Merge branch 'develop' into bwroblew/dl_fails_vec_size

.github/CODEOWNERS

+# Documentation files
+docs/* @saadrahim @LisaDelaney
+*.md  @saadrahim @LisaDelaney
+*.rst  @saadrahim @LisaDelaney
+# Header directory
+library/include/*  @saadrahim @LisaDelaney

CMakeLists.txt

 cmake_minimum_required(VERSION 3.14)
 
+# This has to be initialized before the project() command appears
+# Set the default of CMAKE_BUILD_TYPE to be release, unless user specifies with -D.  MSVC_IDE does not use CMAKE_BUILD_TYPE
+if( NOT MSVC_IDE AND NOT CMAKE_BUILD_TYPE )
+    set( CMAKE_BUILD_TYPE Release CACHE STRING "Choose the type of build, options are: None Debug Release RelWithDebInfo MinSizeRel." )
+endif()
+
+# Default installation path
+if(WIN32)
+    set(CMAKE_INSTALL_PREFIX "/opt/rocm/x86_64-w64-mingw32" CACHE PATH "")
+else()
+    set(CMAKE_INSTALL_PREFIX "/opt/rocm" CACHE PATH "")
+endif()
+
 set(version 1.1.0)
 # Check support for CUDA/HIP in Cmake
 project(composable_kernel VERSION ${version})
@@ -15,6 +28,12 @@ if (DTYPES)
     if (DTYPES MATCHES "fp8")
         add_definitions(-DCK_ENABLE_FP8)
         set(CK_ENABLE_FP8 "ON")
+        add_compile_options(-Wno-bit-int-extension)
+    endif()
+    if (DTYPES MATCHES "bf8")
+        add_definitions(-DCK_ENABLE_BF8)
+        set(CK_ENABLE_BF8 "ON")
+        add_compile_options(-Wno-bit-int-extension)
     endif()
     if (DTYPES MATCHES "fp16")
         add_definitions(-DCK_ENABLE_FP16)
@@ -34,8 +53,9 @@ if (DTYPES)
     endif()
     message("DTYPES macro set to ${DTYPES}")
 else()
-    add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
+    add_definitions(-DCK_ENABLE_INT8 -DCK_ENABLE_FP8 -DCK_ENABLE_BF8 -DCK_ENABLE_FP16 -DCK_ENABLE_FP32 -DCK_ENABLE_FP64 -DCK_ENABLE_BF16)
     set(CK_ENABLE_ALL_DTYPES "ON")
+    add_compile_options(-Wno-bit-int-extension) # enable fp8 and bf8
 endif()
 
 if(DL_KERNELS)
@@ -365,6 +385,10 @@ IF(IS_DIRECTORY "${PROJECT_SOURCE_DIR}/library/src/tensor_operation_instance/gpu
         #message("fp8 instance found!")
         set(add_inst 1)
     endif()
+    if("${cmake_instance}" MATCHES "DTYPES MATCHES \"bf8\" " AND DTYPES MATCHES "bf8")
+        #message("bf8 instance found!")
+        set(add_inst 1)
+    endif()
     if("${cmake_instance}" MATCHES "DTYPES MATCHES \"fp16\"" AND DTYPES MATCHES "fp16")
         #message("fp16 instance found!")
         set(add_inst 1)

Dockerfile

@@ -73,15 +73,8 @@ RUN wget -qO /usr/local/bin/ninja.gz https://github.com/ninja-build/ninja/releas
 RUN gunzip /usr/local/bin/ninja.gz
 RUN chmod a+x /usr/local/bin/ninja
 RUN git clone https://github.com/nico/ninjatracing.git
-RUN apt purge --auto-remove -y cmake
-RUN apt update
-RUN apt install -y software-properties-common lsb-release
-RUN apt clean all
-RUN wget -O - https://apt.kitware.com/keys/kitware-archive-latest.asc 2>/dev/null | gpg --dearmor - | tee /etc/apt/trusted.gpg.d/kitware.gpg >/dev/null
-RUN apt-add-repository "deb https://apt.kitware.com/ubuntu/ $(lsb_release -cs) main"
-RUN apt install -y kitware-archive-keyring
-RUN rm /etc/apt/trusted.gpg.d/kitware.gpg
-RUN apt install -y cmake
+# Update the cmake to the latest version
+RUN pip install --upgrade cmake
 
 # Setup ubsan environment to printstacktrace
 RUN ln -s /usr/bin/llvm-symbolizer-3.8 /usr/local/bin/llvm-symbolizer

Jenkinsfile

@@ -210,6 +210,9 @@ def cmake_build(Map conf=[:]){
     } else{
         setup_args = ' -DBUILD_DEV=On' + setup_args
     }
+    if (params.DL_KERNELS){
+        setup_args = setup_args + " -DDL_KERNELS=ON "
+    }
 
     if(build_type_debug){
         setup_args = " -DCMAKE_BUILD_TYPE=debug -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}'" + setup_args
@@ -367,8 +370,6 @@ def runCKProfiler(Map conf=[:]){
             withDockerContainer(image: image, args: dockerOpts + ' -v=/var/jenkins/:/var/jenkins') {
                 timeout(time: 24, unit: 'HOURS')
                 {
-                    //cmake_build(conf)
-                    //instead of building, just unstash the ckProfiler and install it
                     sh """
                         rm -rf build
                         mkdir build
@@ -614,7 +615,7 @@ 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=5.7;COMPILER_VERSION=rc1
                                               0 21 * * * % ROCMVERSION=5.6;COMPILER_VERSION=;COMPILER_COMMIT=
-                                              0 19 * * * % BUILD_DOCKER=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
+                                              0 19 * * * % BUILD_DOCKER=true;DL_KERNELS=true;COMPILER_VERSION=amd-stg-open;COMPILER_COMMIT=''' : ""
 
 pipeline {
     agent none
@@ -649,6 +650,10 @@ pipeline {
             name: "RUN_FULL_QA",
             defaultValue: false,
             description: "Select whether to run small set of performance tests (default) or full QA")
+        booleanParam(
+            name: "DL_KERNELS",
+            defaultValue: false,
+            description: "Select whether to build DL kernels (default: OFF)")
     }
     environment{
         dbuser = "${dbuser}"
@@ -663,15 +668,12 @@ pipeline {
     }
     stages{
         stage("Build Docker"){
-            //when {
-            //    beforeAgent true
-            //    expression { params.BUILD_DOCKER.toBoolean() }
-            //}
             parallel{
                 stage('Docker /opt/rocm'){
                     agent{ label rocmnode("nogpu") }
                     steps{
                         buildDocker('/opt/rocm')
+                        cleanWs()
                     }
                 }
             }
@@ -693,6 +695,7 @@ pipeline {
                     }
                     steps{
                         buildHipClangJobAndReboot(setup_cmd: "", build_cmd: "", execute_cmd: execute_cmd, no_reboot:true)
+                        cleanWs()
                     }
                 }
             }
@@ -715,6 +718,7 @@ pipeline {
                     }
                     steps{
                         Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
+                        cleanWs()
                     }
                 }
                 stage("Build CK and run Tests on MI100/MI200")
@@ -730,6 +734,7 @@ pipeline {
                     }
                     steps{
                         Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
+                        cleanWs()
                     }
                 }
                 stage("Build CK and run Tests on Navi21")
@@ -742,10 +747,10 @@ pipeline {
                     environment{
                         setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1030" -DDL_KERNELS=ON """ 
                         execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1030" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
-
                     }
                     steps{
                         Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
+                        cleanWs()
                     }
                 }
                 stage("Build CK and run Tests on Navi32")
@@ -756,12 +761,12 @@ pipeline {
                     }
                     agent{ label rocmnode("navi32") }
                     environment{
-                        setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DDTYPES="fp16;fp32;bf16" -DGPU_TARGETS="gfx1101" """
-                        execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1101" -DDTYPES="fp16;fp32;bf16" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
-
+                        setup_args = """ -DCMAKE_INSTALL_PREFIX=../install -DGPU_TARGETS="gfx1101" """
+                        execute_args = """ cd ../client_example && rm -rf build && mkdir build && cd build && cmake -D CMAKE_PREFIX_PATH="${env.WORKSPACE}/install;/opt/rocm" -DGPU_TARGETS="gfx1101" -D CMAKE_CXX_COMPILER="${build_compiler()}" .. && make -j """
                     }
                     steps{
                         Build_CK_and_Reboot(setup_args: setup_args, config_targets: "install", no_reboot:true, build_type: 'Release', execute_cmd: execute_args, prefixpath: '/usr/local')
+                        cleanWs()
                     }
                 }
             }
@@ -784,6 +789,7 @@ pipeline {
                    }
                     steps{
                         runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
+                        cleanWs()
                     }
                 }
                 stage("Run ckProfiler: gfx90a")
@@ -799,6 +805,7 @@ pipeline {
                     }
                     steps{
                         runPerfTest(setup_args:setup_args, config_targets: "ckProfiler", no_reboot:true, build_type: 'Release')
+                        cleanWs()
                     }
                 }
             }
@@ -811,6 +818,7 @@ pipeline {
                     agent { label 'mici' }
                     steps{
                         process_results()
+                        cleanWs()
                     }
                 }
             }

