Merge branch 'develop' into aosewski/gemm_tile_loop

example/42_groupnorm/CMakeLists.txt

-add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
-add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
-add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_groupnorm_sigmoid_mul_fp16 groupnorm_sigmoid_mul_fp16.cpp)
+    add_example_executable(example_groupnorm_splitk_fp16 groupnorm_splitk_fp16.cpp)
+    add_example_executable(example_groupnorm_swish_fp16 groupnorm_swish_fp16.cpp)
+endif()

example/43_splitk_gemm_bias_e_permute/CMakeLists.txt

-add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
-add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp16 splitk_gemm_bias_e_permute_xdl_fp16.cpp)
+endif()
+if(DTYPES MATCHES "fp32" OR NOT DEFINED DTYPES)
+    add_example_executable(example_splitk_gemm_bias_e_permute_xdl_fp32 splitk_gemm_bias_e_permute_xdl_fp32.cpp)
+endif()

example/44_elementwise_permute/CMakeLists.txt

-add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
-add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_elementwise_permute_4D_fp16 elementwise_permute_4D_fp16.cpp)
+    add_example_executable(example_elementwise_permute_4D_fp16_2d elementwise_permute_4D_fp16_2d.cpp)
+endif()

example/46_gemm_add_multiply/CMakeLists.txt

-add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
-add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    if(DL_KERNELS)
+        add_example_executable(example_gemm_add_multiply_dl_fp16 gemm_add_multiply_dl_fp16.cpp)
+    endif()
+    add_example_executable(example_gemm_add_multiply_xdl_fp16 gemm_add_multiply_xdl_fp16.cpp)
+endif()

example/48_pool3d_fwd/CMakeLists.txt

-add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
-
+if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES)
+    add_example_executable(example_pool3d_fwd_fp16 pool3d_fwd_fp16.cpp)
+endif()

example/49_maxpool2d_bwd/CMakeLists.txt

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

example/50_put_element/CMakeLists.txt

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

example/51_avgpool3d_bwd/CMakeLists.txt

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

example/51_avgpool3d_bwd/avgpool3d_bwd_bf16.cpp

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

example/51_avgpool3d_bwd/avgpool3d_bwd_common.hpp

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

example/51_avgpool3d_bwd/avgpool3d_bwd_fp16.cpp

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

example/51_avgpool3d_bwd/avgpool3d_bwd_fp32.cpp

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

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NDimSpatial,
+          typename DOutDataType,
+          typename DInDataType,
+          typename DOutLayout,
+          typename DInLayout>
+struct DeviceAvgPoolBwd : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_dout,
+                        void* p_din,
+                        std::vector<ck::index_t> dout_n_k_wos_lengths,
+                        std::vector<ck::index_t> dout_n_k_wos_strides,
+                        std::vector<ck::index_t> din_n_k_wos_length,
+                        std::vector<ck::index_t> din_n_k_wos_strides,
+                        std::vector<ck::index_t> window_k_c_xs_lengths,
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> window_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -27,15 +27,12 @@ struct DeviceGroupedConvBwdWeight : public BaseOperator
     MakeArgumentPointer(const void* p_in,
                         void* p_wei,
                         const void* p_out,
-                        const ck::index_t G,
-                        const ck::index_t N,
-                        const ck::index_t K,
-                        const ck::index_t C,
-                        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
-                        const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
                         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
                         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
                         const std::array<ck::index_t, NDimSpatial>& input_left_pads,

