Merge branch 'develop' into fp8_gemm_generic_instance

CMakeLists.txt

@@ -449,7 +449,7 @@ rocm_install(FILES
 )
 
 # Install CK version and configuration files
-install(FILES
+rocm_install(FILES
     ${PROJECT_BINARY_DIR}/include/ck/version.h
     ${PROJECT_BINARY_DIR}/include/ck/config.h
     DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/ck/

client_example/20_image_to_column/CMakeLists.txt

+add_executable(client_image_to_column image_to_column.cpp)
+target_link_libraries(client_image_to_column PRIVATE composable_kernel::device_operations)

client_example/20_image_to_column/image_to_column.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include <cstdlib>
+#include <iomanip>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <vector>
+
+#include "ck/ck.hpp"
+#include "ck/library/tensor_operation_instance/gpu/image_to_column.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+using InDataType  = ck::half_t;
+using OutDataType = ck::half_t;
+
+using InLayout = ck::tensor_layout::convolution::GNHWC;
+
+static constexpr ck::index_t NumDimSpatial = 2;
+static constexpr ck::index_t G             = 1;
+static constexpr ck::index_t N             = 32; // batch size
+static constexpr ck::index_t C             = 32; // input channel (per group)
+static constexpr ck::index_t Y             = 3;  // filter H
+static constexpr ck::index_t X             = 3;  // filter W
+static constexpr ck::index_t Hi            = 28; // input H
+static constexpr ck::index_t Wi            = 28; // input W
+static constexpr ck::index_t Ho            = 28; // output H
+static constexpr ck::index_t Wo            = 28; // output W
+
+struct SimpleDeviceMem
+{
+    SimpleDeviceMem() = delete;
+
+    SimpleDeviceMem(std::size_t mem_size) : p_mem_{}
+    {
+        (void)hipMalloc(static_cast<void**>(&p_mem_), mem_size);
+    }
+
+    void* GetDeviceBuffer() { return p_mem_; }
+
+    ~SimpleDeviceMem() { (void)hipFree(p_mem_); }
+
+    void* p_mem_;
+};
+
+int main()
+{
+
+    std::array<ck::index_t, 2> in_spatial_lengths{Hi, Wi};
+    std::array<ck::index_t, 2> wei_spatial_lengths{Y, X};
+    std::array<ck::index_t, 2> out_spatial_lengths{Ho, Wo};
+
+    // We have NHWGC in memory space (G is dummy)
+    // However, CK's API only accept length and stride with order of GNCHW
+    // Hence, we need to adjust the order of stride
+    std::array<ck::index_t, 5> in_strides{C, Hi * Wi * G * C, 1, Wi * G * C, G * C};
+    std::array<ck::index_t, 2> out_strides{Y * X * C, 1};
+
+    std::array<ck::index_t, NumDimSpatial> filter_strides{1, 1};
+    std::array<ck::index_t, NumDimSpatial> filter_dilations{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_left_pads{1, 1};
+    std::array<ck::index_t, NumDimSpatial> input_right_pads{1, 1};
+
+    SimpleDeviceMem in(sizeof(InDataType) * N * Hi * Wi * G * C);
+    SimpleDeviceMem out(sizeof(OutDataType) * N * Ho * Wo * Y * X * C);
+
+    using DeviceOp = ck::tensor_operation::device::
+        DeviceImageToColumn<NumDimSpatial, InLayout, InDataType, OutDataType>;
+
+    // get device op instances
+    const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
+        DeviceOp>::GetInstances();
+
+    std::cout << "found " << op_ptrs.size() << " instances" << std::endl;
+
+    std::string best_op_name;
+    int best_op_id        = -1;
+    float best_avg_time   = std::numeric_limits<float>::max();
+    float best_gb_per_sec = 0;
+
+    // profile device operation instances
+    std::cout << "Run all instances and do timing" << std::endl;
+
+    for(int i = 0; i < op_ptrs.size(); ++i)
+    {
+        auto& op_ptr        = op_ptrs[i];
+        auto argument_ptr   = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        in_strides,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+        auto invoker_ptr    = op_ptr->MakeInvokerPointer();
+        std::string op_name = op_ptr->GetTypeString();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, true});
+
+            std::size_t num_bytes = sizeof(InDataType) * N * Hi * Wi * G * C +
+                                    sizeof(OutDataType) * N * Ho * Wo * Y * X * C;
+
+            float gb_per_sec = num_bytes / 1.E6 / avg_time;
+
+            std::cout << "Perf: " << std::setw(10) << avg_time << " ms, " << gb_per_sec << " GB/s, "
+                      << op_name << std::endl;
+
+            if(avg_time < best_avg_time)
+            {
+                best_op_id      = i;
+                best_op_name    = op_name;
+                best_avg_time   = avg_time;
+                best_gb_per_sec = gb_per_sec;
+            }
+        }
+        else
+        {
+            std::cerr << op_name << " does not support this problem" << std::endl;
+        }
+    }
+
+    if(best_op_id < 0)
+    {
+        std::cerr << "no suitable instance" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    std::cout << "Best Perf: " << std::setw(10) << best_avg_time << " ms, " << best_gb_per_sec
+              << " GB/s, " << best_op_name << std::endl;
+
+    // run the best intance
+    {
+        auto& op_ptr = op_ptrs[best_op_id];
+        std::cout << "Run the best instance without timing: " << op_ptr->GetTypeString()
+                  << std::endl;
+        auto argument_ptr = op_ptr->MakeArgumentPointer(in.GetDeviceBuffer(),
+                                                        out.GetDeviceBuffer(),
+                                                        N,
+                                                        C,
+                                                        in_spatial_lengths,
+                                                        out_spatial_lengths,
+                                                        wei_spatial_lengths,
+                                                        in_strides,
+                                                        out_strides,
+                                                        filter_strides,
+                                                        filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+
+        auto invoker_ptr = op_ptr->MakeInvokerPointer();
+
+        if(op_ptr->IsSupportedArgument(argument_ptr.get()))
+        {
+            invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, false});
+        }
+
+        std::cout << "Done" << std::endl;
+    }
+}

example/52_image_to_column/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_image_to_column)
+   add_example_executable(example_image_to_column_f32 image_to_column_f32.cpp)
+   add_dependencies(example_image_to_column example_image_to_column_f32)
+   set(target 1)
+ endif()
+endforeach()