client_example/20_splitk_gemm/CMakeLists.txt

-add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
-target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_operations)
+if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
+  add_executable(client_splitK_gemm splitK_gemm_fp16_f8.cpp)
+  target_link_libraries(client_splitK_gemm PRIVATE composable_kernel::device_operations)
+endif()

client_example/22_grouped_gemm/CMakeLists.txt

+add_executable(client_grouped_gemm_fixed_nk_fp16 grouped_gemm_fixed_nk_fp16.cpp)
+target_link_libraries(client_grouped_gemm_fixed_nk_fp16 PRIVATE composable_kernel::device_operations)
+
+add_executable(client_grouped_gemm_fixed_nk_fp8 grouped_gemm_fixed_nk_fp8.cpp)
+target_link_libraries(client_grouped_gemm_fixed_nk_fp8 PRIVATE composable_kernel::device_operations)
+
+add_executable(client_grouped_gemm_fixed_nk_i8 grouped_gemm_fixed_nk_i8.cpp)
+target_link_libraries(client_grouped_gemm_fixed_nk_i8 PRIVATE composable_kernel::device_operations)

client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType  = F16;
+using BDataType  = F16;
+using DsDataType = ck::Tuple<>;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    int sum_of_m = 0;
+
+    // Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
+    Ms = {0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0};
+
+    int group_count = Ms.size();
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ns.push_back(768);
+        Ks.push_back(4608);
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+
+        sum_of_m += Ms[i];
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<void*> p_e;
+
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
+
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
+                                             b_dev_bufs[i].GetDeviceBuffer(),
+                                             {},
+                                             e_dev_bufs[i].GetDeviceBuffer(),
+                                             Ms[i],
+                                             Ns[i],
+                                             Ks[i],
+                                             StrideAs[i],
+                                             StrideBs[i],
+                                             {},
+                                             StrideEs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            DsLayout,
+                                                                            ELayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            DsDataType,
+                                                                            EDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    std::vector<const void*> p_a = {}, p_b = {};
+    std::vector<std::array<const void*, 0>> p_ds = {};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem grouped_gemm_kernel_args_dev(
+            op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(),
+                                    grouped_gemm_workspace_dev.GetDeviceBuffer());
+
+        op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                    grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        op_ptr->SetKBatch(argument_ptr.get(), 32);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}

client_example/22_grouped_gemm/grouped_gemm_fixed_nk_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType  = F16;
+using BDataType  = F8;
+using DsDataType = ck::Tuple<>;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    int sum_of_m = 0;
+
+    Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
+
+    int group_count = Ms.size();
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ns.push_back(768);
+        Ks.push_back(4608);
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+
+        sum_of_m += Ms[i];
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<void*> p_e;
+
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
+
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
+                                             b_dev_bufs[i].GetDeviceBuffer(),
+                                             {},
+                                             e_dev_bufs[i].GetDeviceBuffer(),
+                                             Ms[i],
+                                             Ns[i],
+                                             Ks[i],
+                                             StrideAs[i],
+                                             StrideBs[i],
+                                             {},
+                                             StrideEs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            DsLayout,
+                                                                            ELayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            DsDataType,
+                                                                            EDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    std::vector<const void*> p_a = {}, p_b = {};
+    std::vector<std::array<const void*, 0>> p_ds = {};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem grouped_gemm_kernel_args_dev(
+            op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(),
+                                    grouped_gemm_workspace_dev.GetDeviceBuffer());
+
+        op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                    grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        op_ptr->SetKBatch(argument_ptr.get(), 16);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}

client_example/22_grouped_gemm/grouped_gemm_fixed_nk_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iomanip>
+#include <iostream>
+#include <vector>
+#include <random>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/gpu/grouped_gemm_fixed_nk.hpp"
+
+using I8  = int8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType  = F16;
+using BDataType  = I8;
+using DsDataType = ck::Tuple<>;
+using EDataType  = F16;
+
+using ALayout  = Row;
+using BLayout  = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+    std::vector<int> Ms, Ns, Ks, StrideAs, StrideBs, StrideEs;
+
+    int sum_of_m = 0;
+
+    Ms = {167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
+
+    int group_count = Ms.size();
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        Ns.push_back(768);
+        Ks.push_back(4608);
+
+        StrideAs.push_back(std::is_same<Row, ALayout>::value ? Ks[i] : Ms[i]);
+        StrideBs.push_back(std::is_same<Row, BLayout>::value ? Ns[i] : Ks[i]);
+        StrideEs.push_back(std::is_same<Row, ELayout>::value ? Ns[i] : Ms[i]);
+
+        sum_of_m += Ms[i];
+    }
+
+    auto f_matrix_space_size =
+        [](std::size_t nRow, std::size_t nCol, std::size_t stride, auto layout) {
+            using Layout = decltype(layout);
+
+            if constexpr(std::is_same<Layout, ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return (nRow - 1) * stride + nCol;
+            }
+            else
+            {
+                return (nCol - 1) * stride + nRow;
+            }
+        };
+
+    std::vector<SimpleDeviceMem> a_dev_bufs, b_dev_bufs, e_dev_bufs;
+
+    a_dev_bufs.reserve(group_count);
+    b_dev_bufs.reserve(group_count);
+    e_dev_bufs.reserve(group_count);
+
+    std::vector<void*> p_e;
+
+    p_e.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+
+    gemm_descs.reserve(group_count);
+
+    std::vector<ck::tensor_operation::device::GroupedGemmKernelArgument<1>>
+        grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; ++i)
+    {
+        a_dev_bufs.emplace_back(sizeof(ADataType) *
+                                f_matrix_space_size(Ms[i], Ks[i], StrideAs[i], ALayout{}));
+        b_dev_bufs.emplace_back(sizeof(BDataType) *
+                                f_matrix_space_size(Ks[i], Ns[i], StrideBs[i], BLayout{}));
+        e_dev_bufs.emplace_back(sizeof(EDataType) *
+                                f_matrix_space_size(Ms[i], Ns[i], StrideEs[i], ELayout{}));
+
+        gemm_descs.push_back({sum_of_m, Ns[i], Ks[i], 1, StrideBs[i], 1, {0}});
+
+        p_e.push_back(e_dev_bufs[i].GetDeviceBuffer());
+
+        grouped_gemm_kernel_args_.push_back({a_dev_bufs[i].GetDeviceBuffer(),
+                                             b_dev_bufs[i].GetDeviceBuffer(),
+                                             {},
+                                             e_dev_bufs[i].GetDeviceBuffer(),
+                                             Ms[i],
+                                             Ns[i],
+                                             Ks[i],
+                                             StrideAs[i],
+                                             StrideBs[i],
+                                             {},
+                                             StrideEs[i]});
+    }
+
+    using DeviceOp = ck::tensor_operation::device::DeviceGroupedGemmFixedNK<ALayout,
+                                                                            BLayout,
+                                                                            DsLayout,
+                                                                            ELayout,
+                                                                            ADataType,
+                                                                            BDataType,
+                                                                            DsDataType,
+                                                                            EDataType,
+                                                                            AElementOp,
+                                                                            BElementOp,
+                                                                            CDEElementOp>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    const auto a_element_op   = AElementOp{};
+    const auto b_element_op   = BElementOp{};
+    const auto cde_element_op = CDEElementOp{};
+
+    std::string best_op_name;
+    bool found            = false;
+    int best_op_id        = -1;
+    float best_ave_time   = 0;
+    float best_tflops     = 0;
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    std::vector<const void*> p_a = {}, p_b = {};
+    std::vector<std::array<const void*, 0>> p_ds = {};
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr = op_ptrs[i];
+
+        auto argument_ptr = op_ptr->MakeArgumentPointer(
+            p_a, p_b, p_ds, p_e, gemm_descs, a_element_op, b_element_op, cde_element_op);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        SimpleDeviceMem grouped_gemm_kernel_args_dev(
+            op_ptr->GetDeviceKernelArgSize(argument_ptr.get()));
+
+        SimpleDeviceMem grouped_gemm_workspace_dev(op_ptr->GetWorkSpaceSize(argument_ptr.get()));
+
+        std::string op_name = op_ptr->GetTypeString();
+
+        hipGetErrorString(hipMemcpy(grouped_gemm_kernel_args_dev.GetDeviceBuffer(),
+                                    grouped_gemm_kernel_args_.data(),
+                                    op_ptr->GetDeviceKernelArgSize(argument_ptr.get()),
+                                    hipMemcpyHostToDevice));
+
+        op_ptr->SetWorkSpacePointer(argument_ptr.get(),
+                                    grouped_gemm_workspace_dev.GetDeviceBuffer());
+
+        op_ptr->SetDeviceKernelArgs(argument_ptr.get(),
+                                    grouped_gemm_kernel_args_dev.GetDeviceBuffer());
+
+        op_ptr->SetKBatch(argument_ptr.get(), 32);
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t flop = 0, num_btype = 0;
+            for(std::size_t j = 0; j < gemm_descs.size(); ++j)
+            {
+                flop += std::size_t(2) * Ms[j] * Ns[j] * Ks[j];
+
+                num_btype += sizeof(ADataType) * Ms[j] * Ks[j] + sizeof(BDataType) * Ks[j] * Ns[j] +
+                             sizeof(EDataType) * Ms[j] * Ns[j];
+            }
+
+            float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+            float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+            std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
+                      << gb_per_sec << " GB/s, " << op_name << std::endl;
+
+            if(tflops > best_tflops)
+            {
+                found           = true;
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_tflops     = tflops;
+                best_ave_time   = ave_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cout << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
+              << best_gb_per_sec << " GB/s, " << best_op_name << std::endl;
+
+    return 0;
+}

example/01_gemm/CMakeLists.txt

@@ -69,5 +69,7 @@ if(DTYPES MATCHES "fp8" OR NOT DEFINED DTYPES)
   endif()
 endif()
 