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

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 

include/ck/tensor_operation/gpu/device/impl/device_avgpool3d_bwd_ndhwc_ndhwc.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <sstream>
+
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
+#include "ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// In and Din = [N, C, Di, Hi, Wi]
+// Out and Dout = [N, C, Do, Ho, Wo]
+// Out = AvgPoolFwd(In)
+// Din = AvgPoolBwd(Dout)
+// Pooling dimension = D, H, W
+template <typename DOutDataType,
+          typename DInDataType,
+          typename ComputeDataType,
+          ck::index_t BlockSize,
+          ck::index_t MThreadClusterSize,
+          ck::index_t KThreadClusterSize,
+          ck::index_t MThreadSliceSize,
+          ck::index_t KThreadSliceSize,
+          ck::index_t InSrcOutDstVectorSize>
+struct DeviceAvgPool3dBwd_NDHWC_NDHWC : public DeviceAvgPoolBwd<3,
+                                                                DOutDataType,
+                                                                DInDataType,
+                                                                tensor_layout::convolution::NDHWC,
+                                                                tensor_layout::convolution::NDHWC>
+{
+    static constexpr ck::index_t NDimSpatial = 3;
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+
+    static constexpr ck::index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr ck::index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static auto
+    Make3DGridDescriptor_Out_M_K_In_M(const std::vector<ck::index_t>& dout_n_c_wos_lengths,
+                                      const std::vector<ck::index_t>& din_n_c_wos_length,
+                                      const std::vector<ck::index_t>& dout_n_c_wos_strides,
+                                      const std::vector<ck::index_t>& din_n_c_wos_strides,
+                                      const std::vector<ck::index_t>& window_lengths,
+                                      const std::vector<ck::index_t>& window_strides,
+                                      const std::vector<ck::index_t>& window_dilations,
+                                      const std::vector<ck::index_t>& input_left_pads,
+                                      const std::vector<ck::index_t>& input_right_pads,
+                                      const std::vector<ck::index_t>& tildes)
+    {
+        index_t i_ztilde = tildes[0];
+        index_t i_ytilde = tildes[1];
+        index_t i_xtilde = tildes[2];
+
+        const index_t N = dout_n_c_wos_lengths[0];
+        const index_t C = dout_n_c_wos_lengths[1];
+
+        const index_t Di = din_n_c_wos_length[2];
+        const index_t Hi = din_n_c_wos_length[3];
+        const index_t Wi = din_n_c_wos_length[4];
+
+        const index_t Do = dout_n_c_wos_lengths[2];
+        const index_t Ho = dout_n_c_wos_lengths[3];
+        const index_t Wo = dout_n_c_wos_lengths[4];
+
+        const index_t Z = window_lengths[0];
+        const index_t Y = window_lengths[1];
+        const index_t X = window_lengths[2];
+
+        const index_t InLeftPadD = input_left_pads[0];
+        const index_t InLeftPadH = input_left_pads[1];
+        const index_t InLeftPadW = input_left_pads[2];
+
+        const index_t InRightPadD = input_right_pads[0];
+        const index_t InRightPadH = input_right_pads[1];
+        const index_t InRightPadW = input_right_pads[2];
+
+        const index_t ConvStrideD = window_strides[0];
+        const index_t ConvStrideH = window_strides[1];
+        const index_t ConvStrideW = window_strides[2];
+
+        const index_t ConvDilationD = window_dilations[0];
+        const index_t ConvDilationH = window_dilations[1];
+        const index_t ConvDilationW = window_dilations[2];
+
+        const auto out_n_do_ho_wo_c_grid_desc =
+            make_naive_tensor_descriptor(make_tuple(N, Do, Ho, Wo, C),
+                                         make_tuple(dout_n_c_wos_strides[0],
+                                                    dout_n_c_wos_strides[2],
+                                                    dout_n_c_wos_strides[3],
+                                                    dout_n_c_wos_strides[4],
+                                                    dout_n_c_wos_strides[1]));
+
+        const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
+        const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
+        const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
+
+        const auto ZTilde = ConvStrideD / GcdStrideDilationD;
+        const auto YTilde = ConvStrideH / GcdStrideDilationH;
+        const auto XTilde = ConvStrideW / GcdStrideDilationW;
+
+        const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
+        const auto YDot = math::integer_divide_ceil(Y, YTilde);
+        const auto XDot = math::integer_divide_ceil(X, XTilde);
+
+        const auto DTilde = Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
+        const auto HTilde = Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+        const auto WTilde = Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+        // only work on Tildes that contribute to non-padding area of input tensor
+        const auto IDTildeSliceBegin = math::integer_divide_floor(
+            math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
+        const auto IHTildeSliceBegin = math::integer_divide_floor(
+            math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
+        const auto IWTildeSliceBegin = math::integer_divide_floor(
+            math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
+
+        const auto IDTildeSliceEnd =
+            math::min(DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
+        const auto IHTildeSliceEnd =
+            math::min(HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+        const auto IWTildeSliceEnd =
+            math::min(WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+        const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
+        const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
+        const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
+
+        // ReduceK is different for each Reduce
+        const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
+        const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+        const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+        // Problem size of reduction kernel
+        const index_t MRaw = N * DTildeSlice * HTildeSlice * WTildeSlice * C;
+        const index_t MPad = math::integer_least_multiple(MRaw, M_BlockTileSize) - MRaw;
+
+        const index_t KRaw = ZDotSlice * YDotSlice * XDotSlice;
+        const index_t KPad = math::integer_least_multiple(KRaw, K_BlockTileSize) - KRaw;
+
+        // Out[ReduceM, ReduceK]
+        const auto out_n_dop_hop_wop_c_grid_desc = transform_tensor_descriptor(
+            out_n_do_ho_wo_c_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Do, I0, I0),
+                       make_pad_transform(Ho, I0, I0),
+                       make_pad_transform(Wo, I0, I0),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        const auto out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_c_grid_desc =
+            transform_tensor_descriptor(
+                out_n_dop_hop_wop_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(ZDot, DTilde),
+                                         make_tuple(-ConvDilationD / GcdStrideDilationD, I1)),
+                    make_embed_transform(make_tuple(YDot, HTilde),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilde),
+                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+        const auto
+            out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_c_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(ZDot, I0, ZDotSlice),
+                               make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{},
+                               Sequence<6>{},
+                               Sequence<7>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{},
+                               Sequence<6>{},
+                               Sequence<7>{}));
+
+        const auto out_grid_desc_reducemraw_reducekraw = transform_tensor_descriptor(
+            out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_c_grid_desc,
+            make_tuple(
+                make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice, C)),
+                make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice))),
+            make_tuple(Sequence<0, 2, 4, 6, 7>{}, Sequence<1, 3, 5>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto out_grid_desc_reducem_reducek = transform_tensor_descriptor(
+            out_grid_desc_reducemraw_reducekraw,
+            make_tuple(make_right_pad_transform(MRaw, MPad), make_right_pad_transform(KRaw, KPad)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        // In[ReduceM]
+        const auto in_n_di_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor(make_tuple(N, Di, Hi, Wi, C),
+                                         make_tuple(din_n_c_wos_strides[0],
+                                                    din_n_c_wos_strides[2],
+                                                    din_n_c_wos_strides[3],
+                                                    din_n_c_wos_strides[4],
+                                                    din_n_c_wos_strides[1]));
+
+        const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+            in_n_di_hi_wi_c_grid_desc,
+            make_tuple(make_pass_through_transform(N),
+                       make_pad_transform(Di, InLeftPadD, InRightPadD),
+                       make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                       make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+        const auto in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+            transform_tensor_descriptor(
+                in_n_dip_hip_wip_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(XTilde, DTilde),
+                                                make_tuple(ConvDilationD, ConvStrideD)),
+                           make_embed_transform(make_tuple(YTilde, HTilde),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilde, WTilde),
+                                                make_tuple(ConvDilationW, ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+        const auto in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc =
+            transform_tensor_descriptor(
+                in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ztilde),
+                           make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                           make_freeze_transform(i_ytilde),
+                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                           make_freeze_transform(i_xtilde),
+                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{},
+                           Sequence<1>{},
+                           Sequence<2>{},
+                           Sequence<3>{},
+                           Sequence<4>{},
+                           Sequence<5>{},
+                           Sequence<6>{},
+                           Sequence<7>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<>{},
+                           Sequence<1>{},
+                           Sequence<>{},
+                           Sequence<2>{},
+                           Sequence<>{},
+                           Sequence<3>{},
+                           Sequence<4>{}));
+
+        const auto in_grid_desc_reducemraw = transform_tensor_descriptor(
+            in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc,
+            make_tuple(
+                make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice, C))),
+            make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto in_grid_desc_reducem =
+            transform_tensor_descriptor(in_grid_desc_reducemraw,
