Merge remote-tracking branch 'upstream/develop' into ck_host_lib

Jenkinsfile

@@ -735,11 +735,11 @@ def process_results(Map conf=[:]){
 
 //launch develop branch daily at 23:00 UT in FULL_QA mode and at 19:00 UT with latest staging compiler version
 CRON_SETTINGS = BRANCH_NAME == "develop" ? '''0 23 * * * % RUN_FULL_QA=true;ROCMVERSION=6.2;RUN_CK_TILE_FMHA_TESTS=true;RUN_CK_TILE_GEMM_TESTS=true
-                                              0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true
+                                              0 21 * * * % ROCMVERSION=6.2;hipTensor_test=true;RUN_CODEGEN_TESTS=true
                                               0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-staging;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
                                               0 17 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-mainline-open;BUILD_COMPILER=/llvm-project/build/bin/clang++;BUILD_GFX12=true;USE_SCCACHE=false;NINJA_BUILD_TRACE=true
-                                              0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_CODEGEN_TESTS=false;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false
-                                              0 13 * * * % BUILD_LEGACY_OS=true ''' : ""
+                                              0 15 * * * % BUILD_INSTANCES_ONLY=true;RUN_PERFORMANCE_TESTS=false;USE_SCCACHE=false
+                                              0 13 * * * % BUILD_LEGACY_OS=true''' : ""
 
 pipeline {
     agent none
@@ -806,6 +806,10 @@ pipeline {
             name: "RUN_GROUPED_CONV_LARGE_CASES_TESTS",
             defaultValue: false,
             description: "Run the grouped conv large cases tests (default: OFF)")
+        booleanParam(
+            name: "RUN_CODEGEN_TESTS",
+            defaultValue: false,
+            description: "Run codegen tests (default: OFF)")
         booleanParam(
             name: "RUN_CK_TILE_FMHA_TESTS",
             defaultValue: false,
@@ -934,6 +938,29 @@ pipeline {
                 }
             }
         }
+        stage("Run Codegen Tests")
+        {
+            parallel
+            {
+                stage("Run Codegen Tests on gfx90a")
+                {
+                    when {
+                        beforeAgent true
+                        expression { params.RUN_CODEGEN_TESTS.toBoolean() }
+                    }
+                    agent{ label rocmnode("gfx90a")}
+                    environment{
+                        setup_args = "NO_CK_BUILD"
+                        execute_args = """ CXX=/opt/rocm/llvm/bin/clang++ cmake ../codegen && \
+                                           make -j64 check"""
+                    }
+                    steps{
+                        buildHipClangJobAndReboot(setup_args:setup_args, no_reboot:true, build_type: 'Release', execute_cmd: execute_args)
+                        cleanWs()
+                    }
+                }
+            }
+        }
         stage("Run CK_TILE_FMHA Tests")
         {
             parallel

codegen/CMakeLists.txt

+cmake_minimum_required(VERSION 3.16)
+project(composable_kernel_host)
+
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
@@ -5,71 +8,51 @@ set(CMAKE_ARCHIVE_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/lib)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 set(CK_ROOT ${CMAKE_CURRENT_SOURCE_DIR}/..)
 
-add_compile_options(-std=c++17)
-find_package(hip)
-add_custom_target(codegen)
+find_package(ROCM)
+include(ROCMInstallTargets)
+include(ROCMTest)
 
-# add include directories
-include_directories(BEFORE
-            ${PROJECT_BINARY_DIR}/include
-            ${PROJECT_SOURCE_DIR}/include
-            ${PROJECT_SOURCE_DIR}/library/include
-            ${HIP_INCLUDE_DIRS}
-            )
+rocm_setup_version(VERSION 1.0)
 
 list(APPEND CMAKE_MODULE_PATH ${CK_ROOT}/cmake)
 include(Embed)
 file(GLOB_RECURSE KERNEL_FILES CONFIGURE_DEPENDS
     ${CK_ROOT}/include/ck/*.hpp)
-# printouts for debug purposes
-#message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
-#message(STATUS "RELATIVE: ${CK_ROOT}/include")
+# printouts fot debug purposes
+# message(STATUS "KERNEL_FILES: ${KERNEL_FILES}")
+# message(STATUS "RELATIVE: ${CK_ROOT}/include")
 add_embed_library(ck_headers ${KERNEL_FILES} RELATIVE ${CK_ROOT}/include)
 