-add_example_executable(example_gemm_xdl_fp16_f8 gemm_xdl_fp16_f8.cpp)
-add_dependencies(example_gemm_xdl example_gemm_xdl_fp16_f8)
+if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES)
+  add_example_executable(example_gemm_xdl_fp16_f8 gemm_xdl_fp16_f8.cpp)
+  add_dependencies(example_gemm_xdl example_gemm_xdl_fp16_f8)
+endif()

example/15_grouped_gemm/CMakeLists.txt

@@ -25,6 +25,11 @@ if(DTYPES MATCHES "int8" OR NOT DEFINED DTYPES)
   add_example_executable(example_grouped_gemm_xdl_int8 grouped_gemm_xdl_int8.cpp)
   add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int8)
 endif()
+if(DTYPES MATCHES "f8" OR NOT DEFINED DTYPES)
+  add_example_executable(example_grouped_gemm_xdl_fixed_nk_fp8 grouped_gemm_xdl_fixed_nk_fp8.cpp)
+  add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_fixed_nk_fp8)
+endif()
+
 if(USE_BITINT_EXTENSION_INT4)
   add_example_executable(example_grouped_gemm_xdl_int4 grouped_gemm_xdl_int4.cpp)
   add_dependencies(example_grouped_gemm_xdl example_grouped_gemm_xdl_int4)

example/15_grouped_gemm/grouped_gemm_xdl_fixed_nk_fp8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+#include <numeric>
+#include <initializer_list>
+#include <cstdlib>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_gemm.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F8  = ck::f8_t;
+using F16 = ck::half_t;
+using F32 = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using ADataType        = F16;
+using BDataType        = F8;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = F16;
+
+using ALayout  = Row;
+using BLayout  = Col;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = PassThrough;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Fixed_NK
+    // clang-format off
+//######| ALayout| BLayout| DsLayout| ELayout|     AData|     BData|     AccData|         CShuffle|     DsData|     EData|           A|           B|          CDE|           GEMM| NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer| MXdl| NXdl|  ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle| CBlockTransferClusterLengths|  CBlockTransfer|
+//######|        |        |         |        |      Type|      Type|        Type|         DataType|       Type|      Type| Elementwise| Elementwise|  Elementwise| Spacialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
+//######|        |        |         |        |          |          |            |                 |           |          |   Operation|   Operation|    Operation|               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
+//######|        |        |         |        |          |          |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+        < ALayout, BLayout, DsLayout, ELayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   256,    64,   128,    32,   8,   8,   32,   32,    1,    2,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,              3,              8,              8,         1,  S<1, 4, 64, 1>,  S<0, 2, 1, 3>,  S<0, 2, 1, 3>,             3,              8,              8,         1,           1,           1,               S<1, 32, 1, 8>,              8>;
+// clang-format on
+
+struct ProblemSize final
+{
+    std::vector<ck::index_t> Ms;
+    std::vector<ck::index_t> Ns;
+    std::vector<ck::index_t> Ks;
+
+    std::vector<ck::index_t> stride_As;
+    std::vector<ck::index_t> stride_Bs;
+    std::vector<ck::index_t> stride_Cs;
+
+    ck::index_t group_count;
+};
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    int k_batch          = 1;
+    bool time_kernel     = false;
+};
+
+bool run_grouped_gemm(const ProblemSize& problem_size, const ExecutionConfig& config)
+{
+    auto group_count = problem_size.group_count;
+
+    // GEMM shape
+    std::vector<ck::tensor_operation::device::GemmDesc> gemm_descs;
+    std::vector<void*> p_Cs;
+
+    gemm_descs.reserve(group_count);
+
+    int sum_of_m = 0;
+
+    auto f_host_tensor_descriptor =
+        [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
+            using namespace ck::literals;
+
+            if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
+            {
+                return HostTensorDescriptor({row, col}, {stride, 1_uz});
+            }
+            else
+            {
+                return HostTensorDescriptor({row, col}, {1_uz, stride});
+            }
+        };
+
+    std::vector<Tensor<ADataType>> a_tensors;
+    std::vector<Tensor<BDataType>> b_tensors;
+    std::vector<Tensor<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> c_device_tensors;
+
+    a_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    c_host_tensors.reserve(group_count);
+    c_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a_tensors_device, b_tensors_device, c_tensors_device;
+
+    a_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    c_tensors_device.reserve(group_count);
+
+    std::size_t flop = 0, num_btype = 0;
+
+    for(int i = 0; i < group_count; i++)
+    {
+        sum_of_m += problem_size.Ms[i];
+        a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], ALayout{})));
+        b_tensors.push_back(Tensor<BDataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], BLayout{})));
+        c_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        c_device_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        std::cout << "gemm[" << i << "] a_m_k: " << a_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " c_m_n: " << c_device_tensors[i].mDesc
+                  << std::endl;
+
+        flop += std::size_t(2) * problem_size.Ms[i] * problem_size.Ks[i] * problem_size.Ns[i];
+        num_btype += sizeof(ADataType) * a_tensors[i].mDesc.GetElementSize() +
+                     sizeof(BDataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+            break;
+        case 2:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+            break;
+        default:
+            a_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+    }
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::GroupedGemmKernelArgument<>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(ADataType) * sum_of_m * problem_size.Ks[i]));
+
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(BDataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        a_tensors_device[i]->ToDevice(a_tensors[i].mData.data(),
+                                      a_tensors[i].mDesc.GetElementSpaceSize() * sizeof(ADataType));
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data(),
+                                      b_tensors[i].mDesc.GetElementSpaceSize() * sizeof(BDataType));
+        c_tensors_device[i]->SetZero();
+
+        p_Cs.push_back(c_tensors_device[i]->GetDeviceBuffer());
+
+        gemm_descs.push_back({sum_of_m,
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              1,
+                              problem_size.stride_Bs[i],
+                              1,
+                              {}});
+
+        grouped_gemm_kernel_args_.push_back({a_tensors_device[i]->GetDeviceBuffer(),
+                                             b_tensors_device[i]->GetDeviceBuffer(),
+                                             {},
+                                             c_tensors_device[i]->GetDeviceBuffer(),
+                                             problem_size.Ms[i],
+                                             problem_size.Ns[i],
+                                             problem_size.Ks[i],
+                                             problem_size.stride_As[i],
+                                             problem_size.stride_Bs[i],
+                                             {},
+                                             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op = AElementOp{};
+    auto b_element_op = BElementOp{};
+    auto c_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<const void*> p_As                = {};
+    std::vector<const void*> p_Bs                = {};
+    std::vector<std::array<const void*, 0>> p_Ds = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(
+        p_As, p_Bs, p_Ds, p_Cs, gemm_descs, a_element_op, b_element_op, c_element_op);
+
+    DeviceMem gemm_arg_dev_mem(gemm.GetDeviceKernelArgSize(&argument));
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
+
+    hip_check_error(hipMemcpy(gemm_arg_dev_mem.GetDeviceBuffer(),
+                              grouped_gemm_kernel_args_.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    gemm.SetDeviceKernelArgs(argument, gemm_arg_dev_mem.GetDeviceBuffer());
+    gemm.SetKBatch(argument, config.k_batch);
+
+    invoker.Run(argument, StreamConfig{nullptr, false});
+
+    if(config.time_kernel)
+    {
+        float ave_time   = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+        float tflops     = static_cast<float>(flop) / 1.E9 / ave_time;
+        float gb_per_sec = num_btype / 1.E6 / ave_time;
+
+        std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
+                  << " GB/s, " << gemm.GetTypeString() << std::endl;
+    }
+
+    bool pass = true;
+    if(config.do_verification)
+    {
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                AElementOp,
+                                                                                BElementOp,
+                                                                                CDEElementOp>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data(),
+                                            c_device_tensors[i].mDesc.GetElementSize() *
+                                                sizeof(EDataType));
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      a_element_op,
+                                                      b_element_op,
+                                                      c_element_op);
+
+            ref_invoker.Run(ref_argument);
+
+            pass &= ck::utils::check_err(c_device_tensors[i], c_host_tensors[i]);
+        }
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    problem_size.Ms = {
+        167, 183, 177, 181, 153, 139, 156, 173, 163, 150, 204, 184, 168, 156, 168, 148};
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ns.push_back(768);
+        problem_size.Ks.push_back(4608);
+
+        problem_size.stride_As.push_back(problem_size.Ks[i]);
+        problem_size.stride_Bs.push_back(problem_size.Ks[i]);
+        problem_size.stride_Cs.push_back(problem_size.Ns[i]);
+    }
+
+    if(argc == 5)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+        config.k_batch         = std::stoi(argv[4]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=n0, 1=yes)\n");
+        printf("arg4: k_batch (> 0)\n");
+        exit(0);
+    }
+
+    return !run_grouped_gemm(problem_size, config);
+}

