Merge remote-tracking branch 'origin/develop' into migraphx-update

example/49_maxpool2d_bwd/CMakeLists.txt

-if(DTYPES MATCHES "bf16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
-endif()
-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
-endif()
-if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
-    add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)
-endif()
+add_example_executable(example_maxpool2d_bwd_bf16 maxpool2d_bwd_bf16.cpp)
+add_example_executable(example_maxpool2d_bwd_fp16 maxpool2d_bwd_fp16.cpp)
+add_example_executable(example_maxpool2d_bwd_fp32 maxpool2d_bwd_fp32.cpp)

example/49_maxpool2d_bwd/maxpool2d_bwd_common.hpp

@@ -8,7 +8,7 @@
 #include "ck/ck.hpp"
 #include "ck/utility/reduction_enums.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_pool2d_fwd_nhwc_nhwc.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_index_pool_bwd_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_max_pool_bwd_impl.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/utility/check_err.hpp"
@@ -60,7 +60,7 @@ bool maxpool_bwd_test(bool do_verification,
                                                                 1>; // InSrcOutDstVectorSize
 
     using DeviceMaxPoolBwdInstance = ck::tensor_operation::device::
-        DeviceIndexPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 4>;
+        DeviceMaxPoolBwdImpl<DOutDataType, IndexDataType, DInDataType, 4>;
 
     const ck::index_t Ys = (Y - 1) * window_dilation_h + 1;
     const ck::index_t Xs = (X - 1) * window_dilation_w + 1;
@@ -155,7 +155,8 @@ bool maxpool_bwd_test(bool do_verification,
         dout_n_c_ho_wo.mDesc.GetElementSpaceSize(),
         din_n_c_hi_wi_device.mDesc.GetElementSpaceSize(),
         window_spatial_lengths,
-        window_strides);
+        window_strides,
+        window_dilations);
 
     if(!pool_bwd.IsSupportedArgument(pool_bwd_argument_ptr.get()))
     {

example/50_put_element/CMakeLists.txt

-if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
-    add_example_executable(example_put_element_fp16 put_element_fp16.cpp)
-endif()
+add_example_executable(example_put_element_fp16 put_element_fp16.cpp)

example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp

@@ -26,6 +26,8 @@ std::vector<ck::index_t> f_tensor_strides_ncdhw(ck::index_t N_,
         return {C_ * D * H * W, D * H * W, H * W, W, 1_uz};
     else if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NDHWC>::value)
         return {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_};
+    throw std::runtime_error("Avgpool3d_bwd: problem with layout. ");
+    return {0, 0, 0, 0, 0};
 };
 
 template <typename TensorLayout>
@@ -47,6 +49,8 @@ HostTensorDescriptor f_host_tensor_descriptor(std::size_t N_,
         return HostTensorDescriptor({N_, C_, D, H, W},
                                     {D * C_ * H * W, 1_uz, C_ * H * W, W * C_, C_});
     }
+    throw std::runtime_error("Avgpool3d_bwd: problem with layout. ");
+    return HostTensorDescriptor({0, 0, 0, 0, 0}, {0, 0, 0, 0, 0});
 };
 
 template <typename DevicePoolBwdInstance,

example/52_im2col_col2im/CMakeLists.txt

+add_custom_target(example_im2col_col2im)
+
+add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
+add_example_dependencies(example_im2col_col2im example_image_to_column_f32)
+
+add_example_executable(example_column_to_image_f32 column_to_image_f32.cpp)
+add_example_dependencies(example_im2col_col2im example_column_to_image_f32)

example/52_im2col_col2im/column_to_image_f32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = FP32; // ck::bhalf_t;//FP32;
+using OutDataType = FP32; // ck::bhalf_t;//FP32;
+
+using ImLayout        = ck::tensor_layout::convolution::GNHWC;
+using ColumnToImageOp = ck::conv_tensor_rearrange_op::ColumnToImage;
+
+// clang-format off
+using DeviceColToImgInstance = ck::tensor_operation::device::DeviceColumnToImageImpl
+        //#####################|        Num| ImLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+                              < NDimSpatial, ImLayout, InDataType, OutDataType,   256,   128,   128, S<16, 16>,     1>;
+// clang-format on
+
+bool RunColumnToImage(const ExecutionConfig& config, const ck::utils::conv::ConvParam& conv_params)
+{
+    const auto G = conv_params.G_;
+    const auto N = conv_params.N_;
+    const auto C = conv_params.C_;
+
+    const ck::index_t NDoHoWo =
+        N * ck::accumulate_n<ck::index_t>(
+                conv_params.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+    const ck::index_t CZYX =
+        C * ck::accumulate_n<ck::index_t>(
+                conv_params.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+    const auto in_desc = HostTensorDescriptor({G, NDoHoWo, CZYX});
+    const auto out_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(conv_params);
+
+    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> image_g_n_c_wis_strides{};
+    std::array<ck::index_t, 3> gemm_g_m_k_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+    copy(conv_params.input_spatial_lengths_, input_spatial_lengths);
+    copy(conv_params.filter_spatial_lengths_, filter_spatial_lengths);
+    copy(conv_params.output_spatial_lengths_, output_spatial_lengths);
+    copy(in_desc.GetStrides(), gemm_g_m_k_strides);
+    copy(out_desc.GetStrides(), image_g_n_c_wis_strides);
+    copy(conv_params.conv_filter_strides_, conv_filter_strides);
+    copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_params.input_left_pads_, input_left_pads);
+    copy(conv_params.input_right_pads_, input_right_pads);
+
+    Tensor<InDataType> in(in_desc);
+    Tensor<OutDataType> out_device(out_desc);
+    Tensor<OutDataType> out_host(out_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "out: " << out_device.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1: in.GenerateTensorValue(GeneratorTensor_2<InDataType>{1, 2}); break;
+    default: in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+
+    // reset input to zero
+    out_device_buf.SetZero();
+
+    static_assert(std::is_default_constructible_v<DeviceColToImgInstance>);
+
+    // do conv
+    auto col2img  = DeviceColToImgInstance{};
+    auto invoker  = col2img.MakeInvoker();
+    auto argument = col2img.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                         out_device_buf.GetDeviceBuffer(),
+                                         G,
+                                         N,
+                                         C,
+                                         input_spatial_lengths,
+                                         filter_spatial_lengths,
+                                         output_spatial_lengths,
+                                         image_g_n_c_wis_strides,
+                                         gemm_g_m_k_strides,
+                                         conv_filter_strides,
+                                         conv_filter_dilations,
+                                         input_left_pads,
+                                         input_right_pads);
+
+    if(!col2img.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_col2img with the specified compilation parameters does "
+                     "not support this col2img problem"
+                  << std::endl;
+
+        return false;
+    }
+
+    float ave_time        = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+    std::size_t num_btype = G * NDoHoWo * CZYX * (sizeof(OutDataType) + sizeof(InDataType));
+    float gb_per_sec      = num_btype / 1.E6 / ave_time;
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_column_to_image = ck::tensor_operation::host::
+            ReferenceColumnToImage<NDimSpatial, ImLayout, InDataType, OutDataType>();
+
+        auto ref_invoker = ref_column_to_image.MakeInvoker();
+
+        auto ref_argument = ref_column_to_image.MakeArgument(in,
+                                                             out_host,
+                                                             conv_params.filter_spatial_lengths_,
+                                                             conv_params.conv_filter_strides_,
+                                                             conv_params.conv_filter_dilations_,
+                                                             conv_params.input_left_pads_,
+                                                             conv_params.input_right_pads_);
+
+        if(!ref_column_to_image.IsSupportedArgument(&ref_argument))
+        {
+            std::cerr << "wrong! ref_col2img with the specified compilation parameters does "
+                         "not support this col2img problem"
+                      << std::endl;
+            return false;
+        }
+
+        ref_invoker.Run(ref_argument);
+        out_device_buf.FromDevice(out_device.mData.data());
+        return ck::utils::check_err(out_device.mData, out_host.mData);
+    }
+
+    return true;
+}
+
+int RunColumnToImageExample(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatial dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    return !RunColumnToImage(config, conv_params);
+}
+
+int main(int argc, char* argv[]) { return RunColumnToImageExample(argc, argv); }