example/52_image_to_column/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/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"
+
+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     = true;
+};
+
+#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_image_to_column/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 InLayout = ck::tensor_layout::convolution::GNHWC;
+
+// clang-format off
+using DeviceImgToColInstance = ck::tensor_operation::device::DeviceImageToColumnImpl
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+                              < NDimSpatial, InLayout, 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 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<InLayout>(conv_params);
+    const auto out_desc = HostTensorDescriptor({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> input_g_n_c_wis_strides{};
+    std::array<ck::index_t, 2> output_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(), input_g_n_c_wis_strides);
+    copy(out_desc.GetStrides(), output_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(),
+                                         N,
+                                         C,
+                                         input_spatial_lengths,
+                                         filter_spatial_lengths,
+                                         output_spatial_lengths,
+                                         input_g_n_c_wis_strides,
+                                         output_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 = 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, InLayout, 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); }

include/ck/tensor_operation/gpu/device/device_image_to_column.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <array>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+/**
+ * \brief Image to column.
+ *
+ * This Device operator converts image ([G, N, Di, Hi, Wi, C]) to the gemm
+ * problem([N * Do * Ho * Wo, Z *  Y * X * C]). G must be equal to 1.
+ *
+ * \tparam NDimSpatial Number of spatial dimensions.
+ * \tparam InputLayout Input Layout.
+ * \tparam InputDataType Input Data Type.
+ * \tparam OutputDataType Output Data Type.
+ */
+template <index_t NDimSpatial,
+          typename InputLayout,
+          typename InputDataType,
+          typename OutputDataType>
+struct DeviceImageToColumn : public BaseOperator
+{
+
+    /**
+     * \brief Make argument pointer for image to column.
+     *
+     * \param p_in A pointer to the device memory of the input image.
+     * \param p_out A pointer to the device memory of the output.
+     * \param N Convolution batch size.
+     * \param C Convolution number of channels.
+     * \param input_spatial_lengths Input spatial lengths.
+     * \param filter_spatial_lengths Filter spatial lengths.
+     * \param output_spatial_lengths Output spatial lengths.
+     * \param input_g_n_c_wis_strides Input strides in order [G, N, C, D, H, W].
+     * \param output_m_k_strides Output strides.
+     * \param conv_filter_strides Convolution filter strides.
+     * \param conv_filter_dilations Convolution filter dilations.
+     * \param input_left_pads Convolution left pads.
+     * \param input_right_pads Convolution right pads.
+     * \return Pointer to the argument.
+     */
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        void* p_out,
+                        const ck::index_t N,
+                        const ck::index_t C,
+                        const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                        const std::array<index_t, 2>& output_m_k_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp

@@ -280,6 +280,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                                       BK1,
                                       MPerBlock,
                                       NPerBlock,
+                                      KPerBlock,
                                       DoPadGemmM,
                                       DoPadGemmN>{};
 