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_dl_fp16.cpp

@@ -3,7 +3,7 @@
 
 #include "common.hpp"
 
-#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp"
 
 using InDataType  = F16;
 using WeiDataType = F16;
@@ -15,44 +15,55 @@ using WeiElementOp = PassThrough;
 using OutElementOp = PassThrough;
 
 template <ck::index_t NDimSpatial>
-using DeviceConvBwdWeightInstance =
-    ck::tensor_operation::device::DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl<
-        NDimSpatial,          // NDimSpatial
-        InDataType,           // InDataType
-        WeiDataType,          // WeiDataType
-        OutDataType,          // OutDataType
-        AccDataType,          // AccDataType
-        InElementOp,          // InElementwiseOperation
-        WeiElementOp,         // WeiElementwiseOperation
-        OutElementOp,         // OutElementwiseOperation
-        ConvBwdWeightDefault, // ConvBackwardWeightSpecialization
-        256,                  // BlockSize
-        128,                  // MPerBlock
-        128,                  // NPerBlock
-        16,                   // K0PerBlock
-        2,                    // K1
-        4,                    // M1PerThread
-        4,                    // N1PerThread
-        1,                    // KPerThread
-        S<8, 2>,              // M1N1ThreadClusterM1Xs
-        S<8, 2>,              // M1N1ThreadClusterN1Xs
-        S<1, 8, 1, 1, 2>,     // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
-        S<1, 2, 1, 128, 1>,   // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
-        S<0, 2, 3, 1, 4>,     // ABlockTransferThreadClusterArrangeOrder
-        S<0, 2, 3, 1, 4>,     // ABlockTransferSrcAccessOrder
-        S<1, 1, 1, 1, 1>,     // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
-        S<0, 2, 3, 1, 4>,     // ABlockTransferSrcVectorTensorContiguousDimOrder
-        S<1, 1, 1, 1, 1>,     // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
-        S<1, 1, 1, 8, 2>,     // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
-        S<1, 16, 1, 16, 1>,   // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
-        S<0, 1, 4, 2, 3>,     // BBlockTransferThreadClusterArrangeOrder
-        S<0, 1, 4, 2, 3>,     // BBlockTransferSrcAccessOrder
-        S<1, 1, 1, 8, 1>,     // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
-        S<0, 1, 4, 2, 3>,     // BBlockTransferSrcVectorTensorContiguousDimOrder
-        S<1, 1, 1, 1, 2>,     // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
-        S<0, 1, 2, 3, 4, 5>,  // CThreadTransferSrcDstAccessOrder
-        5,                    // CThreadTransferSrcDstVectorDim
-        4>;                   // CThreadTransferDstScalarPerVector
+using DeviceConvBwdWeightInstance = ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Dl<
+    NDimSpatial, // NDimSpatial
+    ck::tuple_element_t<NDimSpatial - 1,
+                        ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                  ck::tensor_layout::convolution::GNHWC,
+                                  ck::tensor_layout::convolution::GNDHWC>>, // InLayout
+    ck::tuple_element_t<NDimSpatial - 1,
+                        ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                  ck::tensor_layout::convolution::GKYXC,
+                                  ck::tensor_layout::convolution::GKZYXC>>, // WeiLayout
+    ck::tuple_element_t<NDimSpatial - 1,
+                        ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                  ck::tensor_layout::convolution::GNHWK,
+                                  ck::tensor_layout::convolution::GNDHWK>>, // OutLayout
+    InDataType,                                                             // InDataType
+    WeiDataType,                                                            // WeiDataType
+    OutDataType,                                                            // OutDataType
+    AccDataType,                                                            // AccDataType
+    InElementOp,          // InElementwiseOperation
+    WeiElementOp,         // WeiElementwiseOperation
+    OutElementOp,         // OutElementwiseOperation
+    ConvBwdWeightDefault, // ConvBackwardWeightSpecialization
+    256,                  // BlockSize
+    128,                  // MPerBlock
+    128,                  // NPerBlock
+    16,                   // K0PerBlock
+    2,                    // K1
+    4,                    // M1PerThread
+    4,                    // N1PerThread
+    1,                    // KPerThread
+    S<8, 2>,              // M1N1ThreadClusterM1Xs
+    S<8, 2>,              // M1N1ThreadClusterN1Xs
+    S<1, 8, 1, 1, 2>,     // ABlockTransferThreadSliceLengths_K0_M0_M1_K1
+    S<1, 2, 1, 128, 1>,   // ABlockTransferThreadClusterLengths_K0_M0_M1_K1
+    S<0, 2, 3, 1, 4>,     // ABlockTransferThreadClusterArrangeOrder
+    S<0, 2, 3, 1, 4>,     // ABlockTransferSrcAccessOrder
+    S<1, 1, 1, 1, 1>,     // ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1
+    S<0, 2, 3, 1, 4>,     // ABlockTransferSrcVectorTensorContiguousDimOrder
+    S<1, 1, 1, 1, 1>,     // ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1
+    S<1, 1, 1, 8, 2>,     // BBlockTransferThreadSliceLengths_K0_N0_N1_K1
+    S<1, 16, 1, 16, 1>,   // BBlockTransferThreadClusterLengths_K0_N0_N1_K1
+    S<0, 1, 4, 2, 3>,     // BBlockTransferThreadClusterArrangeOrder
+    S<0, 1, 4, 2, 3>,     // BBlockTransferSrcAccessOrder
+    S<1, 1, 1, 8, 1>,     // BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1
+    S<0, 1, 4, 2, 3>,     // BBlockTransferSrcVectorTensorContiguousDimOrder
+    S<1, 1, 1, 1, 2>,     // BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1
+    S<0, 1, 2, 3, 4, 5>,  // CThreadTransferSrcDstAccessOrder
+    5,                    // CThreadTransferSrcDstVectorDim
+    4>;                   // CThreadTransferDstScalarPerVector
 
 #include "run_grouped_conv_bwd_weight_example.inc"
 

example/20_grouped_conv_bwd_weight/run_grouped_conv_bwd_weight_example.inc

@@ -14,20 +14,8 @@ template <ck::index_t NDimSpatial>
 bool run_grouped_conv_bwd_weight(const ExecutionConfig& config,
                                  const ck::utils::conv::ConvParam& conv_param)
 {
-    ck::index_t split_k;
-    // Set split_k = 2 for xdl op, split_k = 1 for dl
     // Dl op doesn't support split_k > 1
-    // TODO: Add Dl op split_k > 1 support
-    if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
-         ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
-         ck::get_device_name() == "gfx1102"))
-    {
-        split_k = 2;
-    }
-    else
-    {
-        split_k = 1;
-    }
+    constexpr ck::index_t split_k = 1;
 
     const auto in_g_n_c_wis_desc =
         ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<

example/42_groupnorm/groupnorm_sigmoid_mul_fp16.cpp

@@ -14,18 +14,22 @@ using ComputeDataType = float;
 
 struct YElementOp
 {
-    template <typename T>
-    __host__ __device__ void operator()(T& y, const T& x) const
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
     {
-        static_assert(ck::is_same<T, float>::value || ck::is_same<T, double>::value ||
-                          ck::is_same<T, ck::half_t>::value,
+        static_assert(ck::is_same<X, float>::value || ck::is_same<X, double>::value ||
+                          ck::is_same<X, ck::half_t>::value,
                       "Data type is not supported by this operation!");
 
-        T a;
+        static_assert(ck::is_same<Y, float>::value || ck::is_same<Y, double>::value ||
+                          ck::is_same<Y, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        X a;
 
         ck::tensor_operation::element_wise::Sigmoid{}(a, x);
 
-        y = x * a;
+        y = ck::type_convert<Y>(x * a);
     };
 };
 

include/ck/config.h.in

@@ -43,6 +43,9 @@
 #ifndef CK_ENABLE_FP8
 #define CK_ENABLE_FP8 "ON"
 #endif
+#ifndef CK_ENABLE_BF8
+#define CK_ENABLE_BF8 "ON"
+#endif
 #ifndef CK_ENABLE_FP16
 #define CK_ENABLE_FP16 "ON"
 #endif
@@ -66,6 +69,10 @@
 #cmakedefine CK_ENABLE_FP8 @CK_ENABLE_FP8@
 #endif
 
+#ifndef CK_ENABLE_BF8
+#cmakedefine CK_ENABLE_BF8 @CK_ENABLE_BF8@
+#endif
+
 #ifndef CK_ENABLE_FP16
 #cmakedefine CK_ENABLE_FP16 @CK_ENABLE_FP16@
 #endif