+                                        make_tuple(make_right_pad_transform(MRaw, MPad)),
+                                        make_tuple(Sequence<0>{}),
+                                        make_tuple(Sequence<0>{}));
+
+        return make_tuple(out_grid_desc_reducem_reducek, in_grid_desc_reducem);
+    }
+
+    using DoutDinGridDesc = decltype(Make3DGridDescriptor_Out_M_K_In_M({0, 0, 0, 0, 0},
+                                                                       {0, 0, 0, 0, 0},
+                                                                       {0, 0, 0, 0, 0},
+                                                                       {0, 0, 0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0},
+                                                                       {0, 0, 0}));
+
+    using DoutGridDesc_M_K = remove_cvref_t<tuple_element_t<0, DoutDinGridDesc>>;
+    using DinGridDesc_M    = remove_cvref_t<tuple_element_t<1, DoutDinGridDesc>>;
+
+    // FIXME
+    // for NDHWC, the dim C is the fastest dimension, and is not reduced.
+    // Hence, it is in M dimension for reduction kernel.
+    static constexpr index_t OutSrcInDstVectorDim = 0; // 0: M, 1: K
+
+    using PassThrough = tensor_operation::element_wise::PassThrough;
+    using Div         = tensor_operation::element_wise::UnaryDivide;
+
+    using gridwise_reduce = GridwiseReduction_mk_to_m_threadwise<DOutDataType,
+                                                                 DInDataType,
+                                                                 ComputeDataType,
+                                                                 int,
+                                                                 DoutGridDesc_M_K,
+                                                                 DinGridDesc_M,
+                                                                 reduce::Add,
+                                                                 PassThrough,
+                                                                 Div,
+                                                                 InMemoryDataOperationEnum::Set,
+                                                                 false, // propagate_nan
+                                                                 BlockSize,
+                                                                 MThreadSliceSize,
+                                                                 KThreadSliceSize,
+                                                                 OutSrcInDstVectorDim,
+                                                                 InSrcOutDstVectorSize,
+                                                                 InSrcOutDstVectorSize>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const DOutDataType* p_dout,
+                 DInDataType* p_din,
+                 std::vector<ck::index_t> dout_n_c_wos_lengths,
+                 std::vector<ck::index_t> din_n_c_wos_length,
+                 std::vector<ck::index_t> dout_n_c_wos_strides,
+                 std::vector<ck::index_t> din_n_c_wos_strides,
+                 std::vector<ck::index_t> window_lengths,
+                 std::vector<ck::index_t> window_strides,
+                 std::vector<ck::index_t> window_dilations,
+                 std::vector<ck::index_t> input_left_pads,
+                 std::vector<ck::index_t> input_right_pads)
+            : p_dout_grid_{p_dout},
+              p_din_grid_{p_din},
+              dout_n_c_wos_lengths_{dout_n_c_wos_lengths},
+              din_n_c_wos_length_{din_n_c_wos_length},
+              dout_n_c_wos_strides_{dout_n_c_wos_strides},
+              din_n_c_wos_strides_{din_n_c_wos_strides},
+              num_reduce_{1},
+              div_element_op_{window_lengths[0] * window_lengths[1] * window_lengths[2]}
+        {
+            std::vector<ck::index_t> Tildes(NDimSpatial);
+            for(int i = 0; i < NDimSpatial; ++i)
+            {
+                int GcdStrideDilation = math::gcd(window_strides[i], window_dilations[i]);
+                Tildes[i]             = window_strides[i] / GcdStrideDilation;
+                num_reduce_ *= Tildes[i];
+            }
+
+            for(index_t i_ztilde = 0; i_ztilde < Tildes[0]; ++i_ztilde)
+            {
+                for(index_t i_ytilde = 0; i_ytilde < Tildes[1]; ++i_ytilde)
+                {
+                    for(index_t i_xtilde = 0; i_xtilde < Tildes[2]; ++i_xtilde)
+                    {
+                        // check slice is valid
+                        const auto ZDotSlice =
+                            math::integer_divide_ceil(window_lengths[0] - i_ztilde, Tildes[0]);
+                        const auto YDotSlice =
+                            math::integer_divide_ceil(window_lengths[1] - i_ytilde, Tildes[1]);
+                        const auto XDotSlice =
+                            math::integer_divide_ceil(window_lengths[2] - i_xtilde, Tildes[2]);
+
+                        if(ZDotSlice * YDotSlice * XDotSlice <= 0)
+                        {
+                            continue;
+                        }
+
+                        const auto dout_din_grid_desc =
+                            Make3DGridDescriptor_Out_M_K_In_M(dout_n_c_wos_lengths,
+                                                              din_n_c_wos_length,
+                                                              dout_n_c_wos_strides,
+                                                              din_n_c_wos_strides,
+                                                              window_lengths,
+                                                              window_strides,
+                                                              window_dilations,
+                                                              input_left_pads,
+                                                              input_right_pads,
+                                                              {i_ztilde, i_ytilde, i_xtilde});
+
+                        dout_grid_desc_m_k_container_.push_back(dout_din_grid_desc[I0]);
+                        din_grid_desc_m_container_.push_back(dout_din_grid_desc[I1]);
+                    }
+                }
+            }
+        }
+
+        const DOutDataType* p_dout_grid_;
+        DInDataType* p_din_grid_;
+        std::vector<ck::index_t> dout_n_c_wos_lengths_;
+        std::vector<ck::index_t> din_n_c_wos_length_;
+        std::vector<ck::index_t> dout_n_c_wos_strides_;
+        std::vector<ck::index_t> din_n_c_wos_strides_;
+
+        int num_reduce_;
+        std::vector<DoutGridDesc_M_K> dout_grid_desc_m_k_container_;
+        std::vector<DinGridDesc_M> din_grid_desc_m_container_;
+
+        Div div_element_op_;
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            float ave_time = 0;
+
+            for(index_t i = 0; i < arg.num_reduce_; i++)
+            {
+                const auto kernel = kernel_reduce_threadwise<gridwise_reduce,
+                                                             false,
+                                                             false,
+                                                             false, // don't have index input
+                                                             DOutDataType,
+                                                             DInDataType,
+                                                             ComputeDataType,
+                                                             int,
+                                                             DoutGridDesc_M_K,
+                                                             DinGridDesc_M,
+                                                             PassThrough,
+                                                             Div>;
+
+                ck::index_t M           = arg.dout_grid_desc_m_k_container_[i].GetLength(I0);
+                const index_t grid_size = (M / M_BlockTileSize);
+
+                ave_time += launch_and_time_kernel(stream_config,
+                                                   kernel,
+                                                   dim3(grid_size),
+                                                   dim3(BlockSize),
+                                                   0,
+                                                   arg.dout_grid_desc_m_k_container_[i],
+                                                   arg.din_grid_desc_m_container_[i],
+                                                   PassThrough{},
+                                                   arg.div_element_op_,
+                                                   float(1),
+                                                   arg.p_dout_grid_,
+                                                   nullptr,
+                                                   float(0),
+                                                   arg.p_din_grid_,
+                                                   nullptr);
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        constexpr index_t Rank = NDimSpatial + 2;
+        int doutFastestDim     = -1;
+        int dinFastestDim      = -1;
+
+        for(int i = 0; i < Rank; ++i)
+        {
+            if(arg.dout_n_c_wos_strides_[i] == 1)
+                doutFastestDim = i;
+            if(arg.din_n_c_wos_strides_[i] == 1)
+                dinFastestDim = i;
+        }
+
+        if(doutFastestDim == -1 || dinFastestDim == -1)
+        {
+            if constexpr(InSrcOutDstVectorSize != 1)
+                return false;
+        }
+        else
+        {
+            if(arg.dout_n_c_wos_lengths_[doutFastestDim] % InSrcOutDstVectorSize != 0)
+                return false;
+            if(arg.din_n_c_wos_length_[dinFastestDim] % InSrcOutDstVectorSize != 0)
+                return false;
+        }
+
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_dout,
+                        void* p_din,
+                        std::vector<ck::index_t> dout_n_c_wos_lengths,
+                        std::vector<ck::index_t> din_n_c_wos_length,
+                        std::vector<ck::index_t> dout_n_c_wos_strides,
+                        std::vector<ck::index_t> din_n_c_wos_strides,
+                        std::vector<ck::index_t> window_lengths,
+                        std::vector<ck::index_t> window_strides,
+                        std::vector<ck::index_t> window_dilations,
+                        std::vector<ck::index_t> input_left_pads,
+                        std::vector<ck::index_t> input_right_pads) override
+    {
+        constexpr index_t Rank = NDimSpatial + 2;
+
+        if(dout_n_c_wos_strides.size() != Rank || din_n_c_wos_strides.size() != Rank ||
+           dout_n_c_wos_lengths.size() != Rank || din_n_c_wos_length.size() != Rank)
+            throw std::runtime_error("dimension is incorrect");
+
+        if(window_lengths.size() != NDimSpatial || window_strides.size() != NDimSpatial ||
+           window_dilations.size() != NDimSpatial || input_left_pads.size() != NDimSpatial ||
+           input_right_pads.size() != NDimSpatial)
+            throw std::runtime_error("dimension is incorrect");
+
+        return std::make_unique<Argument>(static_cast<const DOutDataType*>(p_dout),
+                                          static_cast<DInDataType*>(p_din),
+                                          dout_n_c_wos_lengths,
+                                          din_n_c_wos_length,
+                                          dout_n_c_wos_strides,
+                                          din_n_c_wos_strides,
+                                          window_lengths,
+                                          window_strides,
+                                          window_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 << "DeviceAvgPool3dBwd<" << BlockSize << ",";
+        str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
+        str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
+        str <<"InSrcOutDstVectorSize_" << InSrcOutDstVectorSize << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_cgemm_4gemm_xdl_cshuffle.hpp

