device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp Source File

device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp Source File#

Composable Kernel: device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp Source File
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
#pragma once
 
#include <iostream>
#include <sstream>
 
#include "ck/utility/common_header.hpp"
#include "ck/utility/env.hpp"
#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/device_conv_bwd_data.hpp"
#include "ck/tensor_operation/gpu/device/convolution_backward_data_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
 
namespace ck {
namespace tensor_operation {
namespace device {
 
// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
template <typename InDataType,
          typename WeiDataType,
          typename OutDataType,
          typename AccDataType,
          typename InElementwiseOperation,
          typename WeiElementwiseOperation,
          typename OutElementwiseOperation,
          ConvolutionBackwardDataSpecialization ConvBackwardDataSpecialization,
          ck::index_t BlockSize,
          ck::index_t MPerBlock,
          ck::index_t NPerBlock,
          ck::index_t K0PerBlock,
          ck::index_t K1,
          ck::index_t MPerXDL,
          ck::index_t NPerXDL,
          ck::index_t MXdlPerWave,
          ck::index_t NXdlPerWave,
          typename ABlockTransferThreadClusterLengths_K0_M_K1,
          typename ABlockTransferThreadClusterArrangeOrder,
          typename ABlockTransferSrcAccessOrder,
          ck::index_t ABlockTransferSrcVectorDim,
          ck::index_t ABlockTransferSrcScalarPerVector,
          ck::index_t ABlockTransferDstScalarPerVector_K1,
          bool ABlockLdsAddExtraM,
          typename BBlockTransferThreadClusterLengths_K0_N_K1,
          typename BBlockTransferThreadClusterArrangeOrder,
          typename BBlockTransferSrcAccessOrder,
          ck::index_t BBlockTransferSrcVectorDim,
          ck::index_t BBlockTransferSrcScalarPerVector,
          ck::index_t BBlockTransferDstScalarPerVector_K1,
          bool BBlockLdsAddExtraN,
          ck::index_t CThreadTransferSrcDstVectorDim,
          ck::index_t CThreadTransferDstScalarPerVector>
struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
    : public DeviceConvBwdData<2,
                               ck::tensor_layout::convolution::NHWC,
                               ck::tensor_layout::convolution::KYXC,
                               ck::tensor_layout::convolution::NHWK,
                               InDataType,
                               WeiDataType,
                               OutDataType,
                               InElementwiseOperation,
                               WeiElementwiseOperation,
                               OutElementwiseOperation>
{
    using DeviceOp = DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
 
    GET_NXDL_PER_WAVE_IMPL
    static constexpr auto NXdlPerWave64 = GetNXdlPerWave<true>();
    static constexpr auto NXdlPerWave32 = GetNXdlPerWave<false>();
 
    using ADataType = OutDataType;
    using BDataType = WeiDataType;
    using CDataType = InDataType;
 
    // TODO make A/B datatype different
    using ABDataType = InDataType;
 
    static constexpr index_t NDimSpatial = 2;
 
    static constexpr auto I0 = Number<0>{};
    static constexpr auto I1 = Number<1>{};
    static constexpr auto I2 = Number<2>{};
    static constexpr auto I3 = Number<3>{};
    static constexpr auto I4 = Number<4>{};
    static constexpr auto I5 = Number<5>{};
 
    static_assert((K1 % ABlockTransferThreadClusterLengths_K0_M_K1{}[I2]) %
                      ABlockTransferSrcScalarPerVector ==
                  0);
    static_assert((NPerBlock / BBlockTransferThreadClusterLengths_K0_N_K1{}[I1]) %
                      BBlockTransferSrcScalarPerVector ==
                  0);
 
    static constexpr auto K1Number     = Number<K1>{};
    static constexpr auto GemmK1Number = K1Number;
 