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/device_image_to_column.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InputGridDesc,
+          typename InputDataType,
+          typename OutputGridDesc,
+          typename OutputDataType,
+          typename Block2ETileMap,
+          typename GridwiseImageToColumnKernel>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_image_to_column(const InputGridDesc in_grid_desc,
+                               const InputDataType* __restrict__ p_in_global,
+                               const OutputGridDesc out_grid_desc,
+                               OutputDataType* __restrict__ p_out_global,
+                               const Block2ETileMap block_2_tile_map)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||             \
+    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
+    defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx941__) || defined(__gfx942__))
+    GridwiseImageToColumnKernel::Run(
+        in_grid_desc, p_in_global, out_grid_desc, p_out_global, block_2_tile_map);
+#else
+    ignore = in_grid_desc;
+    ignore = p_in_global;
+    ignore = out_grid_desc;
+    ignore = p_out_global;
+    ignore = block_2_tile_map;
+#endif
+}
+
+// Image to column for input layout NDHWC:
+//   input : input image [N, Di, Hi, Wi, C],
+//   output : output image [N * Do * Ho * Wo, Z * Y * X * C]
+template <index_t NDimSpatial,
+          typename InputLayout,
+          typename InputDataType,
+          typename OutputDataType,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t KPerBlock,
+          typename ThreadClusterLengths,
+          index_t ScalarPerVector>
+struct DeviceImageToColumnImpl
+    : public DeviceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>
+{
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvolutionForwardSpecialization::Default>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpecialization::MKPadding, index_t, index_t, index_t>{
+            MPerBlock, 0 /* NPerBlock*/, KPerBlock};
+
+    // Use MakeADescriptor_M_K from grouped convolution forward
+    static auto
+    MakeInputDescriptor_M_K(const ck::index_t N,
+                            const ck::index_t C,
+                            const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                            const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                            const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                            const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                            const std::array<index_t, NDimSpatial>& input_left_pads,
+                            const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths{1};
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths{1};
+        std::array<index_t, NDimSpatial + 3> c_g_n_k_wos_lengths{1};
+
+        auto copy = [](const auto& x, auto& y, index_t dst_offset) {
+            std::copy(x.begin(), x.end(), y.begin() + dst_offset);
+        };
+
+        constexpr index_t spatial_offset = 3;
+
+        copy(input_spatial_lengths, a_g_n_c_wis_lengths, spatial_offset);
+        copy(filter_spatial_lengths, b_g_k_c_xs_lengths, spatial_offset);
+        copy(output_spatial_lengths, c_g_n_k_wos_lengths, spatial_offset);
+
+        // fill only significant values (C and N)
+        a_g_n_c_wis_lengths[I1] = N;
+        a_g_n_c_wis_lengths[I2] = C;
+        b_g_k_c_xs_lengths[I2]  = C;
+        c_g_n_k_wos_lengths[I1] = N;
+
+        const auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<InputLayout>(
+                a_g_n_c_wis_lengths,
+                input_g_n_c_wis_strides,
+                b_g_k_c_xs_lengths,
+                {}, // not needed for A Descriptor
+                c_g_n_k_wos_lengths,
+                {}, // not needed for A Descriptor
+                conv_filter_strides,
+                conv_filter_dilations,
+                input_left_pads,
+                input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+        return in_gemmm_gemmk_desc;
+    }
+
+    static auto
+    MakeOutDescriptor_M_K(const ck::index_t N,
+                          const ck::index_t C,
+                          const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                          const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                          const std::array<index_t, 2>& output_m_k_strides)
+    {
+        const index_t NDoHoWo =
+            N * ck::accumulate_n<index_t>(
+                    output_spatial_lengths.begin(), NDimSpatial, 1, std::multiplies<>());
+        const index_t CZYX =
+            C * ck::accumulate_n<index_t>(
+                    filter_spatial_lengths.begin(), NDimSpatial, 1, std::multiplies<>());
+        const auto desc_mraw_kraw = make_naive_tensor_descriptor(
+            make_tuple(NDoHoWo, CZYX), make_tuple(output_m_k_strides[I0], output_m_k_strides[I1]));
+
+        const auto desc_m_k = matrix_padder.PadADescriptor_M_K(desc_mraw_kraw);
+        return desc_m_k;
+    }
+
+    using InputGridDesc =
+        remove_cvref_t<decltype(MakeInputDescriptor_M_K(1, 1, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using OutputGridDesc = remove_cvref_t<decltype(MakeOutDescriptor_M_K(1, 1, {}, {}, {}))>;
+
+    using Block2ETileMap = remove_cvref_t<decltype(
+        BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, OutputGridDesc>(OutputGridDesc{}))>;
+
+    using GridwiseImageToColumnKernel = GridwiseImageToColumn<InputGridDesc,
+                                                              InputDataType,
+                                                              OutputGridDesc,
+                                                              OutputDataType,
+                                                              BlockSize,
+                                                              MPerBlock,
+                                                              KPerBlock,
+                                                              ThreadClusterLengths,
+                                                              ScalarPerVector,
+                                                              Block2ETileMap>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const void* p_in, // input image
+                 void* p_out,      // output image
+                 const ck::index_t N,
+                 const ck::index_t C,
+                 const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                 const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                 const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                 const std::array<index_t, 2>& output_m_k_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<index_t, NDimSpatial>& input_left_pads,
+                 const std::array<index_t, NDimSpatial>& input_right_pads)
+            : C_(C),
+              X_(filter_spatial_lengths[NDimSpatial - I1]),
+              p_in_{static_cast<const InputDataType*>(p_in)},
+              p_out_{static_cast<OutputDataType*>(p_out)},
+              input_g_n_c_wis_strides_{input_g_n_c_wis_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+
+            in_grid_desc_m_k_ = MakeInputDescriptor_M_K(N,
+                                                        C,
+                                                        input_spatial_lengths,
+                                                        filter_spatial_lengths,
+                                                        output_spatial_lengths,
+                                                        input_g_n_c_wis_strides,
+
+                                                        conv_filter_strides,
+                                                        conv_filter_dilations,
+                                                        input_left_pads,
+                                                        input_right_pads);
+
+            out_grid_desc_m_k_ = MakeOutDescriptor_M_K(
+                N, C, filter_spatial_lengths, output_spatial_lengths, output_m_k_strides);
+        }
+
+        void Print() const
+        {
+            std::cout << in_grid_desc_m_k_ << std::endl;
+            std::cout << out_grid_desc_m_k_ << std::endl;
+        }
+
+        const ck::index_t C_;
+        const ck::index_t X_;
+
+        const InputDataType* p_in_;
+        OutputDataType* p_out_;
+
+        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides_;
+        const std::array<index_t, NDimSpatial>& conv_filter_strides_;
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations_;
+        const std::array<index_t, NDimSpatial>& input_left_pads_;
+        const std::array<index_t, NDimSpatial>& input_right_pads_;
+
+        InputGridDesc in_grid_desc_m_k_;
+        OutputGridDesc out_grid_desc_m_k_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            const auto block_2_tile_map =
+                BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, OutputGridDesc>(
+                    arg.out_grid_desc_m_k_);
+            const index_t grid_size = block_2_tile_map.CalculateGridSize(arg.out_grid_desc_m_k_);
+            const auto kernel       = kernel_image_to_column<InputGridDesc,
+                                                       InputDataType,
+                                                       OutputGridDesc,
+                                                       OutputDataType,
+                                                       Block2ETileMap,
+                                                       GridwiseImageToColumnKernel>;
+
+            float elapsed_time = launch_and_time_kernel(stream_config,
+                                                        kernel,
+                                                        dim3(grid_size),
+                                                        dim3(BlockSize),
+                                                        0,
+                                                        arg.in_grid_desc_m_k_,
+                                                        arg.p_in_,
+                                                        arg.out_grid_desc_m_k_,
+                                                        arg.p_out_,
+                                                        block_2_tile_map);
+            return elapsed_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    bool IsSupportedArgument(const Argument& arg)
+    {
+        using namespace tensor_layout::convolution;
+        if(!(std::is_same_v<InputLayout, GNWC> || std::is_same_v<InputLayout, GNHWC> ||
+             std::is_same_v<InputLayout, GNDHWC>))
+        {
+            return false;
+        }
+        if(!(NDimSpatial >= 1 && NDimSpatial <= 3))
+        {
+            return false;
+        }
+
+        const auto w_pad_left  = arg.input_left_pads_[NDimSpatial - I1];
+        const auto w_pad_right = arg.input_right_pads_[NDimSpatial - I1];
+        const auto dilation_x  = arg.conv_filter_dilations_[NDimSpatial - I1];
+        const auto stride_x    = arg.conv_filter_strides_[NDimSpatial - I1];
+        bool is_w_packed       = arg.input_g_n_c_wis_strides_[NDimSpatial + I2] == arg.C_;
+        bool is_c_packed       = arg.input_g_n_c_wis_strides_[I2] == 1;
+
+        // check vector acces with c not packed
+        if(!is_c_packed && ScalarPerVector != 1)
+            return false;
+        // check vector access of filter window row (only C if C is not packed)
+        if(!is_w_packed && arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of filter window row (X * C)
+        if(arg.X_ * arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of pads (w_pad_left/w_pad_right * C)
+        if(w_pad_left * arg.C_ % ScalarPerVector != 0 ||
+           w_pad_right * arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of with stride and pad
+        if((w_pad_left != 0 || w_pad_right != 0) && stride_x > 1 && arg.C_ % ScalarPerVector != 0)
+            return false;
+        // check vector access of with dilation
+        if(dilation_x > 1 && arg.C_ % ScalarPerVector != 0)
+            return false;
+
+        return GridwiseImageToColumnKernel::CheckValidity(arg.in_grid_desc_m_k_,
+                                                          arg.out_grid_desc_m_k_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const void* p_in, // input image
+                             void* p_out,      // output image
+                             const ck::index_t N,
+                             const ck::index_t C,
+                             const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                             const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                             const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                             const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                             const std::array<index_t, 2>& output_m_k_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                             const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                             const std::array<index_t, NDimSpatial>& input_left_pads,
+                             const std::array<index_t, NDimSpatial>& input_right_pads)
+    {
+        return Argument{static_cast<const InputDataType*>(p_in),
+                        static_cast<OutputDataType*>(p_out),
+                        N,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        input_g_n_c_wis_strides,
+                        output_m_k_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in, // input image
+                        void* p_out,      // output image
+                        const ck::index_t N,
+                        const ck::index_t C,
+                        const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+                        const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+                        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides,
+                        const std::array<index_t, 2>& output_m_k_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+                        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+                        const std::array<index_t, NDimSpatial>& input_left_pads,
+                        const std::array<index_t, NDimSpatial>& input_right_pads) override
+    {
+        return std::make_unique<Argument>(static_cast<const InputDataType*>(p_in),
+                                          static_cast<OutputDataType*>(p_out),
+                                          N,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          input_g_n_c_wis_strides,
+                                          output_m_k_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceImageToColumn"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << KPerBlock << ", "
+            << ScalarPerVector
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp

@@ -268,6 +268,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
         static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
                           (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
                       "Invalid tuning param!");
+        static_assert(KPerBlock % AK1Value == 0 && KPerBlock % BK1Value == 0,
+                      "KPerBlock must be divisible by AK1Value and BK1Value!");
 
         const auto M  = a_grid_desc_m_k.GetLength(I0);
         const auto N  = b_grid_desc_n_k.GetLength(I0);

include/ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename InputGridDesc,
+          typename InputDataType,
+          typename OutputGridDesc,
+          typename OutputDataType,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t KPerBlock,
+          typename ThreadClusterLengths,
+          index_t ScalarPerVector,
+          typename Block2ETileMap>
+struct GridwiseImageToColumn
+{
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    __device__ static void Run(const InputGridDesc& in_grid_desc,
+                               const InputDataType* __restrict__ p_in_global,
+                               const OutputGridDesc& out_grid_desc,
+                               OutputDataType* __restrict__ p_out_global,
+                               const Block2ETileMap& block_2_tile_map)
+    {
+        const auto block_work_idx =
+            block_2_tile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+
+        const index_t k_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * KPerBlock);
+
+        // Global Memory
+        const auto in_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_in_global, in_grid_desc.GetElementSpaceSize());
+        auto out_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_out_global, out_grid_desc.GetElementSpaceSize());
+
+        auto copy_global_to_global = ThreadGroupTensorSliceTransfer_v7<
+            ThisThreadBlock,
+            Tuple<InputDataType>,
+            Tuple<OutputDataType>,
+            decltype(tie(in_grid_desc)),
+            decltype(tie(out_grid_desc)),
+            tensor_operation::element_wise::PassThrough,
+            Sequence<static_cast<index_t>(InMemoryDataOperationEnum::Set)>,
+            Sequence<MPerBlock, KPerBlock>,
+            ThreadClusterLengths,
+            Sequence<0, 1>,
+            Sequence<0, 1>,
+            I1,
+            ScalarPerVector,
+            Sequence<true>,
+            Sequence<true>>{
+            in_grid_desc,
+            make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
+            out_grid_desc,
+            make_tuple(make_multi_index(m_block_data_idx_on_grid, k_block_data_idx_on_grid)),
+            tensor_operation::element_wise::PassThrough{}};
+
+        copy_global_to_global.Run(
+            tie(in_grid_desc), tie(in_global_buf), tie(out_grid_desc), tie(out_global_buf));
+    }
+
+    __host__ static constexpr bool CheckValidity(const InputGridDesc& in_grid_desc,
+                                                 const OutputGridDesc& out_grid_desc)
+    {
+        if(in_grid_desc.GetLength(I0) % MPerBlock != 0 ||
+           in_grid_desc.GetLength(I1) % KPerBlock != 0)
+            return false;
+        if(out_grid_desc.GetLength(I0) % MPerBlock != 0 ||
+           out_grid_desc.GetLength(I1) % KPerBlock != 0)
+            return false;
+        return true;
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp

@@ -164,6 +164,7 @@ template <
     index_t BK1,
     index_t GemmMPerBlock,
     index_t GemmNPerBlock,
+    index_t GemmKPerBlock,
     bool DoPadGemmM,
     bool DoPadGemmN>
 struct TransformConvBwdDataToGemm_v1
@@ -308,9 +309,6 @@ struct TransformConvBwdDataToGemm_v1
             const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
             const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
-            const index_t AK0 =
-                math::integer_divide_ceil(ZDotSlice * YDotSlice * XDotSlice * K, AK1);
-
             if constexpr(NDimSpatial == 2)
             {
                 // A: output tensor
@@ -367,9 +365,11 @@ struct TransformConvBwdDataToGemm_v1
                 const auto out_gemmk_gemmm_padded_grid_desc =
                     ck::tensor_operation::device::PadTensorDescriptor(
                         out_gemmk_gemmmraw_grid_desc,
-                        make_tuple(AK1, GemmMPerBlock),
+                        make_tuple(GemmKPerBlock, GemmMPerBlock),
                         Sequence<true, DoPadGemmM>{});
 
+                const index_t AK0 = out_gemmk_gemmm_padded_grid_desc.GetLength(I0) / AK1;
+
                 const auto out_gemmak0_gemmm_gemmak1_grid_desc = transform_tensor_descriptor(
                     out_gemmk_gemmm_padded_grid_desc,
                     make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
@@ -460,9 +460,11 @@ struct TransformConvBwdDataToGemm_v1
                 const auto out_gemmk_gemmm_padded_grid_desc =
                     ck::tensor_operation::device::PadTensorDescriptor(
                         out_gemmk_gemmmraw_grid_desc,
-                        make_tuple(AK1, GemmMPerBlock),
+                        make_tuple(GemmKPerBlock, GemmMPerBlock),
                         Sequence<true, DoPadGemmM>{});
 