-file(GLOB SOURCES CONFIGURE_DEPENDS src/*.cpp)
+add_compile_options(-std=c++17)
 
-#message(STATUS "SOURCE_FILES: ${SOURCES}")
+file(GLOB SOURCES CONFIGURE_DEPENDS src/*.cpp)
 # TODO: Use object library
 add_library(ck_host STATIC ${SOURCES})
-add_library(composable_kernel::ck_host ALIAS ck_host)
+target_link_libraries(ck_host PRIVATE ck_headers)
 
 set_target_properties(ck_host PROPERTIES 
     LINKER_LANGUAGE CXX
     POSITION_INDEPENDENT_CODE ON)
 
-target_include_directories(ck_host SYSTEM PRIVATE
-    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
-    $<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR}/solution_instances>
-    $<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR}/embed/ck_headers/include>
-)
-
-target_link_libraries(ck_host PRIVATE $<BUILD_INTERFACE:ck_headers>)
-
-target_include_directories(ck_host PUBLIC
-    $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
-    $<INSTALL_INTERFACE:include>
-)
+# target_include_directories(ck_host PUBLIC
+#     $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>
+# )
 
 add_executable(ck-template-driver driver/main.cpp)
 target_link_libraries(ck-template-driver ck_host)
 
-rocm_install(
-    TARGETS ck_host
-    EXPORT ck_hostTargets
+rocm_install_targets(
+    TARGETS ck_host ck_headers
+    EXPORT ck_host_targets
+    INCLUDE include
+    PRIVATE
 )
-rocm_install(EXPORT ck_hostTargets 
-	FILE composable_kernelck_hostTargets.cmake 
+rocm_export_targets(
+    EXPORT ck_host_targets
     NAMESPACE composable_kernel::
-	DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel)
-rocm_install(DIRECTORY include/ck DESTINATION ${CMAKE_INSTALL_INCLUDEDIR})
-
-rocm_install(
-    EXPORT ck_hostTargets
-    FILE composable_kernelck_hostTargets.cmake
-    NAMESPACE composable_kernel::
-    DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
 )
 
-if (NOT CK_BUILD_HOST_LIB AND BUILD_TESTING)
+if(BUILD_TESTING AND NOT CK_BUILD_HOST_LIB)
     add_subdirectory(test)
 endif()
+

codegen/test/CMakeLists.txt

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

codegen/test/rtc/CMakeLists.txt

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

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

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

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

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

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

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

codegen/test/rtc/src/compile_kernel.cpp

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

codegen/test/rtc/src/tmp_dir.cpp

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

example/ck_tile/03_gemm/gemm_basic.cpp

-
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
 
@@ -282,7 +281,11 @@ int main(int argc, char* argv[])
     using CodegenPipelineProblem = ck_tile::
         GemmPipelineProblem<ADataType, BDataType, AccDataType, CodegenGemmShape, CodegenGemmTraits>;
 