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

@@ -144,7 +144,8 @@ template <typename ALayout,
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEBlockTransferScalarPerVector_NPerBlock,
           LoopScheduler LoopSched     = make_default_loop_scheduler(),
-          PipelineVersion PipelineVer = PipelineVersion::v1>
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeDataType    = EDataType>
 struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
                                                                      BLayout,
                                                                      DsLayout,
@@ -243,11 +244,9 @@ struct DeviceGemmMultipleD_Xdl_CShuffle : public DeviceGemmMultipleD<ALayout,
     using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
     using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
 
-    using ComputeDataType = EDataType;
-
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
-        ADataType, // TODO: distinguish A/B datatype
+        ADataType,
         BDataType,
         ComputeDataType,
         AccDataType,

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_gnwc_gkxc_gnwk_dl.hpp → include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp

@@ -14,6 +14,7 @@
 #include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
 #include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 
@@ -72,6 +73,9 @@ __global__ void
             const Block2CTileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__) ||           \
+    defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx940__) || defined(__gfx1100__) || \
+    defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx941__) || defined(__gfx942__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
@@ -96,9 +100,27 @@ __global__ void
         block_2_ctile_map,
         integral_constant<bool, HasMainKBlockLoop>{},
         integral_constant<bool, HasDoubleTailKBlockLoop>{});
+#else
+    ignore = p_a_grid;
+    ignore = p_b_grid;
+    ignore = p_c_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_kbatch_k0_m0_m1_k1;
+    ignore = b_grid_desc_kbatch_k0_n0_n1_k1;
+    ignore = c_grid_desc_m0_m10_m11_n0_n10_n11;
+    ignore = block_2_ctile_map;
+    ignore = compute_ptr_offset_of_batch;
+
+    compute_ptr_offset_of_batch.GetAPtrOffset(0);
+    compute_ptr_offset_of_batch.GetBPtrOffset(0);
+    compute_ptr_offset_of_batch.GetCPtrOffset(0);
+#endif
 }
 
 template <ck::index_t NDimSpatial,
+          typename InLayout,
+          typename WeiLayout,
+          typename OutLayout,
           typename InDataType,
           typename WeiDataType,
           typename OutDataType,
@@ -134,29 +156,46 @@ template <ck::index_t NDimSpatial,
           typename CThreadTransferSrcDstAccessOrder,
           index_t CThreadTransferSrcDstVectorDim,
           index_t CThreadTransferDstScalarPerVector>
-struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
-    : public DeviceGroupedConvBwdWeight<
-          NDimSpatial,
-          ck::tuple_element_t<NDimSpatial - 1,
-                              ck::Tuple<ck::tensor_layout::convolution::GNWC,
-                                        ck::tensor_layout::convolution::GNHWC,
-                                        ck::tensor_layout::convolution::GNDHWC>>,
-          ck::tuple_element_t<NDimSpatial - 1,
-                              ck::Tuple<ck::tensor_layout::convolution::GKXC,
-                                        ck::tensor_layout::convolution::GKYXC,
-                                        ck::tensor_layout::convolution::GKZYXC>>,
-          ck::tuple_element_t<NDimSpatial - 1,
-                              ck::Tuple<ck::tensor_layout::convolution::GNWK,
-                                        ck::tensor_layout::convolution::GNHWK,
-                                        ck::tensor_layout::convolution::GNDHWK>>,
-          InDataType,
-          WeiDataType,
-          OutDataType,
-          InElementwiseOperation,
-          WeiElementwiseOperation,
-          OutElementwiseOperation>
+struct DeviceGroupedConvBwdWeight_Dl : public DeviceGroupedConvBwdWeight<NDimSpatial,
+                                                                         InLayout,
+                                                                         WeiLayout,
+                                                                         OutLayout,
+                                                                         InDataType,
+                                                                         WeiDataType,
+                                                                         OutDataType,
+                                                                         InElementwiseOperation,
+                                                                         WeiElementwiseOperation,
+                                                                         OutElementwiseOperation>
 {
-    using DeviceOp = DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl;
+    // 1d
+    static constexpr bool is_NWGK_GKXC_NWGC =
+        is_same_v<InLayout, tensor_layout::convolution::NWGC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::NWGK>;
+    static constexpr bool is_GNWK_GKXC_GNWC =
+        is_same_v<InLayout, tensor_layout::convolution::GNWC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::GNWK>;
+    // 2d
+    static constexpr bool is_NHWGK_GKYXC_NHWGC =
+        is_same_v<InLayout, tensor_layout::convolution::NHWGC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKYXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::NHWGK>;
+    static constexpr bool is_GNHWK_GKYXC_GNHWC =
+        is_same_v<InLayout, tensor_layout::convolution::GNHWC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKYXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::GNHWK>;
+    // 3d
+    static constexpr bool is_NDHWGK_GKZYXC_NDHWGC =
+        is_same_v<InLayout, tensor_layout::convolution::NDHWGC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKZYXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::NDHWGK>;
+    static constexpr bool is_GNDHWK_GKZYXC_GNDHWC =
+        is_same_v<InLayout, tensor_layout::convolution::GNDHWC> &&
+        is_same_v<WeiLayout, tensor_layout::convolution::GKZYXC> &&
+        is_same_v<OutLayout, tensor_layout::convolution::GNDHWK>;
+
+    using DeviceOp = DeviceGroupedConvBwdWeight_Dl;
 
     using ADataType = OutDataType;
     using BDataType = InDataType;
@@ -176,6 +215,8 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     static constexpr auto I4 = Number<4>{};
     static constexpr auto I5 = Number<5>{};
 
+    static constexpr auto spatial_offset = I3;
+
     static constexpr auto K1Number     = Number<K1>{};
     static constexpr auto GemmK1Number = K1Number;
 
@@ -195,12 +236,12 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
 
     template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
     static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
-        const ck::index_t N,
-        const ck::index_t K,
-        const ck::index_t C,
-        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
         const std::array<ck::index_t, NDimSpatial>& input_left_pads,
@@ -209,90 +250,102 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     {
         using namespace ck;
 
-        const index_t Wi            = input_spatial_lengths[0];
-        const index_t Wo            = output_spatial_lengths[0];
-        const index_t X             = filter_spatial_lengths[0];
-        const index_t InLeftPadW    = input_left_pads[0];
-        const index_t InRightPadW   = input_right_pads[0];
-        const index_t ConvStrideW   = conv_filter_strides[0];
-        const index_t ConvDilationW = conv_filter_dilations[0];
+        const index_t N             = a_g_n_c_wis_lengths[I1];
+        const index_t K             = b_g_k_c_xs_lengths[I1];
+        const index_t C             = a_g_n_c_wis_lengths[I2];
+        const index_t Wi            = a_g_n_c_wis_lengths[spatial_offset];
+        const index_t Wo            = e_g_n_k_wos_lengths[spatial_offset];
+        const index_t X             = b_g_k_c_xs_lengths[spatial_offset];
+        const index_t InLeftPadW    = input_left_pads[I0];
+        const index_t InRightPadW   = input_right_pads[I0];
+        const index_t ConvStrideW   = conv_filter_strides[I0];
+        const index_t ConvDilationW = conv_filter_dilations[I0];
+
+        const auto InNStride  = a_g_n_c_wis_strides[I1];
+        const auto InCStride  = a_g_n_c_wis_strides[I2];
+        const auto InWStride  = a_g_n_c_wis_strides[spatial_offset];
+        const auto WeiKStride = b_g_k_c_xs_strides[I1];
+        const auto WeiCStride = b_g_k_c_xs_strides[I2];
+        const auto OutKStride = e_g_n_k_wos_strides[I2];
+        const auto OutWStride = e_g_n_k_wos_strides[spatial_offset];
 
         const index_t GemmKTotal = N * Wo;
-        const index_t GemmM      = K;
-        const index_t GemmN      = C * X;
-
         const index_t GemmKBatch = batch_k;
         const index_t GemmK0 =
             math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
             K0PerBlock;
-        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
 
         if constexpr(ConvBackwardWeightSpecialization ==
                      ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
         {
             // A: output tensor
-            const auto out_gemmktotal_gemmm_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
+            const auto out_gemmktotal_gemmm_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * Wo, K), make_tuple(OutWStride, OutKStride));
 
-            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
-                out_gemmktotal_gemmm_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmM)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto out_gemmkpad_gemmmpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    out_gemmktotal_gemmm_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, MPerBlock),
+                    Sequence<true, true>{});
 
             const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-                out_gemmkpad_gemmm_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmM)),
+                out_gemmkpad_gemmmpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(out_gemmkpad_gemmmpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
             // B: input tensor
-            const auto in_gemmktotal_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N * Wi, C));
+            const auto in_gemmktotal_gemmn_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * Wi, C), make_tuple(InWStride, InCStride));
 
-            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
-                in_gemmktotal_gemmn_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmM)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto in_gemmkpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    in_gemmktotal_gemmn_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, NPerBlock),
+                    Sequence<true, true>{});
 
             const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-                in_gemmkpad_gemmn_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmM)),
+                in_gemmkpad_gemmnpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(in_gemmkpad_gemmnpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
             // C: weights tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
+            const auto wei_gemmm_gemmn_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(K, X * C), make_tuple(WeiKStride, WeiCStride));
+
+            const auto wei_gemmmpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(wei_gemmm_gemmn_grid_desc,
+                                                                  make_tuple(MPerBlock, NPerBlock),
+                                                                  Sequence<true, true>{});
 
             return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+                              wei_gemmmpad_gemmnpad_grid_desc);
         }
         else
         {
-            const auto out_gemmktotal_gemmm_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N * Wo, K));
-            const auto in_n_wi_c_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+            const auto out_gemmktotal_gemmm_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * Wo, K), make_tuple(OutWStride, OutKStride));
+            const auto in_n_wi_c_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Wi, C), make_tuple(InNStride, InWStride, InCStride));
 
             // A: output tensor
-            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
-                out_gemmktotal_gemmm_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmM)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto out_gemmkpad_gemmmpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    out_gemmktotal_gemmm_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, MPerBlock),
+                    Sequence<true, true>{});
 
             const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-                out_gemmkpad_gemmm_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmM)),
+                out_gemmkpad_gemmmpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(out_gemmkpad_gemmmpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
@@ -321,38 +374,43 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
                                             make_tuple(Sequence<1, 3>{}, Sequence<0, 2>{}),
                                             make_tuple(Sequence<1>{}, Sequence<0>{}));
 
-            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
-                in_gemmktotal_gemmn_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmN)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto in_gemmkpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    in_gemmktotal_gemmn_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, NPerBlock),
+                    Sequence<true, true>{});
 
             const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-                in_gemmkpad_gemmn_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmN)),
+                in_gemmkpad_gemmnpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(in_gemmkpad_gemmnpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
             // C: weight tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
+            const auto wei_gemmm_gemmn_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(K, X * C), make_tuple(WeiKStride, WeiCStride));
+
+            const auto wei_gemmmpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(wei_gemmm_gemmn_grid_desc,
+                                                                  make_tuple(MPerBlock, NPerBlock),
+                                                                  Sequence<true, true>{});
 
             return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+                              wei_gemmmpad_gemmnpad_grid_desc);
         }
 
     } // function end
     template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
     static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
-        const ck::index_t N,
-        const ck::index_t K,
-        const ck::index_t C,
-        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
         const std::array<ck::index_t, NDimSpatial>& input_left_pads,
@@ -361,103 +419,111 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     {
         using namespace ck;
 
-        const index_t Hi = input_spatial_lengths[0];
-        const index_t Wi = input_spatial_lengths[1];
-
-        const index_t Ho = output_spatial_lengths[0];
-        const index_t Wo = output_spatial_lengths[1];
-
-        const index_t Y = filter_spatial_lengths[0];
-        const index_t X = filter_spatial_lengths[1];
-
-        const index_t InLeftPadH = input_left_pads[0];
-        const index_t InLeftPadW = input_left_pads[1];
-
-        const index_t InRightPadH = input_right_pads[0];
-        const index_t InRightPadW = input_right_pads[1];
-
-        const index_t ConvStrideH = conv_filter_strides[0];
-        const index_t ConvStrideW = conv_filter_strides[1];
-
-        const index_t ConvDilationH = conv_filter_dilations[0];
-        const index_t ConvDilationW = conv_filter_dilations[1];
+        const index_t N  = a_g_n_c_wis_lengths[I1];
+        const index_t K  = b_g_k_c_xs_lengths[I1];
+        const index_t C  = a_g_n_c_wis_lengths[I2];
+        const index_t Hi = a_g_n_c_wis_lengths[spatial_offset];
+        const index_t Wi = a_g_n_c_wis_lengths[spatial_offset + I1];
+        const index_t Ho = e_g_n_k_wos_lengths[spatial_offset];
+        const index_t Wo = e_g_n_k_wos_lengths[spatial_offset + I1];
+        const index_t Y  = b_g_k_c_xs_lengths[spatial_offset];
+        const index_t X  = b_g_k_c_xs_lengths[spatial_offset + I1];
+
+        const index_t InLeftPadH    = input_left_pads[I0];
+        const index_t InLeftPadW    = input_left_pads[I1];
+        const index_t InRightPadH   = input_right_pads[I0];
+        const index_t InRightPadW   = input_right_pads[I1];
+        const index_t ConvStrideH   = conv_filter_strides[I0];
+        const index_t ConvStrideW   = conv_filter_strides[I1];
+        const index_t ConvDilationH = conv_filter_dilations[I0];
+        const index_t ConvDilationW = conv_filter_dilations[I1];
+
+        const auto InNStride  = a_g_n_c_wis_strides[I1];
+        const auto InCStride  = a_g_n_c_wis_strides[I2];
+        const auto InHStride  = a_g_n_c_wis_strides[spatial_offset];
+        const auto InWStride  = a_g_n_c_wis_strides[spatial_offset + I1];
+        const auto WeiKStride = b_g_k_c_xs_strides[I1];
+        const auto WeiCStride = b_g_k_c_xs_strides[I2];
+        const auto OutKStride = e_g_n_k_wos_strides[I2];
+        const auto OutWStride = e_g_n_k_wos_strides[spatial_offset + I1];
 
         const index_t GemmKTotal = N * Ho * Wo;
-        const index_t GemmM      = K;
-        const index_t GemmN      = C * X * Y;
-
         const index_t GemmKBatch = batch_k;
         const index_t GemmK0 =
             math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
             K0PerBlock;
-        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
 
         if constexpr(ConvBackwardWeightSpecialization ==
                      ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
         {
             // A: output tensor
-            const auto out_gemmktotal_gemmm_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+            const auto out_gemmktotal_gemmm_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * Ho * Wo, K), make_tuple(OutWStride, OutKStride));
 
-            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
-                out_gemmktotal_gemmm_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmM)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto out_gemmkpad_gemmmpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    out_gemmktotal_gemmm_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, MPerBlock),
+                    Sequence<true, true>{});
 
             const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-                out_gemmkpad_gemmm_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmM)),
+                out_gemmkpad_gemmmpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(out_gemmkpad_gemmmpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
             // B: input tensor
-            const auto in_gemmktotal_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N * Hi * Wi, C));
+            const auto in_gemmktotal_gemmn_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * Hi * Wi, C), make_tuple(InWStride, InCStride));
 
-            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
-                in_gemmktotal_gemmn_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmM)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto in_gemmkpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    in_gemmktotal_gemmn_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, NPerBlock),
+                    Sequence<true, true>{});
 
             const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-                in_gemmkpad_gemmn_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmM)),
+                in_gemmkpad_gemmnpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(in_gemmkpad_gemmnpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
             // C: weight tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+            const auto wei_gemmm_gemmn_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(K, Y * X * C), make_tuple(WeiKStride, WeiCStride));
+
+            const auto wei_gemmmpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(wei_gemmm_gemmn_grid_desc,
+                                                                  make_tuple(MPerBlock, NPerBlock),
+                                                                  Sequence<true, true>{});
 
             return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+                              wei_gemmmpad_gemmnpad_grid_desc);
         }
         else
         {
-            const auto out_gemmktotal_gemmm_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
-            const auto in_n_hi_wi_c_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+            const auto out_gemmktotal_gemmm_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * Ho * Wo, K), make_tuple(OutWStride, OutKStride));
+            const auto in_n_hi_wi_c_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Hi, Wi, C), make_tuple(InNStride, InHStride, InWStride, InCStride));
 
             // A: output tensor
-            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
-                out_gemmktotal_gemmm_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmM)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto out_gemmkpad_gemmmpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    out_gemmktotal_gemmm_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, MPerBlock),
+                    Sequence<true, true>{});
 
             const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-                out_gemmkpad_gemmm_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmM)),
+                out_gemmkpad_gemmmpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(out_gemmkpad_gemmmpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
@@ -488,39 +554,44 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
                                             make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
                                             make_tuple(Sequence<1>{}, Sequence<0>{}));
 
-            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
-                in_gemmktotal_gemmn_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmN)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto in_gemmkpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    in_gemmktotal_gemmn_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, NPerBlock),
+                    Sequence<true, true>{});
 
             const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-                in_gemmkpad_gemmn_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmN)),
+                in_gemmkpad_gemmnpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(in_gemmkpad_gemmnpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
             // C: weight tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+            const auto wei_gemmm_gemmn_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(K, Y * X * C), make_tuple(WeiKStride, WeiCStride));
+
+            const auto wei_gemmmpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(wei_gemmm_gemmn_grid_desc,
+                                                                  make_tuple(MPerBlock, NPerBlock),
+                                                                  Sequence<true, true>{});
 
             return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+                              wei_gemmmpad_gemmnpad_grid_desc);
         }
 
     } // function end
 
     template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
     static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
-        const ck::index_t N,
-        const ck::index_t K,
-        const ck::index_t C,
-        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
         const std::array<ck::index_t, NDimSpatial>& input_left_pads,
@@ -529,110 +600,120 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     {
         using namespace ck;
 