@@ -123,7 +123,8 @@ struct DeviceCGemm_4Gemm_Xdl_CShuffle
         ALayout,
         BLayout,
         CLayout,
-        ADataType, // TODO: distinguish A/B datatype
+        ADataType,
+        BDataType,
         GemmAccDataType,
         CShuffleDataType,
         CDataType,
@@ -284,8 +285,11 @@ struct DeviceCGemm_4Gemm_Xdl_CShuffle
 
             if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
             {
-                const auto kernel =
-                    kernel_gemm_xdl_cshuffle_v1<GridwiseGemm, ADataType, CDataType, true>;
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<GridwiseGemm,
+                                                                ADataType,
+                                                                BDataType,
+                                                                CDataType,
+                                                                true>;
 
                 ave_time += launch_and_time_kernel(stream_config,
                                                    kernel,
@@ -357,8 +361,11 @@ struct DeviceCGemm_4Gemm_Xdl_CShuffle
             }
             else
             {
-                const auto kernel =
-                    kernel_gemm_xdl_cshuffle_v1<GridwiseGemm, ADataType, CDataType, false>;
+                const auto kernel = kernel_gemm_xdl_cshuffle_v1<GridwiseGemm,
+                                                                ADataType,
+                                                                BDataType,
+                                                                CDataType,
+                                                                false>;
 
                 ave_time += launch_and_time_kernel(stream_config,
                                                    kernel,

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

@@ -65,7 +65,8 @@ template <typename ALayout,
           typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
           LoopScheduler LoopSched     = make_default_loop_scheduler(),
-          PipelineVersion PipelineVer = PipelineVersion::v1>
+          PipelineVersion PipelineVer = PipelineVersion::v1,
+          typename ComputeType        = CDataType>
 struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
                                                    BLayout,
                                                    CLayout,
@@ -87,7 +88,8 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
         ALayout,
         BLayout,
         CLayout,
-        ADataType, // TODO: distinguish A/B datatype
+        ADataType,
+        BDataType,
         GemmAccDataType,
         CShuffleDataType,
         CDataType,
@@ -128,7 +130,8 @@ struct DeviceGemm_Xdl_CShuffle : public DeviceGemm<ALayout,
         CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
         CShuffleBlockTransferScalarPerVector_NPerBlock,
         LoopSched,
-        PipelineVer>;
+        PipelineVer,
+        ComputeType>;
 
     using Argument = typename GridwiseGemm::Argument;
 