    static auto
    MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(ck::index_t N,
                                                    ck::index_t K,
                                                    ck::index_t C,
                                                    std::vector<ck::index_t> input_spatial_lengths,
                                                    std::vector<ck::index_t> filter_spatial_lengths,
                                                    std::vector<ck::index_t> output_spatial_lengths,
                                                    std::vector<ck::index_t> conv_filter_strides,
                                                    std::vector<ck::index_t> conv_filter_dilations,
                                                    std::vector<ck::index_t> input_left_pads,
                                                    std::vector<ck::index_t> input_right_pads,
                                                    index_t i_ytilde,
                                                    index_t i_xtilde)
    {
        using namespace ck;
 
        const index_t Hi = input_spatial_lengths[0];
        const index_t Wi = input_spatial_lengths[1];
 
        const index_t Ho = output_spatial_lengths[0];
        const index_t Wo = output_spatial_lengths[1];
 
        const index_t Y = filter_spatial_lengths[0];
        const index_t X = filter_spatial_lengths[1];
 
        const index_t InLeftPadH = input_left_pads[0];
        const index_t InLeftPadW = input_left_pads[1];
 
        const index_t InRightPadH = input_right_pads[0];
        const index_t InRightPadW = input_right_pads[1];
 
        const index_t ConvStrideH = conv_filter_strides[0];
        const index_t ConvStrideW = conv_filter_strides[1];
 
        const index_t ConvDilationH = conv_filter_dilations[0];
        const index_t ConvDilationW = conv_filter_dilations[1];
 
        const auto K0 = K / K1;
 
        const auto out_n_ho_wo_k_grid_desc =
            make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
        const auto wei_k_y_x_c_grid_desc =
            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
        const auto in_n_hi_wi_c_grid_desc =
            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
 
        if constexpr(ConvBackwardDataSpecialization ==
                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
        {
            // A: output tensor
            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
                make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
                make_tuple(make_pass_through_transform(N * Ho * Wo),
                           make_unmerge_transform(make_tuple(K0, K1))),
                make_tuple(Sequence<0>{}, Sequence<1>{}),
                make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
 
            // B: weight tensor
            const auto wei_gemmk0_gemmn_gemmk1_grid_desc =
                transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
                                                       make_pass_through_transform(C)),
                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
            // C: input tensor
            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
                in_n_hi_wi_c_grid_desc,
                make_tuple(make_pass_through_transform(N),
                           make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, ConvStrideW)),
                           make_pass_through_transform(C)),
                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
 
            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
                in_n_y_ho_x_wo_c_grid_desc,
                make_tuple(make_freeze_transform(I0),
                           make_freeze_transform(I0),
                           make_merge_transform(make_tuple(N, Ho, Wo)),
                           make_pass_through_transform(C)),
                make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
 
            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
                              wei_gemmk0_gemmn_gemmk1_grid_desc,
                              in_gemmm_gemmn_grid_desc);
        }
        else
        {
            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
            const auto YTilde = ConvStrideH / GcdStrideDilationH;
            const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
            const auto YDot = math::integer_divide_ceil(Y, YTilde);
            const auto XDot = math::integer_divide_ceil(X, XTilde);
 
            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 HTilde and WTilde that contribute to non-padding area of input tensor
            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 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 HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
            const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
 
            // GemmK is different for each GEMM
            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
            // A: output tensor
            const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
                out_n_ho_wo_k_grid_desc,
                make_tuple(make_pass_through_transform(N),
                           make_pad_transform(Ho, I0, I0),
                           make_pad_transform(Wo, I0, I0),
                           make_pass_through_transform(K)),
                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
 
            const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
                out_n_hop_wop_k_grid_desc,
                make_tuple(
                    make_pass_through_transform(N),
                    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(K)),
                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
 
            const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc =
                transform_tensor_descriptor(
                    out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
                    make_tuple(make_pass_through_transform(N),
                               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_unmerge_transform(make_tuple(K0, K1))),
                    make_tuple(Sequence<0>{},
                               Sequence<1>{},
                               Sequence<2>{},
                               Sequence<3>{},
                               Sequence<4>{},
                               Sequence<5>{}),
                    make_tuple(Sequence<0>{},
                               Sequence<1>{},
                               Sequence<2>{},
                               Sequence<3>{},
                               Sequence<4>{},
                               Sequence<5, 6>{}));
 
            const auto out_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
                out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k0_k1_grid_desc,
                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
                           make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
                           make_pass_through_transform(K1)),
                make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
 