-        const index_t Di = input_spatial_lengths[0];
-        const index_t Hi = input_spatial_lengths[1];
-        const index_t Wi = input_spatial_lengths[2];
-
-        const index_t Do = output_spatial_lengths[0];
-        const index_t Ho = output_spatial_lengths[1];
-        const index_t Wo = output_spatial_lengths[2];
-
-        const index_t Z = filter_spatial_lengths[0];
-        const index_t Y = filter_spatial_lengths[1];
-        const index_t X = filter_spatial_lengths[2];
-
-        const index_t InLeftPadD = input_left_pads[0];
-        const index_t InLeftPadH = input_left_pads[1];
-        const index_t InLeftPadW = input_left_pads[2];
-
-        const index_t InRightPadD = input_right_pads[0];
-        const index_t InRightPadH = input_right_pads[1];
-        const index_t InRightPadW = input_right_pads[2];
-
-        const index_t ConvStrideD = conv_filter_strides[0];
-        const index_t ConvStrideH = conv_filter_strides[1];
-        const index_t ConvStrideW = conv_filter_strides[2];
-
-        const index_t ConvDilationD = conv_filter_dilations[0];
-        const index_t ConvDilationH = conv_filter_dilations[1];
-        const index_t ConvDilationW = conv_filter_dilations[2];
+        const index_t N  = a_g_n_c_wis_lengths[I1];
+        const index_t K  = b_g_k_c_xs_lengths[I1];
+        const index_t C  = a_g_n_c_wis_lengths[I2];
+        const index_t Di = a_g_n_c_wis_lengths[spatial_offset + I0];
+        const index_t Hi = a_g_n_c_wis_lengths[spatial_offset + I1];
+        const index_t Wi = a_g_n_c_wis_lengths[spatial_offset + I2];
+        const index_t Do = e_g_n_k_wos_lengths[spatial_offset + I0];
+        const index_t Ho = e_g_n_k_wos_lengths[spatial_offset + I1];
+        const index_t Wo = e_g_n_k_wos_lengths[spatial_offset + I2];
+        const index_t Z  = b_g_k_c_xs_lengths[spatial_offset + I0];
+        const index_t Y  = b_g_k_c_xs_lengths[spatial_offset + I1];
+        const index_t X  = b_g_k_c_xs_lengths[spatial_offset + I2];
+
+        const index_t InLeftPadD    = input_left_pads[I0];
+        const index_t InLeftPadH    = input_left_pads[I1];
+        const index_t InLeftPadW    = input_left_pads[I2];
+        const index_t InRightPadD   = input_right_pads[I0];
+        const index_t InRightPadH   = input_right_pads[I1];
+        const index_t InRightPadW   = input_right_pads[I2];
+        const index_t ConvStrideD   = conv_filter_strides[I0];
+        const index_t ConvStrideH   = conv_filter_strides[I1];
+        const index_t ConvStrideW   = conv_filter_strides[I2];
+        const index_t ConvDilationD = conv_filter_dilations[I0];
+        const index_t ConvDilationH = conv_filter_dilations[I1];
+        const index_t ConvDilationW = conv_filter_dilations[I2];
+
+        const auto InNStride  = a_g_n_c_wis_strides[I1];
+        const auto InCStride  = a_g_n_c_wis_strides[I2];
+        const auto InDStride  = a_g_n_c_wis_strides[spatial_offset];
+        const auto InHStride  = a_g_n_c_wis_strides[spatial_offset + I1];
+        const auto InWStride  = a_g_n_c_wis_strides[spatial_offset + I2];
+        const auto WeiKStride = b_g_k_c_xs_strides[I1];
+        const auto WeiCStride = b_g_k_c_xs_strides[I2];
+        const auto OutKStride = e_g_n_k_wos_strides[I2];
+        const auto OutWStride = e_g_n_k_wos_strides[spatial_offset + I2];
 
         const index_t GemmKTotal = N * Do * Ho * Wo;
-        const index_t GemmM      = K;
-        const index_t GemmN      = C * Z * X * Y;
-
         const index_t GemmKBatch = batch_k;
         const index_t GemmK0 =
             math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
             K0PerBlock;
-        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
 
         if constexpr(ConvBackwardWeightSpecialization ==
                      ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
         {
             // A: output tensor
-            const auto out_gemmktotal_gemmm_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
+            const auto out_gemmktotal_gemmm_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * Do * Ho * Wo, K), make_tuple(OutWStride, OutKStride));
 
-            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
-                out_gemmktotal_gemmm_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmM)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto out_gemmkpad_gemmmpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    out_gemmktotal_gemmm_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, MPerBlock),
+                    Sequence<true, true>{});
 
             const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-                out_gemmkpad_gemmm_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmM)),
+                out_gemmkpad_gemmmpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(out_gemmkpad_gemmmpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
             // B: input tensor
-            const auto in_gemmktotal_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N * Di * Hi * Wi, C));
+            const auto in_gemmktotal_gemmn_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * Di * Hi * Wi, C), make_tuple(InWStride, InCStride));
 
-            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
-                in_gemmktotal_gemmn_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmM)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto in_gemmkpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    in_gemmktotal_gemmn_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, NPerBlock),
+                    Sequence<true, true>{});
 
             const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-                in_gemmkpad_gemmn_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmM)),
+                in_gemmkpad_gemmnpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(in_gemmkpad_gemmnpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
             // C: weight tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
+            const auto wei_gemmm_gemmn_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(K, Z * Y * X * C), make_tuple(WeiKStride, WeiCStride));
+
+            const auto wei_gemmmpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(wei_gemmm_gemmn_grid_desc,
+                                                                  make_tuple(MPerBlock, NPerBlock),
+                                                                  Sequence<true, true>{});
 
             return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+                              wei_gemmmpad_gemmnpad_grid_desc);
         }
         else
         {
-            const auto out_gemmktotal_gemmm_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
-            const auto in_n_di_hi_wi_c_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+            const auto out_gemmktotal_gemmm_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N * Do * Ho * Wo, K), make_tuple(OutWStride, OutKStride));
+            const auto in_n_di_hi_wi_c_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Di, Hi, Wi, C),
+                make_tuple(InNStride, InDStride, InHStride, InWStride, InCStride));
 
             // A: output tensor
-            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
-                out_gemmktotal_gemmm_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmM)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto out_gemmkpad_gemmmpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    out_gemmktotal_gemmm_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, MPerBlock),
+                    Sequence<true, true>{});
 
             const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
-                out_gemmkpad_gemmm_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmM)),
+                out_gemmkpad_gemmmpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(out_gemmkpad_gemmmpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
@@ -672,27 +753,32 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
                 make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}),
                 make_tuple(Sequence<1>{}, Sequence<0>{}));
 
-            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
-                in_gemmktotal_gemmn_grid_desc,
-                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
-                           make_pass_through_transform(GemmN)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
+            const auto in_gemmkpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(
+                    in_gemmktotal_gemmn_grid_desc,
+                    make_tuple(GemmK1Number * K0PerBlock * GemmKBatch, NPerBlock),
+                    Sequence<true, true>{});
 
             const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
-                in_gemmkpad_gemmn_grid_desc,
-                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
-                           make_pass_through_transform(GemmN)),
+                in_gemmkpad_gemmnpad_grid_desc,
+                make_tuple(
+                    make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                    make_pass_through_transform(in_gemmkpad_gemmnpad_grid_desc.GetLength(I1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
             // C: weight tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
+            const auto wei_gemmm_gemmn_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(K, Z * Y * X * C), make_tuple(WeiKStride, WeiCStride));
+
+            const auto wei_gemmmpad_gemmnpad_grid_desc =
+                ck::tensor_operation::device::PadTensorDescriptor(wei_gemmm_gemmn_grid_desc,
+                                                                  make_tuple(MPerBlock, NPerBlock),
+                                                                  Sequence<true, true>{});
 
             return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+                              wei_gemmmpad_gemmnpad_grid_desc);
         }
 
     } // function end
@@ -701,22 +787,22 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     static auto GetABCGridDesc()
     {
         return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<1>(
-            1, 1, 1, {1}, {1}, {1}, {1}, {1}, {1}, {1}, 1);
+            {1}, {1}, {1}, {1}, {1}, {1}, {1}, {1}, {1}, {1}, 1);
     }
 
     template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
     static auto GetABCGridDesc()
     {
         return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<2>(
-            1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 1);
+            {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 1);
     }
 
     template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
     static auto GetABCGridDesc()
     {
-        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>(1,
-                                                                  1,
-                                                                  1,
+        return MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<3>({1, 1, 1},
+                                                                  {1, 1, 1},
+                                                                  {1, 1, 1},
                                                                   {1, 1, 1},
                                                                   {1, 1, 1},
                                                                   {1, 1, 1},
@@ -785,11 +871,11 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
                  WeiDataType* p_wei_grid,
                  const OutDataType* p_out_grid,
                  const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
-                 const std::array<index_t, NDimSpatial + 3>& /*a_g_n_c_wis_strides*/,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
                  const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
-                 const std::array<index_t, NDimSpatial + 3>& /*b_g_k_c_xs_strides*/,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
                  const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
-                 const std::array<index_t, NDimSpatial + 3>& /*e_g_n_k_wos_strides*/,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
                  const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
                  const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
                  const std::array<ck::index_t, NDimSpatial>& input_left_pads,
@@ -809,38 +895,24 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
               a_element_op_{out_element_op},
               b_element_op_{wei_element_op},
               c_element_op_{in_element_op},
-              Conv_G_{a_g_n_c_wis_lengths[0]},
-              Conv_N_{a_g_n_c_wis_lengths[1]},
-              Conv_K_{b_g_k_c_xs_lengths[1]},
-              Conv_C_{a_g_n_c_wis_lengths[2]},
-              input_spatial_lengths_{},
-              filter_spatial_lengths_{},
-              output_spatial_lengths_{},
+              Conv_G_{a_g_n_c_wis_lengths[I0]},
+              Conv_K_{b_g_k_c_xs_lengths[I1]},
+              Conv_C_{a_g_n_c_wis_lengths[I2]},
+              filter_lengths_{b_g_k_c_xs_lengths},
               conv_filter_strides_{conv_filter_strides},
               conv_filter_dilations_{conv_filter_dilations},
               input_left_pads_{input_left_pads},
               input_right_pads_{input_right_pads},
               k_batch_{split_k}
         {
-            constexpr index_t spatial_offset = 3;
-            std::copy(begin(a_g_n_c_wis_lengths) + spatial_offset,
-                      end(a_g_n_c_wis_lengths),
-                      begin(input_spatial_lengths_));
-            std::copy(begin(b_g_k_c_xs_lengths) + spatial_offset,
-                      end(b_g_k_c_xs_lengths),
-                      begin(filter_spatial_lengths_));
-            std::copy(begin(e_g_n_k_wos_lengths) + spatial_offset,
-                      end(e_g_n_k_wos_lengths),
-                      begin(output_spatial_lengths_));
-
             const auto descs =
                 DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
-                    Conv_N_,
-                    Conv_K_,
-                    Conv_C_,
-                    input_spatial_lengths_,
-                    filter_spatial_lengths_,
-                    output_spatial_lengths_,
+                    a_g_n_c_wis_lengths, // input
+                    a_g_n_c_wis_strides,
+                    b_g_k_c_xs_lengths, // weight
+                    b_g_k_c_xs_strides,
+                    e_g_n_k_wos_lengths, // output
+                    e_g_n_k_wos_strides,
                     conv_filter_strides,
                     conv_filter_dilations,
                     input_left_pads,
@@ -863,24 +935,9 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
                 GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
 
             // A/B/C Batch Stride
-            compute_ptr_offset_of_batch_.BatchStrideA_ =
-                Conv_N_ * Conv_K_ *
-                std::accumulate(begin(output_spatial_lengths_),
-                                end(output_spatial_lengths_),
-                                index_t{1},
-                                std::multiplies<>{});
-            compute_ptr_offset_of_batch_.BatchStrideB_ =
-                Conv_N_ * Conv_C_ *
-                std::accumulate(begin(input_spatial_lengths_),
-                                end(input_spatial_lengths_),
-                                index_t{1},
-                                std::multiplies<>{});
-            compute_ptr_offset_of_batch_.BatchStrideC_ =
-                Conv_K_ * Conv_C_ *
-                std::accumulate(begin(filter_spatial_lengths_),
-                                end(filter_spatial_lengths_),
-                                index_t{1},
-                                std::multiplies<>{});
+            compute_ptr_offset_of_batch_.BatchStrideA_ = e_g_n_k_wos_strides[I0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = a_g_n_c_wis_strides[I0];
+            compute_ptr_offset_of_batch_.BatchStrideC_ = b_g_k_c_xs_strides[I0];
         }
 
         const ADataType* p_a_grid_;
@@ -908,13 +965,10 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
 
         // for checking IsSupportedArgument()
         const index_t Conv_G_;
-        const index_t Conv_N_;
         const index_t Conv_K_;
         const index_t Conv_C_;
 
-        std::array<ck::index_t, NDimSpatial> input_spatial_lengths_;
-        std::array<ck::index_t, NDimSpatial> filter_spatial_lengths_;
-        std::array<ck::index_t, NDimSpatial> output_spatial_lengths_;
+        std::array<ck::index_t, NDimSpatial + 3> filter_lengths_;
         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides_;
         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations_;
         const std::array<ck::index_t, NDimSpatial>& input_left_pads_;
@@ -1036,10 +1090,14 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        // check device
-        if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030" ||
-             ck::get_device_name() == "gfx1100" || ck::get_device_name() == "gfx1101" ||
-             ck::get_device_name() == "gfx1102"))
+
+        // DL version only supports split_k equal to 1
+        if(arg.k_batch_ != 1)
+            return false;
+
+        if constexpr(!((NDimSpatial == 1 && (is_NWGK_GKXC_NWGC || is_GNWK_GKXC_GNWC)) ||
+                       (NDimSpatial == 2 && (is_NHWGK_GKYXC_NHWGC || is_GNHWK_GKYXC_GNHWC)) ||
+                       (NDimSpatial == 3 && (is_NDHWGK_GKZYXC_NDHWGC || is_GNDHWK_GKZYXC_GNDHWC))))
         {
             return false;
         }
@@ -1050,8 +1108,9 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
             // check if it's 1x1, stride=1 pad = 0 conv
             for(int i = 0; i < NDimSpatial; i++)
             {
-                if(!(arg.filter_spatial_lengths_[i] == 1 && arg.conv_filter_strides_[i] == 1 &&
-                     arg.input_left_pads_[i] == 0 && arg.input_right_pads_[i] == 0))
+                if(!(arg.filter_lengths_[spatial_offset + i] == 1 &&
+                     arg.conv_filter_strides_[i] == 1 && arg.input_left_pads_[i] == 0 &&
+                     arg.input_right_pads_[i] == 0))
                 {
                     return false;
                 }
@@ -1206,7 +1265,7 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
         auto str = std::stringstream();
 
         // clang-format off
-        str << "DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl"
+        str << "DeviceGroupedConvBwdWeight_Dl"
             << "<"
             << BlockSize << ", "
             << MPerBlock << ", "

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

@@ -193,6 +193,7 @@ template <typename ALayout,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          typename ComputeType    = ADataType,
           LoopScheduler LoopSched = make_default_loop_scheduler()>
 struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
                                                                         BLayout,
@@ -217,6 +218,8 @@ struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_splitk_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeType,
         AccDataType,
         CShuffleDataType,
         DsDataType,

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

@@ -27,6 +27,12 @@ struct PassThrough
         y = x;
     }
 
+    template <>
+    __host__ __device__ void operator()<float, double>(float& y, const double& x) const
+    {
+        y = type_convert<float>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<float, float>(float& y, const float& x) const
     {
@@ -69,18 +75,36 @@ struct PassThrough
         y = type_convert<bhalf_t>(x);
     }
 
+    template <>
+    __host__ __device__ void operator()<float, half_t>(float& y, const half_t& x) const
+    {
+        y = type_convert<float>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<int8_t, int8_t>(int8_t& y, const int8_t& x) const
     {
         y = x;
     }
 
+    template <>
+    __host__ __device__ void operator()<half_t, int8_t>(half_t& y, const int8_t& x) const
+    {
+        y = type_convert<half_t>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<int8_t, int32_t>(int8_t& y, const int32_t& x) const
     {
         y = type_convert<int8_t>(x);
     }
 
+    template <>
+    __host__ __device__ void operator()<int8_t, float>(int8_t& y, const float& x) const
+    {
+        y = type_convert<int8_t>(x);
+    }
+
 #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
     template <>
     __host__ __device__ void operator()<int4_t, int4_t>(int4_t& y, const int4_t& x) const
@@ -89,6 +113,7 @@ struct PassThrough
     }
 #endif
 
+#if defined CK_ENABLE_FP8
     template <>
     __host__ __device__ void operator()<f8_t, f8_t>(f8_t& y, const f8_t& x) const
     {
@@ -118,6 +143,7 @@ struct PassThrough
     {
         y = type_convert<f8_t>(x);
     }
+#endif
 };
 
 struct UnaryConvert
@@ -146,6 +172,7 @@ struct ConvertBF16RTN
     }
 };
 
+#if defined CK_ENABLE_FP8
 struct ConvertF8SR
 {
     // convert to fp8 using stochastic rounding (SR)
@@ -162,6 +189,7 @@ struct ConvertF8SR
         y = f8_convert_sr<Y>(x);
     }
 };
+#endif
 
 struct Scale
 {
@@ -412,14 +440,19 @@ struct Swish
 {
     Swish(float beta = 1.0f) : beta_(beta) {}
 
-    template <typename T>
-    __host__ __device__ void operator()(T& y, const T& x) const
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
     {
-        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
-                          is_same<T, ck::half_t>::value,
+        static_assert(is_same<X, float>::value || is_same<X, double>::value ||
+                          is_same<X, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        static_assert(is_same<Y, float>::value || is_same<Y, double>::value ||
+                          is_same<Y, ck::half_t>::value,
                       "Data type is not supported by this operation!");
 
-        y = x / (ck::type_convert<T>(1) + ck::math::exp(-beta_ * x));
+        float bx = -beta_ * type_convert<float>(x);
+        y        = type_convert<Y>(x / (1.f + ck::math::exp(bx)));
     };
 
     float beta_ = 1.0f;