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

@@ -784,15 +784,12 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
         Argument(const InDataType* p_in_grid,
                  WeiDataType* p_wei_grid,
                  const OutDataType* p_out_grid,
-                 const ck::index_t G,
-                 const ck::index_t N,
-                 const ck::index_t K,
-                 const ck::index_t C,
-                 const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-                 const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-                 const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
-                 const std::array<ck::index_t, NDimSpatial + 3>& /*input_strides*/,
-                 const std::array<ck::index_t, NDimSpatial + 3>& /*output_strides*/,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                 const std::array<index_t, NDimSpatial + 3>& /*a_g_n_c_wis_strides*/,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                 const std::array<index_t, NDimSpatial + 3>& /*b_g_k_c_xs_strides*/,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                 const std::array<index_t, NDimSpatial + 3>& /*e_g_n_k_wos_strides*/,
                  const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
                  const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
                  const std::array<ck::index_t, NDimSpatial>& input_left_pads,
@@ -812,27 +809,38 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
               a_element_op_{out_element_op},
               b_element_op_{wei_element_op},
               c_element_op_{in_element_op},
-              Conv_G_{G},
-              Conv_N_{N},
-              Conv_K_{K},
-              Conv_C_{C},
-              input_spatial_lengths_{input_spatial_lengths},
-              filter_spatial_lengths_{filter_spatial_lengths},
-              output_spatial_lengths_{output_spatial_lengths},
+              Conv_G_{a_g_n_c_wis_lengths[0]},
+              Conv_N_{a_g_n_c_wis_lengths[1]},
+              Conv_K_{b_g_k_c_xs_lengths[1]},
+              Conv_C_{a_g_n_c_wis_lengths[2]},
+              input_spatial_lengths_{},
+              filter_spatial_lengths_{},
+              output_spatial_lengths_{},
               conv_filter_strides_{conv_filter_strides},
               conv_filter_dilations_{conv_filter_dilations},
               input_left_pads_{input_left_pads},
               input_right_pads_{input_right_pads},
               k_batch_{split_k}
         {
+            constexpr index_t spatial_offset = 3;
+            std::copy(begin(a_g_n_c_wis_lengths) + spatial_offset,
+                      end(a_g_n_c_wis_lengths),
+                      begin(input_spatial_lengths_));
+            std::copy(begin(b_g_k_c_xs_lengths) + spatial_offset,
+                      end(b_g_k_c_xs_lengths),
+                      begin(filter_spatial_lengths_));
+            std::copy(begin(e_g_n_k_wos_lengths) + spatial_offset,
+                      end(e_g_n_k_wos_lengths),
+                      begin(output_spatial_lengths_));
+
             const auto descs =
                 DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
-                    N,
-                    K,
-                    C,
-                    input_spatial_lengths,
-                    filter_spatial_lengths,
-                    output_spatial_lengths,
+                    Conv_N_,
+                    Conv_K_,
+                    Conv_C_,
+                    input_spatial_lengths_,
+                    filter_spatial_lengths_,
+                    output_spatial_lengths_,
                     conv_filter_strides,
                     conv_filter_dilations,
                     input_left_pads,
@@ -856,21 +864,21 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
 
             // A/B/C Batch Stride
             compute_ptr_offset_of_batch_.BatchStrideA_ =
-                N * K *
-                std::accumulate(begin(output_spatial_lengths),
-                                end(output_spatial_lengths),
+                Conv_N_ * Conv_K_ *
+                std::accumulate(begin(output_spatial_lengths_),
+                                end(output_spatial_lengths_),
                                 index_t{1},
                                 std::multiplies<>{});
             compute_ptr_offset_of_batch_.BatchStrideB_ =
-                N * C *
-                std::accumulate(begin(input_spatial_lengths),
-                                end(input_spatial_lengths),
+                Conv_N_ * Conv_C_ *
+                std::accumulate(begin(input_spatial_lengths_),
+                                end(input_spatial_lengths_),
                                 index_t{1},
                                 std::multiplies<>{});
             compute_ptr_offset_of_batch_.BatchStrideC_ =
-                K * C *
-                std::accumulate(begin(filter_spatial_lengths),
-                                end(filter_spatial_lengths),
+                Conv_K_ * Conv_C_ *
+                std::accumulate(begin(filter_spatial_lengths_),
+                                end(filter_spatial_lengths_),
                                 index_t{1},
                                 std::multiplies<>{});
         }
@@ -904,9 +912,9 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
         const index_t Conv_K_;
         const index_t Conv_C_;
 
-        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths_;
-        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths_;
-        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths_;
+        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_;
         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides_;
         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations_;
         const std::array<ck::index_t, NDimSpatial>& input_left_pads_;
@@ -1110,39 +1118,34 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
         return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
     }
 
-    static auto MakeArgument(const InDataType* p_in_grid,
-                             WeiDataType* p_wei_grid,
-                             const OutDataType* p_out_grid,
-                             const ck::index_t G,
-                             const ck::index_t N,
-                             const ck::index_t K,
-                             const ck::index_t C,
-                             const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-                             const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-                             const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
-                             const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
-                             const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
-                             const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
-                             const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
-                             const std::array<ck::index_t, NDimSpatial>& input_left_pads,
-                             const std::array<ck::index_t, NDimSpatial>& input_right_pads,
-                             InElementwiseOperation in_element_op,
-                             WeiElementwiseOperation wei_element_op,
-                             OutElementwiseOperation out_element_op,
-                             ck::index_t split_k)
+    static auto
+    MakeArgument(const InDataType* p_in_grid,
+                 WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                 const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 ck::index_t split_k)
     {
         return Argument{p_in_grid,
                         p_wei_grid,
                         p_out_grid,
-                        G,
-                        N,
-                        K,
-                        C,
-                        input_spatial_lengths,
-                        filter_spatial_lengths,
-                        output_spatial_lengths,
-                        input_strides,
-                        output_strides,
+                        a_g_n_c_wis_lengths, // input
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths, // weight
+                        b_g_k_c_xs_strides,
+                        e_g_n_k_wos_lengths, // output
+                        e_g_n_k_wos_strides,
                         conv_filter_strides,
                         conv_filter_dilations,
                         input_left_pads,
@@ -1159,15 +1162,12 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
     MakeArgumentPointer(const void* p_in_grid,
                         void* p_wei_grid,
                         const void* p_out_grid,
-                        const ck::index_t G,
-                        const ck::index_t N,
-                        const ck::index_t K,
-                        const ck::index_t C,
-                        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
-                        const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
                         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
                         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
                         const std::array<ck::index_t, NDimSpatial>& input_left_pads,
@@ -1180,15 +1180,12 @@ struct DeviceGroupedConvBwdWeightGnwcGkxcGnwk_Dl
         return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
                                           static_cast<WeiDataType*>(p_wei_grid),
                                           static_cast<const OutDataType*>(p_out_grid),
-                                          G,
-                                          N,
-                                          K,
-                                          C,
-                                          input_spatial_lengths,
-                                          filter_spatial_lengths,
-                                          output_spatial_lengths,
-                                          input_strides,
-                                          output_strides,
+                                          a_g_n_c_wis_lengths, // input
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths, // weight
+                                          b_g_k_c_xs_strides,
+                                          e_g_n_k_wos_lengths, // output
+                                          e_g_n_k_wos_strides,
                                           conv_filter_strides,
                                           conv_filter_dilations,
                                           input_left_pads,