            // B weight tensor
            const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
                wei_k_y_x_c_grid_desc,
                make_tuple(make_pass_through_transform(K),
                           make_embed_transform(make_tuple(YDot, YTilde),
                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
                           make_embed_transform(make_tuple(XDot, XTilde),
                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
                           make_pass_through_transform(C)),
                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
 
            const auto wei_k0_k1_ydotslice_xdotslice_c_grid_desc =
                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
                                            make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
                                                       make_slice_transform(YDot, I0, YDotSlice),
                                                       make_slice_transform(XDot, I0, XDotSlice),
                                                       make_freeze_transform(i_ytilde),
                                                       make_freeze_transform(i_xtilde),
                                                       make_pass_through_transform(C)),
                                            make_tuple(Sequence<0>{},
                                                       Sequence<1>{},
                                                       Sequence<3>{},
                                                       Sequence<2>{},
                                                       Sequence<4>{},
                                                       Sequence<5>{}),
                                            make_tuple(Sequence<0, 1>{},
                                                       Sequence<2>{},
                                                       Sequence<3>{},
                                                       Sequence<>{},
                                                       Sequence<>{},
                                                       Sequence<4>{}));
 
            const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
                wei_k0_k1_ydotslice_xdotslice_c_grid_desc,
                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
                           make_pass_through_transform(C),
                           make_pass_through_transform(K1)),
                make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
 
            // C: input tensor
            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
                in_n_hi_wi_c_grid_desc,
                make_tuple(make_pass_through_transform(N),
                           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>{}),
                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
 
            const auto in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc = transform_tensor_descriptor(
                in_n_hip_wip_c_grid_desc,
                make_tuple(make_pass_through_transform(N),
                           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>{}),
                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
 
            const auto in_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
                in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
                make_tuple(make_pass_through_transform(N),
                           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>{}),
                make_tuple(Sequence<0>{},
                           Sequence<>{},
                           Sequence<1>{},
                           Sequence<>{},
                           Sequence<2>{},
                           Sequence<3>{}));
 
            const auto in_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
                in_n_htildeslice_wtildeslice_c_grid_desc,
                make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
                           make_pass_through_transform(C)),
                make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
                make_tuple(Sequence<0>{}, Sequence<1>{}));
 
            return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
                              wei_gemmk0_gemmn_gemmk1_grid_desc,
                              in_gemmm_gemmn_grid_desc);
        }
 
    } // function end
 
    using ABCGridDescs = decltype(MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
        1, 1, 1, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, {1, 1}, 0, 0));
 
    using AGridDesc_K0_M_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I0])>;
    using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
    using CGridDesc_M_N     = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
 
    // GridwiseGemm
    template <index_t NXdlPerWave_>
    using GridwiseGemmBase = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
        BlockSize,
        ABDataType, // TODO: distinguish A/B datatype
        AccDataType,
        CDataType,
        InMemoryDataOperationEnum::Set,
        InElementwiseOperation,
        WeiElementwiseOperation,
        OutElementwiseOperation,
        MPerBlock,
        NPerBlock,
        K0PerBlock,
        MPerXDL,
        NPerXDL,
        K1,
        MXdlPerWave,
        NXdlPerWave_,
        ABlockTransferThreadClusterLengths_K0_M_K1,
        ABlockTransferThreadClusterArrangeOrder,
        ABlockTransferSrcAccessOrder,
        ABlockTransferSrcVectorDim,
        ABlockTransferSrcScalarPerVector,
        ABlockTransferDstScalarPerVector_K1,
        false, // AThreadTransferSrcResetCoordinateAfterRun,
        ABlockLdsAddExtraM,
        BBlockTransferThreadClusterLengths_K0_N_K1,
        BBlockTransferThreadClusterArrangeOrder,
        BBlockTransferSrcAccessOrder,
        BBlockTransferSrcVectorDim,
        BBlockTransferSrcScalarPerVector,
        BBlockTransferDstScalarPerVector_K1,
        false, // BThreadTransferSrcResetCoordinateAfterRun,
        BBlockLdsAddExtraN,
        Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
        7,                                // CThreadTransferSrcDstVectorDim,
        CThreadTransferDstScalarPerVector>;
    using GridwiseGemm64 = GridwiseGemmBase<math::max(NXdlPerWave64, 1)>;
    using GridwiseGemm32 = GridwiseGemmBase<NXdlPerWave32>;
 