+                const index_t AK0 = out_gemmk_gemmm_padded_grid_desc.GetLength(I0) / AK1;
+
                 const auto out_gemmak0_gemmm_gemmak1_grid_desc = transform_tensor_descriptor(
                     out_gemmk_gemmm_padded_grid_desc,
                     make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
@@ -568,9 +570,6 @@ struct TransformConvBwdDataToGemm_v1
             const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
             const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
-            const index_t BK0 =
-                math::integer_divide_ceil(ZDotSlice * YDotSlice * XDotSlice * K, BK1);
-
             // B weight tensor
             if constexpr(NDimSpatial == 2)
             {
@@ -617,9 +616,11 @@ struct TransformConvBwdDataToGemm_v1
                 const auto wei_gemmk_gemmn_padded_grid_desc =
                     ck::tensor_operation::device::PadTensorDescriptor(
                         wei_gemmk_gemmnraw_grid_desc,
-                        make_tuple(BK1, GemmNPerBlock),
+                        make_tuple(GemmKPerBlock, GemmNPerBlock),
                         Sequence<true, DoPadGemmN>{});
 
+                const index_t BK0 = wei_gemmk_gemmn_padded_grid_desc.GetLength(I0) / BK1;
+
                 const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc = transform_tensor_descriptor(
                     wei_gemmk_gemmn_padded_grid_desc,
                     make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
@@ -690,17 +691,19 @@ struct TransformConvBwdDataToGemm_v1
                     make_tuple(Sequence<1, 2, 3, 0>{}, Sequence<4>{}),
                     make_tuple(Sequence<0>{}, Sequence<1>{}));
 
-                const auto wei_gemmk_gemm_padded_grid_desc =
+                const auto wei_gemmk_gemmn_padded_grid_desc =
                     ck::tensor_operation::device::PadTensorDescriptor(
                         wei_gemmk_gemmnraw_grid_desc,
-                        make_tuple(BK1, GemmNPerBlock),
+                        make_tuple(GemmKPerBlock, GemmNPerBlock),
                         Sequence<true, DoPadGemmN>{});
 
+                const index_t BK0 = wei_gemmk_gemmn_padded_grid_desc.GetLength(I0) / BK1;
+
                 const auto wei_gemmbk0_gemm_gemmbk1_grid_desc = transform_tensor_descriptor(
-                    wei_gemmk_gemm_padded_grid_desc,
-                    make_tuple(
-                        make_unmerge_transform(make_tuple(BK0, BK1)),
-                        make_pass_through_transform(wei_gemmk_gemm_padded_grid_desc.GetLength(I1))),
+                    wei_gemmk_gemmn_padded_grid_desc,
+                    make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                               make_pass_through_transform(
+                                   wei_gemmk_gemmn_padded_grid_desc.GetLength(I1))),
                     make_tuple(Sequence<0>{}, Sequence<1>{}),
                     make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 