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

@@ -245,21 +245,10 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                        const ck::index_t K,
                        const std::array<ck::index_t, NDimSpatial + 3>& output_strides)
     {
-        if constexpr(is_GNHWK_GKYXC_GNHWC)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
-        }
-        else if constexpr(is_NHWGK_GKYXC_NHWGC)
-        {
-            const index_t WoStride = output_strides[4];
-            const auto KStride     = Number<1>{};
-            return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, K),
-                                                make_tuple(WoStride, KStride));
-        }
-        else
-        {
-            throw std::runtime_error("wrong! unsupported layout: " + OutLayout::name());
-        }
+        const index_t WoStride = output_strides[4];
+        const auto KStride     = Number<1>{};
+        return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, K),
+                                            make_tuple(WoStride, KStride));
     }
 
     template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
@@ -270,42 +259,36 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                       const ck::index_t C,
                       const std::array<ck::index_t, NDimSpatial + 3>& input_strides)
     {
-        if constexpr(is_GNHWK_GKYXC_GNHWC)
-        {
-            if constexpr(ConvBackwardWeightSpecialization ==
-                         ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
-            {
-                return make_naive_tensor_descriptor_packed(make_tuple(N * Hi * Wi, C));
-            }
-            else
-            {
-                return make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
-            }
-        }
-        else if constexpr(is_NHWGK_GKYXC_NHWGC)
+        const index_t NStride  = input_strides[1];
+        const index_t HiStride = input_strides[3];
+        const index_t WiStride = input_strides[4];
+        const auto CStride     = input_strides[2];
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
         {
-            const index_t NStride  = input_strides[1];
-            const index_t HiStride = input_strides[3];
-            const index_t WiStride = input_strides[4];
-            const auto CStride     = input_strides[2];
-            if constexpr(ConvBackwardWeightSpecialization ==
-                         ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
-            {
-                return make_naive_tensor_descriptor(make_tuple(N * Hi * Wi, C),
-                                                    make_tuple(WiStride, CStride));
-            }
-            else
-            {
-                return make_naive_tensor_descriptor(
-                    make_tuple(N, Hi, Wi, C), make_tuple(NStride, HiStride, WiStride, CStride));
-            }
+            return make_naive_tensor_descriptor(make_tuple(N * Hi * Wi, C),
+                                                make_tuple(WiStride, CStride));
         }
         else
         {
-            throw std::runtime_error("wrong! unsupported layout: " + InLayout::name());
+            return make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
+                                                make_tuple(NStride, HiStride, WiStride, CStride));
         }
     }
 
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
+    constexpr static auto
+    make_wei_grid_desc(const ck::index_t K,
+                       const ck::index_t Y,
+                       const ck::index_t X,
+                       const ck::index_t C,
+                       const std::array<ck::index_t, NDimSpatial + 3>& weights_strides)
+    {
+        const auto CStride = Number<1>{};
+        const auto KStride = weights_strides[1];
+        return make_naive_tensor_descriptor(make_tuple(K, Y * X * C), make_tuple(KStride, CStride));
+    }
+
     template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
     constexpr static auto
     make_out_grid_desc(const ck::index_t N,
@@ -315,21 +298,10 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                        const ck::index_t K,
                        const std::array<ck::index_t, NDimSpatial + 3>& output_strides)
     {
-        if constexpr(is_GNDHWK_GKZYXC_GNDHWC)
-        {
-            return make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
-        }
-        else if constexpr(is_NDHWGK_GKZYXC_NDHWGC)
-        {
-            const index_t WoStride = output_strides[5];
-            const auto KStride     = Number<1>{};
-            return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, K),
-                                                make_tuple(WoStride, KStride));
-        }
-        else
-        {
-            throw std::runtime_error("wrong! unsupported layout: " + OutLayout::name());
-        }
+        const index_t WoStride = output_strides[5];
+        const auto KStride     = Number<1>{};
+        return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, K),
+                                            make_tuple(WoStride, KStride));
     }
 
     template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
@@ -341,44 +313,40 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                       const ck::index_t C,
                       const std::array<ck::index_t, NDimSpatial + 3>& input_strides)
     {
-        if constexpr(is_GNDHWK_GKZYXC_GNDHWC)
+        const index_t NStride  = input_strides[1];
+        const index_t DiStride = input_strides[3];
+        const index_t HiStride = input_strides[4];
+        const index_t WiStride = input_strides[5];
+        const auto CStride     = input_strides[2];
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
         {
-            if constexpr(ConvBackwardWeightSpecialization ==
-                         ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
-            {
-                return make_naive_tensor_descriptor_packed(make_tuple(N * Di * Hi * Wi, C));
-            }
-            else
-            {
-                return make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
-            }
-        }
-        else if constexpr(is_NDHWGK_GKZYXC_NDHWGC)
-        {
-            const index_t NStride  = input_strides[1];
-            const index_t DiStride = input_strides[3];
-            const index_t HiStride = input_strides[4];
-            const index_t WiStride = input_strides[5];
-            const auto CStride     = input_strides[2];
-            if constexpr(ConvBackwardWeightSpecialization ==
-                         ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
-            {
-                return make_naive_tensor_descriptor(make_tuple(N * Di * Hi * Wi, C),
-                                                    make_tuple(WiStride, CStride));
-            }
-            else
-            {
-                return make_naive_tensor_descriptor(
-                    make_tuple(N, Di, Hi, Wi, C),
-                    make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
-            }
+            return make_naive_tensor_descriptor(make_tuple(N * Di * Hi * Wi, C),
+                                                make_tuple(WiStride, CStride));
         }
         else
         {
-            throw std::runtime_error("wrong! unsupported layout: " + InLayout::name());
+            return make_naive_tensor_descriptor(
+                make_tuple(N, Di, Hi, Wi, C),
+                make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
         }
     }
 
+    template <ck::index_t NDim, typename ck::enable_if<NDim == 3, bool>::type = false>
+    constexpr static auto
+    make_wei_grid_desc(const ck::index_t K,
+                       const ck::index_t Z,
+                       const ck::index_t Y,
+                       const ck::index_t X,
+                       const ck::index_t C,
+                       const std::array<ck::index_t, NDimSpatial + 3>& weights_strides)
+    {
+        const auto CStride = Number<1>{};
+        const auto KStride = weights_strides[1];
+        return make_naive_tensor_descriptor(make_tuple(K, Z * Y * X * C),
+                                            make_tuple(KStride, CStride));
+    }
+
     template <ck::index_t NDim, typename ck::enable_if<NDim == 1, bool>::type = false>
     static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
         const ck::index_t N,
@@ -388,6 +356,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
         const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
         const std::array<ck::index_t, NDimSpatial + 3>& /* input_strides */,
+        const std::array<ck::index_t, NDimSpatial + 3>& /* weights_strides */,
         const std::array<ck::index_t, NDimSpatial + 3>& /* output_strides */,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
@@ -409,6 +378,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         const index_t GemmM      = K;
         const index_t GemmN      = C * X;
 
+        const auto PadGemmM = (MPerBlock - GemmM % MPerBlock) % MPerBlock;
+        const auto PadGemmN = (NPerBlock - GemmN % NPerBlock) % NPerBlock;
+
         const index_t GemmKBatch = batch_k;
         const index_t GemmK0 =
             math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
@@ -527,9 +499,37 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
             const auto wei_gemmm_gemmn_grid_desc =
                 make_naive_tensor_descriptor_packed(make_tuple(K, X * C));
 
-            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
-                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+            // Padd
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmM, PadGemmM),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmN, PadGemmN),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto wei_gemmm_gemmn_pad_grid_desc =
+                transform_tensor_descriptor(wei_gemmm_gemmn_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmM, PadGemmM),
+                                                       make_right_pad_transform(GemmN, PadGemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc,
+                              wei_gemmm_gemmn_pad_grid_desc);
         }
     }
 
@@ -542,6 +542,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
         const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
         const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
+        const std::array<ck::index_t, NDimSpatial + 3>& weights_strides,
         const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
@@ -576,6 +577,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         const index_t GemmM      = K;
         const index_t GemmN      = C * X * Y;
 
+        const auto PadGemmM = (MPerBlock - GemmM % MPerBlock) % MPerBlock;
+        const auto PadGemmN = (NPerBlock - GemmN % NPerBlock) % NPerBlock;
+
         const index_t GemmKBatch = batch_k;
         const index_t GemmK0 =
             math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
@@ -584,6 +588,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
 
         const auto out_grid_desc = make_out_grid_desc<NDim>(N, Ho, Wo, K, output_strides);
         const auto in_grid_desc  = make_in_grid_desc<NDim>(N, Hi, Wi, C, input_strides);
+        const auto wei_grid_desc = make_wei_grid_desc<NDim>(K, Y, X, C, weights_strides);
 
         if constexpr(ConvBackwardWeightSpecialization ==
                      ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
@@ -618,13 +623,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
-            // C: weight tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
-
             return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+                              wei_grid_desc);
         }
         else
         {
@@ -684,13 +685,37 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
-            // C: weight tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
-
-            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
-                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+            // Padd
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmM, PadGemmM),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmN, PadGemmN),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto wei_gemmm_gemmn_pad_grid_desc =
+                transform_tensor_descriptor(wei_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmM, PadGemmM),
+                                                       make_right_pad_transform(GemmN, PadGemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc,
+                              wei_gemmm_gemmn_pad_grid_desc);
         }
     }
 
@@ -703,6 +728,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
         const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
         const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
+        const std::array<ck::index_t, NDimSpatial + 3>& weights_strides,
         const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
@@ -744,6 +770,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         const index_t GemmM      = K;
         const index_t GemmN      = C * Z * X * Y;
 
+        const auto PadGemmM = (MPerBlock - GemmM % MPerBlock) % MPerBlock;
+        const auto PadGemmN = (NPerBlock - GemmN % NPerBlock) % NPerBlock;
+
         const index_t GemmKBatch = batch_k;
         const index_t GemmK0 =
             math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
@@ -752,6 +781,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
 
         const auto out_grid_desc = make_out_grid_desc<NDim>(N, Do, Ho, Wo, K, output_strides);
         const auto in_grid_desc  = make_in_grid_desc<NDim>(N, Di, Hi, Wi, C, input_strides);
+        const auto wei_grid_desc = make_wei_grid_desc<NDim>(K, Z, Y, X, C, weights_strides);
 
         if constexpr(ConvBackwardWeightSpecialization ==
                      ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
@@ -786,13 +816,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
-            // C: weight tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
-
             return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
                               in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+                              wei_grid_desc);
         }
         else
         {
@@ -861,13 +887,37 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
 
-            // C: weight tensor
-            const auto wei_gemmm_gemmn_grid_desc =
-                make_naive_tensor_descriptor_packed(make_tuple(K, Z * Y * X * C));
-
-            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
-                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
-                              wei_gemmm_gemmn_grid_desc);
+            // Padd
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmM, PadGemmM),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch),
+                               make_pass_through_transform(GemmK0),
+                               make_right_pad_transform(GemmN, PadGemmN),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto wei_gemmm_gemmn_pad_grid_desc =
+                transform_tensor_descriptor(wei_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmM, PadGemmM),
+                                                       make_right_pad_transform(GemmN, PadGemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc,
+                              wei_gemmm_gemmn_pad_grid_desc);
         }
     } // function end
 
@@ -887,6 +937,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                                                                   lengths,
                                                                   strides,
                                                                   strides,
+                                                                  strides,
                                                                   params,
                                                                   params,
                                                                   params,
@@ -910,6 +961,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                                                                   lengths,
                                                                   strides,
                                                                   strides,
+                                                                  strides,
                                                                   params,
                                                                   params,
                                                                   params,
@@ -933,6 +985,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                                                                   lengths,
                                                                   strides,
                                                                   strides,
+                                                                  strides,
                                                                   params,
                                                                   params,
                                                                   params,
@@ -1051,15 +1104,12 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         Argument(const InDataType* p_in_grid,
                  WeiDataType* p_wei_grid,
                  const OutDataType* p_out_grid,
-                 const ck::index_t G,
-                 const ck::index_t N,
-                 const ck::index_t K,
-                 const ck::index_t C,
-                 const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-                 const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-                 const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
-                 const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
-                 const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
                  const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
                  const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
                  const std::array<ck::index_t, NDimSpatial>& input_left_pads,
@@ -1084,27 +1134,40 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
               a_element_op_{out_element_op},
               b_element_op_{in_element_op},
               c_element_op_{wei_element_op},
-              Conv_G_{G},
-              Conv_N_{N},
-              Conv_K_{K},
-              Conv_C_{C},
-              output_spatial_lengths_{output_spatial_lengths},
-              filter_spatial_lengths_{filter_spatial_lengths},
+              Conv_G_{a_g_n_c_wis_lengths[0]},
+              Conv_N_{a_g_n_c_wis_lengths[1]},
+              Conv_K_{b_g_k_c_xs_lengths[1]},
+              Conv_C_{a_g_n_c_wis_lengths[2]},
+              input_spatial_lengths_{},
+              filter_spatial_lengths_{},
+              output_spatial_lengths_{},
               conv_filter_strides_{conv_filter_strides},
               input_left_pads_{input_left_pads},
               input_right_pads_{input_right_pads},
               k_batch_{split_k}
         {
+            constexpr index_t spatial_offset = 3;
+            std::copy(begin(a_g_n_c_wis_lengths) + spatial_offset,
+                      end(a_g_n_c_wis_lengths),
+                      begin(input_spatial_lengths_));
+            std::copy(begin(b_g_k_c_xs_lengths) + spatial_offset,
+                      end(b_g_k_c_xs_lengths),
+                      begin(filter_spatial_lengths_));
+            std::copy(begin(e_g_n_k_wos_lengths) + spatial_offset,
+                      end(e_g_n_k_wos_lengths),
+                      begin(output_spatial_lengths_));
+
             const auto descs =
                 DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N<NDimSpatial>(
-                    N,
-                    K,
-                    C,
-                    input_spatial_lengths,
-                    filter_spatial_lengths,
-                    output_spatial_lengths,
-                    input_strides,
-                    output_strides,
+                    Conv_N_,
+                    Conv_K_,
+                    Conv_C_,
+                    input_spatial_lengths_,
+                    filter_spatial_lengths_,
+                    output_spatial_lengths_,
+                    a_g_n_c_wis_strides,
+                    b_g_k_c_xs_strides,
+                    e_g_n_k_wos_strides,
                     conv_filter_strides,
                     conv_filter_dilations,
                     input_left_pads,
@@ -1119,12 +1182,12 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                 GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
 
             // A/B/C Batch Stride
-            compute_ptr_offset_of_batch_.BatchStrideA_ = output_strides[0];
-            compute_ptr_offset_of_batch_.BatchStrideB_ = input_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideA_ = e_g_n_k_wos_strides[0];
+            compute_ptr_offset_of_batch_.BatchStrideB_ = a_g_n_c_wis_strides[0];
             compute_ptr_offset_of_batch_.BatchStrideC_ =
-                K * C *
-                std::accumulate(begin(filter_spatial_lengths),
-                                end(filter_spatial_lengths),
+                Conv_K_ * Conv_C_ *
+                std::accumulate(begin(filter_spatial_lengths_),
+                                end(filter_spatial_lengths_),
                                 index_t{1},
                                 std::multiplies<>{});
 
@@ -1163,8 +1226,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         const index_t Conv_N_;
         const index_t Conv_K_;
         const index_t Conv_C_;
-        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths_;
-        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths_;
+        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_;
         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides_;
         const std::array<ck::index_t, NDimSpatial>& input_left_pads_;
         const std::array<ck::index_t, NDimSpatial>& input_right_pads_;
@@ -1339,39 +1403,34 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
     }
 
-    static auto MakeArgument(const InDataType* p_in_grid,
-                             WeiDataType* p_wei_grid,
-                             const OutDataType* p_out_grid,
-                             const ck::index_t G,
-                             const ck::index_t N,
-                             const ck::index_t K,
-                             const ck::index_t C,
-                             const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-                             const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-                             const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
-                             const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
-                             const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
-                             const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
-                             const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
-                             const std::array<ck::index_t, NDimSpatial>& input_left_pads,
-                             const std::array<ck::index_t, NDimSpatial>& input_right_pads,
-                             InElementwiseOperation in_element_op,
-                             WeiElementwiseOperation wei_element_op,
-                             OutElementwiseOperation out_element_op,
-                             const ck::index_t split_k)
+    static auto
+    MakeArgument(const InDataType* p_in_grid,
+                 WeiDataType* p_wei_grid,
+                 const OutDataType* p_out_grid,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
+                 const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
+                 const std::array<ck::index_t, NDimSpatial>& input_left_pads,
+                 const std::array<ck::index_t, NDimSpatial>& input_right_pads,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 const ck::index_t split_k)
     {
         return Argument{p_in_grid,
                         p_wei_grid,
                         p_out_grid,
-                        G,
-                        N,
-                        K,
-                        C,
-                        input_spatial_lengths,
-                        filter_spatial_lengths,
-                        output_spatial_lengths,
-                        input_strides,
-                        output_strides,
+                        a_g_n_c_wis_lengths, // input
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths, // weight
+                        b_g_k_c_xs_strides,
+                        e_g_n_k_wos_lengths, // output
+                        e_g_n_k_wos_strides,
                         conv_filter_strides,
                         conv_filter_dilations,
                         input_left_pads,
@@ -1390,15 +1449,12 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
     MakeArgumentPointer(const void* p_in_grid,
                         void* p_wei_grid,
                         const void* p_out_grid,
-                        const ck::index_t G,
-                        const ck::index_t N,
-                        const ck::index_t K,
-                        const ck::index_t C,
-                        const std::array<ck::index_t, NDimSpatial>& input_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial>& filter_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial>& output_spatial_lengths,
-                        const std::array<ck::index_t, NDimSpatial + 3>& input_strides,
-                        const std::array<ck::index_t, NDimSpatial + 3>& output_strides,
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths, // input
+                        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths, // weight
+                        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths, // output
+                        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides,
                         const std::array<ck::index_t, NDimSpatial>& conv_filter_strides,
                         const std::array<ck::index_t, NDimSpatial>& conv_filter_dilations,
                         const std::array<ck::index_t, NDimSpatial>& input_left_pads,
@@ -1411,15 +1467,12 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
                                           static_cast<WeiDataType*>(p_wei_grid),
                                           static_cast<const OutDataType*>(p_out_grid),
-                                          G,
-                                          N,
-                                          K,
-                                          C,
-                                          input_spatial_lengths,
-                                          filter_spatial_lengths,
-                                          output_spatial_lengths,
-                                          input_strides,
-                                          output_strides,
+                                          a_g_n_c_wis_lengths, // input
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths, // weight
+                                          b_g_k_c_xs_strides,
+                                          e_g_n_k_wos_lengths, // output
+                                          e_g_n_k_wos_strides,
                                           conv_filter_strides,
                                           conv_filter_dilations,
                                           input_left_pads,