-    using CodegenGemmPipeline = ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem>;
+    using CodegenGemmPolicy = ck_tile::
+        UniversalGemmPipelineAgBgCrPolicy<matrix_a_layout, matrix_b_layout, matrix_c_layout>;
+
+    using CodegenGemmPipeline =
+        ck_tile::GemmPipelineAGmemBGmemCRegV1<CodegenPipelineProblem, CodegenGemmPolicy>;
 
     invoke_gemm<ck_tile::half_t,
                 matrix_a_layout,

include/ck_tile/core/config.hpp

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

include/ck_tile/ops/gemm.hpp

@@ -23,6 +23,7 @@
 #include "ck_tile/ops/gemm/block/block_gemm_problem.hpp"
 #include "ck_tile/ops/gemm/kernel/gemm_kernel.hpp"
 #include "ck_tile/ops/gemm/kernel/gemm_tile_partitioner.hpp"
+#include "ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v1_default_policy.hpp"
 #include "ck_tile/ops/gemm/pipeline/gemm_pipeline_agmem_bgmem_creg_v2.hpp"

include/ck_tile/ops/gemm/pipeline/gemm_universal_pipeline_ag_bg_cr_policy.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck_tile/core.hpp"
+#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp"
+
+namespace ck_tile {
+
+// UniversalGemm Policy
+template <typename LayoutA_, typename LayoutB_, typename LayoutC_>
+struct UniversalGemmPipelineAgBgCrPolicy
+{
+    using LayoutA = remove_cvref_t<LayoutA_>;
+    using LayoutB = remove_cvref_t<LayoutB_>;
+    using LayoutC = remove_cvref_t<LayoutC_>;
+
+    static constexpr auto I0 = number<0>{};
+    static constexpr auto I1 = number<1>{};
+    static constexpr auto I2 = number<2>{};
+
+    static constexpr bool TransposeC = true;
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeALdsBlockDescriptor()
+    {
+        using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
+                                                typename Problem::BDataType,
+                                                typename Problem::CDataType,
+                                                Problem::BlockGemmShape::WarpTile::at(I0),
+                                                Problem::BlockGemmShape::WarpTile::at(I1),
+                                                Problem::BlockGemmShape::WarpTile::at(I2),
+                                                TransposeC>;
+
+        using ADataType = remove_cvref_t<typename Problem::ADataType>;
+
+        constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
+        constexpr index_t K1        = WarpGemm::kK;
+        constexpr index_t K0        = KPerBlock / K1;
+
+        if constexpr(std::is_same<tensor_layout::gemm::RowMajor, LayoutA>::value)
+        {
+            constexpr auto MLdsLayer        = 32 * 4 / KPerBlock / sizeof(ADataType) < 1
+                                                  ? 1
+                                                  : 32 * 4 / KPerBlock / sizeof(ADataType);
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor(
+                make_tuple(K0 * number<MLdsLayer>{}, number<MPerBlock / MLdsLayer>{}, K1),
+                make_tuple(K1, number<KPerBlock * MLdsLayer>{}, I1));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(make_xor_transform(make_tuple(number<MPerBlock / MLdsLayer>{},
+                                                         number<K0 * MLdsLayer>{})),
+                           make_pass_through_transform(K1)),
+                make_tuple(sequence<1, 0>{}, sequence<2>{}),
+                make_tuple(sequence<1, 0>{}, sequence<2>{}));
+
+            constexpr auto a_lds_block_desc_ak0_kMLdsLayer_m_ak1 = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(make_unmerge_transform(make_tuple(K0, number<MLdsLayer>{})),
+                           make_pass_through_transform(number<MPerBlock / MLdsLayer>{}),
+                           make_pass_through_transform(K1)),
+                make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
+                make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
+
+            constexpr auto a_lds_block_desc_m_k = transform_tensor_descriptor(
+                a_lds_block_desc_ak0_kMLdsLayer_m_ak1,
+                make_tuple(make_merge_transform_v3_division_mod(make_tuple(K0, K1)),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(number<MPerBlock / MLdsLayer>{}, number<MLdsLayer>{}))),
+                make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
+                make_tuple(sequence<1>{}, sequence<0>{}));
+
+            return a_lds_block_desc_m_k;
