merge develop

example/50_put_element/CMakeLists.txt

+add_example_executable(example_put_element_fp16 put_element_fp16.cpp)

example/50_put_element/put_element_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_put_element_impl.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/device_memory.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+#include "ck/library/utility/host_tensor_generator.hpp"
+
+using XDataType     = ck::half_t;
+using YDataType     = ck::half_t;
+using IndexDataType = int32_t;
+
+using YElementwiseOp = ck::tensor_operation::element_wise::PassThrough;
+
+using DeviceInstance =
+    ck::tensor_operation::device::DevicePutElementImpl<XDataType,     // XDataType
+                                                       IndexDataType, // IndexDataType
+                                                       YDataType,     // YDataType
+                                                       YElementwiseOp,
+                                                       ck::InMemoryDataOperationEnum::Set,
+                                                       1>;
+
+int main()
+{
+    bool do_verification = true;
+    bool time_kernel     = false;
+
+    int N = 1024;
+
+    Tensor<XDataType> x(HostTensorDescriptor{N});
+    Tensor<IndexDataType> indices(HostTensorDescriptor{N});
+    Tensor<YDataType> y(HostTensorDescriptor{N});
+
+    x.GenerateTensorValue(GeneratorTensor_3<XDataType>{-1.0, 1.0});
+    for(int i = 0; i < N; ++i)
+        indices(i) = i;
+
+    DeviceMem x_device_buf(sizeof(XDataType) * x.mDesc.GetElementSpaceSize());
+    DeviceMem y_device_buf(sizeof(YDataType) * y.mDesc.GetElementSpaceSize());
+    DeviceMem indices_device_buf(sizeof(IndexDataType) * indices.mDesc.GetElementSpaceSize());
+
+    x_device_buf.ToDevice(x.mData.data());
+    indices_device_buf.ToDevice(indices.mData.data());
+
+    auto put_instance     = DeviceInstance{};
+    auto put_invoker_ptr  = put_instance.MakeInvokerPointer();
+    auto put_argument_ptr = put_instance.MakeArgumentPointer(
+        static_cast<XDataType*>(x_device_buf.GetDeviceBuffer()),
+        static_cast<IndexDataType*>(indices_device_buf.GetDeviceBuffer()),
+        static_cast<YDataType*>(y_device_buf.GetDeviceBuffer()),
+        N,
+        N,
+        YElementwiseOp{});
+
+    if(!put_instance.IsSupportedArgument(put_argument_ptr.get()))
+    {
+        throw std::runtime_error("argument is not supported!");
+    }
+
+    float ave_time =
+        put_invoker_ptr->Run(put_argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+
+    std::cout << "perf: " << ave_time << " ms" << std::endl;
+
+    bool pass = true;
+    if(do_verification)
+    {
+        Tensor<YDataType> y_host(HostTensorDescriptor{N});
+
+        for(int i = 0; i < N; ++i)
+        {
+            IndexDataType idx = indices(i);
+            y_host(idx)       = x(i);
+        }
+
+        y_device_buf.FromDevice(y.mData.data());
+        pass = ck::utils::check_err(y, y_host);
+    }
+
+    return (pass ? 0 : 1);
+}

example/51_avgpool3d_bwd/CMakeLists.txt

+add_example_executable(example_avgpool3d_bwd_bf16 avgpool3d_bwd_bf16.cpp)
+add_example_executable(example_avgpool3d_bwd_fp16 avgpool3d_bwd_fp16.cpp)
+add_example_executable(example_avgpool3d_bwd_fp32 avgpool3d_bwd_fp32.cpp)

example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+
+#include "avgpool3d_bwd_common.hpp"
+
+using DOutDataType    = ck::bhalf_t;
+using DInDataType     = ck::bhalf_t;
+using ComputeDataType = float;
+
+#if 1
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+#else
+using DOutLayout = ck::tensor_layout::convolution::NCDHW;
+using DInLayout  = ck::tensor_layout::convolution::NCDHW;
+#endif
+
+using DevicePoolBwdInstance =
+    ck::tensor_operation::device::DeviceAvgPool3dBwd_NDHWC_NDHWC<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 64, // BlockSize
+                                                                 64, // ReduceMThreadClusterSize
+                                                                 1,  // ReduceKThreadClusterSize
+                                                                 1,  // ReduceMThreadSliceSize
+                                                                 1,  // ReduceKThreadSliceSize
+                                                                 1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    std::vector<ck::index_t> window_lengths    = {5, 5, 5};
+    std::vector<ck::index_t> window_strides    = {2, 2, 2};
+    std::vector<ck::index_t> window_dilations  = {2, 2, 2};
+    std::vector<ck::index_t> dinput_left_pads  = {0, 0, 0};
+    std::vector<ck::index_t> dinput_right_pads = {0, 0, 0};
+
+    ck::index_t N  = 1;
+    ck::index_t C  = 16;
+    ck::index_t Di = 40;
+    ck::index_t Hi = 40;
+    ck::index_t Wi = 40;
+
+    pool3d_bwd_test<DevicePoolBwdInstance, DOutDataType, DInDataType, DOutLayout, DInLayout>(
+        true,
+        false,
+        N,
+        C,
+        Di,
+        Hi,
+        Wi,
+        window_lengths,
+        window_strides,
+        window_dilations,
+        dinput_left_pads,
+        dinput_right_pads);
+}

example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.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_avgpool_bwd.hpp"
+
+template <typename TensorLayout>
+std::vector<ck::index_t> f_tensor_strides_ncdhw(ck::index_t N_,
+                                                ck::index_t C_,
+                                                ck::index_t D,
+                                                ck::index_t H,
+                                                ck::index_t W,
+                                                TensorLayout layout)
+{
+    using namespace ck::literals;
+    (void)N_;
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
+        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>
+HostTensorDescriptor f_host_tensor_descriptor(std::size_t N_,
+                                              std::size_t C_,
+                                              std::size_t D,
+                                              std::size_t H,
+                                              std::size_t W,
+                                              TensorLayout layout)
+{
+    using namespace ck::literals;
+
+    if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCDHW>::value)
+    {
+        return HostTensorDescriptor({N_, C_, D, H, W}, {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 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,
+          typename DOutDataType,
+          typename DInDataType,
+          typename DOutLayout,
+          typename DInLayout>
+bool pool3d_bwd_test(bool do_verification,
+                     bool time_kernel,
+                     ck::index_t N,
+                     ck::index_t C,
+                     ck::index_t Di,
+                     ck::index_t Hi,
+                     ck::index_t Wi,
+                     std::vector<ck::index_t> window_lengths,
+                     std::vector<ck::index_t> window_strides,
+                     std::vector<ck::index_t> window_dilations,
+                     std::vector<ck::index_t> dinput_left_pads,
+                     std::vector<ck::index_t> dinput_right_pads)
+{
+    auto OutSpatialLength = [&](auto InSpatialLength, int index) {
+        ck::index_t left_pad   = dinput_left_pads[index];
+        ck::index_t right_pad  = dinput_right_pads[index];
+        ck::index_t window_len = window_lengths[index];
+        ck::index_t stride     = window_strides[index];
+        ck::index_t dilation   = window_dilations[index];
+        ck::index_t eff        = (window_len - 1) * dilation + 1;
+        return (InSpatialLength + left_pad + right_pad - eff) / stride + 1;
+    };
+
+    ck::index_t Do = OutSpatialLength(Di, 0);
+    ck::index_t Ho = OutSpatialLength(Hi, 1);
+    ck::index_t Wo = OutSpatialLength(Wi, 2);
+
+    Tensor<DOutDataType> dout(f_host_tensor_descriptor(N, C, Do, Ho, Wo, DOutLayout{}));
+    Tensor<DInDataType> din_dev(f_host_tensor_descriptor(N, C, Di, Hi, Wi, DInLayout{}));
+    Tensor<DInDataType> din_host(f_host_tensor_descriptor(N, C, Di, Hi, Wi, DInLayout{}));
+
+    std::cout << "dout: " << dout.mDesc << std::endl;
+    std::cout << "din_host: " << din_host.mDesc << std::endl;
+
+    dout.GenerateTensorValue(GeneratorTensor_3<DOutDataType>{0.0, 1.0});
+
+    DeviceMem dout_device_buf(sizeof(DOutDataType) * dout.mDesc.GetElementSpaceSize());
+    DeviceMem din_device_buf(sizeof(DInDataType) * din_dev.mDesc.GetElementSpaceSize());
+
+    dout_device_buf.ToDevice(dout.mData.data());
+    din_device_buf.SetZero();
+
+    auto pool        = DevicePoolBwdInstance{};
+    auto invoker_ptr = pool.MakeInvokerPointer();
+    auto argument_ptr =
+        pool.MakeArgumentPointer(static_cast<DOutDataType*>(dout_device_buf.GetDeviceBuffer()),
+                                 static_cast<DInDataType*>(din_device_buf.GetDeviceBuffer()),
+                                 {N, C, Do, Ho, Wo},
+                                 {N, C, Di, Hi, Wi},
+                                 f_tensor_strides_ncdhw(N, C, Do, Ho, Wo, DOutLayout{}),
+                                 f_tensor_strides_ncdhw(N, C, Di, Hi, Wi, DInLayout{}),
+                                 window_lengths,
+                                 window_strides,
+                                 window_dilations,
+                                 dinput_left_pads,
+                                 dinput_right_pads);
+
+    if(!pool.IsSupportedArgument(argument_ptr.get()))
+    {
+        throw std::runtime_error("wrong! device_op with the specified compilation parameters does "
+                                 "not support this problem");
+    }
+
+    float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
+    std::cout << "Perf: " << ave_time << std::endl;
+
+    bool pass = true;
+
+    if(do_verification)
+    {
+        auto ref_pool =
+            ck::tensor_operation::host::ReferenceAvgPoolBwd<3, DInDataType, DOutDataType>();
+
+        auto ref_invoker = ref_pool.MakeInvoker();
+
+        auto ref_argument = ref_pool.MakeArgument(din_host,
+                                                  dout,
+                                                  window_lengths,
+                                                  window_strides,
+                                                  window_dilations,
+                                                  dinput_left_pads,
+                                                  dinput_right_pads);
+
+        ref_invoker.Run(ref_argument);
+
+        din_device_buf.FromDevice(din_dev.mData.data());
+        pass = ck::utils::check_err(din_dev, din_host);
+    }
+
+    return pass;
+}

example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+
+#include "avgpool3d_bwd_common.hpp"
+
+using DOutDataType    = ck::half_t;
+using DInDataType     = ck::half_t;
+using ComputeDataType = float;
+
+#if 1
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+#else
+using DOutLayout = ck::tensor_layout::convolution::NCDHW;
+using DInLayout  = ck::tensor_layout::convolution::NCDHW;
+#endif
+
+using DevicePoolBwdInstance =
+    ck::tensor_operation::device::DeviceAvgPool3dBwd_NDHWC_NDHWC<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 64, // BlockSize
+                                                                 64, // ReduceMThreadClusterSize
+                                                                 1,  // ReduceKThreadClusterSize
+                                                                 1,  // ReduceMThreadSliceSize
+                                                                 1,  // ReduceKThreadSliceSize
+                                                                 1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    std::vector<ck::index_t> window_lengths    = {5, 5, 5};
+    std::vector<ck::index_t> window_strides    = {2, 2, 2};
+    std::vector<ck::index_t> window_dilations  = {2, 2, 2};
+    std::vector<ck::index_t> dinput_left_pads  = {0, 0, 0};
+    std::vector<ck::index_t> dinput_right_pads = {0, 0, 0};
+
+    ck::index_t N  = 1;
+    ck::index_t C  = 16;
+    ck::index_t Di = 40;
+    ck::index_t Hi = 40;
+    ck::index_t Wi = 40;
+
+    pool3d_bwd_test<DevicePoolBwdInstance, DOutDataType, DInDataType, DOutLayout, DInLayout>(
+        true,
+        false,
+        N,
+        C,
+        Di,
+        Hi,
+        Wi,
+        window_lengths,
+        window_strides,
+        window_dilations,
+        dinput_left_pads,
+        dinput_right_pads);
+}

example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <iostream>
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp"
+
+#include "avgpool3d_bwd_common.hpp"
+
+using DOutDataType    = float;
+using DInDataType     = float;
+using ComputeDataType = float;
+
+#if 1
+using DOutLayout = ck::tensor_layout::convolution::NDHWC;
+using DInLayout  = ck::tensor_layout::convolution::NDHWC;
+#else
+using DOutLayout = ck::tensor_layout::convolution::NCDHW;
+using DInLayout  = ck::tensor_layout::convolution::NCDHW;
+#endif
+
+using DevicePoolBwdInstance =
+    ck::tensor_operation::device::DeviceAvgPool3dBwd_NDHWC_NDHWC<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 64, // BlockSize
+                                                                 64, // ReduceMThreadClusterSize
+                                                                 1,  // ReduceKThreadClusterSize
+                                                                 1,  // ReduceMThreadSliceSize
+                                                                 1,  // ReduceKThreadSliceSize
+                                                                 1>; // InSrcOutDstVectorSize
+
+int main()
+{
+    std::vector<ck::index_t> window_lengths    = {5, 5, 5};
+    std::vector<ck::index_t> window_strides    = {2, 2, 2};
+    std::vector<ck::index_t> window_dilations  = {2, 2, 2};
+    std::vector<ck::index_t> dinput_left_pads  = {0, 0, 0};
+    std::vector<ck::index_t> dinput_right_pads = {0, 0, 0};
+
+    ck::index_t N  = 1;
+    ck::index_t C  = 16;
+    ck::index_t Di = 40;
+    ck::index_t Hi = 40;
+    ck::index_t Wi = 40;
+
+    pool3d_bwd_test<DevicePoolBwdInstance, DOutDataType, DInDataType, DOutLayout, DInLayout>(
+        true,
+        false,
+        N,
+        C,
+        Di,
+        Hi,
+        Wi,
+        window_lengths,
+        window_strides,
+        window_dilations,
+        dinput_left_pads,
+        dinput_right_pads);
+}

example/52_im2col_col2im/CMakeLists.txt

+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+    if(gpu IN_LIST gpu_list AND target EQUAL 0)
+        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)
+
+        set(target 1)
+    endif()
+endforeach()

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/60_gemm_multi_ABD/CMakeLists.txt

+list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+    if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
+        add_example_executable(example_gemm_multi_ABD_xdl_fp16 gemm_multi_ABD_xdl_fp16.cpp)
+        set(target 1)
+    endif()
+endforeach()

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 = Col;
+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>,
+    2,
+    8,
+    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 = 4096;
+    ck::index_t StrideB = 4096;
+    ck::index_t StrideD = 4096;
+    ck::index_t StrideE = 4096;
+
+    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/61_contraction_multi_ABD/CMakeLists.txt

+list(APPEND gpu_list2 gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list2 AND target EQUAL 0)
+     add_example_executable(example_contraction_multi_ABD_xdl_fp16 contraction_multi_ABD_xdl_fp16.cpp)
+   set(target 1)
+ endif()
+endforeach()

example/61_contraction_multi_ABD/contraction_multi_ABD_xdl_fp16.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/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_contraction_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_contraction.hpp"
+#include "ck/library/utility/check_err.hpp"
+#include "ck/library/utility/numeric.hpp"
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using F16 = ck::half_t;
+using F32 = float;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+using A0DataType       = F16;
+using A1DataType       = F32;
+using BDataType        = F16;
+using AccDataType      = F32;
+using CShuffleDataType = F32;
+using DDataType        = F16;
+using EDataType        = F16;
+using ComputeDataType  = F16;
+
+static constexpr ck::index_t NumDimM = 2;
+static constexpr ck::index_t NumDimN = 2;
+static constexpr ck::index_t NumDimK = 2;
+
+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_;
+};
+
+struct Multiply
+{
+    __host__ __device__ constexpr void
+    operator()(ck::half_t& a, const ck::half_t& a0, const float& a1) const
+    {
+        a = ck::type_convert<ck::half_t>(ck::type_convert<float>(a0) * a1);
+    }
+};
+
+using AElementOp   = Multiply;
+using BElementOp   = PassThrough;
+using CDEElementOp = AlphaBetaAdd;
+
+static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
+
+using DeviceOpInstance = ck::tensor_operation::device::DeviceContractionMultipleABD_Xdl_CShuffle<
+    NumDimM,
+    NumDimN,
+    NumDimK,
+    ck::Tuple<A0DataType, A1DataType>,
+    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>,
+    2,
+    8,
+    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;
+
+    float alpha = 1.0f;
+    float beta  = 1.0f;
+
+    // A0[M0, M1, K0, K1]
+    std::vector<ck::index_t> a0_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a0_ms_ks_strides{128 * 32 * 64, 32 * 64, 64, 1};
+    // A1[M1, K1] -> A1[M0, M1, K0, K1]
+    std::vector<ck::index_t> a1_ms_ks_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> a1_ms_ks_strides{0, 64, 0, 1};
+    // B[N0, N1, K0, K1]
+    std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
+    std::vector<ck::index_t> b_ns_ks_strides{64 * 32 * 64, 32 * 64, 64, 1};
+    // D[M0, M1, N0, N1]
+    std::vector<ck::index_t> d_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> d_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
+    // E[M0, M1, N0, N1]
+    std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
+    std::vector<ck::index_t> e_ms_ns_strides{128 * 32 * 64, 32 * 64, 64, 1};
+
+    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
+    {
+        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");
+        exit(0);
+    }
+
+    Tensor<A0DataType> a0_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+    Tensor<A1DataType> a1_ms_ks(a1_ms_ks_lengths, a1_ms_ks_strides);
+    Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
+    Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+    Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+    std::cout << "a0_ms_ks: " << a0_ms_ks.mDesc << std::endl;
+    std::cout << "a1_ms_ks: " << a1_ms_ks.mDesc << std::endl;
+    std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
+    std::cout << "d_ms_ns: " << d_ms_ns.mDesc << std::endl;
+    std::cout << "e_ms_ns: " << e_ms_ns_host_result.mDesc << std::endl;
+
+    switch(init_method)
+    {
+    case 0: break;
+    case 1:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_2<A0DataType>{-5, 5});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_2<A1DataType>{-5, 5});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
+        break;
+    default:
+        a0_ms_ks.GenerateTensorValue(GeneratorTensor_3<A0DataType>{0.0, 1.0});
+        a1_ms_ks.GenerateTensorValue(GeneratorTensor_3<A1DataType>{0.0, 1.0});
+        b_ns_ks.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        d_ms_ns.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
+        break;
+    }
+
+    DeviceMem a0_device_buf(sizeof(A0DataType) * a0_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem a1_device_buf(sizeof(A1DataType) * a1_ms_ks.mDesc.GetElementSpaceSize());
+    DeviceMem b_device_buf(sizeof(BDataType) * b_ns_ks.mDesc.GetElementSpaceSize());
+    DeviceMem d_device_buf(sizeof(DDataType) * d_ms_ns.mDesc.GetElementSpaceSize());
+    DeviceMem e_device_buf(sizeof(EDataType) * e_ms_ns_device_result.mDesc.GetElementSpaceSize());
+
+    a0_device_buf.ToDevice(a0_ms_ks.mData.data());
+    a1_device_buf.ToDevice(a1_ms_ks.mData.data());
+    b_device_buf.ToDevice(b_ns_ks.mData.data());
+    d_device_buf.ToDevice(d_ms_ns.mData.data());
+
+    // set zero
+    e_device_buf.SetZero();
+
+    auto a_element_op   = AElementOp{};
+    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(),
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_lengths, a1_ms_ks_lengths},
+        std::array<std::vector<ck::index_t>, 2>{a0_ms_ks_strides, a1_ms_ks_strides},
+        std::array<std::vector<ck::index_t>, 1>{b_ns_ks_lengths},
+        std::array<std::vector<ck::index_t>, 1>{b_ns_ks_strides},
+        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
+        std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
+        e_ms_ns_lengths,
+        e_ms_ns_strides,
+        a_element_op,
+        b_element_op,
+        cde_element_op);
+
+    if(!device_op.IsSupportedArgument(argument))
+    {
+        throw std::runtime_error(
+            "wrong! device_contraction with the specified compilation parameters does "
+            "not support this problem");
+    }
+
+    float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
+
+    if(time_kernel)
+    {
+        ck::index_t M =
+            ck::accumulate_n<ck::index_t>(e_ms_ns_lengths.begin(), NumDimM, 1, std::multiplies<>{});
+
+        ck::index_t N = ck::accumulate_n<ck::index_t>(
+            e_ms_ns_lengths.begin() + NumDimM, NumDimN, 1, std::multiplies<>{});
+
+        ck::index_t K = ck::accumulate_n<ck::index_t>(
+            a0_ms_ks_lengths.begin() + NumDimM, NumDimK, 1, std::multiplies<>{});
+
+        std::size_t flop = std::size_t(2) * M * N * K;
+        std::size_t num_btype =
+            sizeof(A0DataType) * 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;
+    }
+
+    if(do_verification)
+    {
+
+        Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
+
+        Tensor<A0DataType> a_ms_ks(a0_ms_ks_lengths, a0_ms_ks_strides);
+
+        for(size_t m0 = 0; m0 < a_ms_ks.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < a_ms_ks.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t k0 = 0; k0 < a_ms_ks.mDesc.GetLengths()[2]; ++k0)
+                {
+                    for(size_t k1 = 0; k1 < a_ms_ks.mDesc.GetLengths()[3]; ++k1)
+                    {
+                        a_element_op(a_ms_ks(m0, m1, k0, k1),
+                                     a0_ms_ks(m0, m1, k0, k1),
+                                     a1_ms_ks(m0, m1, k0, k1));
+                    }
+                }
+            }
+        }
+
+        using ReferenceOpInstance =
+            ck::tensor_operation::host::ReferenceContraction_M2_N2_K2<NumDimM,
+                                                                      NumDimN,
+                                                                      NumDimK,
+                                                                      A0DataType,
+                                                                      BDataType,
+                                                                      CShuffleDataType,
+                                                                      AccDataType,
+                                                                      ComputeDataType,
+                                                                      PassThrough,
+                                                                      BElementOp>;
+
+        auto ref_op      = ReferenceOpInstance{};
+        auto ref_invoker = ref_op.MakeInvoker();
+
+        Tensor<float> empty_tensor(std::vector<ck::index_t>{}, std::vector<ck::index_t>{});
+        auto ref_argument =
+            ref_op.MakeArgument(a_ms_ks, b_ns_ks, c_ms_ns_host_result, PassThrough{}, b_element_op);
+
+        ref_invoker.Run(ref_argument);
+
+        for(size_t m0 = 0; m0 < e_ms_ns_host_result.mDesc.GetLengths()[0]; ++m0)
+        {
+            for(size_t m1 = 0; m1 < e_ms_ns_host_result.mDesc.GetLengths()[1]; ++m1)
+            {
+                for(size_t n0 = 0; n0 < e_ms_ns_host_result.mDesc.GetLengths()[2]; ++n0)
+                {
+                    for(size_t n1 = 0; n1 < e_ms_ns_host_result.mDesc.GetLengths()[3]; ++n1)
+                    {
+                        cde_element_op(e_ms_ns_host_result(m0, m1, n0, n1),
+                                       c_ms_ns_host_result(m0, m1, n0, n1),
+                                       d_ms_ns(m0, m1, n0, n1));
+                    }
+                }
+            }
+        }
+
+        e_device_buf.FromDevice(e_ms_ns_device_result.mData.data());
+
+        return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
+    }
+
+    return 0;
+}

example/62_convnd_activ/CMakeLists.txt

+add_subdirectory(binary)
+add_subdirectory(multi_AB)
+add_subdirectory(unary)
+
+list(APPEND gpu_list gfx908 gfx90a gfx940 gfx941 gfx942)
+set(target 0)
+foreach(gpu IN LISTS GPU_TARGETS)
+ if(gpu IN_LIST gpu_list AND target EQUAL 0)
+      add_custom_target(example_convnd_activ_xdl)
+      # ScaleAdd ScaleAdd Relu
+      add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_fp16)
+      add_example_executable(example_convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16 convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16.cpp)
+      add_example_dependencies(example_convnd_activ_xdl example_convnd_fwd_xdl_scaleadd_scaleadd_relu_bcasted_bias_fp16)
+   set(target 1)
+ endif()
+endforeach()