example/52_im2col_col2im/common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <cstdlib>
+#include <initializer_list>
+#include <iostream>
+#include <numeric>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_column_to_image_impl.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/utility/algorithm.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp"
+#include "ck/library/utility/convolution_parameter.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/reference_tensor_operation/cpu/reference_image_to_column.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_column_to_image.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+static inline constexpr ck::index_t NDimSpatial = 2;
+
+using FP32 = float;
+
+struct ExecutionConfig final
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+};
+
+#define DefaultConvParams                                                            \
+    ck::utils::conv::ConvParam                                                       \
+    {                                                                                \
+        NDimSpatial, 1, 32, 1, 1, {4, 4}, {64, 64}, {1, 1}, {1, 1}, {0, 0}, { 0, 0 } \
+    }
+
+inline void print_help_msg()
+{
+    std::cerr << "arg1: verification (0=no, 1=yes)\n"
+              << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
+              << "arg3: time kernel (0=no, 1=yes)\n"
+              << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
+}
+
+inline bool parse_cmd_args(int argc,
+                           char* argv[],
+                           ExecutionConfig& config,
+                           ck::utils::conv::ConvParam& conv_params)
+{
+    constexpr int num_execution_config_args =
+        3; // arguments for do_verification, init_method, time_kernel
+    constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
+
+    constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
+    constexpr int threshold_to_catch_all_args =
+        threshold_to_catch_partial_args + num_conv_param_leading_args;
+
+    if(argc == 1)
+    {
+        // use default
+        config = ExecutionConfig{};
+    }
+    // catch only ExecutionConfig arguments
+    else if(argc == threshold_to_catch_partial_args)
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+    }
+    // catch both ExecutionConfig & ConvParam arguments
+    else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
+    {
+        config.do_verification = std::stoi(argv[1]);
+        config.init_method     = std::stoi(argv[2]);
+        config.time_kernel     = std::stoi(argv[3]);
+
+        const ck::index_t num_dim_spatial = std::stoi(argv[4]);
+        conv_params                       = ck::utils::conv::parse_conv_param(
+            num_dim_spatial, threshold_to_catch_partial_args, argv);
+    }
+    else
+    {
+        print_help_msg();
+        return false;
+    }
+
+    return true;
+}

example/52_im2col_col2im/image_to_column_f32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+using InDataType  = FP32;
+using OutDataType = FP32;
+
+using ImLayout        = ck::tensor_layout::convolution::GNHWC;
+using ImageToColumnOp = ck::conv_tensor_rearrange_op::ImageToColumn;
+
+// clang-format off
+using DeviceImgToColInstance = ck::tensor_operation::device::DeviceImageToColumnImpl
+        //#####################|        Num| ImLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+                              < NDimSpatial, ImLayout, InDataType, OutDataType,   256,   128,   128, S<16, 16>,      1>;
+// clang-format on
+
+bool RunImageToColumn(const ExecutionConfig& config, const ck::utils::conv::ConvParam& conv_params)
+{
+    const auto G = conv_params.G_;
+    const auto N = conv_params.N_;
+    const auto C = conv_params.C_;
+
+    const ck::index_t NDoHoWo =
+        N * ck::accumulate_n<ck::index_t>(
+                conv_params.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+    const ck::index_t CZYX =
+        C * ck::accumulate_n<ck::index_t>(
+                conv_params.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+    const auto in_desc =
+        ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed<ImLayout>(conv_params);
+    const auto out_desc = HostTensorDescriptor({G, NDoHoWo, CZYX});
+
+    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> image_g_n_c_wis_strides{};
+    std::array<ck::index_t, 3> gemm_g_m_k_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_strides{};
+    std::array<ck::index_t, NDimSpatial> conv_filter_dilations{};
+    std::array<ck::index_t, NDimSpatial> input_left_pads{};
+    std::array<ck::index_t, NDimSpatial> input_right_pads{};
+
+    auto copy = [](const auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); };
+
+    copy(conv_params.input_spatial_lengths_, input_spatial_lengths);
+    copy(conv_params.filter_spatial_lengths_, filter_spatial_lengths);
+    copy(conv_params.output_spatial_lengths_, output_spatial_lengths);
+    copy(in_desc.GetStrides(), image_g_n_c_wis_strides);
+    copy(out_desc.GetStrides(), gemm_g_m_k_strides);
+    copy(conv_params.conv_filter_strides_, conv_filter_strides);
+    copy(conv_params.conv_filter_dilations_, conv_filter_dilations);
+    copy(conv_params.input_left_pads_, input_left_pads);
+    copy(conv_params.input_right_pads_, input_right_pads);
+
+    Tensor<InDataType> in(in_desc);
+    Tensor<OutDataType> out_device(out_desc);
+    Tensor<OutDataType> out_host(out_desc);
+
+    std::cout << "in: " << in.mDesc << std::endl;
+    std::cout << "out: " << out_device.mDesc << std::endl;
+
+    switch(config.init_method)
+    {
+    case 0: break;
+    case 1: in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}); break;
+    default: in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize());
+    DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize());
+
+    in_device_buf.ToDevice(in.mData.data());
+
+    // reset input to zero
+    out_device_buf.SetZero();
+
+    static_assert(std::is_default_constructible_v<DeviceImgToColInstance>);
+
+    // do conv
+    auto img2col  = DeviceImgToColInstance{};
+    auto invoker  = img2col.MakeInvoker();
+    auto argument = img2col.MakeArgument(in_device_buf.GetDeviceBuffer(),
+                                         out_device_buf.GetDeviceBuffer(),
+                                         G,
+                                         N,
+                                         C,
+                                         input_spatial_lengths,
+                                         filter_spatial_lengths,
+                                         output_spatial_lengths,
+                                         image_g_n_c_wis_strides,
+                                         gemm_g_m_k_strides,
+                                         conv_filter_strides,
+                                         conv_filter_dilations,
+                                         input_left_pads,
+                                         input_right_pads);
+
+    if(!img2col.IsSupportedArgument(argument))
+    {
+        std::cerr << "wrong! device_img2col with the specified compilation parameters does "
+                     "not support this img2col problem"
+                  << std::endl;
+
+        return false;
+    }
+
+    float ave_time        = invoker.Run(argument, StreamConfig{nullptr, config.time_kernel});
+    std::size_t num_btype = G * NDoHoWo * CZYX * (sizeof(OutDataType) + sizeof(InDataType));
+    float gb_per_sec      = num_btype / 1.E6 / ave_time;
+    std::cout << "Perf: " << ave_time << " ms, " << gb_per_sec << " GB/s" << std::endl;
+
+    if(config.do_verification)
+    {
+        auto ref_image_to_column = ck::tensor_operation::host::
+            ReferenceImageToColumn<NDimSpatial, ImLayout, InDataType, OutDataType>();
+
+        auto ref_invoker = ref_image_to_column.MakeInvoker();
+
+        auto ref_argument = ref_image_to_column.MakeArgument(in,
+                                                             out_host,
+                                                             conv_params.filter_spatial_lengths_,
+                                                             conv_params.conv_filter_strides_,
+                                                             conv_params.conv_filter_dilations_,
+                                                             conv_params.input_left_pads_,
+                                                             conv_params.input_right_pads_);
+
+        if(!ref_image_to_column.IsSupportedArgument(&ref_argument))
+        {
+            std::cerr << "wrong! ref_img2col with the specified compilation parameters does "
+                         "not support this img2col problem"
+                      << std::endl;
+            return false;
+        }
+
+        ref_invoker.Run(ref_argument);
+
+        out_device_buf.FromDevice(out_device.mData.data());
+
+        return ck::utils::check_err(out_device.mData, out_host.mData);
+    }
+
+    return true;
+}
+
+int RunImageToColumnExample(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_params = DefaultConvParams;
+
+    if(!parse_cmd_args(argc, argv, config, conv_params))
+    {
+        return EXIT_FAILURE;
+    }
+
+    if(conv_params.num_dim_spatial_ != NDimSpatial)
+    {
+        std::cerr << "unsupported # of spatial dimensions" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    return !RunImageToColumn(config, conv_params);
+}
+
+int main(int argc, char* argv[]) { return RunImageToColumnExample(argc, argv); }

example/53_layernorm2d_bwd/CMakeLists.txt

+add_example_executable(example_layernorm2d_bwd_fp32 layernorm2d_bwd_fp32.cpp)