library/include/ck/library/reference_tensor_operation/cpu/reference_image_to_column.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <type_traits>
+#include <sstream>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/library/utility/host_tensor.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace host {
+
+/**
+ * \brief Reference implementation for image to column.
+ *
+ * Tensor descriptor has [G, N, C, Di, Hi, Wi] data layout.
+ * G must be equal to 1. Memory layout is [G, N, Di, Hi, Wi, C].
+ *
+ * \tparam NDimSpatial Number of spatial dimensions.
+ * \tparam InputLayout Input Layout.
+ * \tparam InDataType Input Data Type.
+ * \tparam OutDataType Output Data Type.
+ */
+template <ck::index_t NDimSpatial,
+          typename InputLayout,
+          typename InDataType,
+          typename OutDataType,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
+struct ReferenceImageToColumn : public device::BaseOperator
+{
+    // Argument
+    struct Argument : public device::BaseArgument
+    {
+        public:
+        Argument(const Tensor<InDataType>& input,
+                 Tensor<OutDataType>& output,
+                 std::vector<ck::index_t> filter_spatial_lengths,
+                 std::vector<ck::index_t> conv_filter_strides,
+                 std::vector<ck::index_t> conv_filter_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads)
+            : input_{input},
+              output_{output},
+              conv_strides_{conv_filter_strides},
+              conv_dilations_{conv_filter_dilations},
+              in_left_pads_{input_left_pads},
+              in_right_pads_{input_right_pads},
+              filter_spatial_lengths_{filter_spatial_lengths}
+        {
+            initOutputSpatialLengths();
+        }
+
+        const Tensor<InDataType>& input_;
+        Tensor<OutDataType>& output_;
+
+        std::vector<index_t> conv_strides_;
+        std::vector<index_t> conv_dilations_;
+        std::vector<index_t> in_left_pads_;
+        std::vector<index_t> in_right_pads_;
+
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
+
+        private:
+        void initOutputSpatialLengths()
+        {
+            constexpr auto input_offset_to_spatial = 3;
+
+            for(ck::index_t i = 0; i < NDimSpatial; ++i)
+            {
+                // XEff = (X - 1) * conv_dilation_w + 1;
+                // Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
+                const ck::index_t x_eff = (filter_spatial_lengths_[i] - 1) * conv_dilations_[i] + 1;
+
+                output_spatial_lengths_.push_back(
+                    (input_.GetLengths()[i + input_offset_to_spatial] + in_left_pads_[i] +
+                     in_right_pads_[i] - x_eff) /
+                        conv_strides_[i] +
+                    1);
+            }
+        }
+    };
+
+    struct Invoker : public device::BaseInvoker
+    {
+        using Argument = ReferenceImageToColumn::Argument;
+
+        float Run(const Argument& arg)
+        {
+            if(!(arg.input_.GetNumOfDimension() == NDimSpatial + 3 &&
+                 arg.output_.GetNumOfDimension() == 2))
+            {
+                throw std::runtime_error("wrong! inconsistent dimension");
+            }
+
+            const index_t N = arg.input_.GetLengths()[1];
+            const index_t C = arg.input_.GetLengths()[2];
+
+            if constexpr(NDimSpatial == 1)
+            {
+                const index_t Wo = arg.output_spatial_lengths_[0];
+                auto func        = [&](auto n, auto wo) {
+                    index_t row    = n * Wo + wo;
+                    index_t column = 0;
+
+                    for(index_t x = 0; x < arg.filter_spatial_lengths_[0]; ++x)
+                    {
+                        auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[0]) +
+                                  static_cast<ck::long_index_t>(x * arg.conv_dilations_[0]) -
+                                  static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                        for(index_t c = 0; c < C; ++c)
+                        {
+                            if(wi >= 0 &&
+                               ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[3])
+                            {
+                                InDataType v_in          = arg.input_(0, n, c, wi);
+                                arg.output_(row, column) = ck::type_convert<OutDataType>(v_in);
+                            }
+                            column++;
+                        }
+                    }
+                };
+
+                make_ParallelTensorFunctor(func, N, Wo)(std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 2)
+            {
+                const index_t Ho = arg.output_spatial_lengths_[0];
+                const index_t Wo = arg.output_spatial_lengths_[1];
+
+                auto func = [&](auto n, auto ho, auto wo) {
+                    index_t row    = n * Ho * Wo + ho * Wo + wo;
+                    index_t column = 0;
+
+                    for(index_t y = 0; y < arg.filter_spatial_lengths_[0]; ++y)
+                    {
+                        auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[0]) +
+                                  static_cast<ck::long_index_t>(y * arg.conv_dilations_[0]) -
+                                  static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+
+                        for(index_t x = 0; x < arg.filter_spatial_lengths_[1]; ++x)
+                        {
+                            auto wi = static_cast<ck::long_index_t>(wo * arg.conv_strides_[1]) +
+                                      static_cast<ck::long_index_t>(x * arg.conv_dilations_[1]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+
+                            for(index_t c = 0; c < C; ++c)
+                            {
+
+                                if(hi >= 0 &&
+                                   ck::type_convert<std::size_t>(hi) < arg.input_.GetLengths()[3] &&
+                                   wi >= 0 &&
+                                   ck::type_convert<std::size_t>(wi) < arg.input_.GetLengths()[4])
+                                {
+                                    InDataType v_in          = arg.input_(0, n, c, hi, wi);
+                                    arg.output_(row, column) = ck::type_convert<OutDataType>(v_in);
+                                }
+                                column++;
+                            }
+                        }
+                    }
+                };
+
+                make_ParallelTensorFunctor(func, N, Ho, Wo)(std::thread::hardware_concurrency());
+
+                return 0;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                const index_t Do = arg.output_spatial_lengths_[0];
+                const index_t Ho = arg.output_spatial_lengths_[1];
+                const index_t Wo = arg.output_spatial_lengths_[2];
+
+                auto func = [&](auto n, auto d_o, auto ho, auto wo) {
+                    index_t row    = n * Do * Ho * Wo + d_o * Ho * Wo + ho * Wo + wo;
+                    index_t column = 0;
+
+                    for(index_t z = 0; z < arg.filter_spatial_lengths_[0]; ++z)
+                    {
+                        auto di = static_cast<ck::long_index_t>(d_o * arg.conv_strides_[0]) +
+                                  static_cast<ck::long_index_t>(z * arg.conv_dilations_[0]) -
+                                  static_cast<ck::long_index_t>(arg.in_left_pads_[0]);
+                        for(index_t y = 0; y < arg.filter_spatial_lengths_[1]; ++y)
+                        {
+                            auto hi = static_cast<ck::long_index_t>(ho * arg.conv_strides_[1]) +
+                                      static_cast<ck::long_index_t>(y * arg.conv_dilations_[1]) -
+                                      static_cast<ck::long_index_t>(arg.in_left_pads_[1]);
+                            for(index_t x = 0; x < arg.filter_spatial_lengths_[2]; ++x)
+                            {
+                                auto wi =
+                                    static_cast<ck::long_index_t>(wo * arg.conv_strides_[2]) +
+                                    static_cast<ck::long_index_t>(x * arg.conv_dilations_[2]) -
+                                    static_cast<ck::long_index_t>(arg.in_left_pads_[2]);
+                                for(index_t c = 0; c < C; ++c)
+                                {
+                                    if(di >= 0 &&
+                                       ck::type_convert<std::size_t>(di) <
+                                           arg.input_.GetLengths()[3] &&
+                                       hi >= 0 &&
+                                       ck::type_convert<std::size_t>(hi) <
+                                           arg.input_.GetLengths()[4] &&
+                                       wi >= 0 &&
+                                       ck::type_convert<std::size_t>(wi) <
+                                           arg.input_.GetLengths()[5])
+                                    {
+                                        InDataType v_in = arg.input_(0, n, c, di, hi, wi);
+                                        arg.output_(row, column) =
+                                            ck::type_convert<OutDataType>(v_in);
+                                    }
+                                    column++;
+                                }
+                            }
+                        }
+                    }
+                };
+
+                make_ParallelTensorFunctor(func, N, Do, Ho, Wo)(
+                    std::thread::hardware_concurrency());
+
+                return 0;
+            }
+        }
+
+        float Run(const device::BaseArgument* p_arg,
+                  const StreamConfig& /*stream_config*/ = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg));
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        using namespace tensor_layout::convolution;
+
+        if constexpr(!(std::is_same_v<InputLayout, GNWC> || std::is_same_v<InputLayout, GNHWC> ||
+                       std::is_same_v<InputLayout, GNDHWC>))
+        {
+            return false;
+        }
+        if constexpr(!(NDimSpatial >= 1 && NDimSpatial <= 3))
+        {
+            return false;
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const Argument& arg)
+    {
+        const ck::index_t G = arg.input_.GetLengths()[0];
+        const ck::index_t N = arg.input_.GetLengths()[1];
+        const ck::index_t C = arg.input_.GetLengths()[2];
+
+        const index_t NDoHoWo =
+            N * ck::accumulate_n<index_t>(
+                    arg.output_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+        const index_t CZYX =
+            C * ck::accumulate_n<index_t>(
+                    arg.filter_spatial_lengths_.begin(), NDimSpatial, 1, std::multiplies<>());
+
+        if(!(arg.output_.GetLengths()[0] == static_cast<std::size_t>(NDoHoWo) &&
+             arg.output_.GetLengths()[1] == static_cast<std::size_t>(CZYX)))
+        {
+            return false;
+        }
+
+        if(G != 1)
+        {
+            return false;
+        }
+        return true;
+    }
+
+    bool IsSupportedArgument(const device::BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(const Tensor<InDataType>& input,
+                             Tensor<OutDataType>& output,
+                             std::vector<ck::index_t> filter_spatial_lengths,
+                             std::vector<ck::index_t> conv_filter_strides,
+                             std::vector<ck::index_t> conv_filter_dilations,
+                             std::vector<ck::index_t> input_left_pads,
+                             std::vector<ck::index_t> input_right_pads)
+    {
+        return Argument{input,
+                        output,
+                        filter_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "ReferenceImageToColumn"
+            << std::endl;
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace host
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp

@@ -53,6 +53,11 @@ using device_grouped_conv2d_fwd_dl_f16_instances = std::tuple<
            // ########################################|     Spatial|   Type|    Type|        Type|    Type|    Type|         |          |    Layout|          | Elementwise|  Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
            // ########################################|            |       |        |            |        |        |         |          |          |          |   Operation|    Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
            // ########################################|            |       |        |            |        |        |         |          |          |          |            |             |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        // generic instances
+        // TODO: Change to ScalarPerVector = 1 when inner_product<half_t, half_t, float> will be supported
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F16,     F16,  DsDatatype,     F16,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough,  PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,     8,    16,     4,     2,  2,          1,          2,      1,       S<4, 2>,       S<1, 1>,      S<2, 1, 2, 2>,        S<1, 1, 8, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 1, 1, 2>,        S<2, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 2>,
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F16,     F16,  DsDatatype,     F16,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough,  PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 2>,
+
         DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F16,     F16,  DsDatatype,     F16,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough,  PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,   256,   128,   128,    16,  2,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>,      S<8, 1, 1, 2>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 2>,      S<1, 2, 0, 3>,       S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
     // clang-format on
     >;
@@ -71,6 +76,10 @@ using device_grouped_conv2d_fwd_dl_f32_instances = std::tuple<
            // ########################################|     Spatial|   Type|    Type|        Type|    Type|    Type|         |          |    Layout|          | Elementwise|  Elementwise|   Elementwise|        Forward|    Spacialization|  Size| Block| Block| Block|   | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths|  ThreadCluster|      SrcAccess|     SrcVectorTensor|    SrcVectorTensor|     DstVectorTensor|        SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
            // ########################################|            |       |        |            |        |        |         |          |          |          |   Operation|    Operation|     Operation| Specialization|                  |      |      |      |      |   |           |           |       |              |              |        K0_M0_M1_K1|          K0_M0_M1_K1|   ArrangeOrder|          Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1|        K0_N0_N1_K1|          K0_N0_N1_K1|   ArrangeOrder|          Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1|               Order|                |                   |
            // ########################################|            |       |        |            |        |        |         |          |          |          |            |             |              |               |                  |      |      |      |      |   |           |           |       |              |              |                   |                     |               |               |                    |                   |                    |                   |                     |               |               |                    |                   |                    |                    |                |                   |
+        // generic instances
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F32,     F32,  DsDatatype,     F32,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough, PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,     8,    16,      4,    2,  1,          1,          2,      1,       S<4, 2>,       S<1, 1>,      S<2, 1, 2, 1>,      S<1, 1, 8, 1>,    S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 1, 1, 1>,        S<2, 1, 4, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 1>,
+        DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F32,     F32,  DsDatatype,     F32,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough, PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 1>,
+
         DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK<           2,    F32,     F32,  DsDatatype,     F32,     F32, InLayout, WeiLayout,  DsLayout, OutLayout, PassThrough, PassThrough,  CDEElementOp,       ConvSpec,    GemmMNKPadding,   256,   128,   128,    16,  1,          4,          4,      1,       S<8, 2>,       S<8, 2>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>,      S<8, 1, 1, 1>,      S<2, 1, 128, 1>,  S<1, 2, 0, 3>,  S<1, 2, 0, 3>,       S<4, 1, 1, 1>,      S<1, 2, 0, 3>,       S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>,               5,                 4>
     // clang-format on
     >;

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp

@@ -145,6 +145,34 @@ void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
                                                               PassThrough,
                                                               PassThrough>>>& instances);
 
+void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
+void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances);
+
 // grouped conv2d forward, NHWGC/GKYXC/NHWGK
 void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
@@ -394,11 +422,13 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                          is_same_v<OutDataType, float>)
             {
                 add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
             }
             else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
                               is_same_v<OutDataType, half_t>)
             {
                 add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
+                add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
             }
             else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
                               is_same_v<WeiDataType, ck::bhalf_t> &&