+        }
+        else // ColumnMajor A
+        {
+            // kfold and mpair dimension is not always required.
+            // more dimension in merge_transform increase the difficulty of generating immarg offset
+            // for compiler.
+            constexpr auto M0 = get_warp_size() * Problem::BlockGemmShape::BlockWarps::at(I0);
+            constexpr auto M1 = MPerBlock / M0;
+
+            constexpr auto KThreadWrite     = Problem::kBlockSize / M0;
+            constexpr auto K0PerThreadWrite = K0 / KThreadWrite;
+            constexpr auto KThreadRead      = 64 / WarpGemm::kM;
+            constexpr auto K0PerThreadRead  = K0 / KThreadRead;
+
+            constexpr auto kfold =
+                (K1 * M0 * sizeof(ADataType) > 128) ? 1 : 128 / (K1 * M0 * sizeof(ADataType));
+            constexpr auto KThreadReadPerm =
+                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
+                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
+                    : KThreadRead;
+
+            // 1<=mpair<=kN0
+            constexpr auto mpair = (K1 * WarpGemm::kM * sizeof(ADataType) > 128)
+                                       ? 1
+                                       : ((128 / (K1 * WarpGemm::kM * sizeof(ADataType))) > M0
+                                              ? M0
+                                              : 128 / (K1 * WarpGemm::kM * sizeof(ADataType)));
+
+            constexpr auto a_lds_block_desc = make_naive_tensor_descriptor_packed(
+                make_tuple(number<KThreadWrite / kfold / KThreadReadPerm>{},
+                           number<K0PerThreadWrite>{},
+                           number<KThreadReadPerm * M1>{},
+                           number<kfold * M0 / mpair>{},
+                           number<mpair>{},
+                           K1));
+
+            constexpr auto a_lds_block_desc_permuted = transform_tensor_descriptor(
+                a_lds_block_desc,
+                make_tuple(
+                    make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(number<K0PerThreadWrite>{}),
+                    make_xor_transform(
+                        make_tuple(number<KThreadReadPerm * M1>{}, number<kfold * M0 / mpair>{})),
+                    make_pass_through_transform(number<mpair>{}),
+                    make_pass_through_transform(K1)),
+                make_tuple(
+                    sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}),
+                make_tuple(
+                    sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}));
+
+            constexpr auto a_lds_block_desc_unmerged = transform_tensor_descriptor(
+                a_lds_block_desc_permuted,
+                make_tuple(
+                    make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(number<K0PerThreadWrite>{}),
+                    make_unmerge_transform(make_tuple(number<KThreadReadPerm>{}, number<M1>{})),
+                    make_unmerge_transform(make_tuple(number<kfold>{}, number<M0 / mpair>{})),
+                    make_pass_through_transform(number<mpair>{}),
+                    make_pass_through_transform(K1)),
+                make_tuple(sequence<0>{},
+                           sequence<1>{},
+                           sequence<2>{},
+                           sequence<3>{},
+                           sequence<4>{},
+                           sequence<5>{}),
+                make_tuple(sequence<1>{},
+                           sequence<2>{},
+                           sequence<0, 3>{},
+                           sequence<4, 5>{},
+                           sequence<6>{},
+                           sequence<7>{}));
+
+            constexpr auto a_lds_block_desc_m_k = transform_tensor_descriptor(
+                a_lds_block_desc_unmerged,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(number<KThreadReadPerm>{},
+                                          number<KThreadWrite / kfold / KThreadReadPerm>{},
+                                          number<kfold>{},
+                                          number<K0PerThreadWrite>{},
+                                          K1)),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(number<M0 / mpair>{}, number<mpair>{}, number<M1>{}))),
+                make_tuple(sequence<0, 1, 4, 2, 7>{}, sequence<5, 6, 3>{}),
+                make_tuple(sequence<1>{}, sequence<0>{}));
+
+            return a_lds_block_desc_m_k;
+        }
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeBLdsBlockDescriptor()
+    {
+        using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
+                                                typename Problem::BDataType,
+                                                typename Problem::CDataType,
+                                                Problem::BlockGemmShape::WarpTile::at(I0),
+                                                Problem::BlockGemmShape::WarpTile::at(I1),
+                                                Problem::BlockGemmShape::WarpTile::at(I2),
+                                                TransposeC>;
+
+        using BDataType = remove_cvref_t<typename Problem::BDataType>;
+
+        constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr index_t K1 = WarpGemm::kK;
+        constexpr index_t K0 = KPerBlock / K1;
+
+        if constexpr(std::is_same<tensor_layout::gemm::ColumnMajor, LayoutB>::value)
+        {
+            // NLdsLayer * K0 as logical Bank
+            constexpr auto NLdsLayer = 32 * 4 / KPerBlock / sizeof(BDataType) < 1
+                                           ? 1
+                                           : 32 * 4 / KPerBlock / sizeof(BDataType);
+            ;
+            constexpr auto b_lds_block_desc = make_naive_tensor_descriptor(
+                make_tuple(K0 * number<NLdsLayer>{}, number<NPerBlock / NLdsLayer>{}, K1),
+                make_tuple(K1, number<KPerBlock * NLdsLayer>{}, I1));
+
+            constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
+                b_lds_block_desc,
+                make_tuple(make_xor_transform(make_tuple(number<NPerBlock / NLdsLayer>{},
+                                                         number<K0 * NLdsLayer>{})),
+                           make_pass_through_transform(K1)),
+                make_tuple(sequence<1, 0>{}, sequence<2>{}),
+                make_tuple(sequence<1, 0>{}, sequence<2>{}));
+
+            constexpr auto b_lds_block_desc_bk0_kNLdsLayer_n_bk1 = transform_tensor_descriptor(
+                b_lds_block_desc_permuted,
+                make_tuple(make_unmerge_transform(make_tuple(K0, number<NLdsLayer>{})),
+                           make_pass_through_transform(number<NPerBlock / NLdsLayer>{}),
+                           make_pass_through_transform(K1)),
+                make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{}),
+                make_tuple(sequence<0, 2>{}, sequence<1>{}, sequence<3>{}));
+
+            constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
+                b_lds_block_desc_bk0_kNLdsLayer_n_bk1,
+                make_tuple(make_merge_transform_v3_division_mod(make_tuple(K0, K1)),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(number<NPerBlock / NLdsLayer>{}, number<NLdsLayer>{}))),
+                make_tuple(sequence<0, 3>{}, sequence<1, 2>{}),
+                make_tuple(sequence<1>{}, sequence<0>{}));
+
+            return b_lds_block_desc_n_k;
+        }
+        else // RowMajor B
+        {
+            constexpr auto N0 = get_warp_size() * Problem::BlockGemmShape::BlockWarps::at(I1);
+            constexpr auto N1 = NPerBlock / N0;
+
+            constexpr auto KThreadWrite     = Problem::kBlockSize / N0;
+            constexpr auto K0PerThreadWrite = K0 / KThreadWrite;
+            constexpr auto KThreadRead      = 64 / WarpGemm::kN;
+            constexpr auto K0PerThreadRead  = K0 / KThreadRead;
+
+            constexpr auto kfold =
+                (K1 * N0 * sizeof(BDataType) > 128) ? 1 : 128 / (K1 * N0 * sizeof(BDataType));
+            constexpr auto KThreadReadPerm =
+                (kfold * K0PerThreadWrite / K0PerThreadRead) > 1
+                    ? KThreadRead / (kfold * K0PerThreadWrite / K0PerThreadRead)
+                    : KThreadRead;
+
+            // 1<=npair<=kN0
+            constexpr auto npair = (K1 * WarpGemm::kN * sizeof(BDataType) > 128)
+                                       ? 1
+                                       : ((128 / (K1 * WarpGemm::kN * sizeof(BDataType))) > N0
+                                              ? N0
+                                              : 128 / (K1 * WarpGemm::kN * sizeof(BDataType)));
+
+            constexpr auto b_lds_block_desc = make_naive_tensor_descriptor_packed(
+                make_tuple(number<KThreadWrite / kfold / KThreadReadPerm>{},
+                           number<K0PerThreadWrite>{},
+                           number<KThreadReadPerm * N1>{},
+                           number<kfold * N0 / npair>{},
+                           number<npair>{},
+                           K1));
+
+            constexpr auto b_lds_block_desc_permuted = transform_tensor_descriptor(
+                b_lds_block_desc,
+                make_tuple(
+                    make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(number<K0PerThreadWrite>{}),
+                    make_xor_transform(
+                        make_tuple(number<KThreadReadPerm * N1>{}, number<kfold * N0 / npair>{})),
+                    make_pass_through_transform(number<npair>{}),
+                    make_pass_through_transform(K1)),
+                make_tuple(
+                    sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}),
+                make_tuple(
+                    sequence<0>{}, sequence<1>{}, sequence<2, 3>{}, sequence<4>{}, sequence<5>{}));
+
+            constexpr auto b_lds_block_desc_unmerged = transform_tensor_descriptor(
+                b_lds_block_desc_permuted,
+                make_tuple(
+                    make_pass_through_transform(number<KThreadWrite / kfold / KThreadReadPerm>{}),
+                    make_pass_through_transform(number<K0PerThreadWrite>{}),
+                    make_unmerge_transform(make_tuple(number<KThreadReadPerm>{}, number<N1>{})),
+                    make_unmerge_transform(make_tuple(number<kfold>{}, number<N0 / npair>{})),
+                    make_pass_through_transform(number<npair>{}),
+                    make_pass_through_transform(K1)),
+                make_tuple(sequence<0>{},
+                           sequence<1>{},
+                           sequence<2>{},
+                           sequence<3>{},
+                           sequence<4>{},
+                           sequence<5>{}),
+                make_tuple(sequence<1>{},
+                           sequence<2>{},
+                           sequence<0, 3>{},
+                           sequence<4, 5>{},
+                           sequence<6>{},
+                           sequence<7>{}));
+
+            constexpr auto b_lds_block_desc_n_k = transform_tensor_descriptor(
+                b_lds_block_desc_unmerged,
+                make_tuple(make_merge_transform_v3_division_mod(
+                               make_tuple(number<KThreadReadPerm>{},
+                                          number<KThreadWrite / kfold / KThreadReadPerm>{},
+                                          number<kfold>{},
+                                          number<K0PerThreadWrite>{},
+                                          K1)),
+                           make_merge_transform_v3_division_mod(
+                               make_tuple(number<N0 / npair>{}, number<npair>{}, number<N1>{}))),
+                make_tuple(sequence<0, 1, 4, 2, 7>{}, sequence<5, 6, 3>{}),
+                make_tuple(sequence<1>{}, sequence<0>{}));
+
+            return b_lds_block_desc_n_k;
+        }
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeA()
+    {
+        constexpr index_t smem_size_a = sizeof(typename Problem::ADataType) *
+                                        MakeALdsBlockDescriptor<Problem>().get_element_space_size();
+        return smem_size_a;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSizeB()
+    {
+        constexpr index_t smem_size_b = sizeof(typename Problem::BDataType) *
+                                        MakeBLdsBlockDescriptor<Problem>().get_element_space_size();
+        return smem_size_b;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize()
+    {
+        constexpr index_t smem_size_a = GetSmemSizeA<Problem>();
+        constexpr index_t smem_size_b = GetSmemSizeB<Problem>();
+        index_t smem_size             = 0;
+        smem_size += smem_size_a + smem_size_b;
+
+        return smem_size;
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeADramTileDistribution()
+    {
+        using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
+                                                typename Problem::BDataType,
+                                                typename Problem::CDataType,
+                                                Problem::BlockGemmShape::WarpTile::at(I0),
+                                                Problem::BlockGemmShape::WarpTile::at(I1),
+                                                Problem::BlockGemmShape::WarpTile::at(I2),
+                                                TransposeC>;
+
+        constexpr index_t BlockSize = Problem::kBlockSize;
+
+        constexpr index_t MPerBlock = Problem::BlockGemmShape::kM;
+        constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr index_t K1 = WarpGemm::kK;
+        constexpr index_t K0 = KPerBlock / K1;
+        constexpr index_t M2 = get_warp_size() / K0;
+
+        constexpr index_t M1 = BlockSize / get_warp_size();
+        static_assert(M2 != 0, "M2 is zero, which will lead to a division by zero error.");
+        static_assert(M1 != 0, "M1 is zero, which will lead to a division by zero error.");
+        constexpr index_t M0 = MPerBlock / (M2 * M1);
+
+        return make_static_tile_distribution(
+            tile_distribution_encoding<sequence<1>,
+                                       tuple<sequence<M0, M1, M2>, sequence<K0, K1>>,
+                                       tuple<sequence<1>, sequence<1, 2>>,
+                                       tuple<sequence<1>, sequence<2, 0>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 1>>{});
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto MakeBDramTileDistribution()
+    {
+        using WarpGemm = WarpGemmMfmaDispatcher<typename Problem::ADataType,
+                                                typename Problem::BDataType,
+                                                typename Problem::CDataType,
+                                                Problem::BlockGemmShape::WarpTile::at(I0),
+                                                Problem::BlockGemmShape::WarpTile::at(I1),
+                                                Problem::BlockGemmShape::WarpTile::at(I2),
+                                                TransposeC>;
+
+        constexpr index_t BlockSize = Problem::kBlockSize;
+
+        constexpr index_t NPerBlock = Problem::BlockGemmShape::kN;
+        constexpr index_t KPerBlock = Problem::BlockGemmShape::kK;
+
+        constexpr index_t K1 = WarpGemm::kK;
+        constexpr index_t K0 = KPerBlock / K1;
+        constexpr index_t N2 = get_warp_size() / K0;
+
+        constexpr index_t N1 = BlockSize / get_warp_size();
+        static_assert(N2 != 0, "M2 is zero, which will lead to a division by zero error.");
+        static_assert(N1 != 0, "M1 is zero, which will lead to a division by zero error.");
+        constexpr index_t N0 = NPerBlock / (N2 * N1);
+
+        return make_static_tile_distribution(
+            tile_distribution_encoding<sequence<1>,
+                                       tuple<sequence<N0, N1, N2>, sequence<K0, K1>>,
+                                       tuple<sequence<1>, sequence<1, 2>>,
+                                       tuple<sequence<1>, sequence<2, 0>>,
+                                       sequence<1, 2>,
+                                       sequence<0, 1>>{});
+    }
+
+    template <typename Problem>
+    CK_TILE_HOST_DEVICE static constexpr auto GetBlockGemm()
+    {
+        using AccDataType     = float;
+        using BlockWarps      = typename Problem::BlockGemmShape::BlockWarps;
+        using WarpTile        = typename Problem::BlockGemmShape::WarpTile;
+        using WarpGemm        = WarpGemmMfmaDispatcher<typename Problem::ADataType,
+                                                typename Problem::BDataType,
+                                                AccDataType,
+                                                WarpTile::at(I0),
+                                                WarpTile::at(I1),
+                                                WarpTile::at(I2),
+                                                TransposeC>;
+        using BlockGemmPolicy = BlockGemmASmemBSmemCRegV1CustomPolicy<typename Problem::ADataType,
+                                                                      typename Problem::BDataType,
+                                                                      typename Problem::CDataType,
+                                                                      BlockWarps,
+                                                                      WarpGemm>;
+        return BlockGemmASmemBSmemCRegV1<Problem, BlockGemmPolicy>{};
+    }
+};
+
+} // namespace ck_tile