example/53_layernorm2d_bwd/layernorm2d_bwd_fp32.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 <getopt.h>
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_bwd_data_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_bwd_gamma_beta_impl.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_layernorm_bwd.hpp"
+
+using DYDataType         = float;
+using XDataType          = float;
+using GammaDataType      = float;
+using MeanInvStdDataType = float;
+using DGammaDataType     = float;
+using DBetaDataType      = float;
+using DXDataType         = float;
+using ComputeDataType    = float;
+
+constexpr int Rank         = 2;
+constexpr int NumReduceDim = 1;
+
+// Layernorm:
+
+// Input shape
+// dy:        [M, N]
+// x:         [M, N]
+// mean:      [M, 1]
+// inv_std:   [M, 1]
+
+// Output shape
+// dx:     [M, N]
+// dgamma: [1, N]
+// dbeta:  [1, N]
+
+// dgamma = reduce_sum(dy * (x - mean) * inv_std, axis=0)
+// dbeta = reduce_sum(dy, axis=0)
+
+// [CAUSION]
+// In DeviceNormalizationBwdDataImpl & DeviceNormalizationBwdGammaBetaImpl, M is Invariant
+// dimension, K is reduced dimension Hence, M in this example and
+// DeviceNormalizationBwdGammaBetaImpl is different
+using XDeviceInstance = ck::tensor_operation::device::DeviceNormalizationBwdDataImpl<
+    DYDataType,
+    XDataType,
+    GammaDataType,
+    MeanInvStdDataType,
+    ComputeDataType,
+    DXDataType,
+    Rank,
+    NumReduceDim,
+    256,   // BlockSize
+    8,     // MThreadClusterSize
+    32,    // KThreadClusterSize
+    1,     // MThreadSliceSize
+    4,     // KThreadSliceSize
+    true,  // IsDYFastestDimReduced
+    4,     // DYSrcVectorSize
+    true,  // IsXFastestDimReduced
+    4,     // XSrcVectorSize
+    true,  // IsGammaFastestDimReduced
+    4,     // GammaSrcVectorSize
+    false, // IsMeanInvStdFastestDimReduced
+    1,     // MeanInvStdSrcVectorSize
+    true,  // IsDXFastestDimReduced
+    4>;    // DXDstVectorSize
+
+using GammaBetaDeviceInstance = ck::tensor_operation::device::DeviceNormalizationBwdGammaBetaImpl<
+    DYDataType,
+    XDataType,
+    MeanInvStdDataType,
+    ComputeDataType,
+    DGammaDataType,
+    DBetaDataType,
+    Rank,
+    NumReduceDim,
+    256,   // BlockSize
+    8,     // MThreadClusterSize
+    32,    // KThreadClusterSize
+    4,     // MThreadSliceSize
+    1,     // KThreadSliceSize
+    false, // IsDYFastestDimReduced
+    4,     // DYSrcVectorSize
+    false, // IsXFastestDimReduced
+    4,     // XSrcVectorSize
+    true,  // IsMeanInvStdFastestDimReduced
+    1,     // MeanInvStdSrcVectorSize
+    4,     // DGammaDstVectorSize
+    4>;    // DBetaDstVectorSize
+
+int main()
+{
+    bool time_kernel = false;
+
+    ck::index_t M = 1024;
+    ck::index_t N = 512;
+
+    Tensor<DYDataType> dy({M, N});
+    Tensor<XDataType> x({M, N});
+    Tensor<GammaDataType> gamma({N});
+    Tensor<MeanInvStdDataType> mean({M});
+    Tensor<MeanInvStdDataType> inv_std({M});
+
+    Tensor<DGammaDataType> dgamma({N});
+    Tensor<DBetaDataType> dbeta({N});
+    Tensor<DXDataType> dx({M, N});
+
+    dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0.0, 1.0});
+    x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+    gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
+    mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
+    inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
+
+    DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
+    DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
+    DeviceMem dx_dev(sizeof(DXDataType) * dx.mDesc.GetElementSpaceSize());
+    DeviceMem dgamma_dev(sizeof(DGammaDataType) * dgamma.mDesc.GetElementSpaceSize());
+    DeviceMem dbeta_dev(sizeof(DBetaDataType) * dbeta.mDesc.GetElementSpaceSize());
+
+    dy_dev.ToDevice(dy.mData.data());
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    mean_dev.ToDevice(mean.mData.data());
+    inv_std_dev.ToDevice(inv_std.mData.data());
+
+    // backward x
+    auto x_device_instance = XDeviceInstance{};
+
+    auto x_argument_ptr = x_device_instance.MakeArgumentPointer({M, N}, // lengths
+                                                                {N, 1}, // dyStrides
+                                                                {N, 1}, // xStrides
+                                                                {0, 1}, // gammaStrides
+                                                                {1, 0}, // meanStrides
+                                                                {1, 0}, // invStdStrides
+                                                                {N, 1}, // dxStrides
+                                                                {1},    // reduceDims
+                                                                dy_dev.GetDeviceBuffer(),
+                                                                x_dev.GetDeviceBuffer(),
+                                                                gamma_dev.GetDeviceBuffer(),
+                                                                mean_dev.GetDeviceBuffer(),
+                                                                inv_std_dev.GetDeviceBuffer(),
+                                                                dx_dev.GetDeviceBuffer());
+
+    if(!x_device_instance.IsSupportedArgument(x_argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported." << __FILE__ << ":" << __LINE__
+                  << std::endl;
+        return 1;
+    };
+
+    auto x_invoker_ptr = x_device_instance.MakeInvokerPointer();
+    x_invoker_ptr->Run(x_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    // backward gamma & beta
+    auto gamma_beta_device_instance = GammaBetaDeviceInstance{};
+    auto gamma_beta_argument_ptr =
+        gamma_beta_device_instance.MakeArgumentPointer({M, N}, // inLengths
+                                                       {N, 1}, // dyStrides
+                                                       {N, 1}, // xStrides
+                                                       {1, 0}, // meanStrides
+                                                       {1, 0}, // invStdStrides
+                                                       {N},    // outLengths
+                                                       {1},    // dgammaStrides
+                                                       {1},    // dbetaStrides
+                                                       {0},    // reduceDims
+                                                       dy_dev.GetDeviceBuffer(),
+                                                       x_dev.GetDeviceBuffer(),
+                                                       mean_dev.GetDeviceBuffer(),
+                                                       inv_std_dev.GetDeviceBuffer(),
+                                                       dgamma_dev.GetDeviceBuffer(),
+                                                       dbeta_dev.GetDeviceBuffer());
+
+    if(!gamma_beta_device_instance.IsSupportedArgument(gamma_beta_argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported." << __FILE__ << ":" << __LINE__
+                  << std::endl;
+        return 1;
+    };
+
+    auto gamma_beta_invoker_ptr = gamma_beta_device_instance.MakeInvokerPointer();
+    gamma_beta_invoker_ptr->Run(gamma_beta_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+    {
+        Tensor<DGammaDataType> host_dgamma({N});
+        Tensor<DBetaDataType> host_dbeta({N});
+        Tensor<DXDataType> host_dx({M, N});
+        using ReferenceInstance =
+            ck::tensor_operation::host::ReferenceLayernormBwd<DYDataType,
+                                                              XDataType,
+                                                              GammaDataType,
+                                                              MeanInvStdDataType,
+                                                              DGammaDataType,
+                                                              DBetaDataType,
+                                                              DXDataType,
+                                                              ComputeDataType>;
+
+        ReferenceInstance ref;
+        auto ref_argument =
+            ref.MakeArgument(dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, {M, N});
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        dgamma_dev.FromDevice(dgamma.mData.data());
+        dbeta_dev.FromDevice(dbeta.mData.data());
+        dx_dev.FromDevice(dx.mData.data());
+
+        pass &= ck::utils::check_err(dgamma, host_dgamma, "Error: Incorrect dgamma", 1e-3, 1e-3);
+        pass &= ck::utils::check_err(dbeta, host_dbeta, "Error: Incorrect dbeta", 1e-3, 1e-3);
+        pass &= ck::utils::check_err(dx, host_dx, "Error: Incorrect dx", 1e-3, 1e-3);
+    }
+
+    return (pass ? 0 : 1);
+}

example/54_groupnorm_bwd/CMakeLists.txt

+add_example_executable(example_groupnorm_bwd_fp32 groupnorm_bwd_fp32.cpp)