library/include/ck/library/tensor_operation_instance/gpu/image_to_column.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+#include <memory>
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/device_image_to_column.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+// nhwc, 1d
+void add_device_image_to_column_nhwc_1d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, BF16, BF16>>>& instances);
+
+void add_device_image_to_column_nhwc_1d_f16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, F16, F16>>>& instances);
+
+void add_device_image_to_column_nhwc_1d_f32_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, F32, F32>>>& instances);
+
+void add_device_image_to_column_nhwc_1d_i8_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<1, GNWC, int8_t, int8_t>>>& instances);
+// nhwc, 2d
+void add_device_image_to_column_nhwc_2d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, BF16, BF16>>>& instances);
+
+void add_device_image_to_column_nhwc_2d_f16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, F16, F16>>>& instances);
+
+void add_device_image_to_column_nhwc_2d_f32_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, F32, F32>>>& instances);
+
+void add_device_image_to_column_nhwc_2d_i8_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<2, GNHWC, int8_t, int8_t>>>& instances);
+// nhwc, 3d
+void add_device_image_to_column_nhwc_3d_bf16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, BF16, BF16>>>& instances);
+
+void add_device_image_to_column_nhwc_3d_f16_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, F16, F16>>>& instances);
+
+void add_device_image_to_column_nhwc_3d_f32_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, F32, F32>>>& instances);
+
+void add_device_image_to_column_nhwc_3d_i8_instances(
+    std::vector<std::unique_ptr<DeviceImageToColumn<3, GNDHWC, int8_t, int8_t>>>& instances);
+
+template <ck::index_t NumDimSpatial, typename InLayout, typename InDataType, typename OutDataType>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::
+        DeviceImageToColumn<NumDimSpatial, InLayout, InDataType, OutDataType>>
+{
+    using DeviceOp = DeviceImageToColumn<NumDimSpatial, InLayout, InDataType, OutDataType>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+
+        if constexpr(NumDimSpatial == 1 && is_same_v<InLayout, GNWC>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
+            {
+                add_device_image_to_column_nhwc_1d_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
+            {
+                add_device_image_to_column_nhwc_1d_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_image_to_column_nhwc_1d_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
+            {
+                add_device_image_to_column_nhwc_1d_i8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
+            {
+                add_device_image_to_column_nhwc_2d_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
+            {
+                add_device_image_to_column_nhwc_2d_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_image_to_column_nhwc_2d_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
+            {
+                add_device_image_to_column_nhwc_2d_i8_instances(op_ptrs);
+            }
+        }
+        else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC>)
+        {
+            if constexpr(is_same_v<InDataType, float> && is_same_v<OutDataType, float>)
+            {
+                add_device_image_to_column_nhwc_3d_f32_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, half_t> && is_same_v<OutDataType, half_t>)
+            {
+                add_device_image_to_column_nhwc_3d_f16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, ck::bhalf_t> &&
+                              is_same_v<OutDataType, ck::bhalf_t>)
+            {
+                add_device_image_to_column_nhwc_3d_bf16_instances(op_ptrs);
+            }
+            else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<OutDataType, int8_t>)
+            {
+                add_device_image_to_column_nhwc_3d_i8_instances(op_ptrs);
+            }
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/image_to_column/device_image_to_column_instance.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using namespace ck::tensor_layout::convolution;
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+template <ck::index_t NDimSpatial, typename InLayout>
+using device_image_to_column_bf16_instances = std::tuple<
+    // clang-format off
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,    64,     8,     8,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,    64,    16,    16,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,    64,    32,    32,   S<8, 8>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,    64,    64,    64,   S<8, 8>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   128,    16,    16,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   128,    64,    64,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   128,    32,    64,  S<8, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   128,    64,   128,  S<8, 16>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,    16,    16, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,    64,    64, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,   128,   128, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,    64,    64, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,   128,   128, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,       BF16,        BF16,   256,   128,   128, S<16, 16>,     8>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial, typename InLayout>
+using device_image_to_column_f16_instances = std::tuple<
+    // clang-format off
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,    64,     8,     8,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,    64,    16,    16,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,    64,    32,    32,   S<8, 8>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,    64,    64,    64,   S<8, 8>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   128,    16,    16,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   128,    64,    64,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   128,    32,    64,  S<8, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   128,    64,   128,  S<8, 16>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,    16,    16, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,    64,    64, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,   128,   128, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,    64,    64, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,   128,   128, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F16,         F16,   256,   128,   128, S<16, 16>,     8>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial, typename InLayout>
+using device_image_to_column_f32_instances = std::tuple<
+    // clang-format off
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,    64,     8,     8,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,    64,    16,    16,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,    64,    32,    32,   S<8, 8>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   128,    16,    16,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   128,    64,    64,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   128,    32,    64,  S<8, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,    16,    16, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,    64,    64, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,   128,   128, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,    64,    64, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,        F32,         F32,   256,   128,   128, S<16, 16>,     4>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial, typename InLayout>
+using device_image_to_column_i8_instances = std::tuple<
+    // clang-format off
+        //#####################|        Num| InLayout| InDataType| OutDataType| Block|  MPer|  KPer|    Thread| Scalar|
+        //#####################|        Dim|         |           |            |  Size| Block| Block|   Cluster|    Per|
+        //#####################|    Spatial|         |           |            |      |      |      |   Lengths| Vector|
+        //#####################|           |         |           |            |      |      |      |          |       |
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,    64,     8,     8,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,    64,    16,    16,   S<8, 8>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,    64,    32,    32,   S<8, 8>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,    64,    64,    64,   S<8, 8>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   128,    16,    16,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   128,    64,    64,  S<8, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   128,    32,    64,  S<8, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   128,    64,   128,  S<8, 16>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,    16,    16, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,    64,    64, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,   128,   128, S<16, 16>,     1>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,    64,    64, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,   128,   128, S<16, 16>,     4>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,   128,   128, S<16, 16>,     8>,
+        DeviceImageToColumnImpl<NDimSpatial, InLayout,     int8_t,      int8_t,   256,   256,   256, S<16, 16>,     16>
+    // clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt

 add_instance_library(device_grouped_conv2d_fwd_instance
+   #xdl
    # GNHWC, GKYXC, GNHWK
    device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp
    device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
@@ -8,6 +9,10 @@ add_instance_library(device_grouped_conv2d_fwd_instance
    device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
    device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
    #dl
+   # GNHWC, GKYXC, GNHWK
    device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp
    device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp
+   # NHWGC, GKYXC, NHWGK
+   device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp
+   device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp
 )

library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F16,
+                                                              F16,
+                                                              Empty_Tuple,
+                                                              F16,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f16_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f16_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f16_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_dl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+void add_device_grouped_conv2d_fwd_dl_nhwgc_gkyxc_nhwgk_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
+                                                              NHWGC,
+                                                              GKYXC,
+                                                              Empty_Tuple,
+                                                              NHWGK,
+                                                              F32,
+                                                              F32,
+                                                              Empty_Tuple,
+                                                              F32,
+                                                              PassThrough,
+                                                              PassThrough,
+                                                              PassThrough>>>& instances)
+{
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f32_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwdDefault>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f32_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1P0>{});
+
+    add_device_operation_instances(instances,
+                                   device_grouped_conv2d_fwd_dl_f32_instances<NHWGC,
+                                                                              GKYXC,
+                                                                              Empty_Tuple,
+                                                                              NHWGK,
+                                                                              Empty_Tuple,
+                                                                              PassThrough,
+                                                                              ConvFwd1x1S1P0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck