -0,0 +1,13 @@
+#!/bin/sh
+EXE="$(find . -name tile_example_gemm_universal -type f | head -n 1)"
+VALID=1
+
+for b_matrix_layout in "C"; do
+    for m in "512" "1024" "2048" "4096"; do
+        for n in "512" "1024" "2048"; do
+            for k in "512" "1024" "2048"; do
+                $EXE -prec=fp8 -m=$m -n=$n -k=$k -a_layout="R" -b_layout="$b_matrix_layout" -c_layout="R" -v=$VALID
+            done
+        done
+    done
+done

example/ck_tile/03_gemm/script/run_full_test.sh

@@ -0,0 +1,45 @@
+#!/bin/bash 
+#
+# in order to run this script you'd first need to build the tile_example_gemm executables in ../build/bin/
+#
+# run the script as "./run_full_test.sh <tag for your test environment> <branch name> <host name> <gpu_arch>
+# input arguments: 
+# environment tag  : a string describing the specifics of your test environment
+# branch name      : name of the branch in git repo (git status | grep -e 'On branch')
+# host name        : $hostname
+# gpu architecture: e.g., gfx90a, or gfx942, etc.
+
+# get the command line arguments:
+export env_type=$1
+echo 'Environment type: ' $env_type
+export branch=$2
+echo 'Branch name: ' $branch
+export host_name=$3
+echo 'Host name: ' $host_name
+export GPU_arch=$4
+echo 'GPU_arch: ' $GPU_arch
+
+function print_log_header(){
+    rm -f $1;
+    echo 'On branch ' $3 &> $1;
+    echo 'Node name: ' $4 >> $1;
+    # get GPU architecture and compute units from rocminfo
+    echo -n "GPU_arch: " >> $1; rocminfo | grep "Name:" | grep "gfx" >> $1;
+    rocminfo | grep "Compute Unit:" >> $1;
+    hipcc --version | grep -e 'HIP version'  >> $1;
+    echo 'Environment type: ' $2 >> $1;
+    /opt/rocm/bin/amdclang++ --version | grep -e 'InstalledDir' >> $1;
+}
+
+# run verification tests
+for dtype in fp16 bf16 fp8 bf8; do
+    example/ck_tile/03_gemm/script/benchmark_basic_$dtype.sh
+done
+example/ck_tile/03_gemm/script/smoke_test_mem_pipeline.sh
+
+# run performance benchmarks
+for dtype in fp16 bf16 fp8 bf8; do
+    export gemm_log="perf_tile_gemm_mem_pipeline_${dtype}_${GPU_arch}.log"
+    print_log_header $gemm_log $env_type $branch $host_name
+    example/ck_tile/03_gemm/script/benchmark_mem_pipeline_$dtype.sh 2>&1 | tee -a $gemm_log
+done

example/ck_tile/03_gemm/script/smoke_test_basic.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+EXE="$(find . -name tile_example_gemm_basic -type f | head -n 1)"
+KNAME=1
+
+export CK_WARMUP=0
+export CK_REPEAT=1
+
+COMMON_ARGS='-v=2 -warmup=0 -repeat=1'
+
+run_tests() {
+    for m in 128 1024; do
+        for n in 128 2048; do
+            for k in 64 128; do
+
+                $EXE -m=$m -n=$n -k=$k -stride_a=0 -stride_b=0 -stride_c=0 -prec=$1 $COMMON_ARGS
+                if [ $? -eq 0 ]; then
+                    echo "Success: Test with m=$m, n=$n, k=$k executed successfully."
+                else
+                    echo "Error: Test with m=$m, n=$n, k=$k failed to execute properly."
+                    # Optionally, exit or break if you need to halt further execution
+                    # exit 1
+                fi
+
+            done
+        done
+    done
+}
+
+set -x
+
+run_tests "fp16"
+run_tests "bf16"
+run_tests "fp8"
+run_tests "bf8"
+
+set +x

example/ck_tile/03_gemm/script/smoke_test_mem_pipeline.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+EXE="$(find . -name tile_example_gemm_universal -type f | head -n 1)"
+KNAME=1
+
+export CK_WARMUP=0
+export CK_REPEAT=1
+
+COMMON_ARGS='-v=2 -warmup=0 -repeat=1'
+
+run_tests() {
+    for m in 512 1024; do
+        for n in 512 2048; do
+            for k in 512 1024; do
+
+                $EXE -m=$m -n=$n -k=$k -stride_a=0 -stride_b=0 -stride_c=0 -prec=$1 $COMMON_ARGS
+                if [ $? -eq 0 ]; then
+                    echo "Success: Test with batch=$batch, m=$m, n=$n, k=$k executed successfully."
+                else
+                    echo "Error: Test with batch=$batch, m=$m, n=$n, k=$k failed to execute properly."
+                    # Optionally, exit or break if you need to halt further execution
+                    # exit 1
+                fi
+
+            done
+        done
+    done
+}
+
+set -x
+
+run_tests "fp16"
+run_tests "bf16"
+run_tests "fp8"
+run_tests "bf8"
+
+set +x

example/ck_tile/03_gemm/stript.sh

@@ -0,0 +1 @@
+for file in gemm_universal_*; do mv "$file" "${file/f16_f16_f16/fp16_fp16_fp16}"; done

example/ck_tile/03_gemm/universal_gemm.cpp

@@ -0,0 +1,360 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <hip/hip_runtime.h>
+
+#include <cstring>
+#include <iostream>
+#include <sstream>
+#include <string>
+#include <tuple>
+
+#include "ck_tile/host.hpp"
+#include "gemm_utils.hpp"
+
+template <typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CDataType,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout>
+float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config& s)
+{
+    using GemmShape = ck_tile::TileGemmShape<
+        ck_tile::sequence<GemmConfig::M_Tile, GemmConfig::N_Tile, GemmConfig::K_Tile>,
+        ck_tile::sequence<GemmConfig::M_Warp, GemmConfig::N_Warp, GemmConfig::K_Warp>,
+        ck_tile::
+            sequence<GemmConfig::M_Warp_Tile, GemmConfig::N_Warp_Tile, GemmConfig::K_Warp_Tile>,
+        GemmConfig::PermuteA,
+        GemmConfig::PermuteB>;
+    using TilePartitioner =
+        ck_tile::GemmSpatiallyLocalTilePartitioner<GemmShape,
+                                                   GemmConfig::TileParitionerGroupNum,
+                                                   GemmConfig::TileParitionerM01>;
+
+    using Traits              = ck_tile::TileGemmTraits<GemmConfig::kPadM,
+                                           GemmConfig::kPadN,
+                                           GemmConfig::kPadK,
+                                           ALayout,
+                                           BLayout,
+                                           CLayout>;
+    using GemmUniversalTraits = ck_tile::TileGemmUniversalTraits<GemmConfig::kPadM,
+                                                                 GemmConfig::kPadN,
+                                                                 GemmConfig::kPadK,
+                                                                 GemmConfig::DoubleSmemBuffer,
+                                                                 ALayout,
+                                                                 BLayout,
+                                                                 CLayout,
+                                                                 GemmConfig::TransposeC,
+                                                                 GemmConfig::UseStructuredSparsity>;
+    using GemmPipelineProblem =
+        ck_tile::GemmPipelineProblem<ADataType, BDataType, AccDataType, GemmShape, Traits>;
+
+    using BaseGemmPipeline = UNIVERSAL_GEMM_PIPELINE<GemmPipelineProblem>;
+
+    const ck_tile::index_t k_grain     = args.k_batch * GemmConfig::K_Tile;
+    const ck_tile::index_t K_split     = (args.K + k_grain - 1) / k_grain * GemmConfig::K_Tile;
+    const ck_tile::index_t num_loop    = TilePartitioner::GetLoopNum(K_split);
+    const bool has_hot_loop            = BaseGemmPipeline::BlockHasHotloop(num_loop);
+    const ck_tile::TailNumber tail_num = BaseGemmPipeline::GetBlockLoopTailNum(num_loop);
+
+    float ave_time{0};
+
+    const auto Run = [&](const auto has_hot_loop_, const auto tail_number_) {
+        constexpr bool has_hot_loop_v = has_hot_loop_.value;
+        constexpr auto tail_number_v  = tail_number_.value;
+        constexpr auto scheduler      = GEMM_PIPELINE_SCHEDULER;
+
+        using UniversalGemmProblem = ck_tile::UniversalGemmPipelineProblem<ADataType,
+                                                                           BDataType,
+                                                                           AccDataType,
+                                                                           GemmShape,
+                                                                           GemmUniversalTraits,
+                                                                           scheduler,
+                                                                           has_hot_loop_v,
+                                                                           tail_number_v>;
+
+        using GemmPipeline = GEMM_PIPELINE<UniversalGemmProblem>;
+        using GemmEpilogue = ck_tile::CShuffleEpilogue<
+            ck_tile::CShuffleEpilogueProblem<ADataType,
+                                             BDataType,
+                                             AccDataType,
+                                             CDataType,
+                                             CLayout,
+                                             GemmPipelineProblem::kBlockSize,
+                                             TilePartitioner::MPerBlock,
+                                             TilePartitioner::NPerBlock,
+                                             GemmConfig::M_Warp,
+                                             GemmConfig::N_Warp,
+                                             GemmConfig::M_Warp_Tile,
+                                             GemmConfig::N_Warp_Tile,
+                                             GemmConfig::K_Warp_Tile,
+                                             UniversalGemmProblem::TransposeC>>;
+        using Kernel = ck_tile::GemmKernel<TilePartitioner, GemmPipeline, GemmEpilogue>;
+        auto kargs   = Kernel::MakeKernelArgs(args);
+
+        const dim3 grids      = Kernel::GridSize(args.M, args.N, args.k_batch);
+        constexpr dim3 blocks = Kernel::BlockSize();
+
+        if(!Kernel::IsSupportedArgument(kargs))
+        {
+            throw std::runtime_error("Wrong! Arguments not supported! Skipping gemm!\n");
+        }
+
+        if(s.log_level_ > 0)
+        {
+            std::cout << "Launching kernel with args:"
+                      << " grid: {" << grids.x << ", " << grids.y << ", " << grids.z << "}"
+                      << ", blocks: {" << blocks.x << ", " << blocks.y << ", " << blocks.z << "}"
+                      << std::endl;
+        }
+
+        ave_time = ck_tile::launch_kernel(s,
+                                          ck_tile::make_kernel<blocks.x, GemmConfig::kBlockPerCu>(
+                                              Kernel{}, grids, blocks, 0, kargs));
+        return ave_time;
+    };
+
+    if(has_hot_loop)
+    {
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
+        if(tail_num == ck_tile::TailNumber::Full)
+        {
+            Run(ck_tile::bool_constant<true>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
+        }
+        else if(tail_num == ck_tile::TailNumber::Odd)
+        {
+            Run(ck_tile::bool_constant<true>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Odd>{});
+        }
+        else if(tail_num == ck_tile::TailNumber::Even)
+        {
+            Run(ck_tile::bool_constant<true>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Even>{});
+        }
+        else
+        {
+            std::ostringstream err;
+            err << "For compute pipeline tail number should always be Full, but have \"" << tail_num
+                << "\" which is not supported! PrefetchStages: " << BaseGemmPipeline::PrefetchStages
+                << "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+            throw std::runtime_error(err.str());
+        }
+#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_MEMORY)
+        // Tail pipeline One to Seven
+        if(tail_num == ck_tile::TailNumber::One)
+        {
+            Run(ck_tile::bool_constant<true>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::One>{});
+        }
+        else if(tail_num == ck_tile::TailNumber::Full)
+        {
+            Run(ck_tile::bool_constant<true>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
+        }
+
+        if constexpr(BaseGemmPipeline::PrefetchStages > 2)
+        {
+            if(tail_num == ck_tile::TailNumber::Two)
+            {
+                Run(ck_tile::bool_constant<true>{},
+                    ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Two>{});
+            }
+        }
+        if constexpr(BaseGemmPipeline::PrefetchStages > 3)
+        {
+            if(tail_num == ck_tile::TailNumber::Three)
+            {
+                Run(ck_tile::bool_constant<true>{},
+                    ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Three>{});
+            }
+        }
+        if constexpr(BaseGemmPipeline::PrefetchStages > 4)
+        {
+            if(tail_num == ck_tile::TailNumber::Four)
+            {
+                Run(ck_tile::bool_constant<true>{},
+                    ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Four>{});
+            }
+        }
+        if constexpr(BaseGemmPipeline::PrefetchStages > 5)
+        {
+            if(tail_num == ck_tile::TailNumber::Five)
+            {
+                Run(ck_tile::bool_constant<true>{},
+                    ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Five>{});
+            }
+        }
+        if constexpr(BaseGemmPipeline::PrefetchStages > 6)
+        {
+            if(tail_num == ck_tile::TailNumber::Six)
+            {
+                Run(ck_tile::bool_constant<true>{},
+                    ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Six>{});
+            }
+        }
+        if constexpr(BaseGemmPipeline::PrefetchStages > 7)
+        {
+            if(tail_num == ck_tile::TailNumber::Seven)
+            {
+                Run(ck_tile::bool_constant<true>{},
+                    ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Seven>{});
+            }
+        }
+#elif(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V4)
+        if(tail_num == ck_tile::TailNumber::Three)
+        {
+            Run(ck_tile::bool_constant<true>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Three>{});
+        }
+        else
+        {
+            Run(ck_tile::bool_constant<true>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Two>{});
+        }
+#endif
+    }
+    else
+    {
+        if(tail_num == ck_tile::TailNumber::Full)
+        {
+            Run(ck_tile::bool_constant<false>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
+        }
+        else if(tail_num == ck_tile::TailNumber::Odd)
+        {
+            Run(ck_tile::bool_constant<false>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Odd>{});
+        }
+        else if(tail_num == ck_tile::TailNumber::Even)
+        {
+            Run(ck_tile::bool_constant<false>{},
+                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Odd>{});
+        }
+        else
+        {
+            std::ostringstream err;
+            err << "Num K loop must be larger than number of prefetech stages."
+                << "\n PrefetchStages: " << BaseGemmPipeline::PrefetchStages
+                << "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
+            throw std::runtime_error(err.str());
+        }
+    }
+
+    return ave_time;
+}
+
+#include "run_gemm_example.inc"
+
+template <typename APrecType, typename BPrecType = APrecType, typename CPrecType = APrecType>
+int run_gemm_example_prec_type(std::string a_layout, std::string b_layout, int argc, char* argv[])
+{
+    using Row = ck_tile::tensor_layout::gemm::RowMajor;
+    using Col = ck_tile::tensor_layout::gemm::ColumnMajor;
+
+    if constexpr(std::is_same_v<BPrecType, ck_tile::pk_int4_t>)
+    {
+        if(a_layout == "R" && b_layout == "C")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Row{}, Col{}, Row{});
+        }
+        else if(a_layout == "C" && b_layout == "C")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Col{}, Col{}, Row{});
+        }
+        else
+        {
+            throw std::runtime_error("Unsupported memory layout for the input matrices when "
+                                     "BPrecType is ck_tile::pk_int4_t!");
+        }
+    }
+    else
+    {
+        if(a_layout == "R" && b_layout == "R")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Row{}, Row{}, Row{});
+        }
+        else if(a_layout == "R" && b_layout == "C")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Row{}, Col{}, Row{});
+        }
+        else if(a_layout == "C" && b_layout == "R")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Col{}, Row{}, Row{});
+        }
+        else if(a_layout == "C" && b_layout == "C")
+        {
+            return run_gemm_example_with_layouts<APrecType, BPrecType, CPrecType>(
+                argc, argv, Col{}, Col{}, Row{});
+        }
+        else
+        {
+            throw std::runtime_error("Unsupported memory layout for the input matrices!");
+        }
+    }
+}
+
+int run_gemm_example(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    std::string data_type = arg_parser.get_str("prec");
+    std::string a_layout  = arg_parser.get_str("a_layout");
+    std::string b_layout  = arg_parser.get_str("b_layout");
+
+    if(data_type == "fp16")
+    {
+        return run_gemm_example_prec_type<ck_tile::half_t>(a_layout, b_layout, argc, argv);
+    }
+    else if(data_type == "bf16")
+    {
+        return run_gemm_example_prec_type<ck_tile::bf16_t>(a_layout, b_layout, argc, argv);
+    }
+    else if(data_type == "fp8")
+    {
+        return run_gemm_example_prec_type<ck_tile::fp8_t, ck_tile::fp8_t, ck_tile::half_t>(
+            a_layout, b_layout, argc, argv);
+    }
+    else if(data_type == "bf8")
+    {
+        return run_gemm_example_prec_type<ck_tile::bf8_t, ck_tile::bf8_t, ck_tile::half_t>(
+            a_layout, b_layout, argc, argv);
+    }
+
+#if(CK_TILE_PIPELINE_DEFAULT == CK_TILE_PIPELINE_COMPUTE_V3)
+    else if(data_type == "pk_int4_t")
+    {
+        // TODO: Add support for bhalf_t ADataType
+        return run_gemm_example_prec_type<ck_tile::half_t, ck_tile::pk_int4_t, ck_tile::half_t>(
+            a_layout, b_layout, argc, argv);
+    }
+#endif
+    else
+    {
+        throw std::runtime_error("Unsupported data type for this operation !!!");
+    }
+}
+
+int main(int argc, char* argv[])
+{
+    try
+    {
+        run_gemm_example(argc, argv);
+    }
+    catch(const std::runtime_error& e)
+    {
+        std::cerr << "Caught runtime error: " << e.what() << '\n';
+        // Return a non-zero code to indicate failure
+        return EXIT_FAILURE;
+    }
+    return EXIT_SUCCESS;
+}

example/ck_tile/04_img2col/CMakeLists.txt

@@ -0,0 +1,4 @@
+# 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 image_to_column.cpp)
+rocm_install(TARGETS tile_example_img2col COMPONENT examples)

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/05_reduce/CMakeLists.txt

@@ -0,0 +1,21 @@
+set(EXAMPLE_REDUCE "tile_example_reduce")
+# 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"
+message("adding example ${EXAMPLE_REDUCE}")
+
+add_executable(${EXAMPLE_REDUCE} reduce.cpp)
+rocm_install(TARGETS ${EXAMPLE_REDUCE} COMPONENT examples)
+
+target_include_directories(${EXAMPLE_REDUCE} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+set(EXAMPLE_REDUCE_COMPILE_OPTIONS)
+
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+list(APPEND EXAMPLE_REDUCE_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
+
+target_compile_options(${EXAMPLE_REDUCE} PRIVATE ${EXAMPLE_REDUCE_COMPILE_OPTIONS})
+
+# TODO: we have to turn off this global prop, otherwise the progress bar generated
+# by cmake will print too many files, execvp: /bin/sh: Argument list too long
+# however, this property may affect global
+# TODO: consider codegen a makefile by us
+set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)

example/ck_tile/05_reduce/reduce.cpp

@@ -0,0 +1,115 @@
+#include "ck_tile/host.hpp"
+#include "reduce.hpp"
+#include <cstring>
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("m", "3328", "m dimension")
+        .insert("n", "4096", "n dimension")
+        .insert("v", "1", "cpu validation or not")
+        .insert("prec", "fp16", "precision")
+        .insert("warmup", "5", "cold iter")
+        .insert("repeat", "20", "hot iter");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename DataType>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    using XDataType       = DataType;
+    using ComputeDataType = float;
+    using YDataType       = DataType;
+
+    ck_tile::index_t m = arg_parser.get_int("m");
+    ck_tile::index_t n = arg_parser.get_int("n");
+    int do_validation  = arg_parser.get_int("v");
+    int warmup         = arg_parser.get_int("warmup");
+    int repeat         = arg_parser.get_int("repeat");
+
+    ck_tile::HostTensor<XDataType> x_host({m, n});
+    ck_tile::HostTensor<YDataType> y_host_ref({m});
+    ck_tile::HostTensor<YDataType> y_host_dev({m});
+
+    ck_tile::FillUniformDistribution<XDataType>{-5.f, 5.f}(x_host);
+
+    ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_buf(y_host_dev.get_element_space_size_in_bytes());
+
+    x_buf.ToDevice(x_host.data());
+
+    using ReduceOp   = ck_tile::ReduceOp::Add;
+    using BlockWarps = ck_tile::sequence<4, 1>;
+    using BlockTile  = ck_tile::sequence<128, 128>;
+    using WarpTile   = ck_tile::sequence<32, 128>;
+    using Vector     = ck_tile::sequence<8, 8>;
+
+    // cross warp-reduce
+    // using BlockWarps = ck_tile::sequence<2, 2>;
+    // using BlockTile  = ck_tile::sequence<2, 1024>;
+    // using WarpTile   = ck_tile::sequence<1, 512>;
+    // using Vector = ck_tile::sequence<1, 8>;
+
+    constexpr ck_tile::index_t kBlockSize  = 256;
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+    ck_tile::index_t kGridSize             = (m / BlockTile::at(ck_tile::number<0>{}));
+    std::cout << "grid size " << kGridSize << std::endl;
+
+    using Shape = ck_tile::Reduce2dShape<BlockWarps, BlockTile, WarpTile, Vector>;
+    using Porblem =
+        ck_tile::Reduce2dProblem<XDataType, ComputeDataType, YDataType, Shape, ReduceOp>;
+
+    using Kernel = ck_tile::Reduce<Porblem>;
+
+    float ave_time = launch_kernel(ck_tile::stream_config{nullptr, true, 0, warmup, repeat},
+                                   ck_tile::make_kernel<kBlockSize, kBlockPerCu>(
+                                       Kernel{},
+                                       kGridSize,
+                                       kBlockSize,
+                                       0,
+                                       static_cast<XDataType*>(x_buf.GetDeviceBuffer()),
+                                       static_cast<YDataType*>(y_buf.GetDeviceBuffer()),
+                                       m,
+                                       n));
+
+    std::size_t num_btype = sizeof(XDataType) * m * n + sizeof(YDataType) * m;
+
+    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(do_validation)
+    {
+        // reference
+        ck_tile::reference_reduce<XDataType, ComputeDataType, YDataType>(
+            x_host, y_host_ref, ReduceOp{});
+        y_buf.FromDevice(y_host_dev.mData.data());
+        pass = ck_tile::check_err(y_host_dev, y_host_ref);
+
+        std::cout << "valid:" << (pass ? "y" : "n") << std::flush << std::endl;
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    const std::string data_type = arg_parser.get_str("prec");
+
+    if(data_type == "fp16")
+    {
+        return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
+    }
+    // else if(data_type == "bf16")
+    // {
+    //     return run<ck_tile::bf16_t>(arg_parser) ? 0 : -2;
+    // }
+}

example/ck_tile/05_reduce/reduce.hpp

@@ -0,0 +1,164 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/common.hpp"
+#include "ck_tile/ops/reduce/block/block_reduce.hpp"
+#include "ck_tile/ops/reduce/block/block_reduce2d_default_policy.hpp"
+
+namespace ck_tile {
+
+template <typename BlockWarps, // num warps along seq<M, N>
+          typename BlockTile,  // block size, seq<M, N>
+          typename WarpTile,   // warp size, seq<M, N>
+          typename Vector>     // contiguous pixels(vector size) along seq<M, N>
+struct Reduce2dShape
+{
+    static constexpr index_t Block_M = BlockTile::at(number<0>{});
+    static constexpr index_t Block_N = BlockTile::at(number<1>{});
+
+    static constexpr index_t Warp_M = WarpTile::at(number<0>{});
+    static constexpr index_t Warp_N = WarpTile::at(number<1>{});
+
+    static constexpr index_t Vector_M = Vector::at(number<0>{});
+    static constexpr index_t Vector_N = Vector::at(number<1>{});
+
+    static constexpr index_t WarpPerBlock_M = BlockWarps::at(number<0>{});
+    static constexpr index_t WarpPerBlock_N = BlockWarps::at(number<1>{});
+
+    static constexpr index_t ThreadPerWarp_M = Warp_M / Vector_M;
+    static constexpr index_t ThreadPerWarp_N = Warp_N / Vector_N;
+
+    static constexpr index_t Repeat_M = Block_M / (WarpPerBlock_M * Warp_M);
+    static constexpr index_t Repeat_N = Block_N / (WarpPerBlock_N * Warp_N);
+
+    static constexpr index_t BlockSize =
+        warpSize * reduce_on_sequence(BlockWarps{}, multiplies{}, number<1>{});
+};
+
+template <typename XDataType_,
+          typename ComputeDataType_,
+          typename YDataType_,
+          typename BlockShape_,
+          typename ReduceOp_>
+struct Reduce2dProblem
+{
+    using XDataType       = remove_cvref_t<XDataType_>;
+    using ComputeDataType = remove_cvref_t<ComputeDataType_>;
+    using YDataType       = remove_cvref_t<YDataType_>;
+    using BlockShape      = remove_cvref_t<BlockShape_>;
+    using ReduceOp        = ReduceOp_;
+
+    static constexpr bool kNeedCrossLaneSync = BlockShape::ThreadPerWarp_N > 1;
+    static constexpr bool kNeedCrossWarpSync = BlockShape::WarpPerBlock_N > 1;
+};
+
+template <typename Problem_, typename Policy_ = BlockReduce2dDefaultPolicy>
+struct Reduce
+{
+    using Problem = ck_tile::remove_cvref_t<Problem_>;
+    using Policy  = ck_tile::remove_cvref_t<Policy_>;
+
+    using XDataType       = ck_tile::remove_cvref_t<typename Problem::XDataType>;
+    using ComputeDataType = ck_tile::remove_cvref_t<typename Problem::ComputeDataType>;
+    using YDataType       = ck_tile::remove_cvref_t<typename Problem::YDataType>;
+
+#if 0
+    CK_TILE_DEVICE void operator()(const XDataType* p_x, YDataType* p_y, index_t M, index_t N)
+    const
+    {
+        using S = typename Problem::BlockShape;
+
+        const auto x_m_n = make_naive_tensor_view<address_space_enum::global>(
+            p_x, make_tuple(M, N), make_tuple(N, 1), number<S::Vector_N>{}, number<1>{});
+
+        const auto y_m = make_naive_tensor_view_packed<address_space_enum::global>(
+            p_y, make_tuple(M), number<1>{});
+
+        const auto iM = get_block_id() * S::Block_M;
+
+        auto x_window = make_tile_window(x_m_n,
+                                         make_tuple(number<S::Block_M>{}, number<S::Block_N>{}),
+                                         {iM, 0},
+                                         Policy::template MakeXBlockTileDistribution<Problem>());
+
+        auto y_window = make_tile_window(y_m, make_tuple(number<S::Block_M>{}), {iM});
+
+        const auto f_reduce = [](const auto& v0, const auto& v1) { return v0 + v1; };
+
+        const XDataType reduce_init_value = 0;
+
+        constexpr auto reduce_dims = sequence<1>{};
+
+        auto y_compute = decltype(block_tile_reduce<ComputeDataType>(
+            load_tile(x_window), reduce_dims, f_reduce, reduce_init_value)){};
+
+        set_tile(y_compute, reduce_init_value);
+
+        index_t num_n_tile_iteration =
+            __builtin_amdgcn_readfirstlane(integer_divide_ceil(N, S::Block_N));
+
+        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
+        {
+            const auto x = load_tile(x_window);
+            block_tile_reduce(y_compute, x, reduce_dims, f_reduce);
+            move_tile_window(x_window, {0, S::Block_N});
+        }
+
+        block_tile_reduce_sync(y_compute, f_reduce);
+
+        store_tile(y_window, cast_tile<YDataType>(y_compute));
+    }
+#else
+    CK_TILE_DEVICE void operator()(const XDataType* p_x, YDataType* p_y, index_t M, index_t N) const
+    {
+        using S = typename Problem::BlockShape;
+
+        const auto x_m_n = make_naive_tensor_view<address_space_enum::global>(
+            p_x, make_tuple(M, N), make_tuple(N, 1), number<S::Vector_N>{}, number<1>{});
+
+        const auto y_m = make_naive_tensor_view_packed<address_space_enum::global>(
+            p_y, make_tuple(M), number<1>{});
+
+        const auto iM = get_block_id() * S::Block_M;
+
+        auto x_window = make_tile_window(x_m_n,
+                                         make_tuple(number<S::Block_M>{}, number<S::Block_N>{}),
+                                         {iM, 0},
+                                         Policy::template MakeXBlockTileDistribution<Problem>());
+
+        auto y_window = make_tile_window(y_m, make_tuple(number<S::Block_M>{}), {iM});
+
+        __shared__ char smem[Policy::template GetSmemSize<Problem>()];
+
+        index_t num_n_tile_iteration =
+            __builtin_amdgcn_readfirstlane(integer_divide_ceil(N, S::Block_N));
+
+        auto reduce_func         = typename Problem::ReduceOp{};
+        auto block_reduce2d      = Policy::template GetBlockReduce2d<Problem>();
+        auto block_reduce2d_sync = Policy::template GetBlockReduce2dSync<Problem>();
+        auto block_reduce2d_cross_warp_sync =
+            Policy::template GetBlockReduce2dCrossWarpSync<Problem>();
+
+        using XTensorType = decltype(load_tile(x_window));
+        auto y_compute    = block_reduce2d.template MakeYBlockTile<XTensorType>();
+        set_tile(y_compute, reduce_func.template GetIdentityValue<ComputeDataType>());
+
+        for(int iN = __builtin_amdgcn_readfirstlane(0); iN < num_n_tile_iteration; ++iN)
+        {
+            const auto x = load_tile(x_window);
+            block_reduce2d(x, y_compute, reduce_func);
+            move_tile_window(x_window, {0, S::Block_N});
+        }
+
+        block_reduce2d_sync(y_compute, reduce_func);
+        block_reduce2d_cross_warp_sync(y_compute, smem, reduce_func);
+
+        store_tile(y_window, cast_tile<YDataType>(y_compute));
+    }
+#endif
+};
+
+} // namespace ck_tile

example/ck_tile/06_permute/CMakeLists.txt

@@ -0,0 +1,14 @@
+# 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_permute permute.cpp)
+rocm_install(TARGETS tile_example_permute COMPONENT examples)
+
+if(NOT DEFINED PERMUTE_USE_ALTERNATIVE_IMPL)
+# set(PERMUTE_USE_ALTERNATIVE_IMPL false)
+set(PERMUTE_USE_ALTERNATIVE_IMPL true)
+endif()
+if(PERMUTE_USE_ALTERNATIVE_IMPL)
+target_compile_options(tile_example_permute PRIVATE -DPERMUTE_USE_ALTERNATIVE_IMPL)
+target_sources(tile_example_permute PRIVATE alternative_impl/matrix_core_swizzle.cpp)
+endif()
+# target_compile_options(tile_example_permute PRIVATE -v --save-temps -Wno-gnu-line-marker)

example/ck_tile/06_permute/README.md

@@ -0,0 +1,46 @@
+# permute
+
+This folder contains example for permute kernel, which is similiar to [torch.permute](https://pytorch.org/docs/stable/generated/torch.permute.html) (combined with [torch.contiguous](https://pytorch.org/docs/stable/generated/torch.Tensor.contiguous.html)). Currently we implement a generic permute kernel that support up to rank 8 arbitrary permutation with a single kernel instance. Performance is not the first consideration, we prefer a simple and general kernel implementation using `ck_tile` in this example.
+
+
+```
+args:
+          -v    weather do CPU validation or not (default:1)
+       -prec    data type. fp16/bf16/fp32 (default:fp16)
+      -shape    the shape of the input tensor (default:2,3,4)
+       -perm    permute perm (default:2,1,0)
+```
+
+## build
+```
+# 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...
+make tile_example_permute -j
+```
+This will result in an executable `build/bin/tile_example_permute`
+
+
+## some examples
+```
+# torch
+x=torch.randn(2,3,4,6)
+y=x.permute(0,3,2,1).contiguous()
+
+# ck_tile
+./build/bin/tile_example_permute -shape=2,3,4,6 -perm=0,3,2,1
+```
+
+or you can try the smoke_test
+```
+# in the root of ck_tile, after you build this example
+sh example/ck_tile/06_permute/script/smoke_test.sh
+```
+
+### alternative implementation
+we have an alternative implementation under `alternative_impl/` folder, that can swizzle the tensor to be more friendly for data loading for matrix core layout. This can be enabled when dealing with a `rank-7` tensor, with a fixed pattern of either `0,1,4,2,5,3,6` or `0,1,2,4,5,3,6`. There are other shape limitation of this implementation, check the source code of `permute.cpp` for detail.
+```
+# example
+./build/bin/tile_example_permute -shape=3,6,4,32,16,2,8 -perm=0,1,4,2,5,3,6 # b_n0_k0_n1_k1_n2_k2
+./build/bin/tile_example_permute -shape=3,8,4,16,16,4,8 -perm=0,1,2,4,5,3,6 # b_n0_n1_k0_k1_n2_k2
+```

example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle.cpp

@@ -0,0 +1,98 @@
+#include "matrix_core_swizzle.hpp"
+#include "matrix_core_swizzle_kernel.hpp"
+
+float matrix_core_swizzle(matrix_core_swizzle_traits t,
+                          matrix_core_swizzle_args a,
+                          const ck_tile::stream_config& s)
+{
+    if(t.data_type.compare("fp16") == 0)
+    {
+        if(t.inst.compare("32x32x8") == 0)
+        {
+            constexpr int BLOCK_SIZE             = 256;
+            constexpr int NPerBlock              = 256;
+            constexpr int KPerBlock              = 128;
+            constexpr matrix_core_inst_enum Inst = matrix_core_inst_enum::MFMA_32x32x8_F16;
+            if(t.permute.compare("0,1,4,2,5,3,6") == 0)
+            {
+                constexpr matrix_core_permute_style pstyle =
+                    matrix_core_permute_style::permute_b_n0_k0_n1_k1_n2_k2;
+                using Kernel =
+                    matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;
+
+                auto k         = Kernel(a);
+                float ave_time = ck_tile::launch_kernel(s, k);
+
+                return ave_time;
+            }
+            else if(t.permute.compare("0,1,2,4,5,3,6") == 0)
+            {
+                constexpr matrix_core_permute_style pstyle =
+                    matrix_core_permute_style::permute_b_n0_n1_k0_k1_n2_k2;
+                using Kernel =
+                    matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;
+
+                auto k         = Kernel(a);
+                float ave_time = ck_tile::launch_kernel(s, k);
+
+                return ave_time;
+            }
+            else if(t.permute.compare("0,1,3,4,2,5") == 0)
+            {
+                constexpr matrix_core_permute_style pstyle =
+                    matrix_core_permute_style::b_nr_kr_kw_nw_kv;
+                using Kernel =
+                    matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;
+
+                auto k         = Kernel(a);
+                float ave_time = ck_tile::launch_kernel(s, k);
+
+                return ave_time;
+            }
+        }
+        else if(t.inst.compare("16x16x16") == 0)
+        {
+            constexpr int BLOCK_SIZE             = 256;
+            constexpr int NPerBlock              = 256;
+            constexpr int KPerBlock              = 128;
+            constexpr matrix_core_inst_enum Inst = matrix_core_inst_enum::MFMA_16x16x16_F16;
+            if(t.permute.compare("0,1,4,2,5,3,6") == 0)
+            {
+                constexpr matrix_core_permute_style pstyle =
+                    matrix_core_permute_style::permute_b_n0_k0_n1_k1_n2_k2;
+                using Kernel =
+                    matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;
+
+                auto k         = Kernel(a);
+                float ave_time = ck_tile::launch_kernel(s, k);
+
+                return ave_time;
+            }
+            else if(t.permute.compare("0,1,2,4,5,3,6") == 0)
+            {
+                constexpr matrix_core_permute_style pstyle =
+                    matrix_core_permute_style::permute_b_n0_n1_k0_k1_n2_k2;
+                using Kernel =
+                    matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;
+
+                auto k         = Kernel(a);
+                float ave_time = ck_tile::launch_kernel(s, k);
+
+                return ave_time;
+            }
+            else if(t.permute.compare("0,1,3,4,2,5") == 0)
+            {
+                constexpr matrix_core_permute_style pstyle =
+                    matrix_core_permute_style::b_nr_kr_kw_nw_kv;
+                using Kernel =
+                    matrix_core_swizzle_kernel<BLOCK_SIZE, NPerBlock, KPerBlock, pstyle, Inst>;
+
+                auto k         = Kernel(a);
+                float ave_time = ck_tile::launch_kernel(s, k);
+
+                return ave_time;
+            }
+        }
+    }
+    return -1;
+}

example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle.hpp

@@ -0,0 +1,20 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+#include "matrix_core_swizzle_kernel.hpp"
+#include <string>
+
+struct matrix_core_swizzle_traits
+{
+    std::string data_type; // fp16 only
+    std::string inst;      // 32x32x8, 16x16x16
+    std::string permute;   //
+};
+
+using matrix_core_swizzle_args = matrix_core_swizzle_host_args;
+
+// host API
+float matrix_core_swizzle(matrix_core_swizzle_traits,
+                          matrix_core_swizzle_args,
+                          const ck_tile::stream_config&);

example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle_kernel.hpp

@@ -0,0 +1,413 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/host.hpp"
+#include "ck_tile/ops/gemm.hpp"
+
+// if set to 1, slightly more instructions generated to calculate address
+#ifndef MERGE_2D_013425
+#define MERGE_2D_013425 0
+#endif
+
+enum class matrix_core_inst_enum
+{
+    MFMA_32x32x8_F16  = 0,
+    MFMA_16x16x16_F16 = 1,
+};
+
+namespace detail {
+template <matrix_core_inst_enum>
+struct to_warp_gemm;
+
+template <>
+struct to_warp_gemm<matrix_core_inst_enum::MFMA_32x32x8_F16>
+{
+    using type = ck_tile::WarpGemmMfmaF16F16F32M32N32K8;
+};
+
+template <>
+struct to_warp_gemm<matrix_core_inst_enum::MFMA_16x16x16_F16>
+{
+    using type = ck_tile::WarpGemmMfmaF16F16F32M16N16K16;
+};
+} // namespace detail
+template <matrix_core_inst_enum Inst>
+using to_warp_gemm_t = typename detail::to_warp_gemm<Inst>::type;
+
+// TODO: in below permute pattern, the last 3 dim is within wave
+enum class matrix_core_permute_style
+{
+    permute_b_n0_k0_n1_k1_n2_k2 = 0, // 0,1,4,2,5,3,6
+    permute_b_n0_n1_k0_k1_n2_k2 = 1, // 0,1,2,4,5,3,6
+    b_nr_kr_kw_nw_kv            = 2, // 0,1,3,4,2,5
+    b_nr_kr_waveflatten         = b_nr_kr_kw_nw_kv,
+};
+
+// assume this is B matrix, originally we have batch*n*k
+// now batch* n0*n1*n2*k0*k1*k2 -> batch* n0*k0*n1*k1*n2*k2
+// assume using 32x32x8-f16, 4 waves and extend the KPerLane to 8xfp16(dwordx4)
+//
+//                                      4(waves)  32(mfma_m lane)
+//                                          |      |
+// batch* n0*n1*n2*k0*k1*k2 -> batch* n0*k0*n1*k1*n2*k2 -> 8(thread loading)
+//                                    nr  kr    |
+//        nr  4  32 kr 2  8                     2(klane)
+//
+// permute: 0,1,4,2,5,3,6
+// or
+// batch* n0*n1*n2*k0*k1*k2 -> batch* n0*n1*k0*k1*n2*k2 -> 8(thread loading)
+// permute: 0,1,2,4,5,3,6
+//
+// this kernel only deal with fp16/bf16 data(16bit), and use 2d block size to do the swizzling
+// for simplicity, only consider n/k is multiple of block-size
+
+// independend host arg with no template
+struct matrix_core_swizzle_host_args
+{
+    const void* p_src;
+    void* p_dst;
+    int32_t batch;
+    int32_t n;
+    int32_t k;
+};
+
+// NOTE: this kernel could follow the style of generic permute kernel
+// but here we pass in fixed layout as template arg and generate different kernel instance
+// purposely
+template <int BLOCK_SIZE_ = 256,
+          int NPerBlock_  = 256,
+          int KPerBlock_  = 128,
+          matrix_core_permute_style pstyle_ =
+              matrix_core_permute_style::permute_b_n0_k0_n1_k1_n2_k2,
+          matrix_core_inst_enum Inst_ = matrix_core_inst_enum::MFMA_32x32x8_F16>
+struct matrix_core_swizzle_kernel
+{
+    using karg = matrix_core_swizzle_host_args;
+    using harg = matrix_core_swizzle_host_args;
+
+    static constexpr int BLOCK_SIZE      = BLOCK_SIZE_;
+    static constexpr int WavesPerBlock_N = 4;
+    static constexpr int WavesPerBlock_K = 1;
+    static_assert(WavesPerBlock_N * WavesPerBlock_K * 64 == BLOCK_SIZE);
+    static constexpr int NPerBlock                    = NPerBlock_;
+    static constexpr int KPerBlock                    = KPerBlock_;
+    static constexpr matrix_core_permute_style pstyle = pstyle_;
+    static constexpr matrix_core_inst_enum Inst       = Inst_;
+
+    static constexpr ck_tile::index_t Alignment = 8;
+    karg a;
+    dim3 grids;
+
+    using WarpGemm = to_warp_gemm_t<Inst>;
+
+    __host__ matrix_core_swizzle_kernel(harg h)
+    {
+        a                   = h;
+        ck_tile::index_t ns = (h.n + NPerBlock - 1) / NPerBlock;
+        ck_tile::index_t ks = (h.k + KPerBlock - 1) / KPerBlock;
+        grids               = dim3(ks, ns, h.batch);
+    }
+
+    __host__ bool is_applicable(harg h) { return h.n % NPerBlock == 0 && h.k % KPerBlock == 0; }
+
+    __host__ void operator()(const ck_tile::stream_config& s) const
+    {
+        ck_tile::kentry<BLOCK_SIZE, 1, kernel><<<grids, BLOCK_SIZE, 0, s.stream_id_>>>(a);
+    }
+
+    struct kernel
+    {
+        __device__ static constexpr auto get_src_dist()
+        {
+            using namespace ck_tile;
+            constexpr index_t K2 = Alignment;
+            constexpr index_t N2 = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
+            constexpr index_t K1 = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
+            constexpr index_t N1 = BLOCK_SIZE / get_warp_size();
+
+            static_assert(NPerBlock % (N1 * N2) == 0);
+            static_assert(KPerBlock % (K1 * K2) == 0);
+
+            constexpr index_t K0 = KPerBlock / (K1 * K2);
+            constexpr index_t N0 = NPerBlock / (N1 * N2);
+
+            // clang-format off
+            return make_static_tile_distribution(
+                tile_distribution_encoding<
+                    sequence<1>,// 0
+                    //             1              2            3             4             5             6
+                    tuple<sequence<N0>, sequence<N1>, sequence<N2>, sequence<K0>, sequence<K1>, sequence<K2>>,
+
+                    //            N1           K1  N2
+                    tuple<sequence<2>, sequence<5, 3>>,
+                    tuple<sequence<0>, sequence<0, 0>>,
+
+                    //       N0 K0 K2
+                    sequence<1, 4, 6>,
+                    sequence<0, 0, 0>>{});
+            // clang-format on
+        }
+        __device__ static constexpr auto get_dst_dist()
+        {
+            using namespace ck_tile;
+            constexpr index_t K2 = Alignment;
+            constexpr index_t N2 = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
+            constexpr index_t K1 = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
+            constexpr index_t N1 = BLOCK_SIZE / get_warp_size();
+
+            static_assert(NPerBlock % (N1 * N2) == 0);
+            static_assert(KPerBlock % (K1 * K2) == 0);
+
+            constexpr index_t K0 = KPerBlock / (K1 * K2);
+            constexpr index_t N0 = NPerBlock / (N1 * N2);
+
+            if constexpr(pstyle == matrix_core_permute_style::permute_b_n0_k0_n1_k1_n2_k2)
+            {
+                // clang-format off
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<
+                        sequence<1>,// 0
+                        //             1              2            3             4             5             6
+                        tuple<sequence<N0>, sequence<K0>, sequence<N1>, sequence<K1>, sequence<N2>, sequence<K2>>,
+
+                        //            N1           K1  N2
+                        tuple<sequence<3>, sequence<4, 5>>,
+                        tuple<sequence<0>, sequence<0, 0>>,
+
+                        //       N0 K0 K2
+                        sequence<1, 2, 6>,
+                        sequence<0, 0, 0>>{});
+                // clang-format on
+            }
+            else if constexpr(pstyle == matrix_core_permute_style::permute_b_n0_n1_k0_k1_n2_k2)
+            {
+                // clang-format off
+                return make_static_tile_distribution(
+                    tile_distribution_encoding<
+                        sequence<1>,// 0
+                        //             1              2            3             4             5             6
+                        tuple<sequence<N0>, sequence<N1>, sequence<K0>, sequence<K1>, sequence<N2>, sequence<K2>>,
+
+                        //            N1           K1  N2
+                        tuple<sequence<2>, sequence<4, 5>>,
+                        tuple<sequence<0>, sequence<0, 0>>,
+
+                        //       N0 K0 K2
+                        sequence<1, 3, 6>,
+                        sequence<0, 0, 0>>{});
+                // clang-format on
+            }
+            else
+            {
+                // clang-format off
+                // b_nr_kr_kw_nw_kv or b_nr_kr_waveflatten
+                constexpr index_t Kv = Alignment;
+                constexpr index_t Nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
+                constexpr index_t Kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
+
+                static_assert(KPerBlock % (K1 * K2) == 0);
+                constexpr index_t Nr = NPerBlock / Nw;
+                constexpr index_t Kr = KPerBlock / (Kv * Kw);
+
+                constexpr index_t Nr_p = WavesPerBlock_N;
+                constexpr index_t Kr_p = WavesPerBlock_K;
+                constexpr index_t Nr_y = Nr / Nr_p;
+                constexpr index_t Kr_y = Kr / Kr_p;
+
+                return make_static_tile_distribution(
+#if MERGE_2D_013425
+                    tile_distribution_encoding<
+                        sequence<1>,// 0    R
+                        // major       1                         2
+                        // minor       0     1     2             0     1     2   3
+                        tuple<sequence<Nr_y, Nr_p, Nw>, sequence<Kr_y, Kr_p, Kw, Kv>>,    // H
+
+                        //            Nr_p, Kr_p         Kw Nw
+                        tuple<sequence<1  , 2>, sequence<2, 1>>,    // p major
+                        tuple<sequence<1  , 1>, sequence<2, 2>>,    // p minor
+
+                        //       Nr_y Kr_y Kv
+                        sequence<1,   2,   2>,          // Y major
+                        sequence<0,   0,   3>>{});      // y minor
+#else
+                    tile_distribution_encoding<
+                        sequence<1>,// 0    R
+                        // major       1                     2                     3
+                        // minor       0     1               0     1               0   1   2
+                        tuple<sequence<Nr_y, Nr_p>, sequence<Kr_y, Kr_p>, sequence<Kw, Nw, Kv>>,    // H
+
+                        //            Nr_p, Kr_p         Kw Nw
+                        tuple<sequence<1  , 2>, sequence<3, 3>>,    // p major
+                        tuple<sequence<1  , 1>, sequence<0, 1>>,    // p minor
+
+                        //       Nr_y Kr_y Kv
+                        sequence<1,   2,   3>,          // Y major
+                        sequence<0,   0,   2>>{});      // y minor
+#endif
+                // clang-format on
+            }
+        }
+
+        __device__ void operator()(karg a_)
+        {
+            using namespace ck_tile;
+            index_t i_k = blockIdx.x;
+            index_t i_n = blockIdx.y;
+            index_t i_b = blockIdx.z;
+
+            constexpr index_t k2 = Alignment;
+            constexpr index_t n2 = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
+            constexpr index_t k1 = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
+            constexpr index_t n1 = BLOCK_SIZE / get_warp_size();
+            const index_t k0     = a_.k / (k1 * k2);
+            const index_t n0     = a_.n / (n1 * n2);
+
+            constexpr index_t k2_tile = Alignment;
+            constexpr index_t n2_tile = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
+            constexpr index_t k1_tile = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
+            constexpr index_t n1_tile = BLOCK_SIZE / get_warp_size();
+            constexpr index_t k0_tile = KPerBlock / (k1_tile * k2_tile);
+            constexpr index_t n0_tile = NPerBlock / (n1_tile * n2_tile);
+
+            const fp16_t* p_src = reinterpret_cast<const fp16_t*>(a_.p_src) + i_b * a_.k * a_.n;
+            fp16_t* p_dst       = reinterpret_cast<fp16_t*>(a_.p_dst) + i_b * a_.k * a_.n;
+
+            const auto src_view = [&]() {
+                const auto tmp = make_naive_tensor_view_packed<address_space_enum::global>(
+                    p_src,
+                    make_tuple(n0, n1, n2, k0, k1, k2),
+                    number<Alignment>{}); // control vector load
+                return tmp;
+            }();
+
+            const auto src_window = make_tile_window(src_view,
+                                                     make_tuple(number<n0_tile>{},
+                                                                number<n1_tile>{},
+                                                                number<n2_tile>{},
+                                                                number<k0_tile>{},
+                                                                number<k1_tile>{},
+                                                                number<k2_tile>{}),
+                                                     {i_n * n0_tile, 0, 0, i_k * k0_tile, 0, 0},
+                                                     get_src_dist());
+
+            auto dst_view = [&]() {
+                if constexpr(pstyle == matrix_core_permute_style::permute_b_n0_k0_n1_k1_n2_k2)
+                {
+                    auto tmp = make_naive_tensor_view_packed<address_space_enum::global>(
+                        p_dst,
+                        make_tuple(n0, k0, n1, k1, n2, k2),
+                        number<Alignment>{}); // control vector load
+                    return tmp;
+                }
+                else if constexpr(pstyle == matrix_core_permute_style::permute_b_n0_n1_k0_k1_n2_k2)
+                {
+                    auto tmp = make_naive_tensor_view_packed<address_space_enum::global>(
+                        p_dst,
+                        make_tuple(n0, n1, k0, k1, n2, k2),
+                        number<Alignment>{}); // control vector load
+                    return tmp;
+                }
+                else
+                {
+#if MERGE_2D_013425
+                    constexpr index_t kv = Alignment;
+                    constexpr index_t nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
+                    constexpr index_t kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
+                    // constexpr index_t waveflatten = kw*nw*kv;
+                    const index_t kr = a_.k / (k1 * k2);
+                    const index_t nr = a_.n / nw;
+                    auto tmp         = make_naive_tensor_view_packed<address_space_enum::global>(
+                        p_dst,
+                        make_tuple(nr, kr, number<kw>{}, number<nw>{}, number<kv>{}),
+                        number<Alignment>{}); // control vector load
+                    auto tmp_1 = transform_tensor_view(
+                        tmp,
+                        make_tuple(
+                            make_merge_transform(make_tuple(nr, number<nw>{})),
+                            make_merge_transform(make_tuple(kr, number<kw>{}, number<kv>{}))),
+                        make_tuple(sequence<0, 3>{}, sequence<1, 2, 4>{}),
+                        make_tuple(sequence<0>{}, sequence<1>{}));
+                    return tmp_1;
+#else
+                    // b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv,
+                    constexpr index_t kv = Alignment;
+                    constexpr index_t nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
+                    constexpr index_t kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
+                    constexpr index_t waveflatten = kw * nw * kv;
+                    const index_t kr = a_.k / (k1 * k2);
+                    const index_t nr = a_.n / nw;
+                    auto tmp = make_naive_tensor_view_packed<address_space_enum::global>(
+                        p_dst,
+                        make_tuple(nr, kr, waveflatten),
+                        number<Alignment>{}); // control vector load
+                    return tmp;
+#endif
+                }
+            }();
+
+            auto dst_window = [&]() {
+                if constexpr(pstyle == matrix_core_permute_style::permute_b_n0_k0_n1_k1_n2_k2)
+                {
+                    return make_tile_window(dst_view,
+                                            make_tuple(number<n0_tile>{},
+                                                       number<k0_tile>{},
+                                                       number<n1_tile>{},
+                                                       number<k1_tile>{},
+                                                       number<n2_tile>{},
+                                                       number<k2_tile>{}),
+                                            {i_n * n0_tile, i_k * k0_tile, 0, 0, 0, 0},
+                                            get_dst_dist());
+                }
+                else if constexpr(pstyle == matrix_core_permute_style::permute_b_n0_n1_k0_k1_n2_k2)
+                {
+                    return make_tile_window(dst_view,
+                                            make_tuple(number<n0_tile>{},
+                                                       number<n1_tile>{},
+                                                       number<k0_tile>{},
+                                                       number<k1_tile>{},
+                                                       number<n2_tile>{},
+                                                       number<k2_tile>{}),
+                                            {i_n * n0_tile, 0, i_k * k0_tile, 0, 0, 0},
+                                            get_dst_dist());
+                }
+                else
+                {
+#if MERGE_2D_013425
+                    // b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv
+                    return make_tile_window(dst_view,
+                                            make_tuple(number<NPerBlock>{}, number<KPerBlock>{}),
+                                            {i_n * NPerBlock, i_k * KPerBlock},
+                                            get_dst_dist());
+#else
+                    // b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv
+                    constexpr index_t kv = Alignment;
+                    constexpr index_t nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane;
+                    constexpr index_t kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane;
+                    constexpr index_t waveflatten_tile = kw * nw * kv;
+                    constexpr index_t nr_tile = NPerBlock / nw;
+                    constexpr index_t kr_tile = KPerBlock / (kw * kv);
+                    return make_tile_window(dst_view,
+                                            make_tuple(number<nr_tile>{},
+                                                       number<kr_tile>{},
+                                                       number<waveflatten_tile>{}),
+                                            {i_n * nr_tile, i_k * kr_tile, 0},
+                                            get_dst_dist());
+#endif
+                }
+            }();
+
+            // actual load store
+            auto src_tile = load_tile(src_window);
+
+            // now we only swap the distribution from src to dst, no extra movement occurs
+            auto dst_tile                = make_static_distributed_tensor<fp16_t>(get_dst_dist());
+            dst_tile.get_thread_buffer() = src_tile.get_thread_buffer();
+
+            // final store
+            store_tile(dst_window, dst_tile);
+        }
+    };
+};