example/54_groupnorm_bwd/groupnorm_bwd_fp32.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 <getopt.h>
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_common_util.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+#include "ck/library/utility/literals.hpp"
+
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_bwd_data_impl.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_normalization_bwd_gamma_beta_impl.hpp"
+#include "ck/library/reference_tensor_operation/cpu/reference_groupnorm_bwd.hpp"
+
+using DYDataType         = float;
+using XDataType          = float;
+using GammaDataType      = float;
+using MeanInvStdDataType = float;
+using DGammaDataType     = float;
+using DBetaDataType      = float;
+using DXDataType         = float;
+using ComputeDataType    = float;
+
+constexpr int Rank         = 5;
+constexpr int NumReduceDim = 3;
+
+// Grouprnorm
+// kernel 1:                  M    , K
+// dy:     N, H, W, G, C -> N * G, H * W * C
+// x:      N, H, W, G, C -> N * G, H * W * C
+// gamma:  1, 1, 1, G, C -> 1 * G, 1 * 1 * C
+// mean:   N, 1, 1, G, 1 -> N * G, 1 * 1 * 1
+// rstd:   N, 1, 1, G, 1 -> N * G, 1 * 1 * 1
+
+// dx:     N, H, W, G, C -> N * G, H * W * C
+
+using XDeviceInstance = ck::tensor_operation::device::DeviceNormalizationBwdDataImpl<
+    DYDataType,
+    XDataType,
+    GammaDataType,
+    MeanInvStdDataType,
+    ComputeDataType,
+    DXDataType,
+    Rank,
+    NumReduceDim,
+    256,   // BlockSize
+    8,     // MThreadClusterSize
+    32,    // KThreadClusterSize
+    1,     // MThreadSliceSize
+    4,     // KThreadSliceSize
+    true,  // IsDYFastestDimReduced
+    4,     // DYSrcVectorSize
+    true,  // IsXFastestDimReduced
+    4,     // XSrcVectorSize
+    true,  // IsGammaFastestDimReduced
+    4,     // GammaSrcVectorSize
+    false, // IsMeanInvStdFastestDimReduced
+    1,     // MeanInvStdSrcVectorSize
+    true,  // IsDXFastestDimReduced
+    4>;    // DXDstVectorSize
+
+// kernel 2:                  M    , K
+// dy:     N, H, W, G, C -> G * C, N * H * W
+// x:      N, H, W, G, C -> G * C, N * H * W
+// mean:   N, 1, 1, G, 1 -> G * 1, N * 1 * 1
+// rstd:   N, 1, 1, G, 1 -> G * 1, N * 1 * 1
+
+// dgamma: 1, 1, 1, G, C -> G * C
+// dbeta:  1, 1, 1, G, C -> G * C
+
+// reduced axis: 0, 1, 2
+
+using GammaBetaDeviceInstance = ck::tensor_operation::device::DeviceNormalizationBwdGammaBetaImpl<
+    DYDataType,
+    XDataType,
+    MeanInvStdDataType,
+    ComputeDataType,
+    DGammaDataType,
+    DBetaDataType,
+    Rank,
+    NumReduceDim,
+    256,   // BlockSize
+    8,     // ClusterInvariant
+    32,    // ClusterReduce
+    4,     // SliceInvariant
+    1,     // SliceReduce
+    false, // IsDYFastestDimReduced
+    4,     // DYSrcVectorSize
+    false, // IsXFastestDimReduced
+    4,     // XSrcVectorSize
+    false, // IsMeanInvStdFastestDimReduced
+    1,     // MeanInvStdSrcVectorSize
+    4,     // DGammaDstVectorSize
+    4>;    // DBetaDstVectorSize
+
+int main()
+{
+    bool time_kernel = false;
+
+    ck::index_t N = 16;
+    ck::index_t H = 16;
+    ck::index_t W = 16;
+    ck::index_t G = 32;
+    ck::index_t C = 64;
+
+    Tensor<DYDataType> dy({N, H, W, G, C});
+    Tensor<XDataType> x({N, H, W, G, C});
+    Tensor<GammaDataType> gamma({G, C});
+    Tensor<MeanInvStdDataType> mean({N, G});
+    Tensor<MeanInvStdDataType> inv_std({N, G});
+
+    Tensor<DGammaDataType> dgamma({G, C});
+    Tensor<DBetaDataType> dbeta({G, C});
+    Tensor<DXDataType> dx({N, H, W, G, C});
+
+    dy.GenerateTensorValue(GeneratorTensor_3<DYDataType>{0.0, 1.0});
+    x.GenerateTensorValue(GeneratorTensor_3<XDataType>{0.0, 1.0});
+    gamma.GenerateTensorValue(GeneratorTensor_3<GammaDataType>{0.0, 1.0});
+    mean.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
+    inv_std.GenerateTensorValue(GeneratorTensor_3<MeanInvStdDataType>{0.0, 1.0});
+
+    DeviceMem dy_dev(sizeof(DYDataType) * dy.mDesc.GetElementSpaceSize());
+    DeviceMem x_dev(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem gamma_dev(sizeof(GammaDataType) * gamma.mDesc.GetElementSpaceSize());
+    DeviceMem mean_dev(sizeof(MeanInvStdDataType) * mean.mDesc.GetElementSpaceSize());
+    DeviceMem inv_std_dev(sizeof(MeanInvStdDataType) * inv_std.mDesc.GetElementSpaceSize());
+    DeviceMem dx_dev(sizeof(DXDataType) * dx.mDesc.GetElementSpaceSize());
+    DeviceMem dgamma_dev(sizeof(DGammaDataType) * dgamma.mDesc.GetElementSpaceSize());
+    DeviceMem dbeta_dev(sizeof(DBetaDataType) * dbeta.mDesc.GetElementSpaceSize());
+
+    dy_dev.ToDevice(dy.mData.data());
+    x_dev.ToDevice(x.mData.data());
+    gamma_dev.ToDevice(gamma.mData.data());
+    mean_dev.ToDevice(mean.mData.data());
+    inv_std_dev.ToDevice(inv_std.mData.data());
+
+    std::vector<ck::index_t> dyStrides{dy.mDesc.GetStrides().begin(), dy.mDesc.GetStrides().end()};
+    std::vector<ck::index_t> xStrides{x.mDesc.GetStrides().begin(), x.mDesc.GetStrides().end()};
+    std::vector<ck::index_t> gammaStrides  = {0, 0, 0, C, 1};
+    std::vector<ck::index_t> meanStrides   = {G, 0, 0, 1, 0};
+    std::vector<ck::index_t> invStdStrides = {G, 0, 0, 1, 0};
+    std::vector<ck::index_t> dxStrides{dx.mDesc.GetStrides().begin(), dx.mDesc.GetStrides().end()};
+
+    // backward x
+    auto x_device_instance = XDeviceInstance{};
+
+    auto x_argument_ptr = x_device_instance.MakeArgumentPointer({N, H, W, G, C}, // lengths
+                                                                dyStrides,       // dyStrides
+                                                                xStrides,        // xStrides
+                                                                gammaStrides,    // gammaStrides
+                                                                meanStrides,     // meanStrides
+                                                                invStdStrides,   // invStdStrides
+                                                                dxStrides,       // dxStrides
+                                                                {1, 2, 4},       // reduceDims
+                                                                dy_dev.GetDeviceBuffer(),
+                                                                x_dev.GetDeviceBuffer(),
+                                                                gamma_dev.GetDeviceBuffer(),
+                                                                mean_dev.GetDeviceBuffer(),
+                                                                inv_std_dev.GetDeviceBuffer(),
+                                                                dx_dev.GetDeviceBuffer());
+
+    if(!x_device_instance.IsSupportedArgument(x_argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported." << __FILE__ << ":" << __LINE__
+                  << std::endl;
+        return 1;
+    };
+
+    auto x_invoker_ptr = x_device_instance.MakeInvokerPointer();
+    x_invoker_ptr->Run(x_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    // backward gamma & beta
+
+    auto gamma_beta_device_instance = GammaBetaDeviceInstance{};
+    auto gamma_beta_argument_ptr =
+        gamma_beta_device_instance.MakeArgumentPointer({N, H, W, G, C}, // inLengths
+                                                       dyStrides,       // dyStrides
+                                                       xStrides,        // xStrides
+                                                       meanStrides,     // meanStrides
+                                                       invStdStrides,   // invStdStrides
+                                                       {G, C},          // outLengths
+                                                       {C, 1},          // dgammaStrides
+                                                       {C, 1},          // dbetaStrides
+                                                       {0, 1, 2},       // reduceDims
+                                                       dy_dev.GetDeviceBuffer(),
+                                                       x_dev.GetDeviceBuffer(),
+                                                       mean_dev.GetDeviceBuffer(),
+                                                       inv_std_dev.GetDeviceBuffer(),
+                                                       dgamma_dev.GetDeviceBuffer(),
+                                                       dbeta_dev.GetDeviceBuffer());
+
+    if(!gamma_beta_device_instance.IsSupportedArgument(gamma_beta_argument_ptr.get()))
+    {
+        std::cout << "The runtime parameters are not supported." << __FILE__ << ":" << __LINE__
+                  << std::endl;
+        return 1;
+    };
+
+    auto gamma_beta_invoker_ptr = gamma_beta_device_instance.MakeInvokerPointer();
+    gamma_beta_invoker_ptr->Run(gamma_beta_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    bool pass = true;
+    {
+        Tensor<DGammaDataType> host_dgamma({G, C});
+        Tensor<DBetaDataType> host_dbeta({G, C});
+        Tensor<DXDataType> host_dx({N, H, W, G, C});
+        using ReferenceInstance =
+            ck::tensor_operation::host::ReferenceGroupnormBwd<DYDataType,
+                                                              XDataType,
+                                                              GammaDataType,
+                                                              MeanInvStdDataType,
+                                                              DGammaDataType,
+                                                              DBetaDataType,
+                                                              DXDataType,
+                                                              ComputeDataType>;
+
+        ReferenceInstance ref;
+        auto ref_argument = ref.MakeArgument(
+            dy, x, gamma, mean, inv_std, host_dgamma, host_dbeta, host_dx, {N, H, W, G, C});
+        auto ref_invoker = ref.MakeInvoker();
+        ref_invoker.Run(ref_argument);
+
+        dgamma_dev.FromDevice(dgamma.mData.data());
+        dbeta_dev.FromDevice(dbeta.mData.data());
+        dx_dev.FromDevice(dx.mData.data());
+
+        pass &= ck::utils::check_err(dgamma, host_dgamma, "Error: Incorrect dgamma", 1e-3, 1e-3);
+        pass &= ck::utils::check_err(dbeta, host_dbeta, "Error: Incorrect dbeta", 1e-3, 1e-3);
+        pass &= ck::utils::check_err(dx, host_dx, "Error: Incorrect dx", 1e-3, 1e-3);
+    }
+
+    return (pass ? 0 : 1);
+}

example/59_grouped_gemm_multi_ABD/CMakeLists.txt

+add_custom_target(example_grouped_gemm_xdl_multi_abd)
+
+add_example_executable(example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16 grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp)
+add_example_dependencies(example_grouped_gemm_xdl_multi_abd example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16)
+
+add_example_executable(example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8 grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp)
+add_example_dependencies(example_grouped_gemm_xdl_multi_abd example_grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8)

example/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_bf16_i8.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 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_multi_abd_xdl_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
+#include "ck/tensor_operation/gpu/element/unary_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 BF16 = ck::bhalf_t;
+using I8   = int8_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 Add         = ck::tensor_operation::element_wise::Add;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Col;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using DsLayout = ck::Tuple<Row>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+using CDEElementOp = AddFastGelu;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Multi_ABD_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|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   128,    16,   128,    32,   8,   8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               1>;
+
+// 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;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+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::GemmMultiABDDesc> gemm_descs;
+
+    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<A0DataType>> a0_tensors;
+    std::vector<Tensor<B1DataType>> b_tensors;
+    std::vector<Tensor<B0DataType>> b0_tensors;
+    std::vector<Tensor<B1DataType>> b1_tensors;
+    std::vector<Tensor<D0DataType>> d0_tensors;
+    std::vector<Tensor<EDataType>> c_host_tensors;
+    std::vector<Tensor<EDataType>> c_device_tensors;
+
+    a0_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    b0_tensors.reserve(group_count);
+    b1_tensors.reserve(group_count);
+    d0_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> a0_tensors_device, b0_tensors_device, b1_tensors_device,
+        d0_tensors_device, c_tensors_device;
+
+    a0_tensors_device.reserve(group_count);
+    b0_tensors_device.reserve(group_count);
+    b1_tensors_device.reserve(group_count);
+    d0_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];
+
+        a0_tensors.push_back(Tensor<A0DataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], A0Layout{})));
+
+        b_tensors.push_back(Tensor<B1DataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], B0Layout{})));
+        b0_tensors.push_back(Tensor<B0DataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], B0Layout{})));
+        b1_tensors.push_back(Tensor<B1DataType>(
+            f_host_tensor_descriptor(problem_size.Ks[i], problem_size.Ns[i], 0, B1Layout{})));
+
+        d0_tensors.push_back(Tensor<D0DataType>(
+            f_host_tensor_descriptor(problem_size.Ms[i], problem_size.Ns[i], 0, ELayout{})));
+
+        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: " << a0_tensors[i].mDesc
+                  << " b_k_n: " << b0_tensors[i].mDesc << " d_m_n: " << d0_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(A0DataType) * a0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(B0DataType) * b0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(B1DataType) * b1_tensors[i].mDesc.GetElementSize() +
+                     sizeof(D0DataType) * d0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * c_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+            b0_tensors[i].GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+            b1_tensors[i].GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
+            break;
+        case 2:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+            b0_tensors[i].GenerateTensorValue(GeneratorTensor_3<B0DataType>{-5, 5});
+            b1_tensors[i].GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
+            break;
+        default:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+            b1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+
+        d0_tensors[i].GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::
+        GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a0_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
+
+        b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        b1_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(B1DataType) * problem_size.Ns[i]));
+
+        d0_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        a0_tensors_device[i]->ToDevice(a0_tensors[i].mData.data(),
+                                       a0_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(A0DataType));
+
+        b0_tensors_device[i]->ToDevice(b0_tensors[i].mData.data(),
+                                       b0_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(B0DataType));
+
+        b1_tensors_device[i]->ToDevice(b1_tensors[i].mData.data(),
+                                       b1_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(B1DataType));
+
+        d0_tensors_device[i]->ToDevice(d0_tensors[i].mData.data());
+        c_tensors_device[i]->SetZero();
+
+        gemm_descs.push_back(
+            {sum_of_m, problem_size.Ns[i], problem_size.Ks[i], {1}, {1, 1}, {0}, 1});
+
+        grouped_gemm_kernel_args_.push_back(
+            {std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumBTensor>{b0_tensors_device[i]->GetDeviceBuffer(),
+                                                 b1_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumDTensor>{d0_tensors_device[i]->GetDeviceBuffer()},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             std::array<ck::index_t, NumATensor>{problem_size.stride_As[i]},
+             std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i], 0},
+             std::array<ck::index_t, NumDTensor>{0},
+             problem_size.stride_Cs[i]});
+    }
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, NumATensor>> p_As = {};
+    std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
+    std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
+    std::vector<void*> p_Cs                               = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
+
+    DeviceMem gemm_kernel_args_dev(gemm.GetDeviceKernelArgSize(&argument));
+    hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                              grouped_gemm_kernel_args_.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+
+    gemm.SetDeviceKernelArgs(argument, gemm_kernel_args_dev.GetDeviceBuffer());
+    gemm.SetKBatch(argument, config.k_batch);
+
+    gemm.SetElementwiseOps(argument, a_element_op, b_element_op, cde_element_op);
+
+    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<A0DataType,
+                                                                                B1DataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            for(int n = 0; n < problem_size.Ns[i]; ++n)
+            {
+                for(int k = 0; k < problem_size.Ks[i]; ++k)
+                {
+                    b_element_op(b_tensors[i](k, n), b0_tensors[i](k, n), b1_tensors[i](k, n));
+                }
+            }
+
+            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(a0_tensors[i],
+                                                      b_tensors[i],
+                                                      c_host_tensors[i],
+                                                      PassThrough{},
+                                                      PassThrough{},
+                                                      PassThrough{});
+
+            ref_invoker.Run(ref_argument);
+
+            for(int m = 0; m < problem_size.Ms[i]; ++m)
+            {
+                for(int n = 0; n < problem_size.Ns[i]; ++n)
+                {
+                    cde_element_op(
+                        c_host_tensors[i](m, n), c_host_tensors[i](m, n), d0_tensors[i](m, n));
+                }
+            }
+
+            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;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(32 + rand() % 32);
+        problem_size.Ns.push_back(1024);
+        problem_size.Ks.push_back(512);
+
+        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/59_grouped_gemm_multi_ABD/grouped_gemm_multi_abd_xdl_fixed_nk_bias_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2024, 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_multi_abd_xdl_fixed_nk.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_gemm_multi_abd.hpp"
+#include "ck/tensor_operation/gpu/element/combined_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 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 Add         = ck::tensor_operation::element_wise::Add;
+using Scale       = ck::tensor_operation::element_wise::Scale;
+using AddScale = ck::tensor_operation::element_wise::BinaryWithUnaryCombinedOp<Add, Scale, Scale>;
+
+using A0DataType       = F16;
+using A1DataType       = F32;
+using AsDataType       = ck::Tuple<A0DataType, A1DataType>;
+using B0DataType       = F16;
+using BsDataType       = ck::Tuple<B0DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = F32;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = F32;
+
+using A0Layout = Row;
+using A1Layout = Row;
+using AsLayout = ck::Tuple<A0Layout, A1Layout>;
+using B0Layout = Col;
+using BsLayout = ck::Tuple<B0Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using AElementOp = AddScale;
+using BElementOp = PassThrough;
+
+using CDEElementOp = Add;
+
+static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceGemmInstance = ck::tensor_operation::device::DeviceGroupedGemm_Xdl_Multi_ABD_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|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,    GemmDefault,        1,   128,    16,   128,    32,   8,   8,   16,   16,    1,    4,     S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              1,              1,         1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              1,         1,           1,           1,               S<1, 16, 1, 8>,               1,  ck::half_t>;
+
+// 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;
+    bool time_kernel     = false;
+    int k_batch          = 1;
+};
+
+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::GemmMultiABDDesc> gemm_descs;
+
+    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<A0DataType>> a0_tensors;
+    std::vector<Tensor<A1DataType>> a1_tensors;
+    std::vector<Tensor<B0DataType>> b_tensors;
+    std::vector<Tensor<D0DataType>> d0_tensors;
+    std::vector<Tensor<EDataType>> e_host_tensors;
+    std::vector<Tensor<EDataType>> e_device_tensors;
+
+    a0_tensors.reserve(group_count);
+    a1_tensors.reserve(group_count);
+    b_tensors.reserve(group_count);
+    d0_tensors.reserve(group_count);
+    e_host_tensors.reserve(group_count);
+    e_device_tensors.reserve(group_count);
+
+    using DeviceMemPtr = std::unique_ptr<DeviceMem>;
+
+    std::vector<DeviceMemPtr> a0_tensors_device, a1_tensors_device, b_tensors_device,
+        d0_tensors_device, c_tensors_device;
+
+    a0_tensors_device.reserve(group_count);
+    a1_tensors_device.reserve(group_count);
+    b_tensors_device.reserve(group_count);
+    d0_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];
+        a0_tensors.push_back(Tensor<A0DataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], A0Layout{})));
+        a1_tensors.push_back(Tensor<A1DataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ks[i], problem_size.stride_As[i], A1Layout{})));
+        b_tensors.push_back(Tensor<B0DataType>(f_host_tensor_descriptor(
+            problem_size.Ks[i], problem_size.Ns[i], problem_size.stride_Bs[i], B0Layout{})));
+        d0_tensors.push_back(Tensor<D0DataType>(
+            f_host_tensor_descriptor(problem_size.Ms[i], problem_size.Ns[i], 0, ELayout{})));
+        e_host_tensors.push_back(Tensor<EDataType>(f_host_tensor_descriptor(
+            problem_size.Ms[i], problem_size.Ns[i], problem_size.stride_Cs[i], ELayout{})));
+        e_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: " << a0_tensors[i].mDesc
+                  << " b_k_n: " << b_tensors[i].mDesc << " d_m_n: " << d0_tensors[i].mDesc
+                  << " c_m_n: " << e_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(A0DataType) * a0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(A1DataType) * a1_tensors[i].mDesc.GetElementSize() +
+                     sizeof(B0DataType) * b_tensors[i].mDesc.GetElementSize() +
+                     sizeof(D0DataType) * d0_tensors[i].mDesc.GetElementSize() +
+                     sizeof(EDataType) * e_device_tensors[i].mDesc.GetElementSize();
+
+        switch(config.init_method)
+        {
+        case 0: break;
+        case 1:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+            a1_tensors[i].GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+            break;
+        case 2:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+            a1_tensors[i].GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_3<B0DataType>{-0.5, 0.5});
+            break;
+        default:
+            a0_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            a1_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<0>{});
+            b_tensors[i].GenerateTensorValue(GeneratorTensor_Sequential<1>{});
+        }
+
+        d0_tensors[i].GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    constexpr ck::index_t NumATensor = 2;
+    constexpr ck::index_t NumBTensor = 1;
+    constexpr ck::index_t NumDTensor = 1;
+
+    using GroupedGemmKernelArgument = ck::tensor_operation::device::
+        GroupedGemmMultiABDKernelArgument<NumATensor, NumBTensor, NumDTensor>;
+
+    std::vector<GroupedGemmKernelArgument> grouped_gemm_kernel_args_;
+    grouped_gemm_kernel_args_.reserve(group_count);
+
+    for(int i = 0; i < group_count; i++)
+    {
+        a0_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(A0DataType) * sum_of_m * problem_size.Ks[i]));
+
+        a1_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(A1DataType) * sum_of_m * problem_size.Ks[i]));
+
+        b_tensors_device.emplace_back(std::make_unique<DeviceMem>(
+            sizeof(B0DataType) * problem_size.Ns[i] * problem_size.Ks[i]));
+
+        d0_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(D0DataType) * problem_size.Ns[i]));
+
+        c_tensors_device.emplace_back(
+            std::make_unique<DeviceMem>(sizeof(EDataType) * sum_of_m * problem_size.Ns[i]));
+
+        a0_tensors_device[i]->ToDevice(a0_tensors[i].mData.data(),
+                                       a0_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(A0DataType));
+
+        a1_tensors_device[i]->ToDevice(a1_tensors[i].mData.data(),
+                                       a1_tensors[i].mDesc.GetElementSpaceSize() *
+                                           sizeof(A1DataType));
+        b_tensors_device[i]->ToDevice(b_tensors[i].mData.data(),
+                                      b_tensors[i].mDesc.GetElementSpaceSize() *
+                                          sizeof(B0DataType));
+        d0_tensors_device[i]->ToDevice(d0_tensors[i].mData.data());
+        c_tensors_device[i]->SetZero();
+
+        gemm_descs.push_back({sum_of_m,
+                              problem_size.Ns[i],
+                              problem_size.Ks[i],
+                              {1, 1},
+                              {problem_size.stride_Bs[i]},
+                              {0},
+                              1});
+
+        grouped_gemm_kernel_args_.push_back(
+            {std::array<const void*, NumATensor>{a0_tensors_device[i]->GetDeviceBuffer(),
+                                                 a1_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumBTensor>{b_tensors_device[i]->GetDeviceBuffer()},
+             std::array<const void*, NumDTensor>{d0_tensors_device[i]->GetDeviceBuffer()},
+             c_tensors_device[i]->GetDeviceBuffer(),
+             problem_size.Ms[i],
+             problem_size.Ns[i],
+             problem_size.Ks[i],
+             std::array<ck::index_t, NumATensor>{problem_size.stride_As[i],
+                                                 problem_size.stride_As[i]},
+             std::array<ck::index_t, NumBTensor>{problem_size.stride_Bs[i]},
+             std::array<ck::index_t, NumDTensor>{0},
+             problem_size.stride_Cs[i]});
+    }
+
+    constexpr float scale = 1.f;
+    auto a_element_op     = AElementOp{Add{}, Scale{scale}, Scale{scale}};
+    auto b_element_op     = BElementOp{};
+    auto cde_element_op   = CDEElementOp{};
+
+    auto gemm    = DeviceGemmInstance{};
+    auto invoker = gemm.MakeInvoker();
+
+    std::vector<std::array<const void*, NumATensor>> p_As = {};
+    std::vector<std::array<const void*, NumBTensor>> p_Bs = {};
+    std::vector<std::array<const void*, NumDTensor>> p_Ds = {};
+    std::vector<void*> p_Cs                               = {};
+
+    // do GEMM
+    auto argument = gemm.MakeArgument(p_As, p_Bs, p_Ds, p_Cs, gemm_descs);
+
+    if(!gemm.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    DeviceMem gemm_workspace_dev(gemm.GetWorkSpaceSize(&argument));
+    gemm.SetWorkSpacePointer(&argument, gemm_workspace_dev.GetDeviceBuffer());
+
+    DeviceMem gemm_kernel_args_dev(gemm.GetDeviceKernelArgSize(&argument));
+    hip_check_error(hipMemcpy(gemm_kernel_args_dev.GetDeviceBuffer(),
+                              grouped_gemm_kernel_args_.data(),
+                              gemm.GetDeviceKernelArgSize(&argument),
+                              hipMemcpyHostToDevice));
+
+    gemm.SetDeviceKernelArgs(argument, gemm_kernel_args_dev.GetDeviceBuffer());
+    gemm.SetKBatch(argument, config.k_batch);
+
+    gemm.SetElementwiseOps(argument, a_element_op, b_element_op, cde_element_op);
+
+    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<A0DataType,
+                                                                                B0DataType,
+                                                                                EDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+
+        for(std::size_t i = 0; i < gemm_descs.size(); i++)
+        {
+            for(int m = 0; m < problem_size.Ms[i]; ++m)
+            {
+                for(int k = 0; k < problem_size.Ks[i]; ++k)
+                {
+                    a_element_op(a0_tensors[i](m, k), a0_tensors[i](m, k), a1_tensors[i](m, k));
+                }
+            }
+
+            c_tensors_device[i]->FromDevice(e_device_tensors[i].mData.data(),
+                                            e_device_tensors[i].mDesc.GetElementSize() *
+                                                sizeof(EDataType));
+
+            auto ref_gemm    = ReferenceGemmInstance{};
+            auto ref_invoker = ref_gemm.MakeInvoker();
+
+            auto ref_argument = ref_gemm.MakeArgument(a0_tensors[i],
+                                                      b_tensors[i],
+                                                      e_host_tensors[i],
+                                                      PassThrough{},
+                                                      b_element_op,
+                                                      PassThrough{});
+
+            ref_invoker.Run(ref_argument);
+
+            for(int m = 0; m < problem_size.Ms[i]; ++m)
+            {
+                for(int n = 0; n < problem_size.Ns[i]; ++n)
+                {
+                    cde_element_op(
+                        e_host_tensors[i](m, n), e_host_tensors[i](m, n), d0_tensors[i](m, n));
+                }
+            }
+
+            pass &= ck::utils::check_err(e_device_tensors[i], e_host_tensors[i]);
+        }
+    }
+
+    return pass;
+}
+
+int main(int argc, char* argv[])
+{
+    ProblemSize problem_size;
+    ExecutionConfig config;
+
+    problem_size.group_count = 16;
+
+    for(int i = 0; i < problem_size.group_count; i++)
+    {
+        problem_size.Ms.push_back(32 + rand() % 32);
+        problem_size.Ns.push_back(1024);
+        problem_size.Ks.push_back(512);
+
+        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/60_gemm_multi_ABD/CMakeLists.txt

+add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
+add_example_executable(example_gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8.cpp)
+add_example_executable(example_gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8 gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8.cpp)
+add_example_executable(example_gemm_multi_ABD_xdl_fastgelu_bf16_i8 gemm_multi_ABD_xdl_fastgelu_bf16_i8.cpp)

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_bias_fastgelu_bf16_i8.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/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = BF16;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<D0Layout>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using AddFastGelu = ck::tensor_operation::element_wise::AddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+using CDEElementOp = AddFastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
+    // 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|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,               S<1, 32, 1, 8>,               8,  ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = 0;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
+        exit(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});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
+    Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    b1_device_buf.ToDevice(b1_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 1;
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                                   b1_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumDTensor>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               std::array<ck::index_t, NumATensor>{StrideA},
+                               std::array<ck::index_t, NumBTensor>{StrideB, 0},
+                               std::array<ck::index_t, NumDTensor>{StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+    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"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        Tensor<A0DataType> a_m_k({M, K});
+
+        Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
+            }
+        }
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B1DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fastgelu_bf16_i8.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/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType, B1DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout, B1Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<>;
+using ELayout  = Row;
+
+using Multiply    = ck::tensor_operation::element_wise::Multiply;
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+using FastGelu    = ck::tensor_operation::element_wise::FastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = Multiply;
+using CDEElementOp = FastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
+    // 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|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,               S<1, 32, 1, 8>,               8,  ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = 0;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
+        exit(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});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
+    Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    b1_device_buf.ToDevice(b1_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 2;
+    constexpr ck::index_t NumDTensor = 0;
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer(),
+                                                                   b1_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumDTensor>{},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               std::array<ck::index_t, NumATensor>{StrideA},
+                               std::array<ck::index_t, NumBTensor>{StrideB, 0},
+                               std::array<ck::index_t, NumDTensor>{},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+    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"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        Tensor<A0DataType> a_m_k({M, K});
+
+        Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
+            }
+        }
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B1DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_fp16.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/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
+#include "ck/library/utility/check_err.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+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 AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F16;
+using EDataType        = F16;
+
+using ALayout = Row;
+using BLayout = Row;
+using DLayout = Row;
+using ELayout = Row;
+
+struct AddScale
+{
+    static constexpr auto I0 = ck::Number<0>{};
+    static constexpr auto I1 = ck::Number<1>{};
+    static constexpr auto I2 = ck::Number<2>{};
+    static constexpr auto I3 = ck::Number<3>{};
+
+    __host__ __device__ constexpr void
+    operator()(ck::half4_t& a, const ck::half4_t& a0, const ck::half4_t& a1) const
+    {
+        const auto a0_v_t = ck::vector_type<ck::half_t, 4>{a0};
+        const auto a1_v_t = ck::vector_type<ck::half_t, 4>{a1};
+
+        auto r_v_t = ck::vector_type<ck::half_t, 4>{};
+
+        r_v_t.AsType<ck::half_t>()(I0) =
+            scale * (a0_v_t.AsType<ck::half_t>()[I0] + a1_v_t.AsType<ck::half_t>()[I0]);
+        r_v_t.AsType<ck::half_t>()(I1) =
+            scale * (a0_v_t.AsType<ck::half_t>()[I1] + a1_v_t.AsType<ck::half_t>()[I1]);
+        r_v_t.AsType<ck::half_t>()(I2) =
+            scale * (a0_v_t.AsType<ck::half_t>()[I2] + a1_v_t.AsType<ck::half_t>()[I2]);
+        r_v_t.AsType<ck::half_t>()(I3) =
+            scale * (a0_v_t.AsType<ck::half_t>()[I3] + a1_v_t.AsType<ck::half_t>()[I3]);
+
+        a = r_v_t.AsType<ck::half4_t>()[I0];
+    }
+
+    __host__ __device__ constexpr void
+    operator()(ck::half_t& a, const ck::half_t& a0, const ck::half_t& a1) const
+    {
+        a = scale * (a0 + a1);
+    }
+
+    // this attribute controls the copy_function applying element_wise_op with
+    // pack4_data
+    constexpr const static bool is_pack4_invocable = true;
+
+    float scale = 1.0;
+};
+
+struct AlphaBetaAdd
+{
+    AlphaBetaAdd(float alpha, float beta) : alpha_(alpha), beta_(beta){};
+
+    template <typename E, typename C, typename D>
+    __host__ __device__ constexpr void operator()(E& e, const C& c, const D& d) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<ck::half_t, float, ck::half_t>(
+        ck::half_t& e, const float& c, const ck::half_t& d) const
+    {
+        e = ck::type_convert<ck::half_t>(alpha_ * c + beta_ * ck::type_convert<float>(d));
+    };
+
+    float alpha_;
+    float beta_;
+};
+
+using AElementOp   = AddScale;
+using BElementOp   = PassThrough;
+using CDEElementOp = AlphaBetaAdd;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle<
+    ck::Tuple<ALayout, ALayout>,
+    ck::Tuple<BLayout>,
+    ck::Tuple<DLayout>,
+    ELayout,
+    ck::Tuple<ADataType, ADataType>,
+    ck::Tuple<BDataType>,
+    AccDataType,
+    CShuffleDataType,
+    ck::Tuple<DDataType>,
+    EDataType,
+    AElementOp,
+    BElementOp,
+    CDEElementOp,
+    GemmSpec,
+    1,
+    256,
+    256,
+    128,
+    32,
+    8,
+    8,
+    32,
+    32,
+    4,
+    2,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    2,
+    8,
+    8,
+    1,
+    S<4, 64, 1>,
+    S<1, 0, 2>,
+    S<1, 0, 2>,
+    1,
+    2,
+    8,
+    1,
+    1,
+    1,
+    S<1, 32, 1, 8>,
+    8>;
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 3840;
+    ck::index_t N = 4096;
+    ck::index_t K = 4096;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = N;
+    ck::index_t StrideE = N;
+
+    float alpha = 1.0f;
+    float beta  = 1.0f;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 6)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        alpha = std::stof(argv[4]);
+        beta  = std::stof(argv[5]);
+    }
+    else if(argc == 13)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+
+        alpha = std::stof(argv[11]);
+        beta  = std::stof(argv[12]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE, alpha, "
+               "beta\n");
+        exit(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});
+            }
+        };
+
+    Tensor<ADataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<ADataType> a1_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
+    Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
+    Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, DLayout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "a1_m_k: " << a1_m_k.mDesc << std::endl;
+    std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        a1_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
+        b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<DDataType>{-5, 5});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        a1_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
+        b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(ADataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem a1_device_buf(sizeof(ADataType) * a1_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    a1_device_buf.ToDevice(a1_m_k.mData.data());
+    b_device_buf.ToDevice(b_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{0.2};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{alpha, beta};
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(std::array<const void*, 2>{a0_device_buf.GetDeviceBuffer(),
+                                                          a1_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, 1>{b_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, 1>{d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               std::array<ck::index_t, 2>{StrideA, StrideA},
+                               std::array<ck::index_t, 1>{StrideB},
+                               std::array<ck::index_t, 1>{StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
+
+    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"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        Tensor<ADataType> a_m_k({M, K});
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                a_element_op(a_m_k(m, k), a0_m_k(m, k), a1_m_k(m, k));
+            }
+        }
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
+                                                                                BDataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                BElementOp,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument =
+            ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, PassThrough{}, b_element_op, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/60_gemm_multi_ABD/gemm_multi_ABD_xdl_multiply_bias_fastgelu_bf16_i8.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/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.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"
+#include "ck/library/utility/check_err.hpp"
+
+#include "ck/utility/blkgemmpipe_scheduler.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16  = ck::half_t;
+using BF16 = ck::bhalf_t;
+using I8   = int8_t;
+using F32  = float;
+
+using Row = ck::tensor_layout::gemm::RowMajor;
+using Col = ck::tensor_layout::gemm::ColumnMajor;
+
+using A0DataType       = BF16;
+using AsDataType       = ck::Tuple<A0DataType>;
+using B0DataType       = I8;
+using B1DataType       = BF16;
+using BsDataType       = ck::Tuple<B0DataType>;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using D0DataType       = BF16;
+using DsDataType       = ck::Tuple<B1DataType, D0DataType>;
+using EDataType        = BF16;
+
+using A0Layout = Row;
+using AsLayout = ck::Tuple<A0Layout>;
+using B0Layout = Row;
+using B1Layout = B0Layout;
+using BsLayout = ck::Tuple<B0Layout>;
+using D0Layout = Row;
+using DsLayout = ck::Tuple<B1Layout, D0Layout>;
+using ELayout  = Row;
+
+using PassThrough         = ck::tensor_operation::element_wise::PassThrough;
+using MultiplyAddFastGelu = ck::tensor_operation::element_wise::MultiplyAddFastGelu;
+
+using AElementOp   = PassThrough;
+using BElementOp   = PassThrough;
+using CDEElementOp = MultiplyAddFastGelu;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::Default;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceGemmMultipleABD_Xdl_CShuffle
+    // 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|
+///######|         |         |         |        |           |           |            |                 |           |          |            |            |             |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+         < AsLayout, BsLayout, DsLayout, ELayout, AsDataType, BsDataType, AccDataType, CShuffleDataType, DsDataType, EDataType,  AElementOp,  BElementOp, CDEElementOp,       GemmSpec,        1,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,               8,              4,          0,          1,           1,               S<1, 32, 1, 8>,               8,  ck::BlockGemmPipelineScheduler::Intrawave, ck::BlockGemmPipelineVersion::v4>;
+// clang-format on
+
+int main(int argc, char* argv[])
+{
+    bool do_verification = true;
+    int init_method      = 1;
+    bool time_kernel     = false;
+
+    // GEMM shape
+    ck::index_t M = 4096;
+    ck::index_t N = 768;
+    ck::index_t K = 6144;
+
+    ck::index_t StrideA = K;
+    ck::index_t StrideB = N;
+    ck::index_t StrideD = 0;
+    ck::index_t StrideE = N;
+
+    if(argc == 1)
+    {
+        // use default case
+    }
+    else if(argc == 4)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+    }
+    else if(argc == 11)
+    {
+        do_verification = std::stoi(argv[1]);
+        init_method     = std::stoi(argv[2]);
+        time_kernel     = std::stoi(argv[3]);
+
+        M = std::stoi(argv[4]);
+        N = std::stoi(argv[5]);
+        K = std::stoi(argv[6]);
+
+        StrideA = std::stoi(argv[7]);
+        StrideB = std::stoi(argv[8]);
+        StrideD = std::stoi(argv[9]);
+        StrideE = std::stoi(argv[10]);
+    }
+    else
+    {
+        printf("arg1: verification (0=no, 1=yes)\n");
+        printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
+        printf("arg3: time kernel (0=no, 1=yes)\n");
+        printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideD, StrideE\n");
+        exit(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});
+            }
+        };
+
+    Tensor<A0DataType> a0_m_k(f_host_tensor_descriptor(M, K, StrideA, A0Layout{}));
+    Tensor<B0DataType> b0_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+    Tensor<B1DataType> b1_k_n(f_host_tensor_descriptor(K, N, 0, B1Layout{}));
+    Tensor<D0DataType> d_m_n(f_host_tensor_descriptor(M, N, StrideD, D0Layout{}));
+    Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+    Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
+
+    std::cout << "a0_m_k: " << a0_m_k.mDesc << std::endl;
+    std::cout << "b0_k_n: " << b0_k_n.mDesc << std::endl;
+    std::cout << "b1_k_n: " << b1_k_n.mDesc << std::endl;
+    std::cout << "d_m_n: " << d_m_n.mDesc << std::endl;
+    std::cout << "e_m_n: " << e_m_n_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_2<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
+        break;
+    default:
+        a0_m_k.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        b0_k_n.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-5, 5});
+        b1_k_n.GenerateTensorValue(GeneratorTensor_3<B1DataType>{0, 5});
+        d_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{-0.5, 0.5});
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_m_k.mDesc.GetElementSpaceSize());
+    DeviceMem b0_device_buf(sizeof(B0DataType) * b0_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem b1_device_buf(sizeof(B1DataType) * b1_k_n.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(D0DataType) * d_m_n.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_m_k.mData.data());
+    b0_device_buf.ToDevice(b0_k_n.mData.data());
+    b1_device_buf.ToDevice(b1_k_n.mData.data());
+    d_device_buf.ToDevice(d_m_n.mData.data());
+    e_device_buf.ToDevice(e_m_n_device_result.mData.data());
+
+    auto a_element_op   = AElementOp{};
+    auto b_element_op   = BElementOp{};
+    auto cde_element_op = CDEElementOp{};
+
+    constexpr ck::index_t NumATensor = 1;
+    constexpr ck::index_t NumBTensor = 1;
+    constexpr ck::index_t NumDTensor = 2;
+
+    // do GEMM
+    auto device_op = DeviceOpInstance{};
+    auto invoker   = device_op.MakeInvoker();
+    auto argument =
+        device_op.MakeArgument(std::array<const void*, NumATensor>{a0_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumBTensor>{b0_device_buf.GetDeviceBuffer()},
+                               std::array<const void*, NumDTensor>{b1_device_buf.GetDeviceBuffer(),
+                                                                   d_device_buf.GetDeviceBuffer()},
+                               e_device_buf.GetDeviceBuffer(),
+                               M,
+                               N,
+                               K,
+                               std::array<ck::index_t, NumATensor>{StrideA},
+                               std::array<ck::index_t, NumBTensor>{StrideB},
+                               std::array<ck::index_t, NumDTensor>{0, StrideD},
+                               StrideE,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_gemm with the specified compilation parameters does "
+            "not support this GEMM problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    std::size_t flop = std::size_t(2) * M * N * K;
+    std::size_t num_btype =
+        sizeof(A0DataType) * M * K + sizeof(B0DataType) * K * N + sizeof(EDataType) * M * N;
+
+    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"
+              << std::endl;
+
+    e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+    if(do_verification)
+    {
+        Tensor<CShuffleDataType> c_m_n({M, N});
+
+        Tensor<A0DataType> a_m_k({M, K});
+
+        Tensor<B1DataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, B0Layout{}));
+
+#if 0
+        for(int n = 0; n < N; ++n)
+        {
+            for(int k = 0; k < K; ++k)
+            {
+                b_element_op(b_k_n(k, n), b0_k_n(k, n), b1_k_n(k, n));
+            }
+        }
+#endif
+
+        using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<A0DataType,
+                                                                                B0DataType,
+                                                                                CShuffleDataType,
+                                                                                AccDataType,
+                                                                                PassThrough,
+                                                                                PassThrough,
+                                                                                PassThrough>;
+        auto ref_gemm               = ReferenceGemmInstance{};
+        auto ref_invoker            = ref_gemm.MakeInvoker();
+
+        auto ref_argument = ref_gemm.MakeArgument(
+            a0_m_k, b0_k_n, c_m_n, PassThrough{}, PassThrough{}, PassThrough{});
+
+        ref_invoker.Run(ref_argument);
+
+        for(int m = 0; m < M; ++m)
+        {
+            for(int n = 0; n < N; ++n)
+            {
+                cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), b1_k_n(0, n), d_m_n(m, n));
+            }
+        }
+
+        e_device_buf.FromDevice(e_m_n_device_result.mData.data());
+
+        return ck::utils::check_err(e_m_n_device_result, e_m_n_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}