example/ck_tile/06_permute/permute.cpp

@@ -0,0 +1,411 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "permute.hpp"
+#include "ck_tile/host.hpp"
+
+#include <array>
+#include <cstring>
+#include <functional>
+#include <numeric>
+#include <ostream>
+#include <string>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+#ifdef PERMUTE_USE_ALTERNATIVE_IMPL
+#include "alternative_impl/matrix_core_swizzle.hpp"
+#endif
+
+namespace detail {
+template <int bytes>
+struct to_integer_type;
+
+template <>
+struct to_integer_type<4>
+{
+    using type = int32_t;
+};
+template <>
+struct to_integer_type<2>
+{
+    using type = int16_t;
+};
+template <>
+struct to_integer_type<1>
+{
+    using type = int8_t;
+};
+} // namespace detail
+
+template <int bytes>
+using to_integer_type = typename detail::to_integer_type<bytes>::type;
+
+// host API (shoule come from codegen)
+float permute(permute_traits t, permute_args a, const ck_tile::stream_config& s)
+{
+    if(t.data_type.compare("fp8") == 0)
+    {
+        using DataType        = ck_tile::fp8_t;
+        using PipelineProblem = ck_tile::GenericPermuteProblem<DataType>;
+        using Kernel          = ck_tile::GenericPermute<PipelineProblem>;
+
+        auto kargs = Kernel::MakeKargs(a);
+
+        const dim3 grids      = Kernel::GridSize(a);
+        constexpr dim3 blocks = Kernel::BlockSize();
+
+        float ave_time = ck_tile::launch_kernel(
+            s, ck_tile::make_kernel<blocks.x, 1>(Kernel{}, grids, blocks, 0, kargs));
+
+        return ave_time;
+    }
+    else if(t.data_type.compare("fp16") == 0)
+    {
+        using DataType        = ck_tile::half_t;
+        using PipelineProblem = ck_tile::GenericPermuteProblem<DataType>;
+        using Kernel          = ck_tile::GenericPermute<PipelineProblem>;
+
+        auto kargs = Kernel::MakeKargs(a);
+
+        const dim3 grids      = Kernel::GridSize(a);
+        constexpr dim3 blocks = Kernel::BlockSize();
+
+        float ave_time = ck_tile::launch_kernel(
+            s, ck_tile::make_kernel<blocks.x, 1>(Kernel{}, grids, blocks, 0, kargs));
+
+        return ave_time;
+    }
+    else if(t.data_type.compare("fp32") == 0)
+    {
+        using DataType        = float;
+        using PipelineProblem = ck_tile::GenericPermuteProblem<DataType>;
+        using Kernel          = ck_tile::GenericPermute<PipelineProblem>;
+
+        auto kargs = Kernel::MakeKargs(a);
+
+        const dim3 grids      = Kernel::GridSize(a);
+        constexpr dim3 blocks = Kernel::BlockSize();
+
+        float ave_time = ck_tile::launch_kernel(
+            s, ck_tile::make_kernel<blocks.x, 1>(Kernel{}, grids, blocks, 0, kargs));
+
+        return ave_time;
+    }
+
+    return 0;
+}
+
+template <typename T>
+std::ostream& operator<<(std::ostream& os, const std::vector<T>& v)
+{
+    using size_type = typename std::vector<T>::size_type;
+
+    os << "[";
+    for(size_type idx = 0; idx < v.size(); ++idx)
+    {
+        if(0 < idx)
+        {
+            os << ", ";
+        }
+        os << v[idx];
+    }
+    return os << "]";
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("v", "1", "weather do CPU validation or not")
+        .insert("prec", "fp16", "data type. fp8/fp16/fp32 (representing 8/16/32 bit data)")
+        .insert("shape", "2,3,4", "the shape of the input tensor")
+        .insert("perm", "2,1,0", "permute perm")
+        .insert("kname", "0", "t to 1 will print kernel name")
+        .insert("seed",
+                "11939",
+                "random seed used for initializing input tensors. 0 for "
+                "non-deterministic seed")
+        .insert("warmup", "5", "number of iterations before benchmark the kernel")
+        .insert("repeat", "20", "number of iterations to benchmark the kernel");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+// different threshold for different dtype
+template <typename DataType>
+auto get_elimit(std::string /*init_method*/)
+{
+    double rtol = 1e-3;
+    double atol = 1e-3;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::bf16_t>(std::string /*init_method*/)
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::fp8_t>(std::string init_method)
+{
+    if(init_method == "ui" || init_method == "ni")
+    {
+        unsigned max_rounding_point_distance = 0;
+        double atol                          = 2e-3;
+        return ck_tile::make_tuple(max_rounding_point_distance, atol);
+    }
+    else
+    {
+        unsigned max_rounding_point_distance = 1;
+        double atol                          = 0.0625;
+        return ck_tile::make_tuple(max_rounding_point_distance, atol);
+    }
+}
+
+// "1,2,3,4" -> vector{1,2,3,4}
+std::vector<ck_tile::index_t> decode_vec(std::string q_val)
+{
+#define _S2I_(str_) static_cast<ck_tile::index_t>(std::atoi((str_).c_str()))
+    std::string::size_type pos = 0;
+    std::vector<ck_tile::index_t> v;
+    while(true)
+    {
+        auto found = q_val.find(',', pos);
+        ck_tile::index_t n =
+            _S2I_(q_val.substr(pos, found == std::string::npos ? found : found - pos));
+        v.push_back(n);
+        if(found == std::string::npos)
+        {
+            break;
+        }
+        pos = found + 1;
+    }
+    return v;
+#undef _S2I_
+}
+
+template <typename DataType>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    std::string data_type = arg_parser.get_str("prec");
+    int do_validation     = arg_parser.get_int("v");
+
+    auto shape        = decode_vec(arg_parser.get_str("shape"));
+    auto perm         = decode_vec(arg_parser.get_str("perm"));
+    int stream_warmup = arg_parser.get_int("warmup");
+    int stream_repeat = arg_parser.get_int("repeat");
+    bool kname        = arg_parser.get_bool("kname");
+    int seed          = arg_parser.get_int("seed");
+
+    assert(shape.size() == perm.size());
+    ck_tile::index_t rank = perm.size();
+    if(rank > ck_tile::GenericPermuteHostArgs::kMaxRanks)
+    {
+        printf("rank %d permute is not support yet\n", rank);
+        return false;
+    }
+
+    ck_tile::HostTensor<DataType> x(shape);
+    ck_tile::FillUniformDistributionIntegerValue<DataType>{-15, 15, seed}(x);
+
+    std::vector<ck_tile::index_t> y_shape = [&]() {
+        std::vector<ck_tile::index_t> tmp(rank, 0);
+        // std::cout << "@@@@" << tmp << std::endl;
+        for(int i = 0; i < static_cast<int>(rank); i++)
+        {
+            // std::cout << "  i:" << i << ", perm:" << perm[i] << ", rak:" <<
+            // static_cast<int>(rank)
+            // << std::endl;
+            tmp[i] = shape[perm[i]];
+        }
+        // std::cout << "@@@" << tmp << std::endl;
+        return tmp;
+    }();
+
+    ck_tile::HostTensor<DataType> y(y_shape);
+
+    ck_tile::DeviceMem x_buf(x.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_buf(y.get_element_space_size_in_bytes());
+
+    x_buf.ToDevice(x.data());
+
+    std::cout << "[" << data_type << "] shape:" << shape << "->" << y_shape << ", permute:" << perm
+              << std::flush;
+
+    ck_tile::stream_config stream_config{nullptr,
+                                         true,
+                                         /* log_level = */ (kname ? 1 : 0),
+                                         stream_warmup,
+                                         stream_repeat};
+    float ave_time   = 0.f;
+    auto run_permute = [&]() {
+        permute_traits t;
+        t.data_type = data_type;
+
+        permute_args a;
+        a.p_src = x_buf.GetDeviceBuffer();
+        a.p_dst = y_buf.GetDeviceBuffer();
+        a.rank  = rank;
+        std::copy(shape.begin(), shape.end(), a.shape);
+        std::copy(perm.begin(), perm.end(), a.perm);
+
+        return permute(t, a, stream_config);
+    };
+#ifdef PERMUTE_USE_ALTERNATIVE_IMPL
+    // batch* n0*n1*n2*k0*k1*k2 -> batch* n0*k0*n1*k1*n2*k2
+    if((arg_parser.get_str("perm") == std::string("0,1,4,2,5,3,6") ||
+        arg_parser.get_str("perm") == std::string("0,1,2,4,5,3,6") ||
+        arg_parser.get_str("perm") == std::string("0,1,3,4,2,5")))
+    {
+        if(arg_parser.get_str("perm") == std::string("0,1,3,4,2,5"))
+        {
+            // b_nr_kr_kw_nw_kv = 2,   // 0,1,3,4,2,5
+            matrix_core_swizzle_traits t;
+            t.data_type = data_type;
+            t.permute   = arg_parser.get_str("perm");
+
+            matrix_core_swizzle_args a;
+            a.p_src = x_buf.GetDeviceBuffer();
+            a.p_dst = y_buf.GetDeviceBuffer();
+            a.batch = shape[0];
+
+            auto nr = shape[1];
+            auto nw = shape[2];
+            auto kr = shape[3];
+            auto kw = shape[4];
+            auto kv = shape[5];
+            a.n     = nr * nw;
+            a.k     = kr * kw * kv;
+            if(kv == 8 && kw == 4 && nw == 16 && nr % 4 == 0 && kr % 8 == 0)
+            {
+                t.inst = "16x16x16";
+                std::cout << ", matrix_core_swizzle_waveflatten_" << t.inst << std::flush;
+
+                ave_time = matrix_core_swizzle(t, a, stream_config);
+            }
+            else if(kv == 8 && kw == 2 && nw == 32 && nr % 4 == 0 && kr % 8 == 0)
+            {
+                t.inst = "32x32x8";
+                std::cout << ", matrix_core_swizzle_waveflatten_" << t.inst << std::flush;
+
+                ave_time = matrix_core_swizzle(t, a, stream_config);
+            }
+            else
+            {
+                ave_time = run_permute();
+            }
+        }
+        else
+        {
+            matrix_core_swizzle_traits t;
+            t.data_type = data_type;
+            t.permute   = arg_parser.get_str("perm");
+
+            matrix_core_swizzle_args a;
+            a.p_src = x_buf.GetDeviceBuffer();
+            a.p_dst = y_buf.GetDeviceBuffer();
+            a.batch = shape[0];
+            a.n     = shape[1] * shape[2] * shape[3];
+            a.k     = shape[4] * shape[5] * shape[6];
+            if(shape[6] == 8 && shape[3] == 32 && shape[5] == 2 && shape[2] == 4 &&
+               shape[4] % 8 == 0 && shape[1] % 2 == 0)
+            {
+                // 32x32x8 inst
+                // perm=0,1,4,2,5,3,6
+                // y_shape=*,2x,8x,4,2,32,8 (3,6,16,4,2,32,8)
+                // shape = *,2x,4,32,8x,2,8 (3,6,4,32,16,2,8)
+
+                t.inst = "32x32x8";
+                std::cout << ", matrix_core_swizzle_" << t.inst << std::flush;
+
+                ave_time = matrix_core_swizzle(t, a, stream_config);
+            }
+            else if(shape[6] == 8 && shape[3] == 16 && shape[5] == 4 && shape[2] == 4 &&
+                    shape[4] % 4 == 0 && shape[1] % 4 == 0)
+            {
+                // 16x16x16 inst
+                // perm=0,1,4,2,5,3,6
+                // y_shape=*,4x,4x,4,4,16,8
+                // shape = *,4x,4,16,4x,4,8 (3,8,4,16,16,4,8)
+                t.inst = "16x16x16";
+                std::cout << ", matrix_core_swizzle_" << t.inst << std::flush;
+
+                ave_time = matrix_core_swizzle(t, a, stream_config);
+            }
+            else
+            {
+                ave_time = run_permute();
+            }
+        }
+    }
+    else
+#endif
+    {
+        ave_time = run_permute();
+    }
+    std::cout << ", time:" << ave_time << "ms" << std::flush;
+
+    bool pass = true;
+    if(do_validation)
+    {
+        reference_permute(x, y, perm);
+#if 0
+        if constexpr (std::is_same_v<float, DataType>){
+            // using itype = to_integer_type<sizeof(DataType)>;
+            fflush(stdout);
+            for(int zz = 0; zz < static_cast<int>(x.get_element_size()); zz++   ) {
+                printf("%3.0f ", x.mData[zz]);
+            }
+            printf("->\n");
+            for(int zz = 0; zz < static_cast<int>(x.get_element_size()); zz++   ) {
+                printf("%3.0f ", y.mData[zz]);
+            }
+            fflush(stdout);
+        }
+#endif
+        ck_tile::HostTensor<DataType> y_dev(y.get_lengths());
+
+        y_buf.FromDevice(y_dev.data());
+
+        pass = std::equal(
+            y_dev.begin(), y_dev.end(), y.begin(), [&](const DataType& d, const DataType& h) {
+                using itype = to_integer_type<sizeof(DataType)>;
+                itype i_d   = ck_tile::bit_cast<itype>(d);
+                itype i_h   = ck_tile::bit_cast<itype>(h);
+                return i_d == i_h;
+            });
+        std::cout << ", valid:" << (pass ? "y" : "n") << std::flush;
+    }
+
+    std::cout << std::endl;
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    const std::string data_type = arg_parser.get_str("prec");
+    if(data_type == "fp8")
+    {
+        return run<ck_tile::fp8_t>(arg_parser) ? 0 : -2;
+    }
+    else if(data_type == "fp16")
+    {
+        return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
+    }
+    else if(data_type == "fp32")
+    {
+        return run<float>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/06_permute/permute.hpp

@@ -0,0 +1,19 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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/permute.hpp"
+#include <string>
+
+struct permute_traits
+{
+    std::string data_type;
+};
+
+using permute_args = ck_tile::GenericPermuteHostArgs;
+
+// host API
+float permute(permute_traits, permute_args, const ck_tile::stream_config&);

example/ck_tile/06_permute/script/smoke_test.sh

@@ -0,0 +1,34 @@
+#!/bin/sh
+# TODO: run this script from CK root
+BUILD=build
+EXE=$BUILD/bin/tile_example_permute
+COMMON_ARGS='-v=1 -warmup=0 -repeat=1'
+# mode=0
+# export HIP_VISIBLE_DEVICES=4
+if [ $# -ge 1 ] ; then
+    set -x
+fi
+
+$EXE -prec=fp16 -shape=3,6,4,32,16,2,8 -perm=0,1,4,2,5,3,6  $COMMON_ARGS
+$EXE -prec=fp16 -shape=5,10,4,32,8,2,8 -perm=0,1,4,2,5,3,6  $COMMON_ARGS
+$EXE -prec=fp16 -shape=3,8,4,16,16,4,8 -perm=0,1,4,2,5,3,6  $COMMON_ARGS
+$EXE -prec=fp16 -shape=3,6,4,32,16,2,8 -perm=0,1,2,4,5,3,6  $COMMON_ARGS
+$EXE -prec=fp16 -shape=5,10,4,32,8,2,8 -perm=0,1,2,4,5,3,6  $COMMON_ARGS
+$EXE -prec=fp16 -shape=3,8,4,16,16,4,8 -perm=0,1,2,4,5,3,6  $COMMON_ARGS
+$EXE -prec=fp16 -shape=2,8,16,8,4,8 -perm=0,1,3,4,2,5  $COMMON_ARGS
+$EXE -prec=fp16 -shape=1,24,32,16,2,8 -perm=0,1,3,4,2,5  $COMMON_ARGS
+
+echo "------------------------------------------------------------------"
+
+for prec in "fp8" "fp16" "fp32" ; do
+
+$EXE -prec=$prec -shape=3,8 -perm=1,0 $COMMON_ARGS
+$EXE -prec=$prec -shape=48,6,8 -perm=2,1,0  $COMMON_ARGS
+$EXE -prec=$prec -shape=24,128,3 -perm=0,2,1  $COMMON_ARGS
+$EXE -prec=$prec -shape=4,10,7,6 -perm=0,2,3,1  $COMMON_ARGS
+$EXE -prec=$prec -shape=8,24,36,10 -perm=3,1,2,0  $COMMON_ARGS
+$EXE -prec=$prec -shape=8,1,36,4 -perm=2,1,0,3  $COMMON_ARGS
+$EXE -prec=$prec -shape=5,10,16,2,36,4 -perm=4,5,2,1,0,3  $COMMON_ARGS
+$EXE -prec=$prec -shape=2,32,8,3,6,2,5,4 -perm=5,2,4,7,1,6,3,0  $COMMON_ARGS
+echo "------------------------------------------------------------------"
+done

example/ck_tile/09_topk_softmax/CMakeLists.txt

@@ -0,0 +1,9 @@
+add_executable(tile_example_topk_softmax topk_softmax.cpp topk_softmax_api.cpp)
+rocm_install(TARGETS tile_example_topk_softmax COMPONENT examples)
+
+target_include_directories(tile_example_topk_softmax PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/)
+set(EXAMPLE_TOPK_SOFTMAX_COMPILE_OPTIONS)
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+list(APPEND EXAMPLE_TOPK_SOFTMAX_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
+# list(APPEND EXAMPLE_TOPK_SOFTMAX_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker)
+target_compile_options(tile_example_topk_softmax PRIVATE ${EXAMPLE_TOPK_SOFTMAX_COMPILE_OPTIONS})

example/ck_tile/09_topk_softmax/README.md

@@ -0,0 +1,28 @@
+# topk-softmax
+
+This folder contains example for topk-softmax kernel using ck_tile tile-programming implementation. This kernel is often used in Moe model, before launching the fused-moe-gemm block. The input is a `token*expert` 2d matrix. The op will do a softmax per row(`expert`), then find the `topk` value for each row. Output is a `token*topk`  weight(usually fp32) and index(int32) 2d tensor.
+
+## build
+```
+# 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...
+make tile_example_topk_softmax -j
+```
+This will result in an executable `build/bin/tile_example_topk_softmax`
+
+## example
+```
+args:
+          -v    weather do CPU validation or not (default:1)
+       -pr_i    input data type. fp16/fp32 (representing 8/16/32 bit data) (default:fp16)
+       -pr_w    output weight data type(currently only fp32 supported now) (default:fp32)
+          -t    number of input tokens (default:32)
+          -e    number of experts (default:8)
+          -k    topk (default:2)
+       -st_i    row stride of input, -1 means same as experts (default:-1)
+       -st_o    row stride of output/indices, -1 means same as topk (default:-1)
+       -seed    seed to be used, -1 means random every time (default:-1)
+      -kname    when set to 1 it will print kernel name (default:0)
+
+```

example/ck_tile/09_topk_softmax/script/smoke_test.sh

@@ -0,0 +1,22 @@
+#!/bin/sh
+
+EXE=./build/bin/tile_example_topk_softmax
+
+for pr_i in "fp16" "bf16" ; do
+$EXE -pr_i=$pr_i -t=80 -e=17
+$EXE -pr_i=$pr_i -t=111 -e=117
+$EXE -pr_i=$pr_i -t=1000 -e=55
+$EXE -pr_i=$pr_i -t=99 -e=180
+$EXE -pr_i=$pr_i -t=175 -e=64 -k=8
+$EXE -pr_i=$pr_i -t=65 -e=8 -k=2
+$EXE -pr_i=$pr_i -t=1 -e=25
+$EXE -pr_i=$pr_i -t=31 -e=19 -k=15
+$EXE -pr_i=$pr_i -t=81 -e=37 -k=7
+$EXE -pr_i=$pr_i -t=199 -e=128 -k=13
+$EXE -pr_i=$pr_i -t=23 -e=1 -k=1
+$EXE -pr_i=$pr_i -t=127 -e=99 -k=19 -st_i=233 -st_o=31
+$EXE -pr_i=$pr_i -t=71 -e=11 -k=11 -st_i=30 -st_o=12
+$EXE -pr_i=$pr_i -t=1 -e=1 -k=1
+$EXE -pr_i=$pr_i -t=99 -e=2 -k=1 -st_i=11 -st_o=5
+$EXE -pr_i=$pr_i -t=333 -e=99 -k=13 -st_i=191 -st_o=17
+done

example/ck_tile/09_topk_softmax/topk_softmax.cpp

@@ -0,0 +1,299 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <vector>
+#include <iostream>
+#include <numeric>
+#include <cassert>
+#include <cstdlib>
+#include <iostream>
+#include <time.h>
+#include <unordered_set>
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/reduce.hpp"
+#include "topk_softmax_api.hpp"
+
+#if 0
+template <typename T>
+void dump_host_tensor_2d(const ck_tile::HostTensor<T>& x)
+{
+    auto len = x.get_lengths();
+    assert(len.size() == 2);
+    std::cout << "[";
+    for(size_t i = 0; i < len[0]; i++)
+    {
+        std::cout << i << ": [";
+        for(size_t j = 0; j < len[1]; j++)
+        {
+            if constexpr(std::is_same_v<T, ck_tile::fp16_t>)
+            {
+                auto v = ck_tile::type_convert<float>(x(i, j));
+
+                std::cout << v;
+                if(j != len[1] - 1)
+                    std::cout << ",";
+            }
+            else
+            {
+                std::cout << x(i, j) << " ";
+            }
+        }
+        std::cout << "]";
+        if(i != len[0] - 1)
+            std::cout << ",";
+        else
+            std::cout << "]";
+        std::cout << std::endl;
+    }
+    std::cout << "--------------------" << std::endl;
+}
+#endif
+
+// CPU reference
+template <typename InputType, typename WeightType, typename IndexType = ck_tile::index_t>
+auto reference_topk_softmax(const ck_tile::HostTensor<InputType>& x,
+                            ck_tile::index_t k,
+                            ck_tile::index_t dim = -1,
+                            bool largest         = true,
+                            bool sorted          = true)
+{
+    using namespace ck_tile;
+
+    auto y = reference_softmax<InputType, WeightType, WeightType>(x, dim);
+
+    auto [y_values, y_indices] = reference_topk(y, k, dim, largest, sorted);
+
+    return ck_tile::make_tuple(y_values, y_indices);
+}
+
+template <typename InputType, typename WeightType, typename IndexType = ck_tile::index_t>
+auto reference_topk_softmax(const ck_tile::HostTensor<InputType>& x,
+                            ck_tile::HostTensor<WeightType>& y_values,
+                            ck_tile::HostTensor<IndexType>& y_indices,
+                            ck_tile::index_t k,
+                            ck_tile::index_t dim = -1,
+                            bool largest         = true,
+                            bool sorted          = true)
+{
+    using namespace ck_tile;
+
+    auto y = reference_softmax<InputType, WeightType, WeightType>(x, dim);
+    reference_topk(y, y_values, y_indices, k, dim, largest, sorted);
+}
+
+// different threshold for different dtype
+template <typename DataType>
+auto get_elimit(std::string /*init_method*/)
+{
+    double rtol = 1e-3;
+    double atol = 1e-3;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::bf16_t>(std::string /*init_method*/)
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::fp8_t>(std::string init_method)
+{
+    if(init_method == "ui" || init_method == "ni")
+    {
+        unsigned max_rounding_point_distance = 0;
+        double atol                          = 2e-3;
+        return ck_tile::make_tuple(max_rounding_point_distance, atol);
+    }
+    else
+    {
+        unsigned max_rounding_point_distance = 1;
+        double atol                          = 0.0625;
+        return ck_tile::make_tuple(max_rounding_point_distance, atol);
+    }
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("v", "1", "weather do CPU validation or not")
+        .insert("pr_i", "fp16", "input data type. fp16/fp32 (representing 8/16/32 bit data)")
+        .insert("pr_w", "fp32", "output weight data type(currently only fp32 supported now)")
+        .insert("t", "32", "number of input tokens")
+        .insert("e", "8", "number of experts")
+        .insert("k", "2", "topk")
+        .insert("st_i", "-1", "row stride of input, -1 means same as experts")
+        .insert("st_o", "-1", "row stride of output/indices, -1 means same as topk")
+        .insert("seed", "-1", "seed to be used, -1 means random every time")
+        .insert("kname", "0", "when set to 1 it will print kernel name")
+        .insert("warmup", "5", "number of iterations before benchmark the kernel")
+        .insert("repeat", "20", "number of iterations to benchmark the kernel");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename InputType, typename WeightType, typename IndexType = ck_tile::index_t>
+bool test_topk_softmax(ck_tile::ArgParser args)
+{
+    int validate            = args.get_int("v");
+    std::string input_prec  = args.get_str("pr_i");
+    std::string weight_prec = args.get_str("pr_w");
+    int tokens              = args.get_int("t");
+    int experts             = args.get_int("e");
+    int topk                = args.get_int("k");
+    int seed                = args.get_int("seed");
+    int stride_input        = args.get_int("st_i");
+    int stride_output       = args.get_int("st_o");
+    int kname               = args.get_int("kname");
+    int warmup              = args.get_int("warmup");
+    int repeat              = args.get_int("repeat");
+
+    if(stride_input < 0)
+    {
+        stride_input = experts;
+    }
+    if(stride_output < 0)
+    {
+        stride_output = topk;
+    }
+    assert(stride_input >= experts);
+    assert(stride_output >= topk);
+
+    if(seed < 0)
+    {
+        seed = std::time(nullptr);
+    }
+
+    if(topk > experts)
+    {
+        printf("topk:%d value should be smaller than, or equal to number of experts:%d\n",
+               topk,
+               experts);
+        return false;
+    }
+
+    // tokens already considered batch size
+    ck_tile::HostTensor<InputType> x_host({tokens, experts}, {stride_input, 1});
+    ck_tile::HostTensor<WeightType> value_host({tokens, topk}, {stride_output, 1});
+    ck_tile::HostTensor<IndexType> index_host({tokens, topk}, {stride_output, 1});
+
+    {
+        // random require per-row unique
+        auto rand_gen = ck_tile::FillUniformDistribution_Unique<InputType>{
+            -5.f, 5.f, static_cast<uint32_t>(seed)};
+
+        for(int i_t = 0; i_t < tokens; i_t++)
+        {
+            ck_tile::HostTensor<InputType> x_row({experts});
+            rand_gen(x_row);
+            std::copy(x_row.begin(), x_row.end(), x_host.begin() + i_t * stride_input);
+            rand_gen.clear();
+        }
+    }
+
+    ck_tile::DeviceMem x_dev(x_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem value_dev(value_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem index_dev(index_host.get_element_space_size_in_bytes());
+
+    x_dev.ToDevice(x_host.data());
+
+    topk_softmax_trait trait{input_prec, weight_prec, experts};
+
+    topk_softmax_kargs karg{x_dev.GetDeviceBuffer(),
+                            value_dev.GetDeviceBuffer(),
+                            index_dev.GetDeviceBuffer(),
+                            tokens,
+                            experts,
+                            topk,
+                            stride_input,
+                            stride_output};
+
+    ck_tile::stream_config sc{nullptr,
+                              true,
+                              /* log_level = */ (kname ? 1 : 0),
+                              warmup,
+                              repeat};
+    auto ms = topk_softmax(trait, karg, sc);
+    printf("[%s|%s]tokens:%d, experts:%d, topk:%d, st_i:%d, st_o:%d, ms:%f, ",
+           input_prec.c_str(),
+           weight_prec.c_str(),
+           tokens,
+           experts,
+           topk,
+           stride_input,
+           stride_output,
+           ms);
+    if(ms < 0)
+        printf("not supported\n");
+    fflush(stdout);
+    if(ms < 0)
+    {
+        return false;
+    }
+
+    value_dev.FromDevice(value_host.data());
+    index_dev.FromDevice(index_host.data());
+
+    bool rtn = true;
+    if(validate)
+    {
+        ck_tile::HostTensor<WeightType> value_ref({tokens, topk}, {stride_output, 1});
+        ck_tile::HostTensor<IndexType> index_ref({tokens, topk}, {stride_output, 1});
+
+        reference_topk_softmax<InputType, WeightType, IndexType>(
+            x_host, value_ref, index_ref, topk);
+
+        auto [rtol, atol] = get_elimit<InputType>("");
+        for(int i_t = 0; i_t < tokens; i_t++)
+        {
+            auto s_begin = std::vector<size_t>{static_cast<size_t>(i_t), static_cast<size_t>(0)};
+            auto s_end =
+                std::vector<size_t>{static_cast<size_t>(i_t + 1), static_cast<size_t>(topk)};
+            auto s_value_host = value_host.slice(s_begin, s_end);
+            auto s_value_ref  = value_ref.slice(s_begin, s_end);
+            rtn &= ck_tile::check_err(s_value_host,
+                                      s_value_ref,
+                                      std::string("[") + std::to_string(i_t) +
+                                          std::string("] Value Error:"),
+                                      rtol,
+                                      atol);
+            auto s_index_host = index_host.slice(s_begin, s_end);
+            auto s_index_ref  = index_ref.slice(s_begin, s_end);
+            rtn &= ck_tile::check_err(s_index_host,
+                                      s_index_ref,
+                                      std::string("[") + std::to_string(i_t) +
+                                          std::string("] Index Error:"),
+                                      rtol,
+                                      atol);
+        }
+    }
+
+    printf("valid:%s\n", rtn ? "y" : "n");
+    fflush(stdout);
+    return rtn;
+}
+
+int main(int argc, char** argv)
+{
+    auto [result, args] = create_args(argc, argv);
+    if(!result)
+        return -1;
+    std::string input_prec  = args.get_str("pr_i");
+    std::string weight_prec = args.get_str("pr_w");
+
+    bool r = true;
+    if(input_prec.compare("fp16") == 0 && weight_prec.compare("fp32") == 0)
+    {
+        r &= test_topk_softmax<ck_tile::fp16_t, float, ck_tile::index_t>(args);
+    }
+    else if(input_prec.compare("bf16") == 0 && weight_prec.compare("fp32") == 0)
+    {
+        r &= test_topk_softmax<ck_tile::bf16_t, float, ck_tile::index_t>(args);
+    }
+
+    return r ? 0 : -1;
+}

example/ck_tile/09_topk_softmax/topk_softmax_api.cpp

@@ -0,0 +1,96 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "topk_softmax_api.hpp"
+
+#define TOPK_SOFTMAX_DISPATCH(experts_)                                                         \
+    constexpr ck_tile::index_t ts_experts = experts_;                                           \
+    using ts_problem                      = ck_tile::                                           \
+        TopkSoftmaxWarpPerRowProblem<ts_input_type, ts_weight_type, ts_index_type, ts_experts>; \
+    using ts_pipeline = ck_tile::TopkSoftmaxWarpPerRowPipeline<ts_problem>;                     \
+                                                                                                \
+    using kernel = ck_tile::TopkSoftmaxKernel<ts_pipeline>;                                     \
+                                                                                                \
+    auto kargs = kernel::MakeKargs(a);                                                          \
+                                                                                                \
+    const dim3 grids      = kernel::GridSize(a);                                                \
+    constexpr dim3 blocks = kernel::BlockSize();                                                \
+                                                                                                \
+    float ave_time = ck_tile::launch_kernel(                                                    \
+        s, ck_tile::make_kernel<blocks.x, 1>(kernel{}, grids, blocks, 0, kargs));               \
+                                                                                                \
+    return ave_time;
+
+float topk_softmax(topk_softmax_trait t, topk_softmax_kargs a, ck_tile::stream_config s)
+{
+    if(t.input_type == "fp16" && t.weight_type == "fp32")
+    {
+        using ts_input_type  = ck_tile::fp16_t;
+        using ts_weight_type = float;
+        using ts_index_type  = ck_tile::index_t;
+#if 1
+        if(t.experts <= 8)
+        {
+            TOPK_SOFTMAX_DISPATCH(8)
+        }
+        else if(t.experts <= 16)
+        {
+            TOPK_SOFTMAX_DISPATCH(16)
+        }
+        else if(t.experts <= 32)
+        {
+            TOPK_SOFTMAX_DISPATCH(32)
+        }
+        else if(t.experts <= 64)
+        {
+            TOPK_SOFTMAX_DISPATCH(64)
+        }
+        else if(t.experts <= 128)
+        {
+            TOPK_SOFTMAX_DISPATCH(128)
+        }
+        else if(t.experts <= 192)
+        {
+            TOPK_SOFTMAX_DISPATCH(192)
+        }
+#else
+        if(t.experts <= 128)
+        {
+            TOPK_SOFTMAX_DISPATCH(128)
+        }
+#endif
+    }
+    else if(t.input_type == "bf16" && t.weight_type == "fp32")
+    {
+#if 1
+        using ts_input_type  = ck_tile::bf16_t;
+        using ts_weight_type = float;
+        using ts_index_type  = ck_tile::index_t;
+        if(t.experts <= 8)
+        {
+            TOPK_SOFTMAX_DISPATCH(8)
+        }
+        else if(t.experts <= 16)
+        {
+            TOPK_SOFTMAX_DISPATCH(16)
+        }
+        else if(t.experts <= 32)
+        {
+            TOPK_SOFTMAX_DISPATCH(32)
+        }
+        else if(t.experts <= 64)
+        {
+            TOPK_SOFTMAX_DISPATCH(64)
+        }
+        else if(t.experts <= 128)
+        {
+            TOPK_SOFTMAX_DISPATCH(128)
+        }
+        else if(t.experts <= 192)
+        {
+            TOPK_SOFTMAX_DISPATCH(192)
+        }
+#endif
+    }
+    return -1;
+}

example/ck_tile/09_topk_softmax/topk_softmax_api.hpp

@@ -0,0 +1,21 @@
+// 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.hpp"
+#include "ck_tile/ops/topk_softmax.hpp"
+#include <string>
+
+struct topk_softmax_trait
+{
+    std::string input_type;
+    std::string weight_type; // currently always float
+    int experts;
+};
+
+struct topk_softmax_kargs : public ck_tile::TopkSoftmaxHostArgs
+{
+};
+
+float topk_softmax(topk_softmax_trait t, topk_softmax_kargs a, ck_tile::stream_config s);

example/ck_tile/10_rmsnorm2d/CMakeLists.txt

@@ -0,0 +1,50 @@
+set(RMSNORM2D_FWD_KNOWN_APIS "fwd;bwd")
+set(RMSNORM2D_FWD_ENABLE_APIS  "fwd" CACHE STRING
+    "semicolon-separated list of APIs to generate (${RMSNORM2D_FWD_KNOWN_APIS}) & link, or \"all\".")
+if(RMSNORM2D_FWD_ENABLE_APIS  STREQUAL "all")
+  set(RMSNORM2D_FWD_ENABLE_APIS  ${RMSNORM2D_FWD_KNOWN_APIS})
+endif()
+
+# generate a list of kernels, but not actually emit files at config sta
+execute_process(
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${RMSNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --list_blobs
+  RESULT_VARIABLE ret
+)
+if(ret AND NOT ret EQUAL 0)
+  message( FATAL_ERROR "Fail to generate kernels via Python. ${ret}")
+endif()
+
+file(STRINGS ${CMAKE_CURRENT_BINARY_DIR}/rmsnorm2d_fwd_blobs.txt RMSNORM2D_FWD_GEN_BLOBS)
+
+add_custom_command(
+  OUTPUT ${RMSNORM2D_FWD_GEN_BLOBS}
+  COMMAND ${Python3_EXECUTABLE} ${CMAKE_CURRENT_LIST_DIR}/generate.py
+  --api ${RMSNORM2D_FWD_ENABLE_APIS} --working_path ${CMAKE_CURRENT_BINARY_DIR} --gen_blobs
+)
+
+set(TILE_RMSNORM2D_FWD "tile_rmsnorm2d_fwd")
+
+message("adding ${TILE_RMSNORM2D_FWD}")
+add_executable(${TILE_RMSNORM2D_FWD} rmsnorm2d_fwd.cpp)
+rocm_install(TARGETS ${TILE_RMSNORM2D_FWD} COMPONENT examples)
+target_include_directories(${TILE_RMSNORM2D_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+target_sources(${TILE_RMSNORM2D_FWD} PRIVATE ${RMSNORM2D_FWD_GEN_BLOBS})
+
+set(TILE_RMSNORM2D_FWD_COMPILE_OPTIONS)
+
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+list(APPEND TILE_RMSNORM2D_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal --offload-compress)
+
+target_compile_options(${TILE_RMSNORM2D_FWD} PRIVATE ${TILE_RMSNORM2D_FWD_COMPILE_OPTIONS})
+
+set(EXAMPLE_RMSNORM2D_FWD "tile_example_rmsnorm2d_fwd")
+add_executable(${EXAMPLE_RMSNORM2D_FWD} example_rmsnorm2d_fwd.cpp)
+rocm_install(TARGETS ${EXAMPLE_RMSNORM2D_FWD} COMPONENT examples)
+target_compile_options(${EXAMPLE_RMSNORM2D_FWD} PRIVATE ${TILE_RMSNORM2D_FWD_COMPILE_OPTIONS})
+
+# TODO: we have to turn off this global prop, otherwise the progress bar generated
+# by cmake will print too many files, execvp: /bin/sh: Argument list too long
+# however, this property may affect global
+# TODO: consider codegen a makefile by us
+set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)

example/ck_tile/10_rmsnorm2d/README.md

@@ -0,0 +1,22 @@
+# Rmsnorm2D forward
+
+This folder contains example for Rmsnorm2D forward using ck_tile tile-programming implementation.
+
+## build
+```
+# 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...
+make tile_rmsnorm2d_fwd -j
+```
+This will result in an executable `build/bin/tile_rmsnorm2d_fwd`
+
+## cmdline
+```
+args:
+          -m    m dimension (default:3328)
+          -n    m dimension (default:4096)
+          -e    epsilon (default:1e-5)
+          -v    cpu validation or not (default:1)
+       -prec    precision (default:fp16)
+```

example/ck_tile/10_rmsnorm2d/example_rmsnorm2d_fwd.cpp

@@ -0,0 +1,194 @@
+#include "ck_tile/host.hpp"
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/kernel_launch.hpp"
+#include "ck_tile/ops/epilogue.hpp"
+#include "ck_tile/ops/rmsnorm2d.hpp"
+#include <cstring>
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("m", "3328", "m dimension")
+        .insert("n", "4096", "n dimension")
+        .insert("stride", "-1", "stride per row, if -1 then equal to n")
+        .insert("e", "1e-5", "epsilon")
+        .insert("v", "1", "cpu validation or not")
+        .insert("prec", "fp16", "precision")
+        .insert("warmup", "0", "cold iter")
+        .insert("repeat", "1", "hot iter");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename DataType>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    ck_tile::index_t m      = arg_parser.get_int("m");
+    ck_tile::index_t n      = arg_parser.get_int("n");
+    ck_tile::index_t stride = arg_parser.get_int("stride");
+    if(stride < 0)
+        stride = n;
+    float epsilon         = arg_parser.get_float("e");
+    std::string data_type = arg_parser.get_str("prec");
+    int do_validation     = arg_parser.get_int("v");
+    int warmup            = arg_parser.get_int("warmup");
+    int repeat            = arg_parser.get_int("repeat");
+
+    assert(stride >= n);
+
+    using XDataType           = DataType;
+    using YDataType           = DataType;
+    using GammaDataType       = DataType;
+    using InvRmsDataType      = ck_tile::null_type;
+    using UnquantYDataType    = ck_tile::null_type;
+    using SmoothScaleDataType = ck_tile::null_type;
+    using YScaleDataType      = ck_tile::null_type;
+
+    using ComputeDataType = float;
+
+    // host verify
+    ck_tile::HostTensor<XDataType> x_host({m, n}, {stride, 1});
+    ck_tile::HostTensor<GammaDataType> gamma_host({n});
+
+    ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {stride, 1});
+    ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {stride, 1});
+
+    ck_tile::HostTensor<InvRmsDataType> invRms_host_ref({m});
+
+    ck_tile::HostTensor<UnquantYDataType> unquant_y_host_ref({m, n}, {stride, 1});
+
+    ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
+    ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host);
+
+    ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem gamma_buf(gamma_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_buf(y_host_dev.get_element_space_size_in_bytes());
+
+    x_buf.ToDevice(x_host.data());
+    gamma_buf.ToDevice(gamma_host.data());
+
+    constexpr bool kTwoPass = true;
+
+    using BlockWarps = ck_tile::sequence<2, 2>;
+    using BlockTile  = ck_tile::sequence<2, 128>;
+    using WarpTile   = ck_tile::sequence<1, 64>;
+    using Vector     = ck_tile::sequence<1, 1>;
+    using Shape      = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
+
+    using PipelineTraits =
+        ck_tile::Rmsnorm2dFwdTraits<true,  // kPadN
+                                    false, // kSaveInvRms
+                                    false, // kSaveUnquant
+                                    kTwoPass,
+                                    ck_tile::Rmsnorm2dFusedAddEnum::NO_ADD,      // fuse add
+                                    ck_tile::Rmsnorm2dFusedQuantEnum::NO_SWEEP>; // fuse quant
+
+    using Problem = ck_tile::Rmsnorm2dFwdPipelineProblem<XDataType,
+                                                         GammaDataType,
+                                                         ComputeDataType,
+                                                         YDataType,
+                                                         InvRmsDataType,
+                                                         UnquantYDataType,
+                                                         SmoothScaleDataType,
+                                                         YScaleDataType,
+                                                         Shape,
+                                                         PipelineTraits>;
+
+    using OnePassPipeline = ck_tile::Rmsnorm2dFwdPipelineOnePass<Problem>;
+    using TwoPassPipeline = ck_tile::Rmsnorm2dFwdPipelineTwoPass<Problem>;
+    using Pipeline        = std::conditional_t<kTwoPass, TwoPassPipeline, OnePassPipeline>;
+
+    using Default2DEpilogueProblem = ck_tile::
+        Default2DEpilogueProblem<ComputeDataType, YDataType, false, PipelineTraits::kPadN, false>;
+    using Default2DEpilogue = ck_tile::Default2DEpilogue<Default2DEpilogueProblem>;
+
+    using Kernel = ck_tile::Rmsnorm2dFwd<Pipeline, Default2DEpilogue>;
+
+    ck_tile::Rmsnorm2dFwdHostArgs args{x_buf.GetDeviceBuffer(),
+                                       nullptr,
+                                       nullptr,
+                                       gamma_buf.GetDeviceBuffer(),
+                                       y_buf.GetDeviceBuffer(),
+                                       nullptr,
+                                       nullptr,
+                                       nullptr,
+                                       nullptr,
+                                       epsilon,
+                                       m,
+                                       n,
+                                       stride,
+                                       stride,
+                                       stride,
+                                       stride};
+
+    auto kargs = Kernel::MakeKargs(args);
+
+    const dim3 grids                       = Kernel::GridSize(args);
+    constexpr dim3 blocks                  = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+    auto s = ck_tile::stream_config{nullptr, true, 0, warmup, repeat};
+
+    ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+
+    bool pass = true;
+
+    if(do_validation)
+    {
+        // reference
+        ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                         GammaDataType,
+                                         ComputeDataType,
+                                         YDataType,
+                                         InvRmsDataType,
+                                         UnquantYDataType>(
+            x_host, gamma_host, y_host_ref, invRms_host_ref, unquant_y_host_ref, epsilon);
+
+        y_buf.FromDevice(y_host_dev.data());
+
+        auto [rtol, atol] = ck_tile::make_tuple(1e-3, 1e-3);
+        if(stride == n)
+        {
+            pass = ck_tile::check_err(
+                y_host_dev, y_host_ref, std::string("OUT Error: Incorrect results!"), rtol, atol);
+        }
+        else
+        {
+            for(int i_r = 0; i_r < m; i_r++)
+            {
+                std::vector<YDataType> y_host_dev_row(y_host_dev.begin() + i_r * stride,
+                                                      y_host_dev.begin() + i_r * stride + n);
+                std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * stride,
+                                                      y_host_ref.begin() + i_r * stride + n);
+                pass &= ck_tile::check_err(y_host_dev_row,
+                                           y_host_ref_row,
+                                           std::string("OUT[") + std::to_string(i_r) +
+                                               std::string("] Error: Incorrect results!"),
+                                           rtol,
+                                           atol);
+            }
+        }
+
+        std::cout << "[" << data_type << "]"
+                  << " m:" << m << ", n:" << n << ", stride:" << stride
+                  << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    const std::string data_type = arg_parser.get_str("prec");
+    if(data_type == "fp16")
+    {
+        return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/10_rmsnorm2d/generate.py

@@ -0,0 +1,715 @@
+# SPDX-License-Identifier: MIT
+# Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+# generate kernel instances to speed up compilation
+
+import argparse
+from enum import IntEnum
+from pathlib import Path
+import sys
+from typing import List, Optional, Any
+import functools
+import itertools
+import copy
+from dataclasses import dataclass
+
+
+def get_if_str(idx, total, lase_else = True):
+    if idx == 0:
+        return 'if'
+    elif idx < total - 1:
+        return 'else if'
+    else:
+        if lase_else:
+            return 'else'
+        else:
+            return 'else if'
+
+FUSED_ADD_ENUM_STR_MAP = [
+    'no',
+    'pras',      # pre-norm
+    'pra' ]      # post-norm
+
+FUSED_FUSED_SWEEP_STR_MAP = [
+    'no',
+    'sdquant',  # smooth dynamic quant
+    'dquant' ]  # dynamic quant (without sm_scale)
+
+DATA_TYPE_MAP = {'fp32' : 'float',
+                 'fp16' : 'ck_tile::fp16_t',
+                 'bf16' : 'ck_tile::bf16_t',
+                 'int8' : 'ck_tile::int8_t',
+                 'fp8'  : 'ck_tile::fp8_t'}
+
+def BOOL_MAP(b_) -> str:
+    if b_:
+        return 'true'
+    else:
+        return 'false'
+
+
+class rmsnorm_fwd_codegen:
+    API_TRAITS_DEFINE = """
+// this is used to pattern-match internl kernel implementation, not to instantiate kernel
+template <typename XDataType_,
+          typename YDataType_,
+          typename SmoothScaleDataType_,
+          typename YScaleDataType_,
+          typename UnquantYDataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kSaveInvRms_,
+          bool kSaveUnquant_,
+          bool kTwoPass_,
+          ck_tile::index_t kFusedAdd_ = 0,
+          ck_tile::index_t kFusedQuant_ = 0>
+struct rmsnorm2d_fwd_traits_
+{
+    using XDataType           = ck_tile::remove_cvref_t<XDataType_>;
+    using YDataType           = ck_tile::remove_cvref_t<YDataType_>;
+    using SmoothScaleDataType = ck_tile::remove_cvref_t<SmoothScaleDataType_>;
+    using YScaleDataType      = ck_tile::remove_cvref_t<YScaleDataType_>;
+    using UnquantYDataType    = ck_tile::remove_cvref_t<UnquantYDataType_>;
+
+    static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize;
+    static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % warpSize == 0);
+    static constexpr ck_tile::index_t total_warps =
+        (ThreadPerBlock_M_ * ThreadPerBlock_N_) / warpSize;
+
+    // num of warps along m
+    static constexpr ck_tile::index_t BlockWarps_M = []() {
+        if constexpr(is_warp_per_row)
+        {
+            static_assert(warpSize % ThreadPerBlock_N_ == 0);
+            return total_warps * (warpSize / ThreadPerBlock_N_);
+        }
+        else
+        {
+            // static_assert(warpSize % ThreadPerBlock_M_ == 0);
+            return total_warps / (ThreadPerBlock_N_ / warpSize);
+        }
+    }();
+
+    // num of warps along n
+    static constexpr ck_tile::index_t BlockWarps_N = []() {
+        if constexpr(is_warp_per_row)
+        {
+            static_assert(warpSize % ThreadPerBlock_N_ == 0);
+            return 1;
+        }
+        else
+        {
+            static_assert(ThreadPerBlock_N_ % warpSize == 0);
+            return ThreadPerBlock_N_ / warpSize;
+        }
+    }();
+
+    static constexpr ck_tile::index_t Repeat_M = Repeat_M_;
+    static constexpr ck_tile::index_t Repeat_N = Repeat_N_;
+
+    static constexpr ck_tile::index_t Block_M = Repeat_M_ * ThreadPerBlock_M_;
+    static constexpr ck_tile::index_t Block_N = Repeat_N_ * ThreadPerBlock_N_ * Vector_N_;
+
+    static constexpr ck_tile::index_t Warp_M = ThreadPerBlock_M_ / BlockWarps_M;
+    static constexpr ck_tile::index_t Warp_N = ThreadPerBlock_N_ / BlockWarps_N * Vector_N_;
+
+    using BlockTile  = ck_tile::sequence<Block_M, Block_N>;
+    using BlockWarps = ck_tile::sequence<BlockWarps_M, BlockWarps_N>;
+    using WarpTile   = ck_tile::sequence<Warp_M, Warp_N>;
+    using Vector     = ck_tile::sequence<1, Vector_N_>;
+
+    using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
+
+    static constexpr bool kPadN        = kPadN_;
+    static constexpr bool kSaveInvRms  = kSaveInvRms_;
+    static constexpr bool kSaveUnquant = kSaveUnquant_;
+    static constexpr bool kTwoPass     = kTwoPass_;
+    static constexpr ck_tile::index_t kFusedAdd = kFusedAdd_;
+    static constexpr ck_tile::index_t kFusedQuant = kFusedQuant_;
+};
+
+template <typename XDataType_,
+          typename YDataType_,
+          typename SmoothScaleDataType_,
+          typename YScaleDataType_,
+          typename UnquantYDataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kSaveInvRms_,
+          bool kSaveUnquant_,
+          bool kTwoPass_,
+          int  kFusedAdd_,
+          int  kFusedQuant_>
+using traits_ = rmsnorm2d_fwd_traits_<XDataType_,
+                                      YDataType_,
+                                      SmoothScaleDataType_,
+                                      YScaleDataType_,
+                                      UnquantYDataType_,
+                                      Repeat_M_,
+                                      Repeat_N_,
+                                      ThreadPerBlock_M_,
+                                      ThreadPerBlock_N_,
+                                      Vector_N_,
+                                      kPadN_,
+                                      kSaveInvRms_,
+                                      kSaveUnquant_,
+                                      kTwoPass_,
+                                      kFusedAdd_,
+                                      kFusedQuant_>;
+"""
+
+    API_COMMON_HEADER = """
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "rmsnorm2d_fwd.hpp"
+#include <ck_tile/ops/epilogue.hpp>
+#include <iostream>
+
+#pragma once
+
+using S = ck_tile::stream_config;
+using A = rmsnorm2d_fwd_args;
+
+{F_traits_define}
+
+template <typename Traits_>
+float rmsnorm2d_fwd_(const S& s, A a)
+{{
+    using XDataType           = typename Traits_::XDataType;
+    using YDataType           = typename Traits_::YDataType;
+    using SmoothScaleDataType = typename Traits_::SmoothScaleDataType;
+    using YScaleDataType      = typename Traits_::YScaleDataType;
+    using UnquantYDataType    = typename Traits_::UnquantYDataType;
+    using ComputeDataType     = typename RmsnormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::ComputeDataType;
+
+    using PipelineTraits =
+        ck_tile::Rmsnorm2dFwdTraits<Traits_::kPadN,
+                                    Traits_::kSaveInvRms,
+                                    Traits_::kSaveUnquant,
+                                    Traits_::kTwoPass,
+                                    static_cast<ck_tile::Rmsnorm2dFusedAddEnum>(Traits_::kFusedAdd),
+                                    static_cast<ck_tile::Rmsnorm2dFusedQuantEnum>(Traits_::kFusedQuant)>;
+
+    using PipelineProblem =
+        ck_tile::Rmsnorm2dFwdPipelineProblem<typename RmsnormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::XDataType,
+                                             typename RmsnormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::GammaDataType,
+                                             typename RmsnormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::ComputeDataType,
+                                             typename RmsnormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::YDataType,
+                                             typename RmsnormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::InvRmsDataType,
+                                             typename RmsnormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::UnquantYDataType,
+                                             typename RmsnormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::SmoothScaleDataType,
+                                             typename RmsnormTypeConfig<XDataType, YDataType, SmoothScaleDataType, YScaleDataType>::YScaleDataType,
+                                             typename Traits_::Shape,
+                                             PipelineTraits>;
+
+    using OnePassPipeline = ck_tile::Rmsnorm2dFwdPipelineOnePass<PipelineProblem>;
+    using TwoPassPipeline = ck_tile::Rmsnorm2dFwdPipelineTwoPass<PipelineProblem>;
+    using Pipeline        = std::conditional_t<Traits_::kTwoPass, TwoPassPipeline, OnePassPipeline>;
+
+    using Default2DEpilogueProblem = ck_tile::Default2DEpilogueProblem<ComputeDataType, YDataType, false, Traits_::kPadN, false>;
+    using Default2DEpilogue = ck_tile::Default2DEpilogue<Default2DEpilogueProblem>;
+
+    static constexpr bool UseSmoothInputScale = Traits_::kFusedQuant == 1;
+    using DynamicQuantEpilogueProblem = ck_tile::DynamicQuantEpilogueProblem<ComputeDataType, SmoothScaleDataType, YScaleDataType, YDataType, typename Traits_::Shape,
+            ck_tile::DynamicQuantEpilogueTraits<false, Traits_::kPadN, UseSmoothInputScale, false,  true/*max3*/>>;
+
+    using DynamicQuantEpilogue = ck_tile::DynamicQuantEpilogue<DynamicQuantEpilogueProblem>;
+
+    using Default2DAndDynamicQuantEpilogueProblem = ck_tile::Default2DAndDynamicQuantEpilogueProblem<
+        ComputeDataType, SmoothScaleDataType, YScaleDataType, YDataType, UnquantYDataType, typename Traits_::Shape,
+        ck_tile::Default2DAndDynamicQuantEpilogueTraits<false, Traits_::kPadN, UseSmoothInputScale, false,  true/*max3*/>>;
+    using Default2DAndDynamicQuantEpilogue = ck_tile::Default2DAndDynamicQuantEpilogue<Default2DAndDynamicQuantEpilogueProblem>;
+
+    using Epilogue = std::conditional_t<Traits_::kFusedQuant != 0,
+                                        std::conditional_t<Traits_::kSaveUnquant,
+                                                           Default2DAndDynamicQuantEpilogue,
+                                                           DynamicQuantEpilogue>,
+                                        Default2DEpilogue>;
+
+    using Kernel = ck_tile::Rmsnorm2dFwd<Pipeline, Epilogue>;
+
+    const dim3 grids                       = Kernel::GridSize(a);
+    constexpr dim3 blocks                  = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+
+    auto kargs = Kernel::MakeKargs(a);
+    if(s.log_level_ > 0)
+        std::cout << ", " << Kernel::GetName() << std::flush;
+
+    return ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{{}}, grids, blocks, 0, kargs));
+}}
+
+"""
+
+    API_BASE = """
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "rmsnorm2d_fwd.hpp"
+
+{F_traits_define}
+
+// Note: this internal API only declare, not define here, otherwise will block `make -j`
+template <typename Traits_>
+float rmsnorm2d_fwd_(const ck_tile::stream_config& s, rmsnorm2d_fwd_args a);
+
+float rmsnorm2d_fwd(rmsnorm2d_fwd_traits t,
+                    rmsnorm2d_fwd_args a,
+                    const ck_tile::stream_config& s)
+{{
+    float r = -1;
+{F_dispatch}
+    return r;
+}}
+
+"""
+
+    INSTANCE_BASE = """
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "rmsnorm2d_fwd_api_common.hpp"
+
+// clang-format off
+//                                                                                    rm  rn  tm    tn  vn    pd     rms     2p
+{F_instance_def}
+// clang-format on
+
+"""
+
+    API_PER_DTYPE = """
+    {F_if}(t.prec_i == \"{F_i_type}\" && t.prec_o == \"{F_o_type}\"){{
+{F_per_n_case}
+    }}
+"""
+    API_PER_N_CASE = """
+        {F_if} {F_N_COND} {{
+{F_inner_dispatch}
+        }}
+"""
+    API_INNER_CASE = """
+            {F_if} {F_VEC_COND}
+                r={F_instance_func}(s, a);
+"""
+
+    def __init__(self, working_path, kernel_filter):
+        self.working_path = working_path
+        self.kernel_filter = kernel_filter
+
+    class k_fuesd_add_enum(IntEnum):
+        F_NO_ADD = 0
+        F_PRE_ADD = 1
+        F_PRE_ADD_STORE_RESIDUAL = 2
+
+    class k_fused_sweep_enum(IntEnum):
+        F_NO_SWEEP = 0
+        F_RENORM = 1
+        F_DYNAMIC_QUANT = 2
+
+    @dataclass
+    class k_traits:
+        F_kPadN : bool
+        F_kSaveMeanInvStd : bool
+        F_kTwoPass : bool
+        F_kFusedAdd : Any
+        F_kFusedQuant : Any
+
+    @dataclass
+    class k_shape:
+        F_BlockTile    : List[int]
+        F_WarpPerBlock : List[int]
+        F_WarpTile     : List[int]
+        F_Vector_      : List[int]
+        @property
+        def F_BlockSize(self) -> int:
+            return functools.reduce(lambda a, b: a*b, self.F_WarpTile)
+
+    @dataclass
+    class k_problem:
+        F_XDataType       : str
+        F_GammaDataType   : str
+        F_ComputeDataType : str
+        F_YDataType       : str
+        F_InvRmsDataType  : str
+        F_BlockShape      : str
+        F_Traits          : Any #k_traits
+
+    @dataclass
+    class k_pipeline_one_pass:
+        F_Problem         : Any #k_problem
+
+    @dataclass
+    class k_pipeline_two_pass:
+        F_Problem         : Any #k_problem
+
+    @dataclass
+    class default_2d_epilogue_problem:
+        F_AccDataType : str
+        F_ODataType : str
+        F_kPadM : bool
+        F_kPadN : bool
+
+    @dataclass
+    class default_2d_epilogue:
+        F_problem : Any
+
+    @dataclass
+    class k_kernel:
+        F_pipeline : Any
+        F_epilogue : Any
+
+    @dataclass
+    class h_traits:
+        F_XDataType : str
+        F_YDataType : str
+        F_SmoothScaleDataType : str
+        F_YScaleDataType : str
+        F_UnquantYDataType : str
+        F_Repeat_M : int
+        F_Repeat_N : int
+        F_ThreadPerBlock_M : int
+        F_ThreadPerBlock_N : int
+        F_Vector_N : int
+        F_kPadN : bool
+        F_kSaveInvRms : bool
+        F_kSaveUnquant: bool
+        F_kTwoPass : bool
+        F_kFusedAdd : int
+        F_kFusedQuant : int
+
+        @property
+        def trait_name(self) ->str:
+            t_ = f'{DATA_TYPE_MAP[self.F_XDataType]}, {DATA_TYPE_MAP[self.F_YDataType]}, {DATA_TYPE_MAP[self.F_SmoothScaleDataType]}, {DATA_TYPE_MAP[self.F_YScaleDataType]}, {DATA_TYPE_MAP[self.F_UnquantYDataType]}, {self.F_Repeat_M:2}, {self.F_Repeat_N:2}, {self.F_ThreadPerBlock_M:2}, {self.F_ThreadPerBlock_N:4}'
+            t_ += f', {self.F_Vector_N:2}, {BOOL_MAP(self.F_kPadN):5}, {BOOL_MAP(self.F_kSaveInvRms):5}, {BOOL_MAP(self.F_kSaveUnquant):5}'
+            t_ += f', {BOOL_MAP(self.F_kTwoPass):5}, {self.F_kFusedAdd:4}, {self.F_kFusedQuant:4}'
+            return t_
+
+        # string when calling this kernel
+        @property
+        def call_name(self) -> str:
+            return f'rmsnorm2d_fwd_<traits_<{self.trait_name}>>'
+
+        # string when define this kernel
+        @property
+        def def_name(self) -> str:
+            return f'template float rmsnorm2d_fwd_<traits_<{self.trait_name}>>(const S&, A);'
+
+    # this class hold kernel under same source file
+    @dataclass
+    class h_instance:
+        F_DataTypePair : str
+        F_N : str
+        F_add : int
+        F_sweep : int
+        F_saveunquant : bool
+        instance_list : List[Any] # List[h_traits]
+
+        @property
+        def name(self) -> str:
+            prec_i, prec_o = self.F_DataTypePair.split(',')
+            dtype_str = f'{prec_i}' if prec_i == prec_o else f'{prec_i}_{prec_o}'
+            nnn = f'rmsnorm2d_fwd_{dtype_str}_n{self.F_N}'
+            if self.F_add != 0:
+                nnn = nnn + '_' + FUSED_ADD_ENUM_STR_MAP[self.F_add]
+            if self.F_sweep != 0:
+                nnn = nnn + '_' + FUSED_FUSED_SWEEP_STR_MAP[self.F_sweep]
+            if self.F_saveunquant:
+                nnn = nnn + '_saveunquant'
+            return nnn
+
+        @property
+        def instance_name(self) ->str:
+            return self.name
+
+        @property
+        def content(self) ->str:
+            instance_defs = ''
+            for ins in self.instance_list:
+                instance_defs += ins.def_name + '\n'
+            return rmsnorm_fwd_codegen.INSTANCE_BASE.format(F_instance_def=instance_defs)
+
+    @property
+    def name_api(self) -> str:
+        return 'rmsnorm2d_fwd_api'
+
+    @property
+    def name_common_header(self) -> str:
+        return 'rmsnorm2d_fwd_api_common'
+
+    @property
+    def content_api(self) -> str:
+        # 1 sort based on dtype
+        t_dtype_dict = dict()
+        blobs = self.get_blobs()
+        for blob in blobs:
+            if blob.F_DataTypePair not in t_dtype_dict:
+                t_dtype_dict[blob.F_DataTypePair] = {}
+            if blob.F_N not in t_dtype_dict[blob.F_DataTypePair]:
+                t_dtype_dict[blob.F_DataTypePair][blob.F_N] = []
+            t_dtype_dict[blob.F_DataTypePair][blob.F_N].append(blob)
+
+        d_str = ''
+        for i_d, dtype_ in enumerate(t_dtype_dict):
+            blob_per_t = t_dtype_dict[dtype_]
+            n_str = ''
+            for i_n, n_ in enumerate(blob_per_t):
+                blob_per_n = blob_per_t[n_]
+                inner_str = ""
+                for i_b, b_ in enumerate(blob_per_n):
+                    # generate single kernel instance file
+                    #vec_str = ""
+                    for i_ins, ins in enumerate(b_.instance_list):
+                        idx_in_n = i_b * len(b_.instance_list) + i_ins
+                        len_in_n = len(blob_per_n) * len(b_.instance_list)
+                        # _if = 'if' if i_ins == 0 else 'else if'
+                        if ins.F_kFusedQuant == 0:
+                            _sweep_cond = 't.fused_quant == {f_fused_sweep}'.format(f_fused_sweep = ins.F_kFusedQuant)
+                        elif ins.F_kFusedQuant == 1:
+                            _sweep_cond = 't.fused_quant == {f_fused_sweep} && (t.prec_sm == \"{f_sx_type}\" && t.prec_sy == \"{f_sy_type}\" && t.save_unquant == {f_suq})'.format(
+                                f_fused_sweep = ins.F_kFusedQuant, f_sx_type=ins.F_SmoothScaleDataType, f_sy_type=ins.F_YScaleDataType, f_suq=BOOL_MAP(ins.F_kSaveUnquant))
+                        elif ins.F_kFusedQuant == 2:
+                            _sweep_cond = 't.fused_quant == {f_fused_sweep} && (t.prec_sy == \"{f_sy_type}\" && t.save_unquant == {f_suq})'.format(
+                                f_fused_sweep = ins.F_kFusedQuant, f_sy_type=ins.F_YScaleDataType, f_suq=BOOL_MAP(ins.F_kSaveUnquant))
+                        _cond = '((a.n % {f_vec_n} == 0) && (t.fused_add == {f_fused_add}) && ({f_sweep_cond}))'.format(
+                                        f_vec_n = ins.F_Vector_N, f_fused_add = ins.F_kFusedAdd,
+                                        f_sweep_cond = _sweep_cond)
+                        inner_str += self.API_INNER_CASE.format(F_if = get_if_str(idx_in_n, len_in_n, False),
+                                            F_VEC_COND = _cond, F_instance_func=ins.call_name)
+                    #inner_str = inner_str + vec_str
+                n_cnd = f'(a.n <= {n_})' if (i_n < len(blob_per_t) - 1) else ''
+                n_str += self.API_PER_N_CASE.format(F_if = get_if_str(i_n, len(blob_per_t)), F_N_COND=n_cnd, F_inner_dispatch=inner_str)
+            prec_i, prec_o = dtype_.split(',')
+            d_str += self.API_PER_DTYPE.format(F_if = get_if_str(i_d, len(t_dtype_dict), False), F_i_type=prec_i, F_o_type=prec_o, F_per_n_case=n_str)
+
+        api_base = self.API_BASE.format(F_traits_define=self.API_TRAITS_DEFINE, F_dispatch=d_str)
+        return api_base
+
+    @property
+    def content_common_header(self) -> str:
+        return self.API_COMMON_HEADER.format(F_traits_define=self.API_TRAITS_DEFINE)
+
+    def get_blobs(self):
+        h_traits = rmsnorm_fwd_codegen.h_traits
+        h_instance = rmsnorm_fwd_codegen.h_instance
+
+        dynamic_quant_out_dtype = ['int8', 'fp8']
+        # some predefined support range
+        # (prec_i,prec_o) for simplicity this string will be used as key for dict
+        scale_list = [('fp32,fp32')]
+        dtype_list = [('fp16,fp16'), ('bf16,bf16'),
+                        ('fp16,int8'), ('bf16,int8'),
+                        ('fp16,fp8'), ('bf16,fp8')] # NOTE: only fused-dynamic-quant use int8 out
+        #fused_add_list = [0, 1, 2]
+        #fused_sweep_list = [0, 1, 2] # NOTE: only single pass can use fused (smooth) dynamic quant
+        fused_add_list = [0, 1]
+        fused_sweep_list = [0, 1, 2] # NOTE: only single pass can use fused (smooth) dynamic quant
+        bool_list = [False, True]
+
+        #                                                              rm  rn  tm   tn  vn  pd     mv     unquant  2p     add    sweep
+        h_trait_dict = {'64'  : [ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  1,  8,  8,  8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  1,  4,  16, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  1,  4,  64, 1,  True,  False, False,   False,   0,    0)],
+                        '128' : [ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  1,  4,  16, 8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  1,  4,  64, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  2,  4,  64, 1,  True,  False, False,   False,   0,    0)],
+                        '256' : [ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  1,  4,  64, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  2,  4,  64, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  4,  64, 1,  True,  False, False,   False,   0,    0)],
+                        '512' : [ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  1,  4,  64, 8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  2,  4,  64, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  4,  64, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  8,  4,  64, 1,  True,  False, False,   False,   0,    0)],
+                        '640' : [ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  5,  4,  64, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  5,  4, 128, 1,  True,  False, False,   False,   0,    0)],
+                        '768' : [ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  4,  64, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  6,  4,  64, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 12,  4,  64, 1,  True,  False, False,   False,   0,    0)],
+                        '1024' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  2,  2,  64, 8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  2,  64, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  8,  2,  64, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  1, 256, 1,  True,  False, False,   False,   0,    0)],
+                        '1536' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  4,  64, 8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  2, 128, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  1, 256, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  6,  1, 256, 1,  True,  False, False,   False,   0,    0)],
+                        '2048' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  1,  1, 256, 8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  2,  1, 256, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  1, 256, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  8,  1, 256, 1,  True,  False, False,   False,   0,    0)],
+                        '3072' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  1, 128, 8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  1, 256, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  6,  1, 256, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  1,1024, 1,  True,  False, False,   False,   0,    0)],
+                        '4096' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  2,  1, 256, 8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  1, 256, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  2,  1,1024, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  1,1024, 1,  True,  False, False,   False,   0,    0)],
+                        '6144' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  1, 256, 8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  1, 512, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  3,  1,1024, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  6,  1,1024, 1,  True,  False, False,   False,   0,    0)],
+                        '8192' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  1, 256, 8,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  1, 512, 4,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  1,1024, 2,  True,  False, False,   False,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  8,  1,1024, 1,  True,  False, False,   False,   0,    0)],
+                        'big'  :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  1,  1,1024, 8,  True,  False, False,    True,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  1, 256, 4,  True,  False, False,    True,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 12,  1, 256, 2,  True,  False, False,    True,   0,    0),
+                                  h_traits('x', 'y', 'xs', 'ys', 'uqy', 1,  4,  1,1024, 1,  True,  False, False,    True,   0,    0)]}
+        total_blob = list()
+        for hs_key in h_trait_dict:
+            hs = h_trait_dict[hs_key]
+            current_n = hs[0].F_Repeat_N * hs[0].F_ThreadPerBlock_N * hs[0].F_Vector_N
+            for dtype, scale_type, fused_add, fused_quant, save_unquant in itertools.product(dtype_list, scale_list, fused_add_list, fused_sweep_list, bool_list):
+                prec_i, prec_o = dtype.split(',')
+                scale_sm, scale_y = scale_type.split(',')
+                if prec_o in dynamic_quant_out_dtype and fused_quant != 1 and fused_quant != 2:
+                    continue # skip non dynamic quant case
+                if (fused_quant == 1 or fused_quant == 2) and hs_key == 'big':
+                    continue
+                if (fused_quant == 0 and save_unquant == True):
+                    continue # save_unquant should always be false when there is no quant enabled
+                current_hs = list()
+                for chs_ in hs:
+                    h_ = copy.copy(chs_) # copy the base instance out
+                    h_.F_XDataType = prec_i
+                    h_.F_YDataType = prec_o
+                    h_.F_SmoothScaleDataType = scale_sm
+                    h_.F_YScaleDataType = scale_y
+                    h_.F_UnquantYDataType = prec_i
+                    h_.F_kFusedAdd = fused_add
+                    h_.F_kFusedQuant = fused_quant
+                    h_.F_kSaveUnquant = save_unquant
+                    current_hs.append(h_) # + "\n"
+                #f.write(str(f.parent / GEN_DIR / (blobs.api_common_header_
+                current_n_str = 'big' if hs_key == 'big' else current_n
+                total_blob.append(h_instance(dtype, current_n_str, fused_add, fused_quant, save_unquant, current_hs))
+        return total_blob
+
+    def list_blobs(self) -> None:
+        w_p = Path(self.working_path)
+        list_p = w_p / 'rmsnorm2d_fwd_blobs.txt'
+        blobs = self.get_blobs()
+        with list_p.open('w') as list_f:
+            # api related file
+            list_f.write(str(w_p / (self.name_api + ".cpp"))  + "\n")
+            list_f.write(str(w_p / (self.name_common_header + ".hpp"))  + "\n")
+            # kernel instance file
+            for b in blobs:
+                list_f.write(str(w_p / (b.name + ".cpp")) + "\n")
+
+    def gen_blobs(self) -> None:
+        w_p = Path(self.working_path)
+        (w_p / (self.name_api + ".cpp")).write_text(self.content_api)
+        (w_p / (self.name_common_header + ".hpp")).write_text(self.content_common_header)
+        blobs = self.get_blobs()
+        for b in blobs:
+            (w_p / (b.name + ".cpp")).write_text(b.content)
+
+
+def list_blobs(args):
+    api_list = args.api.split(',')
+    for api in api_list:
+        if api == 'fwd':
+            rmsnorm_fwd_codegen(args.working_path, args.filter).list_blobs()
+
+
+def gen_blobs(args):
+    api_list = args.api.split(',')
+    for api in api_list:
+        if api == 'fwd':
+            rmsnorm_fwd_codegen(args.working_path, args.filter).gen_blobs()
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(
+        prog="generate",
+        description="gen API for CK rmsnorm kernel",
+    )
+    parser.add_argument(
+        "-a",
+        "--api",
+        default='fwd[all]',
+        required=False,
+        help="supply API(s) to generate (default: fwd). separated by comma."
+    )
+
+    # the directory for list_blobs/gen_blobs to write files into
+    parser.add_argument(
+        "-w",
+        "--working_path",
+        default="./",
+        required=False,
+        help="the path where all the blobs are going to be generated"
+    )
+
+    # this script have 2 modes
+    # 1) list_blobs mode, will generate a txt file with all the files going to be generated.
+    #    this is useful in build system like cmake to construct source code dependency, by
+    #    reading the content out of this file
+    # 2) gen_blobs mode, will generate the actuall kernel instance and api. If in framework
+    #    like FA, only need to use this mode
+    parser.add_argument(
+        "-l",
+        "--list_blobs",
+        action='store_true',
+        help="list all the kernels to a file, "
+    )
+
+    parser.add_argument(
+        "-g",
+        "--gen_blobs",
+        action='store_true',
+        help="generate all kernels into different tile"
+    )
+
+    # TODO: if using filter, must apply same value to output_dir and list_blobs
+    parser.add_argument(
+        "-f",
+        "--filter",
+        required=False,
+        help="filter out kernels that need to generate, using fnmatch module"
+    )
+
+    parser.add_argument(
+        "-t",
+        "--traits",
+        default="all",
+        required=False,
+        help="enable/disable some feature. default generate all"
+    )
+
+    parser.add_argument(
+        "-r",
+        "--receipt",
+        default=0,
+        required=False,
+        help="codegen receipt."
+    )
+
+    args = parser.parse_args()
+
+    # print(f'{args.list_blobs}-{args.gen_blobs}')
+    if (args.gen_blobs and args.list_blobs) or ((not args.gen_blobs) and (not args.list_blobs)):
+        print('gen_blobs/list_blobs must specify only one option')
+        sys.exit()
+
+    p = Path(args.working_path)
+    if not p.exists():
+        p.mkdir()
+
+    if args.list_blobs:
+        list_blobs(args)
+    else:
+        gen_blobs(args)

example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.cpp

@@ -0,0 +1,536 @@
+#include "ck_tile/host.hpp"
+#include "rmsnorm2d_fwd.hpp"
+#include <cstring>
+
+// different threshold for different dtype
+template <typename DataType>
+auto get_elimit()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::bf16_t>()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::int8_t>()
+{
+    double rtol = 1e-02;
+    double atol = 1.0;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("m", "3328", "m dimension")
+        .insert("n", "4096", "n dimension")
+        .insert("x_stride", "-1", "x row_stride, if -1 then equal to n")
+        .insert("xr_stride", "-1", "x residule row_stride, if -1 then equal to n")
+        .insert("y_stride", "-1", "y row_stride, if -1 then equal to n")
+        .insert("yr_stride", "-1", "y residule row_stride, if -1 then equal to n")
+        .insert("e", "1e-5", "epsilon")
+        .insert("save_rms", "0", "save rms(invrms) or not. set to 1 in training case")
+        .insert("save_unquant", "0", "save result before quant")
+        .insert("v", "1", "cpu validation or not")
+        .insert("kname", "1", "print kernel name or not")
+        .insert("prec_i", "fp16", "input precision")
+        .insert("prec_o", "auto", "output precision, set auto will be the same as input")
+        .insert("prec_sm",
+                "auto",
+                "output quant scale type, set auto will use fp32. used when fquant=1")
+        .insert("prec_sy",
+                "auto",
+                "output quant scale type, set auto will use fp32. used when fquant=1 or 2")
+        .insert("fadd", "0", "fused-add, 0:no fused add, 1:preadd+store, 2:preadd only")
+        .insert("fquant", "0", "fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant")
+        .insert("warmup", "5", "cold iter")
+        .insert("repeat", "20", "hot iter");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename InDataType,
+          typename OutDataType,
+          typename SmoothScaleDataType,
+          typename YScaleDataType,
+          bool SaveRms,
+          bool SaveUnquant>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    ck_tile::index_t m = arg_parser.get_int("m");
+    ck_tile::index_t n = arg_parser.get_int("n");
+    float epsilon      = arg_parser.get_float("e");
+    int kname          = arg_parser.get_int("kname");
+    int do_validation  = arg_parser.get_int("v");
+    int fused_add      = arg_parser.get_int("fadd");
+    int fused_quant    = arg_parser.get_int("fquant");
+    int warmup         = arg_parser.get_int("warmup");
+    int repeat         = arg_parser.get_int("repeat");
+
+    ck_tile::index_t x_stride = arg_parser.get_int("x_stride");
+    if(x_stride < 0)
+        x_stride = n;
+    ck_tile::index_t xr_stride = arg_parser.get_int("xr_stride");
+    if(xr_stride < 0)
+        xr_stride = n;
+    ck_tile::index_t y_stride = arg_parser.get_int("y_stride");
+    if(y_stride < 0)
+        y_stride = n;
+    ck_tile::index_t yr_stride = arg_parser.get_int("yr_stride");
+    if(yr_stride < 0)
+        yr_stride = n;
+    assert(x_stride >= n);
+
+    std::string prec_i  = arg_parser.get_str("prec_i");
+    std::string prec_o  = arg_parser.get_str("prec_o");
+    std::string prec_sm = arg_parser.get_str("prec_sm");
+    std::string prec_sy = arg_parser.get_str("prec_sy");
+    if(prec_o == "auto")
+    {
+        prec_o = prec_i;
+    }
+    if(prec_sm == "auto")
+    {
+        prec_sm = "fp32";
+    }
+    if(prec_sy == "auto")
+    {
+        prec_sy = "fp32";
+    }
+
+    if((fused_quant == 1 || fused_quant == 2) && prec_o != "int8" && prec_o != "fp8")
+    {
+        std::cout
+            << "if fused_quant is 1 or 2, only support \"-prec_o=int8\" or \"-prec_o=fp8\" cases."
+            << std::endl;
+        return false;
+    }
+
+    if((fused_quant == 0) && SaveUnquant)
+    {
+        std::cout
+            << "save_unquant should be 0 if quant output is not enabled because it is meaningless. "
+            << "Output Y is what wanted." << std::endl;
+        return false;
+    }
+
+    using TypeConfig =
+        RmsnormTypeConfig<InDataType, OutDataType, SmoothScaleDataType, YScaleDataType>;
+
+    using XDataType         = typename TypeConfig::XDataType;
+    using YDataType         = typename TypeConfig::YDataType;
+    using GammaDataType     = typename TypeConfig::GammaDataType;
+    using XResidualDataType = XDataType;
+    using YResidualDataType = XDataType;
+
+    using InvRmsDataType =
+        std::conditional_t<SaveRms, typename TypeConfig::InvRmsDataType, ck_tile::null_type>;
+    using UnquantYDataType =
+        std::conditional_t<SaveUnquant, typename TypeConfig::UnquantYDataType, ck_tile::null_type>;
+
+    using ComputeDataType = typename TypeConfig::ComputeDataType;
+
+    // host verify
+    ck_tile::HostTensor<XDataType> x_host({m, n}, {x_stride, 1});
+    ck_tile::HostTensor<GammaDataType> gamma_host({n});
+    ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host({n});
+    ck_tile::HostTensor<SmoothScaleDataType> sm_scale_host_dev({n});
+
+    ck_tile::HostTensor<XResidualDataType> x_residual_host({m, n}, {xr_stride, 1});
+    ck_tile::HostTensor<YResidualDataType> y_residual_host({m, n}, {yr_stride, 1});
+
+    ck_tile::HostTensor<YDataType> y_host_ref({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<YDataType> y_host_dev({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<YScaleDataType> y_scale_host_ref({m});
+    ck_tile::HostTensor<YScaleDataType> y_scale_host_dev({m});
+
+    ck_tile::HostTensor<InvRmsDataType> invRms_host_ref({m});
+
+    ck_tile::HostTensor<UnquantYDataType> unquant_y_host_ref({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<UnquantYDataType> unquant_y_host_dev({m, n}, {y_stride, 1});
+    ck_tile::HostTensor<ck_tile::null_type> unquant_y_null({1});
+
+    ck_tile::FillUniformDistribution<XDataType>{-.5f, .5f}(x_host);
+    ck_tile::FillUniformDistribution<XResidualDataType>{-.5f, .5f}(x_residual_host);
+    ck_tile::FillUniformDistribution<SmoothScaleDataType>{-1.f, 1.f}(sm_scale_host);
+    ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host);
+
+    ck_tile::DeviceMem x_buf(x_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem gamma_buf(gamma_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_buf(y_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_scale_buf(y_scale_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem sm_scale_buf(sm_scale_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem x_residual_buf(x_residual_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem y_residual_buf(y_residual_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem unquant_y_buf(unquant_y_host_dev.get_element_space_size_in_bytes());
+
+    x_buf.ToDevice(x_host.data());
+    gamma_buf.ToDevice(gamma_host.data());
+    x_residual_buf.ToDevice(x_residual_host.data());
+    sm_scale_buf.ToDevice(sm_scale_host.data());
+
+    auto prec_str = [&]() {
+        auto base_str = prec_i;
+        if(prec_i != prec_o)
+        {
+            base_str += "|" + prec_o;
+        }
+        if(fused_quant == 1)
+        {
+            base_str += std::string("(") + prec_sy + ")";
+        }
+        return base_str;
+    }();
+
+    std::cout << "[" << prec_str << "]"
+              << " m:" << m << ", n:" << n << ", x_stride:" << x_stride
+              << ", xr_stride:" << xr_stride << ", y_stride:" << y_stride
+              << ", yr_stride:" << yr_stride << std::flush;
+
+    rmsnorm2d_fwd_traits traits{
+        prec_i, prec_o, prec_sm, prec_sy, SaveRms, SaveUnquant, fused_add, fused_quant};
+
+    rmsnorm2d_fwd_args args{x_buf.GetDeviceBuffer(),
+                            fused_add != 0 ? x_residual_buf.GetDeviceBuffer() : nullptr,
+                            fused_quant == 1 ? sm_scale_buf.GetDeviceBuffer() : nullptr,
+                            gamma_buf.GetDeviceBuffer(),
+                            y_buf.GetDeviceBuffer(),
+                            fused_add == 1 ? y_residual_buf.GetDeviceBuffer() : nullptr,
+                            fused_quant != 0 ? y_scale_buf.GetDeviceBuffer() : nullptr,
+                            nullptr, // p_invRms, unsupported yet
+                            SaveUnquant ? unquant_y_buf.GetDeviceBuffer() : nullptr,
+                            epsilon,
+                            m,
+                            n,
+                            x_stride,   // x row_stride
+                            xr_stride,  // x residule row stride
+                            y_stride,   // y row stride
+                            yr_stride}; // y residule row stride
+
+    float ave_time = rmsnorm2d_fwd(
+        traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
+
+    std::size_t num_byte =
+        sizeof(XDataType) * m * n + sizeof(GammaDataType) * n + sizeof(YDataType) * m * n;
+    num_byte += SaveRms ? sizeof(InvRmsDataType) * m * n : 0;
+    num_byte += SaveUnquant ? sizeof(UnquantYDataType) * m * n : 0;
+    num_byte += fused_add ? sizeof(XResidualDataType) * m * n : 0;
+    num_byte += ((fused_quant == 1) || (fused_quant == 2)) ? sizeof(YScaleDataType) * m : 0;
+    num_byte += (fused_quant == 1) ? sizeof(SmoothScaleDataType) * n : 0;
+
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+    std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::flush;
+
+    bool pass = true;
+
+    if(do_validation)
+    {
+        // reference
+        if(fused_add != 0)
+        {
+            // fused pre_add/pre_add_store
+            // TODO we accumulate directly to x_host for simplcity here...
+            std::transform(x_host.mData.cbegin(),
+                           x_host.mData.cend(),
+                           x_residual_host.mData.cbegin(),
+                           x_host.mData.begin(),
+                           [](auto x_, auto r_) {
+                               auto o_ = ck_tile::type_convert<ComputeDataType>(x_) +
+                                         ck_tile::type_convert<ComputeDataType>(r_);
+                               return ck_tile::type_convert<XDataType>(o_);
+                           });
+        }
+
+        if(fused_quant != 0)
+        {
+            auto dquant_functor = [&](int m_, auto& o_, auto& acc_) {
+                int N_ = acc_.mDesc.get_lengths()[1];
+                if(fused_quant == 1)
+                {
+                    for(int n_ = 0; n_ < N_; n_++)
+                    {
+                        // input smooth outlier
+                        acc_(m_, n_) = acc_(m_, n_) *
+                                       ck_tile::type_convert<ComputeDataType>(sm_scale_host(n_));
+                    }
+                }
+                ComputeDataType absmax = static_cast<ComputeDataType>(0);
+                for(int n_ = 0; n_ < N_; n_++)
+                {
+                    const auto a = ck_tile::abs(acc_(m_, n_));
+                    absmax       = a > absmax ? a : absmax;
+                }
+                // printf("cpu:absmax:%f\n", absmax);
+                constexpr ComputeDataType kMaxY =
+                    std::is_same<YDataType, ck_tile::fp8_t>::value    ? 240.0
+                    : std::is_same<YDataType, ck_tile::int8_t>::value ? 127.0
+                                                                      : 0.0;
+                ComputeDataType y_scale = absmax / kMaxY;
+                y_scale_host_ref(m_)    = ck_tile::type_convert<YScaleDataType>(y_scale);
+                for(int n_ = 0; n_ < N_; n_++)
+                {
+                    o_(m_, n_) = ck_tile::type_convert<YDataType>(acc_(m_, n_) / y_scale);
+                }
+            };
+
+            auto default_and_dquant_functor = [&](int m_, auto& o_unquant_, auto& o_, auto& acc_) {
+                const int N = acc_.mDesc.get_lengths()[1];
+                for(int n_ = 0; n_ < N; ++n_)
+                {
+                    o_unquant_(m_, n_) = ck_tile::type_convert<OutDataType>(acc_(m_, n_));
+                }
+
+                dquant_functor(m_, o_, acc_);
+            };
+
+            if constexpr(SaveUnquant)
+            {
+                ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                                 GammaDataType,
+                                                 ComputeDataType,
+                                                 YDataType,
+                                                 InvRmsDataType,
+                                                 UnquantYDataType>(x_host,
+                                                                   gamma_host,
+                                                                   y_host_ref,
+                                                                   invRms_host_ref,
+                                                                   unquant_y_host_ref,
+                                                                   epsilon,
+                                                                   default_and_dquant_functor);
+            }
+            else
+            {
+                ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                                 GammaDataType,
+                                                 ComputeDataType,
+                                                 YDataType,
+                                                 InvRmsDataType,
+                                                 UnquantYDataType>(x_host,
+                                                                   gamma_host,
+                                                                   y_host_ref,
+                                                                   invRms_host_ref,
+                                                                   unquant_y_host_ref,
+                                                                   epsilon,
+                                                                   dquant_functor);
+            }
+        }
+        else
+        {
+            assert(SaveUnquant == false);
+            ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                             GammaDataType,
+                                             ComputeDataType,
+                                             YDataType,
+                                             InvRmsDataType,
+                                             ck_tile::null_type>(
+                x_host, gamma_host, y_host_ref, invRms_host_ref, unquant_y_null, epsilon);
+        }
+
+        y_buf.FromDevice(y_host_dev.data());
+
+        ck_tile::HostTensor<YResidualDataType> y_residual_host_dev({m, n}, {yr_stride, 1});
+        if(fused_add == 1)
+        {
+            y_residual_buf.FromDevice(y_residual_host_dev.data());
+        }
+
+        auto [rtol, atol] = get_elimit<YDataType>();
+        if(x_stride == n)
+        {
+            pass = ck_tile::check_err(
+                y_host_dev, y_host_ref, std::string("\nOUT Error: Incorrect results!"), rtol, atol);
+
+            if constexpr(SaveUnquant)
+            {
+                pass &= ck_tile::check_err(unquant_y_host_dev,
+                                           unquant_y_host_ref,
+                                           std::string("\n OUT ERROR: Incorrect unquant results!"),
+                                           rtol,
+                                           atol);
+            }
+
+            if(fused_add == 1)
+            {
+                pass &= ck_tile::check_err(y_residual_host_dev,
+                                           x_host,
+                                           std::string("\nADD Error: Incorrect results!"),
+                                           rtol,
+                                           atol);
+            }
+        }
+        else
+        {
+            for(int i_r = 0; i_r < m; i_r++)
+            {
+                std::vector<YDataType> y_host_dev_row(y_host_dev.begin() + i_r * y_stride,
+                                                      y_host_dev.begin() + i_r * y_stride + n);
+                std::vector<YDataType> y_host_ref_row(y_host_ref.begin() + i_r * y_stride,
+                                                      y_host_ref.begin() + i_r * y_stride + n);
+                pass &= ck_tile::check_err(y_host_dev_row,
+                                           y_host_ref_row,
+                                           std::string("\nOUT[") + std::to_string(i_r) +
+                                               std::string("] Error: Incorrect results!"),
+                                           rtol,
+                                           atol);
+
+                if(fused_add == 1)
+                {
+                    std::vector<YResidualDataType> y_residual_host_dev_row(
+                        y_residual_host_dev.begin() + i_r * yr_stride,
+                        y_residual_host_dev.begin() + i_r * yr_stride + n);
+                    std::vector<YResidualDataType> y_residual_host_ref_row(
+                        x_host.begin() + i_r * yr_stride, x_host.begin() + i_r * yr_stride + n);
+                    pass &= ck_tile::check_err(y_residual_host_dev_row,
+                                               y_residual_host_ref_row,
+                                               std::string("\nADD[") + std::to_string(i_r) +
+                                                   std::string("] Error: Incorrect results!"),
+                                               rtol,
+                                               atol);
+                }
+
+                if constexpr(SaveUnquant)
+                {
+                    std::vector<UnquantYDataType> unquant_y_host_dev_row(
+                        unquant_y_host_dev.begin() + i_r * y_stride,
+                        unquant_y_host_dev.begin() + i_r * y_stride + n);
+                    std::vector<UnquantYDataType> unquant_y_host_ref_row(
+                        unquant_y_host_ref.begin() + i_r * y_stride,
+                        unquant_y_host_ref.begin() + i_r * y_stride + n);
+                    pass &=
+                        ck_tile::check_err(unquant_y_host_dev_row,
+                                           unquant_y_host_ref_row,
+                                           std::string("\nOUT[") + std::to_string(i_r) +
+                                               std::string("] Error: Incorrect unquant y results!"),
+                                           rtol,
+                                           atol);
+                }
+            }
+        }
+
+        if(fused_quant == 1)
+        {
+            y_scale_buf.FromDevice(y_scale_host_dev.data());
+            pass &= ck_tile::check_err(y_scale_host_dev,
+                                       y_scale_host_ref,
+                                       std::string("\nSCALE Error: Incorrect results!"),
+                                       rtol,
+                                       atol);
+        }
+
+        std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
+    }
+
+    return pass;
+}
+
+bool is_quant_data_type(const std::string& prec) { return (prec == "int8") || (prec == "fp8"); }
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    std::string prec_i  = arg_parser.get_str("prec_i");
+    std::string prec_o  = arg_parser.get_str("prec_o");
+    std::string prec_sm = arg_parser.get_str("prec_sm");
+    std::string prec_sy = arg_parser.get_str("prec_sy");
+    if(prec_o == "auto")
+    {
+        prec_o = prec_i;
+    }
+    if(prec_sm == "auto")
+    {
+        prec_sm = "fp32";
+    }
+    if(prec_sy == "auto")
+    {
+        prec_sy = "fp32";
+    }
+
+    int save_rms    = arg_parser.get_int("save_rms");
+    int fused_quant = arg_parser.get_int("fquant");
+    int save_unquant =
+        arg_parser.get_int("save_unquant") && is_quant_data_type(prec_o) && (fused_quant != 0);
+
+    if(prec_i == "fp16" && prec_o == "fp16" && prec_sm == "fp32" && prec_sy == "fp32" && save_rms)
+    {
+        return run<ck_tile::half_t, ck_tile::half_t, float, float, true, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "fp16" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms)
+    {
+        return run<ck_tile::half_t, ck_tile::half_t, float, float, false, false>(arg_parser) ? 0
+                                                                                             : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "bf16" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            save_rms)
+    {
+        return run<ck_tile::bf16_t, ck_tile::bf16_t, float, float, true, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "bf16" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms)
+    {
+        return run<ck_tile::bf16_t, ck_tile::bf16_t, float, float, false, false>(arg_parser) ? 0
+                                                                                             : -2;
+    }
+
+    // dynamic quant case, only in inference
+    else if(prec_i == "fp16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && !save_unquant)
+    {
+        return run<ck_tile::half_t, ck_tile::int8_t, float, float, true, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && !save_unquant)
+    {
+        return run<ck_tile::bf16_t, ck_tile::int8_t, float, float, true, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && !save_unquant)
+    {
+        return run<ck_tile::half_t, ck_tile::fp8_t, float, float, false, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && !save_unquant)
+    {
+        return run<ck_tile::bf16_t, ck_tile::fp8_t, float, float, false, false>(arg_parser) ? 0
+                                                                                            : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && save_unquant)
+    {
+        return run<ck_tile::half_t, ck_tile::int8_t, float, float, true, true>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "int8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && save_unquant)
+    {
+        return run<ck_tile::bf16_t, ck_tile::int8_t, float, float, true, true>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "fp16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && save_unquant)
+    {
+        return run<ck_tile::half_t, ck_tile::fp8_t, float, float, false, true>(arg_parser) ? 0 : -2;
+    }
+    else if(prec_i == "bf16" && prec_o == "fp8" && prec_sm == "fp32" && prec_sy == "fp32" &&
+            !save_rms && save_unquant)
+    {
+        return run<ck_tile::bf16_t, ck_tile::fp8_t, float, float, false, true>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/10_rmsnorm2d/rmsnorm2d_fwd.hpp

@@ -0,0 +1,69 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2025, 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/rmsnorm2d.hpp"
+#include <string>
+
+template <typename InType,
+          typename OutType,
+          typename SmoothScaleDataType_,
+          typename YScaleDataType_>
+struct RmsnormTypeConfig;
+
+template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
+struct RmsnormTypeConfig<ck_tile::half_t, OutType, SmoothScaleDataType_, YScaleDataType_>
+{
+    using XDataType           = ck_tile::half_t;
+    using YDataType           = OutType;
+    using GammaDataType       = ck_tile::half_t;
+    using InvRmsDataType      = ck_tile::half_t;
+    using UnquantYDataType    = ck_tile::half_t;
+    using ComputeDataType     = float;
+    using SmoothScaleDataType = SmoothScaleDataType_;
+    using YScaleDataType      = YScaleDataType_;
+};
+
+template <typename OutType, typename SmoothScaleDataType_, typename YScaleDataType_>
+struct RmsnormTypeConfig<ck_tile::bf16_t, OutType, SmoothScaleDataType_, YScaleDataType_>
+{
+    using XDataType           = ck_tile::bf16_t;
+    using YDataType           = OutType;
+    using GammaDataType       = ck_tile::bf16_t;
+    using InvRmsDataType      = ck_tile::bf16_t;
+    using UnquantYDataType    = ck_tile::bf16_t;
+    using ComputeDataType     = float;
+    using SmoothScaleDataType = SmoothScaleDataType_;
+    using YScaleDataType      = YScaleDataType_;
+};
+
+// runtime args
+struct rmsnorm2d_fwd_args : public ck_tile::Rmsnorm2dFwdHostArgs
+{
+};
+
+template <typename Traits_>
+float rmsnorm2d_fwd_(const ck_tile::stream_config& s, rmsnorm2d_fwd_args a);
+
+// This is the public API, will be generated by script
+struct rmsnorm2d_fwd_traits
+{
+    std::string prec_i; // input precision
+    std::string prec_o; // output precision
+
+    // if fused_quant == 1, need set prec_sm/prec_sy to proper string, otherwise can set
+    // arbitrary(will skip check) if fused_quant == 2, need set prec_sy to proper string, otherwise
+    // can set arbitrary(will skip check)
+    std::string prec_sm; // x-scale, used for [1*N] input smooth quant
+    std::string prec_sy; // y-scale, used for [M*1] output for next layer
+
+    bool save_rms;
+    bool save_unquant;
+    int fused_add;   // 0:no-add, 1:pre-add-store, 2:pre-add
+    int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant
+};
+
+float rmsnorm2d_fwd(rmsnorm2d_fwd_traits, rmsnorm2d_fwd_args, const ck_tile::stream_config&);

example/ck_tile/10_rmsnorm2d/script/perf_test.sh

@@ -0,0 +1,37 @@
+#!/bin/sh
+EXE="$(find . -name tile_rmsnorm2d_fwd -type f | head -n 1)"
+
+$EXE -m=1 -n=1 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=80 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec=bf16 -repeat=1000
+
+$EXE -m=700 -n=80 -e=1e-12 -v=1  -prec=fp16 -repeat=1000
+$EXE -m=700 -n=128 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=144 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=168 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=184 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=256 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=288 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=344 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=376 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=448 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=512 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=924 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=1024 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=1078 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=1996 -e=1e-12 -v=1 -prec=fp16 -repeat=1000
+$EXE -m=700 -n=4080 -e=1e-12 -v=1 -prec=fp16 -repeat=1000

example/ck_tile/10_rmsnorm2d/script/smoke_test.sh

@@ -0,0 +1,43 @@
+#!/bin/sh
+EXE="$(find . -name tile_rmsnorm2d_fwd -type f | head -n 1)"
+
+for fquant in "" "-fquant=1 -prec_o=int8" "-fquant=2 -prec_o=int8" "-fquant=1 -prec_o=fp8" "-fquant=2 -prec_o=fp8"\
+  "-fquant=1 -prec_o=int8 -save_unquant=1" "-fquant=2 -prec_o=int8 -save_unquant=1" "-fquant=1 -prec_o=fp8 -save_unquant=1" "-fquant=2 -prec_o=fp8 -save_unquant=1"; do
+for pr_i in "fp16" "bf16" ; do
+for fadd in "0" "1"; do
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=99  -n=13
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=17  -n=16
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=100
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=4   -n=128
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=80  -n=127
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=22  -n=255 -stride=256
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7   -n=599
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=19  -n=512
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=33  -n=313 -stride=1000
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=11  -n=510
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=171 -n=676 -stride=818
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=91  -n=636
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=12  -n=768 -stride=800
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=100 -n=766 -stride=812
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=31  -n=1024
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=64  -n=1000 -stride=1004
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=8   -n=1501
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=1826
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=5   -n=2040
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=7   -n=2734
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=3182
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=9   -n=4096
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=8192
+done
+done
+done
+
+# The following cases uses two pass pipeline which doesn't support quant epilogue. 
+for fquant in ""
+for pr_i in "fp16" "bf16" ; do
+for fadd in "0" "1"; do
+$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=1   -n=10547
+#$EXE -prec_i=$pr_i -fadd=$fadd $fquant -m=3   -n=17134
+done
+done
+done

example/ck_tile/11_add_rmsnorm2d_rdquant/CMakeLists.txt

@@ -0,0 +1,27 @@
+set(TILE_ADD_RMSNORM2D_RDQUANT_FWD "tile_add_rmsnorm2d_rdquant_fwd")
+# 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"
+message("adding ${TILE_ADD_RMSNORM2D_RDQUANT_FWD}")
+file(GLOB INSTANCE_SRCS instances/*.cpp)
+add_executable(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} add_rmsnorm2d_rdquant_fwd.cpp)
+rocm_install(TARGETS ${TILE_ADD_RMSNORM2D_RDQUANT_FWD} COMPONENT examples)
+target_include_directories(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${CMAKE_CURRENT_LIST_DIR})
+target_sources(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${INSTANCE_SRCS})
+
+set(TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS)
+
+# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations
+list(APPEND TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal)
+
+target_compile_options(${TILE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS})
+
+set(EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD "tile_example_add_rmsnorm2d_rdquant_fwd")
+add_executable(${EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD} example_add_rmsnorm2d_rdquant_fwd.cpp)
+rocm_install(TARGETS ${EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD} COMPONENT examples)
+target_compile_options(${EXAMPLE_ADD_RMSNORM2D_RDQUANT_FWD} PRIVATE ${TILE_ADD_RMSNORM2D_RDQUANT_FWD_COMPILE_OPTIONS})
+
+# TODO: we have to turn off this global prop, otherwise the progress bar generated
+# by cmake will print too many files, execvp: /bin/sh: Argument list too long
+# however, this property may affect global
+# TODO: consider codegen a makefile by us
+set_property(GLOBAL PROPERTY RULE_MESSAGES OFF)

example/ck_tile/11_add_rmsnorm2d_rdquant/README.md

@@ -0,0 +1,22 @@
+# Add + Rmsnorm2D + rowwise dynamic quantization forward
+
+This folder contains example for add + Rmsnorm2D + rowwise dynamic quantization forward using ck_tile tile-programming implementation. Rdquant is short for rowwise dynamic quantization here.
+
+## build
+```
+# 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...
+make tile_add_rmsnorm2d_rdquant_fwd -j
+```
+This will result in an executable `build/bin/tile_add_rmsnorm2d_rdquant_fwd`
+
+## cmdline
+```
+args:
+          -m    m dimension (default:3328)
+          -n    m dimension (default:4096)
+          -e    epsilon (default:1e-5)
+          -v    cpu validation or not (default:1)
+       -prec    precision (default:fp16)
+```

example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.cpp

@@ -0,0 +1,309 @@
+#include "ck_tile/host.hpp"
+#include "add_rmsnorm2d_rdquant_fwd.hpp"
+#include <cstring>
+
+// different threshold for different dtype
+template <typename InputDataType>
+auto get_elimit()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::bf16_t>()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::int8_t>()
+{
+    // due to rounding, int8 quantization might have 1 abs error
+    double rtol = 1;
+    double atol = 1;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("m", "3328", "m dimension")
+        .insert("n", "4096", "n dimension")
+        .insert("stride", "-1", "stride per row, if -1 then equal to n")
+        .insert("e", "1e-5", "epsilon")
+        .insert("save_x", "1", "save rms(invrms) or not. set to 1 in training case")
+        .insert("v", "1", "cpu validation or not")
+        .insert("kname", "1", "print kernel name or not")
+        .insert("prec", "fp16", "precision")
+        .insert("quant", "int8", "precision")
+        .insert("warmup", "5", "cold iter")
+        .insert("repeat", "20", "hot iter");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename InputDataType, typename QuantizedDataType, bool SaveX>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    ck_tile::index_t m      = arg_parser.get_int("m");
+    ck_tile::index_t n      = arg_parser.get_int("n");
+    ck_tile::index_t stride = arg_parser.get_int("stride");
+    if(stride < 0)
+        stride = n;
+    float epsilon                   = arg_parser.get_float("e");
+    std::string input_data_type     = arg_parser.get_str("prec");
+    std::string quantized_data_type = arg_parser.get_str("quant");
+    int kname                       = arg_parser.get_int("kname");
+    int do_validation               = arg_parser.get_int("v");
+    int warmup                      = arg_parser.get_int("warmup");
+    int repeat                      = arg_parser.get_int("repeat");
+
+    assert(stride >= n);
+
+    using TypeConfig = AddRmsnormRdquantTypeConfig<InputDataType, QuantizedDataType>;
+
+    using ADataType        = typename TypeConfig::ADataType;
+    using BDataType        = typename TypeConfig::BDataType;
+    using GammaDataType    = typename TypeConfig::GammaDataType;
+    using XDataType        = typename TypeConfig::XDataType;
+    using UnquantYDataType = ck_tile::null_type;
+    using YScaleDataType   = typename TypeConfig::YScaleDataType;
+    using QYDataType       = typename TypeConfig::QYDataType;
+    using ComputeDataType  = float;
+
+    // host verify
+    ck_tile::HostTensor<ADataType> a_host({m, n}, {stride, 1});
+    ck_tile::HostTensor<BDataType> b_host({m, n}, {stride, 1});
+    ck_tile::HostTensor<GammaDataType> gamma_host({n});
+
+    ck_tile::HostTensor<XDataType> x_host_ref({m, n}, {stride, 1});
+    ck_tile::HostTensor<XDataType> x_host_dev({m, n}, {stride, 1});
+
+    ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
+    ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
+
+    ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {stride, 1});
+    ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {stride, 1});
+    ck_tile::HostTensor<UnquantYDataType> unquant_y_host_ref({m, n}, {stride, 1});
+
+    ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_host);
+    ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_host);
+    ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host);
+
+    ck_tile::DeviceMem a_buf(a_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem b_buf(b_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem gamma_buf(gamma_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem x_buf(x_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
+
+    a_buf.ToDevice(a_host.data());
+    b_buf.ToDevice(b_host.data());
+    gamma_buf.ToDevice(gamma_host.data());
+
+    std::cout << "[" << input_data_type << ", " << quantized_data_type << "]"
+              << " m:" << m << ", n:" << n << ", stride:" << stride << std::flush;
+
+    add_rmsnorm2d_rdquant_fwd_traits traits{input_data_type, quantized_data_type, SaveX};
+
+    add_rmsnorm2d_rdquant_fwd_args args{a_buf.GetDeviceBuffer(),
+                                        b_buf.GetDeviceBuffer(),
+                                        gamma_buf.GetDeviceBuffer(),
+                                        x_buf.GetDeviceBuffer(),
+                                        yscale_buf.GetDeviceBuffer(),
+                                        qy_buf.GetDeviceBuffer(),
+                                        epsilon,
+                                        m,
+                                        n,
+                                        stride};
+
+    float ave_time = add_rmsnorm2d_rdquant_fwd(
+        traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat});
+
+    std::size_t num_byte = sizeof(ADataType) * m * n + sizeof(BDataType) * m * n +
+                           sizeof(GammaDataType) * n + sizeof(YScaleDataType) * m +
+                           sizeof(QYDataType) * m * n;
+
+    if constexpr(SaveX)
+        num_byte += sizeof(XDataType) * m * n;
+
+    float gb_per_sec = num_byte / 1.E6 / ave_time;
+    std::cout << ", " << ave_time * 1.E3 << " us, " << gb_per_sec << " GB/s" << std::endl;
+
+    bool pass = true;
+
+    if(do_validation)
+    {
+        using YDataType      = ComputeDataType;
+        using InvRmsDataType = InputDataType;
+
+        // Add
+        {
+            auto op = [](const auto& v0, const auto& v1) { return v0 + v1; };
+            ck_tile::reference_binary_elementwise<ADataType, BDataType, XDataType, ComputeDataType>(
+                a_host, b_host, x_host_ref, op);
+
+            if constexpr(SaveX)
+            {
+                x_buf.FromDevice(x_host_dev.data());
+
+                auto [rtol, atol] = get_elimit<XDataType>();
+                if(stride == n)
+                {
+                    pass = ck_tile::check_err(x_host_dev,
+                                              x_host_ref,
+                                              std::string("x Error: Incorrect results!"),
+                                              rtol,
+                                              atol);
+                }
+                else
+                {
+                    for(int i_r = 0; i_r < m; i_r++)
+                    {
+                        std::vector<QYDataType> x_host_dev_row(x_host_dev.begin() + i_r * stride,
+                                                               x_host_dev.begin() + i_r * stride +
+                                                                   n);
+                        std::vector<QYDataType> x_host_ref_row(x_host_ref.begin() + i_r * stride,
+                                                               x_host_ref.begin() + i_r * stride +
+                                                                   n);
+                        pass &= ck_tile::check_err(x_host_dev_row,
+                                                   x_host_ref_row,
+                                                   std::string("x[") + std::to_string(i_r) +
+                                                       std::string("] Error: Incorrect results!"),
+                                                   rtol,
+                                                   atol);
+                    }
+                }
+            }
+        }
+
+        ck_tile::HostTensor<YDataType> y_host({m, n});
+        // Rmsnorm2d
+        {
+            ck_tile::HostTensor<InvRmsDataType> invRms_host_ref({m});
+
+            // CAUSION: kernel use ComputeDataType version of x, but we use XDataType here for
+            // simplicity
+            ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                             GammaDataType,
+                                             ComputeDataType,
+                                             YDataType,
+                                             InvRmsDataType,
+                                             UnquantYDataType>(
+                x_host_ref, gamma_host, y_host, invRms_host_ref, unquant_y_host_ref, epsilon);
+        }
+
+        // yscale
+        {
+            ck_tile::HostTensor<YDataType> y_rowwise_amax_host({m});
+
+            using ReduceAmax = ck_tile::ReduceOp::AbsMax;
+            ck_tile::reference_reduce<YDataType, ComputeDataType, YDataType>(
+                y_host, y_rowwise_amax_host, ReduceAmax{});
+
+            auto op = [](const auto& v0) {
+                return v0 /
+                       ck_tile::type_convert<ComputeDataType>(ck_tile::numeric<QYDataType>::max());
+            };
+            ck_tile::reference_unary_elementwise<YDataType, YScaleDataType, ComputeDataType>(
+                y_rowwise_amax_host, yscale_host_ref, op);
+
+            yscale_buf.FromDevice(yscale_host_dev.mData.data());
+
+            auto [rtol, atol] = get_elimit<YScaleDataType>();
+            pass &= ck_tile::check_err(yscale_host_dev,
+                                       yscale_host_ref,
+                                       std::string("yscale Error: Incorrect results!"),
+                                       rtol,
+                                       atol);
+        }
+
+        // rowwise quantization
+        {
+            ck_tile::reference_rowwise_quantization2d<YDataType, YScaleDataType, QYDataType>(
+                y_host, yscale_host_ref, qy_host_ref);
+
+            qy_buf.FromDevice(qy_host_dev.data());
+            auto [rtol, atol] = get_elimit<QYDataType>();
+
+            if(stride == n)
+            {
+                pass = ck_tile::check_err(qy_host_dev,
+                                          qy_host_ref,
+                                          std::string("qy Error: Incorrect results!"),
+                                          rtol,
+                                          atol);
+            }
+            else
+            {
+                for(int i_r = 0; i_r < m; i_r++)
+                {
+                    std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride,
+                                                            qy_host_dev.begin() + i_r * stride + n);
+                    std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride,
+                                                            qy_host_ref.begin() + i_r * stride + n);
+                    pass &= ck_tile::check_err(qy_host_dev_row,
+                                               qy_host_ref_row,
+                                               std::string("qy[") + std::to_string(i_r) +
+                                                   std::string("] Error: Incorrect results!"),
+                                               rtol,
+                                               atol);
+                }
+            }
+        }
+
+        std::cout << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    const std::string input_data_type     = arg_parser.get_str("prec");
+    const std::string quantized_data_type = arg_parser.get_str("quant");
+    int save_x                            = arg_parser.get_int("save_x");
+    if(input_data_type == "fp16" && quantized_data_type == "int8" && save_x)
+    {
+        return run<ck_tile::half_t, ck_tile::int8_t, true>(arg_parser) ? 0 : -2;
+    }
+    else if(input_data_type == "fp16" && quantized_data_type == "int8" && !save_x)
+    {
+        return run<ck_tile::half_t, ck_tile::int8_t, false>(arg_parser) ? 0 : -2;
+    }
+    else if(input_data_type == "bf16" && quantized_data_type == "int8" && save_x)
+    {
+        return run<ck_tile::bf16_t, ck_tile::int8_t, true>(arg_parser) ? 0 : -2;
+    }
+    else if(input_data_type == "bf16" && quantized_data_type == "int8" && !save_x)
+    {
+        return run<ck_tile::bf16_t, ck_tile::int8_t, true>(arg_parser) ? 0 : -2;
+    }
+    else if(input_data_type == "fp16" && quantized_data_type == "fp8" && save_x)
+    {
+        return run<ck_tile::half_t, ck_tile::fp8_t, true>(arg_parser) ? 0 : -2;
+    }
+    else if(input_data_type == "fp16" && quantized_data_type == "fp8" && !save_x)
+    {
+        return run<ck_tile::half_t, ck_tile::fp8_t, false>(arg_parser) ? 0 : -2;
+    }
+    else if(input_data_type == "bf16" && quantized_data_type == "fp8" && save_x)
+    {
+        return run<ck_tile::bf16_t, ck_tile::fp8_t, true>(arg_parser) ? 0 : -2;
+    }
+    else if(input_data_type == "bf16" && quantized_data_type == "fp8" && !save_x)
+    {
+        return run<ck_tile::bf16_t, ck_tile::fp8_t, true>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/11_add_rmsnorm2d_rdquant/add_rmsnorm2d_rdquant_fwd.hpp

@@ -0,0 +1,150 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, 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/add_rmsnorm2d_rdquant.hpp"
+#include <string>
+
+template <typename InputDataType, typename QuantizedDataType>
+struct AddRmsnormRdquantTypeConfig;
+
+template <>
+struct AddRmsnormRdquantTypeConfig<ck_tile::half_t, ck_tile::int8_t>
+{
+    using ADataType       = ck_tile::half_t;
+    using BDataType       = ck_tile::half_t;
+    using GammaDataType   = ck_tile::half_t;
+    using XDataType       = ck_tile::half_t;
+    using YScaleDataType  = float;
+    using QYDataType      = ck_tile::int8_t;
+    using ComputeDataType = float;
+};
+
+template <>
+struct AddRmsnormRdquantTypeConfig<ck_tile::bf16_t, ck_tile::int8_t>
+{
+    using ADataType       = ck_tile::bf16_t;
+    using BDataType       = ck_tile::bf16_t;
+    using GammaDataType   = ck_tile::bf16_t;
+    using XDataType       = ck_tile::bf16_t;
+    using YScaleDataType  = float;
+    using QYDataType      = ck_tile::int8_t;
+    using ComputeDataType = float;
+};
+
+template <>
+struct AddRmsnormRdquantTypeConfig<ck_tile::half_t, ck_tile::fp8_t>
+{
+    using ADataType       = ck_tile::half_t;
+    using BDataType       = ck_tile::half_t;
+    using GammaDataType   = ck_tile::half_t;
+    using XDataType       = ck_tile::half_t;
+    using YScaleDataType  = float;
+    using QYDataType      = ck_tile::fp8_t;
+    using ComputeDataType = float;
+};
+
+template <>
+struct AddRmsnormRdquantTypeConfig<ck_tile::bf16_t, ck_tile::fp8_t>
+{
+    using ADataType       = ck_tile::bf16_t;
+    using BDataType       = ck_tile::bf16_t;
+    using GammaDataType   = ck_tile::bf16_t;
+    using XDataType       = ck_tile::bf16_t;
+    using YScaleDataType  = float;
+    using QYDataType      = ck_tile::fp8_t;
+    using ComputeDataType = float;
+};
+
+// runtime args
+struct add_rmsnorm2d_rdquant_fwd_args : public ck_tile::AddRmsnorm2dRdquantFwdHostArgs
+{
+};
+
+// this is used to pattern-match internl kernel implementation, not to instantiate kernel
+template <typename InputDataType_,
+          typename QuantizedDataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kSaveX_,
+          bool kThreePass_>
+struct add_rmsnorm2d_rdquant_fwd_traits_
+{
+    using InputDataType     = ck_tile::remove_cvref_t<InputDataType_>;
+    using QuantizedDataType = ck_tile::remove_cvref_t<QuantizedDataType_>;
+
+    static constexpr bool is_warp_per_row = ThreadPerBlock_N_ <= warpSize;
+    static_assert((ThreadPerBlock_M_ * ThreadPerBlock_N_) % warpSize == 0);
+    static constexpr ck_tile::index_t total_warps =
+        (ThreadPerBlock_M_ * ThreadPerBlock_N_) / warpSize;
+
+    // num of warps along m
+    static constexpr ck_tile::index_t BlockWarps_M = []() {
+        if constexpr(is_warp_per_row)
+        {
+            static_assert(warpSize % ThreadPerBlock_N_ == 0);
+            return total_warps * (warpSize / ThreadPerBlock_N_);
+        }
+        else
+        {
+            // static_assert(warpSize % ThreadPerBlock_M_ == 0);
+            return total_warps / (ThreadPerBlock_N_ / warpSize);
+        }
+    }();
+
+    // num of warps along n
+    static constexpr ck_tile::index_t BlockWarps_N = []() {
+        if constexpr(is_warp_per_row)
+        {
+            static_assert(warpSize % ThreadPerBlock_N_ == 0);
+            return 1;
+        }
+        else
+        {
+            static_assert(ThreadPerBlock_N_ % warpSize == 0);
+            return ThreadPerBlock_N_ / warpSize;
+        }
+    }();
+
+    static constexpr ck_tile::index_t Repeat_M = Repeat_M_;
+    static constexpr ck_tile::index_t Repeat_N = Repeat_N_;
+
+    static constexpr ck_tile::index_t Block_M = Repeat_M_ * ThreadPerBlock_M_;
+    static constexpr ck_tile::index_t Block_N = Repeat_N_ * ThreadPerBlock_N_ * Vector_N_;
+
+    static constexpr ck_tile::index_t Warp_M = ThreadPerBlock_M_ / BlockWarps_M;
+    static constexpr ck_tile::index_t Warp_N = ThreadPerBlock_N_ / BlockWarps_N * Vector_N_;
+
+    using BlockTile  = ck_tile::sequence<Block_M, Block_N>;
+    using BlockWarps = ck_tile::sequence<BlockWarps_M, BlockWarps_N>;
+    using WarpTile   = ck_tile::sequence<Warp_M, Warp_N>;
+    using Vector     = ck_tile::sequence<1, Vector_N_>;
+
+    using Shape = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
+
+    static constexpr bool kPadN      = kPadN_;
+    static constexpr bool kSaveX     = kSaveX_;
+    static constexpr bool kThreePass = kThreePass_;
+};
+
+template <typename Traits_>
+float add_rmsnorm2d_rdquant_fwd_(const ck_tile::stream_config& s, add_rmsnorm2d_rdquant_fwd_args a);
+
+// This is the public API, will be generated by script
+struct add_rmsnorm2d_rdquant_fwd_traits
+{
+    std::string input_data_type;
+    std::string quantized_data_type;
+    bool save_x;
+};
+
+float add_rmsnorm2d_rdquant_fwd(add_rmsnorm2d_rdquant_fwd_traits,
+                                add_rmsnorm2d_rdquant_fwd_args,
+                                const ck_tile::stream_config&);

example/ck_tile/11_add_rmsnorm2d_rdquant/example_add_rmsnorm2d_rdquant_fwd.cpp

@@ -0,0 +1,283 @@
+#include "ck_tile/host.hpp"
+#include "ck_tile/core.hpp"
+#include "ck_tile/host/kernel_launch.hpp"
+#include "ck_tile/ops/add_rmsnorm2d_rdquant.hpp"
+#include <cstring>
+
+// different threshold for different dtype
+template <typename DataType>
+auto get_elimit()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::bf16_t>()
+{
+    double rtol = 1e-2;
+    double atol = 1e-2;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+template <>
+auto get_elimit<ck_tile::int8_t>()
+{
+    // due to rounding, int8 quantization might have 1 abs error
+    double rtol = 1;
+    double atol = 1;
+    return ck_tile::make_tuple(rtol, atol);
+}
+
+auto create_args(int argc, char* argv[])
+{
+    ck_tile::ArgParser arg_parser;
+    arg_parser.insert("m", "3328", "m dimension")
+        .insert("n", "4096", "n dimension")
+        .insert("stride", "-1", "stride per row, if -1 then equal to n")
+        .insert("e", "1e-5", "epsilon")
+        .insert("v", "1", "cpu validation or not")
+        .insert("prec", "fp16", "precision")
+        .insert("warmup", "0", "cold iter")
+        .insert("repeat", "1", "hot iter");
+
+    bool result = arg_parser.parse(argc, argv);
+    return std::make_tuple(result, arg_parser);
+}
+
+template <typename DataType>
+bool run(const ck_tile::ArgParser& arg_parser)
+{
+    ck_tile::index_t m      = arg_parser.get_int("m");
+    ck_tile::index_t n      = arg_parser.get_int("n");
+    ck_tile::index_t stride = arg_parser.get_int("stride");
+    if(stride < 0)
+        stride = n;
+    float epsilon         = arg_parser.get_float("e");
+    std::string data_type = arg_parser.get_str("prec");
+    int do_validation     = arg_parser.get_int("v");
+    int warmup            = arg_parser.get_int("warmup");
+    int repeat            = arg_parser.get_int("repeat");
+
+    assert(stride >= n);
+
+    using ADataType        = DataType;
+    using BDataType        = DataType;
+    using GammaDataType    = DataType;
+    using XDataType        = DataType;
+    using UnquantYDataType = ck_tile::null_type;
+    using YScaleDataType   = float;
+    using QYDataType       = ck_tile::int8_t;
+    using ComputeDataType  = float;
+
+    // host verify
+    ck_tile::HostTensor<ADataType> a_host({m, n}, {stride, 1});
+    ck_tile::HostTensor<BDataType> b_host({m, n}, {stride, 1});
+    ck_tile::HostTensor<GammaDataType> gamma_host({n});
+
+    ck_tile::HostTensor<XDataType> x_host_ref({m, n}, {stride, 1});
+    ck_tile::HostTensor<XDataType> x_host_dev({m, n}, {stride, 1});
+    ck_tile::HostTensor<YScaleDataType> yscale_host_ref({m}, {1});
+    ck_tile::HostTensor<YScaleDataType> yscale_host_dev({m}, {1});
+    ck_tile::HostTensor<QYDataType> qy_host_ref({m, n}, {stride, 1});
+    ck_tile::HostTensor<QYDataType> qy_host_dev({m, n}, {stride, 1});
+    ck_tile::HostTensor<UnquantYDataType> unquant_y_host_ref({m, n}, {stride, 1});
+
+    ck_tile::FillUniformDistribution<ADataType>{-.5f, .5f}(a_host);
+    ck_tile::FillUniformDistribution<BDataType>{-.5f, .5f}(b_host);
+    ck_tile::FillUniformDistribution<GammaDataType>{-.5f, .5f}(gamma_host);
+
+    ck_tile::DeviceMem a_buf(a_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem b_buf(b_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem gamma_buf(gamma_host.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem x_buf(x_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem yscale_buf(yscale_host_dev.get_element_space_size_in_bytes());
+    ck_tile::DeviceMem qy_buf(qy_host_dev.get_element_space_size_in_bytes());
+
+    a_buf.ToDevice(a_host.data());
+    b_buf.ToDevice(b_host.data());
+    gamma_buf.ToDevice(gamma_host.data());
+
+    constexpr bool kThreePass = true;
+
+    using BlockWarps = ck_tile::sequence<4, 1>;
+    using BlockTile  = ck_tile::sequence<4, 128>;
+    using WarpTile   = ck_tile::sequence<1, 64>;
+    using Vector     = ck_tile::sequence<1, 1>;
+
+    using Shape   = ck_tile::Generic2dBlockShape<BlockTile, BlockWarps, WarpTile, Vector>;
+    using Problem = ck_tile::AddRmsnorm2dRdquantFwdPipelineProblem<ADataType,
+                                                                   BDataType,
+                                                                   GammaDataType,
+                                                                   ComputeDataType,
+                                                                   XDataType,
+                                                                   YScaleDataType,
+                                                                   QYDataType,
+                                                                   Shape,
+                                                                   true, // kPadN
+                                                                   true, // kSaveX
+                                                                   kThreePass>;
+
+    using OnePassPipeline   = ck_tile::AddRmsnorm2dRdquantFwdPipelineOnePass<Problem>;
+    using ThreePassPipeline = ck_tile::AddRmsnorm2dRdquantFwdPipelineThreePass<Problem>;
+    using Pipeline          = std::conditional_t<kThreePass, ThreePassPipeline, OnePassPipeline>;
+    using Kernel            = ck_tile::AddRmsnorm2dRdquantFwd<Pipeline>;
+
+    ck_tile::AddRmsnorm2dRdquantFwdHostArgs args{a_buf.GetDeviceBuffer(),
+                                                 b_buf.GetDeviceBuffer(),
+                                                 gamma_buf.GetDeviceBuffer(),
+                                                 x_buf.GetDeviceBuffer(),
+                                                 yscale_buf.GetDeviceBuffer(),
+                                                 qy_buf.GetDeviceBuffer(),
+                                                 epsilon,
+                                                 m,
+                                                 n,
+                                                 stride};
+
+    auto kargs = Kernel::MakeKargs(args);
+
+    const dim3 grids                       = Kernel::GridSize(args);
+    constexpr dim3 blocks                  = Kernel::BlockSize();
+    constexpr ck_tile::index_t kBlockPerCu = 1;
+    auto s = ck_tile::stream_config{nullptr, true, 0, warmup, repeat};
+
+    ck_tile::launch_kernel(
+        s, ck_tile::make_kernel<blocks.x, kBlockPerCu>(Kernel{}, grids, blocks, 0, kargs));
+
+    bool pass = true;
+
+    if(do_validation)
+    {
+        using YDataType      = ComputeDataType;
+        using InvRmsDataType = DataType;
+
+        // Add
+        {
+            auto op = [](const auto& v0, const auto& v1) { return v0 + v1; };
+            ck_tile::reference_binary_elementwise<ADataType, BDataType, XDataType, ComputeDataType>(
+                a_host, b_host, x_host_ref, op);
+
+            x_buf.FromDevice(x_host_dev.data());
+
+            auto [rtol, atol] = get_elimit<XDataType>();
+            if(stride == n)
+            {
+                pass = ck_tile::check_err(
+                    x_host_dev, x_host_ref, std::string("x Error: Incorrect results!"), rtol, atol);
+            }
+            else
+            {
+                for(int i_r = 0; i_r < m; i_r++)
+                {
+                    std::vector<QYDataType> x_host_dev_row(x_host_dev.begin() + i_r * stride,
+                                                           x_host_dev.begin() + i_r * stride + n);
+                    std::vector<QYDataType> x_host_ref_row(x_host_ref.begin() + i_r * stride,
+                                                           x_host_ref.begin() + i_r * stride + n);
+                    pass &= ck_tile::check_err(x_host_dev_row,
+                                               x_host_ref_row,
+                                               std::string("x[") + std::to_string(i_r) +
+                                                   std::string("] Error: Incorrect results!"),
+                                               rtol,
+                                               atol);
+                }
+            }
+        }
+
+        ck_tile::HostTensor<YDataType> y_host({m, n});
+        // Rmsnorm2d
+        {
+            ck_tile::HostTensor<InvRmsDataType> invRms_host_ref({m});
+
+            // CAUSION: kernel use ComputeDataType version of x, but we use XDataType here for
+            // simplicity
+            ck_tile::reference_rmsnorm2d_fwd<XDataType,
+                                             GammaDataType,
+                                             ComputeDataType,
+                                             YDataType,
+                                             InvRmsDataType,
+                                             UnquantYDataType>(
+                x_host_ref, gamma_host, y_host, invRms_host_ref, unquant_y_host_ref, epsilon);
+        }
+
+        // yscale
+        {
+            ck_tile::HostTensor<YDataType> y_rowwise_amax_host({m});
+
+            using ReduceAmax = ck_tile::ReduceOp::AbsMax;
+            ck_tile::reference_reduce<YDataType, ComputeDataType, YDataType>(
+                y_host, y_rowwise_amax_host, ReduceAmax{});
+
+            auto op = [](const auto& v0) {
+                return v0 /
+                       ck_tile::type_convert<ComputeDataType>(ck_tile::numeric<QYDataType>::max());
+            };
+            ck_tile::reference_unary_elementwise<YDataType, YScaleDataType, ComputeDataType>(
+                y_rowwise_amax_host, yscale_host_ref, op);
+
+            yscale_buf.FromDevice(yscale_host_dev.mData.data());
+
+            auto [rtol, atol] = get_elimit<YScaleDataType>();
+            pass &= ck_tile::check_err(yscale_host_dev,
+                                       yscale_host_ref,
+                                       std::string("yscale Error: Incorrect results!"),
+                                       rtol,
+                                       atol);
+        }
+
+        // rowwise quantization
+        {
+            ck_tile::reference_rowwise_quantization2d<YDataType, YScaleDataType, QYDataType>(
+                y_host, yscale_host_ref, qy_host_ref);
+
+            qy_buf.FromDevice(qy_host_dev.data());
+            auto [rtol, atol] = get_elimit<QYDataType>();
+
+            if(stride == n)
+            {
+                pass = ck_tile::check_err(qy_host_dev,
+                                          qy_host_ref,
+                                          std::string("qy Error: Incorrect results!"),
+                                          rtol,
+                                          atol);
+            }
+            else
+            {
+                for(int i_r = 0; i_r < m; i_r++)
+                {
+                    std::vector<QYDataType> qy_host_dev_row(qy_host_dev.begin() + i_r * stride,
+                                                            qy_host_dev.begin() + i_r * stride + n);
+                    std::vector<QYDataType> qy_host_ref_row(qy_host_ref.begin() + i_r * stride,
+                                                            qy_host_ref.begin() + i_r * stride + n);
+                    pass &= ck_tile::check_err(qy_host_dev_row,
+                                               qy_host_ref_row,
+                                               std::string("qy[") + std::to_string(i_r) +
+                                                   std::string("] Error: Incorrect results!"),
+                                               rtol,
+                                               atol);
+                }
+            }
+        }
+
+        std::cout << "[" << data_type << "]"
+                  << " m:" << m << ", n:" << n << ", stride:" << stride
+                  << ", valid:" << (pass ? "y" : "n") << std::flush << std::endl;
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    auto [result, arg_parser] = create_args(argc, argv);
+    if(!result)
+        return -1;
+
+    const std::string data_type = arg_parser.get_str("prec");
+    if(data_type == "fp16")
+    {
+        return run<ck_tile::half_t>(arg_parser) ? 0 : -2;
+    }
+
+    return -3;
+}

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_api.cpp

@@ -0,0 +1,227 @@
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <ck_tile/core.hpp>
+#include "add_rmsnorm2d_rdquant_fwd.hpp"
+
+template <typename InputDataType_,
+          typename QuantizedDataType_,
+          ck_tile::index_t Repeat_M_,         // each thread repeat along M
+          ck_tile::index_t Repeat_N_,         // each thread repeat along N
+          ck_tile::index_t ThreadPerBlock_M_, // num threads along M
+          ck_tile::index_t ThreadPerBlock_N_, // num threads along N
+          ck_tile::index_t Vector_N_,         // vector size along N
+          bool kPadN_,
+          bool kSaveX_,
+          bool kThreePass_>
+using trait_ = add_rmsnorm2d_rdquant_fwd_traits_<InputDataType_,
+                                                 QuantizedDataType_,
+                                                 Repeat_M_,
+                                                 Repeat_N_,
+                                                 ThreadPerBlock_M_,
+                                                 ThreadPerBlock_N_,
+                                                 Vector_N_,
+                                                 kPadN_,
+                                                 kSaveX_,
+                                                 kThreePass_>;
+
+template <typename input_data_type, typename quantized_data_type>
+float add_rmsnorm2d_rdquant_fwd_b16_(add_rmsnorm2d_rdquant_fwd_traits t,
+                                     add_rmsnorm2d_rdquant_fwd_args a,
+                                     const ck_tile::stream_config& s)
+{
+    float r = -1;
+    // clang-format off
+    //                                                      rm  rn  tm   tn  vn   pd     x      3p
+    if(a.n <= 64) {
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1,  1,  4,  64, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 128) {
+        if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1,  1,  4,  64, 2,  true,  true, false>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type, 1,  2,  4,  64, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 256) {
+        if (a.n % 4 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 1,  4,  64, 4,  true,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2,  4,  64, 2,  true,  true, false>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4,  4,  64, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 512) {
+        if (a.n % 8 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 1,  4,  64, 8,  true,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2,  4,  64, 4,  true,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4,  4,  64, 2,  true,  true, false>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 8,  4,  64, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 768) {
+        if (a.n % 4 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 3,  4,  64, 4,  true,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 6,  4,  64, 2,  true,  true, false>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1,12,  4,  64, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 1024) {
+        if (a.n % 8 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 1, 2,  128, 8,  true,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2, 2,  128, 4,  true,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 2,  128, 2,  true,  true, false>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1,  256, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 1536) {
+        if (a.n % 8 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 3, 4,   64, 8,  true,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 3, 2,  128, 4,  true,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 3, 1,  256, 2,  true,  true, false>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 6, 1,  256, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 2048) {
+        if (a.n % 8 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 1, 1,  256, 8,  true,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2, 1,  256, 4,  true,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1,  256, 2,  true,  true, false>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 8, 1,  256, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 3072) {
+        if (a.n % 8 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 3, 1,  128, 8,  true,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 3, 1,  256, 4,  true,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 6, 1,  256, 2,  true,  true, false>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 3, 1, 1024, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 4096) {
+        if (a.n % 8 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2, 1,  256, 8,  true,  true, false>>(s, a);
+        else if (a.n % 4 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1,  256, 4,  true,  true, false>>(s, a);
+        else if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2, 1, 1024, 2,  true,  true, false>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1, 1024, 1,  true,  true, false>>(s, a);
+    }
+    else if(a.n <= 8192) {
+        if(a.n<8192){
+            if(t.save_x){
+                if (a.n % 8 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2, 1,  512, 8,  true,  true, false>>(s, a);
+                else if (a.n % 4 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1,  512, 4,  true,  true, false>>(s, a);
+                else if (a.n % 2 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1, 1024, 2,  true,  true, false>>(s, a);
+                else
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 8, 1, 1024, 1,  true,  true, false>>(s, a);
+            }
+            else{
+                if (a.n % 8 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2, 1,  512, 8,  true,  false, false>>(s, a);
+                else if (a.n % 4 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1,  512, 4,  true,  false, false>>(s, a);
+                else if (a.n % 2 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1, 1024, 2,  true,  false, false>>(s, a);
+                else
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 8, 1, 1024, 1,  true,  false, false>>(s, a);
+            }
+        }
+        else{
+            if(t.save_x){
+                if (a.n % 8 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2, 1,  512, 8,  false,  true, false>>(s, a);
+                else if (a.n % 4 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1,  512, 4,  false,  true, false>>(s, a);
+                else if (a.n % 2 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1, 1024, 2,  false,  true, false>>(s, a);
+                else
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 8, 1, 1024, 1,  false,  true, false>>(s, a);
+            }
+            else{
+                if (a.n % 8 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2, 1,  512, 8,  false,  false, false>>(s, a);
+                else if (a.n % 4 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1,  512, 4,  false,  false, false>>(s, a);
+                else if (a.n % 2 == 0)
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1, 1024, 2,  false,  false, false>>(s, a);
+                else
+                    r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 8, 1, 1024, 1,  false,  false, false>>(s, a);
+            }
+        }
+    }
+    else if(a.n > 8192) {
+        if (a.n % 8 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 2, 1,  512, 8,  true,  true, true>>(s, a);
+        else if (a.n % 4 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1,  512, 4,  true,  true, true>>(s, a);
+        else if (a.n % 2 == 0)
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 4, 1, 1024, 2,  true,  true, true>>(s, a);
+        else
+            r = add_rmsnorm2d_rdquant_fwd_<trait_<input_data_type, quantized_data_type,  1, 8, 1, 1024, 1,  true,  true, true>>(s, a);
+    }
+    return r;
+    // clang-format on
+}
+
+float add_rmsnorm2d_rdquant_fwd(add_rmsnorm2d_rdquant_fwd_traits t,
+                                add_rmsnorm2d_rdquant_fwd_args a,
+                                const ck_tile::stream_config& s)
+{
+    if(t.input_data_type.compare("fp16") == 0 && t.quantized_data_type.compare("int8") == 0 &&
+       t.save_x)
+    {
+        return add_rmsnorm2d_rdquant_fwd_b16_<ck_tile::fp16_t, ck_tile::int8_t>(t, a, s);
+    }
+    else if(t.input_data_type.compare("fp16") == 0 && t.quantized_data_type.compare("int8") == 0 &&
+            !t.save_x)
+    {
+        return add_rmsnorm2d_rdquant_fwd_b16_<ck_tile::fp16_t, ck_tile::int8_t>(t, a, s);
+    }
+    else if(t.input_data_type.compare("bf16") == 0 && t.quantized_data_type.compare("int8") == 0 &&
+            t.save_x)
+    {
+        return add_rmsnorm2d_rdquant_fwd_b16_<ck_tile::bf16_t, ck_tile::int8_t>(t, a, s);
+    }
+    else if(t.input_data_type.compare("bf16") == 0 && t.quantized_data_type.compare("int8") == 0 &&
+            !t.save_x)
+    {
+        return add_rmsnorm2d_rdquant_fwd_b16_<ck_tile::bf16_t, ck_tile::int8_t>(t, a, s);
+    }
+    else if(t.input_data_type.compare("fp16") == 0 && t.quantized_data_type.compare("fp8") == 0 &&
+            t.save_x)
+    {
+        return add_rmsnorm2d_rdquant_fwd_b16_<ck_tile::fp16_t, ck_tile::fp8_t>(t, a, s);
+    }
+    else if(t.input_data_type.compare("fp16") == 0 && t.quantized_data_type.compare("fp8") == 0 &&
+            !t.save_x)
+    {
+        return add_rmsnorm2d_rdquant_fwd_b16_<ck_tile::fp16_t, ck_tile::fp8_t>(t, a, s);
+    }
+    else if(t.input_data_type.compare("bf16") == 0 && t.quantized_data_type.compare("fp8") == 0 &&
+            t.save_x)
+    {
+        return add_rmsnorm2d_rdquant_fwd_b16_<ck_tile::bf16_t, ck_tile::fp8_t>(t, a, s);
+    }
+    else if(t.input_data_type.compare("bf16") == 0 && t.quantized_data_type.compare("fp8") == 0 &&
+            !t.save_x)
+    {
+        return add_rmsnorm2d_rdquant_fwd_b16_<ck_tile::bf16_t, ck_tile::fp8_t>(t, a, s);
+    }
+    else
+        throw std::runtime_error("Without supported instances!");
+}

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1024_instance.cpp

@@ -0,0 +1,26 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                                rm  rn  tm  tn  vn  pd      x     3p
+#if 0
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  2,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  4,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  8,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1, 16,  4,  64, 1,  true , true, false>>(const S&, A);
+
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, 1,  1,  1, 256, 4,  true , true, false>>(const S&, A);
+#endif
+
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 1, 2,  128, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 2, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 2,  128, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1,  256, 1,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 1, 2,  128, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 2, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 2,  128, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1,  256, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n1536_instance.cpp

@@ -0,0 +1,17 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 3, 4,   64, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 3, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 3, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 6, 1,  256, 1,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 3, 4,   64, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 3, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 3, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 6, 1,  256, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n2048_instance.cpp

@@ -0,0 +1,18 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 1, 1,  256, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 2, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 8, 1,  256, 1,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 1, 1,  256, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 2, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 8, 1,  256, 1,  true,  true, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n256_instance.cpp

@@ -0,0 +1,15 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  1,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  2,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  4,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  1,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  2,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  4,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n3072_instance.cpp

@@ -0,0 +1,17 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 3, 1,  128, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 3, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 6, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 3, 1, 1024, 1,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 3, 1,  128, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 3, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 6, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 3, 1, 1024, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n4096_instance.cpp

@@ -0,0 +1,17 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 2, 1,  256, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 2, 1, 1024, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1, 1024, 1,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 2, 1,  256, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1,  256, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 2, 1, 1024, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1, 1024, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n512_instance.cpp

@@ -0,0 +1,17 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  1,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  2,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  4,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  8,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  1,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  2,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  4,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  8,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n64_n128_instance.cpp

@@ -0,0 +1,15 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn    pd     x      3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  1,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  1,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  2,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  1,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  1,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  2,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n768_instance.cpp

@@ -0,0 +1,15 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  3,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1,  6,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 12,  4,  64, 1,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  3,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1,  6,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 12,  4,  64, 1,  true , true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n8192_instance.cpp

@@ -0,0 +1,42 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 2, 1,  512, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1,  512, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1, 1024, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 8, 1, 1024, 1,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 2, 1,  512, 8,  true,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1,  512, 4,  true,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1, 1024, 2,  true,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 8, 1, 1024, 1,  true,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 2, 1,  512, 8,  false,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1,  512, 4,  false,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1, 1024, 2,  false,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 8, 1, 1024, 1,  false,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 2, 1,  512, 8,  false,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1,  512, 4,  false,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 4, 1, 1024, 2,  false,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::int8_t, 1, 8, 1, 1024, 1,  false,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 2, 1,  512, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1,  512, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1, 1024, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 8, 1, 1024, 1,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 2, 1,  512, 8,  true,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1,  512, 4,  true,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1, 1024, 2,  true,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 8, 1, 1024, 1,  true,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 2, 1,  512, 8,  false,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1,  512, 4,  false,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1, 1024, 2,  false,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 8, 1, 1024, 1,  false,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 2, 1,  512, 8,  false,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1,  512, 4,  false,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 4, 1, 1024, 2,  false,  false, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t,  ck_tile::fp8_t, 1, 8, 1, 1024, 1,  false,  false, false>>(const S&, A);
+
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_bf16_n8192_tp_instance.cpp

@@ -0,0 +1,17 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 2, 1,  512, 8,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1,  512, 4,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 4, 1, 1024, 2,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::int8_t, 1, 8, 1, 1024, 1,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 2, 1,  512, 8,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1,  512, 4,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 4, 1, 1024, 2,  true,  true, true>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::bf16_t, ck_tile::fp8_t, 1, 8, 1, 1024, 1,  true,  true, true>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1024_instance.cpp

@@ -0,0 +1,26 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                                rm  rn  tm  tn  vn  pd      x     3p
+#if 0
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  2,  4,  64, 8,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  4,  4,  64, 4,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  8,  4,  64, 2,  true , true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1, 16,  4,  64, 1,  true , true, false>>(const S&, A);
+
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t, 1,  1,  1, 256, 4,  true , true, false>>(const S&, A);
+#endif
+
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::int8_t, 1, 1, 2,  128, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::int8_t, 1, 2, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::int8_t, 1, 4, 2,  128, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::int8_t, 1, 4, 1,  256, 1,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::fp8_t, 1, 1, 2,  128, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::fp8_t, 1, 2, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::fp8_t, 1, 4, 2,  128, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::fp8_t, 1, 4, 1,  256, 1,  true,  true, false>>(const S&, A);
+// clang-format on

example/ck_tile/11_add_rmsnorm2d_rdquant/instances/add_rmsnorm2d_rdquant_fwd_fp16_n1536_instance.cpp

@@ -0,0 +1,17 @@
+
+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "add_rmsnorm2d_rdquant_fwd_instance_common.hpp"
+
+// clang-format off
+//                                                               rm  rn  tm  tn  vn     pd    x     3p
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::int8_t, 1, 3, 4,   64, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::int8_t, 1, 3, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::int8_t, 1, 3, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::int8_t, 1, 6, 1,  256, 1,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::fp8_t, 1, 3, 4,   64, 8,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::fp8_t, 1, 3, 2,  128, 4,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::fp8_t, 1, 3, 1,  256, 2,  true,  true, false>>(const S&, A);
+template float add_rmsnorm2d_rdquant_fwd_<trait_<ck_tile::fp16_t,  ck_tile::fp8_t, 1, 6, 1,  256, 1,  true,  true, false>>(const S&, A);
+// clang-format on