    // Argument
    struct Argument : public BaseArgument
    {
        Argument(InDataType* p_in_grid,
                 const WeiDataType* p_wei_grid,
                 const OutDataType* p_out_grid,
                 ck::index_t N,
                 ck::index_t K,
                 ck::index_t C,
                 std::vector<ck::index_t> input_spatial_lengths,
                 std::vector<ck::index_t> filter_spatial_lengths,
                 std::vector<ck::index_t> output_spatial_lengths,
                 std::vector<ck::index_t> conv_filter_strides,
                 std::vector<ck::index_t> conv_filter_dilations,
                 std::vector<ck::index_t> input_left_pads,
                 std::vector<ck::index_t> input_right_pads)
            : p_a_grid_{p_out_grid},
              p_b_grid_{p_wei_grid},
              p_c_grid_{p_in_grid},
              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_filter_strides_{conv_filter_strides},
              conv_filter_dilations_{conv_filter_dilations},
              input_left_pads_{input_left_pads},
              input_right_pads_{input_right_pads}
        {
            const index_t ConvStrideH = conv_filter_strides[0];
            const index_t ConvStrideW = conv_filter_strides[1];
 
            const index_t ConvDilationH = conv_filter_dilations[0];
            const index_t ConvDilationW = conv_filter_dilations[1];
 
            const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
            const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
            const auto YTilde = ConvStrideH / GcdStrideDilationH;
            const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
            for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
            {
                for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
                {
                    // check slice is valid
                    const index_t Y      = filter_spatial_lengths_[0];
                    const index_t X      = filter_spatial_lengths_[1];
                    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
                    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
                    if(YDotSlice * XDotSlice <= 0)
                    {
                        continue;
                    }
 
                    const auto descs = DeviceOp::MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
                        N,
                        K,
                        C,
                        input_spatial_lengths,
                        filter_spatial_lengths,
                        output_spatial_lengths,
                        conv_filter_strides,
                        conv_filter_dilations,
                        input_left_pads,
                        input_right_pads,
                        i_ytilde,
                        i_xtilde);
                    a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
                    b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
                    c_grid_desc_m_n_container_.push_back(descs[I2]);
                }
            }
        }
 
        const ADataType* p_a_grid_;
        const BDataType* p_b_grid_;
        CDataType* p_c_grid_;
        std::vector<AGridDesc_K0_M_K1> a_grid_desc_k0_m_k1_container_;
        std::vector<BGridDesc_K0_N_K1> b_grid_desc_k0_n_k1_container_;
        std::vector<CGridDesc_M_N> c_grid_desc_m_n_container_;
        // for checking IsSupportedArgument()
        index_t Conv_N_;
        index_t Conv_K_;
        index_t Conv_C_;
 
        std::vector<ck::index_t> input_spatial_lengths_;
        std::vector<ck::index_t> filter_spatial_lengths_;
        std::vector<ck::index_t> output_spatial_lengths_;
        std::vector<ck::index_t> conv_filter_strides_;
        std::vector<ck::index_t> conv_filter_dilations_;
        std::vector<ck::index_t> input_left_pads_;
        std::vector<ck::index_t> input_right_pads_;
    };
 
    // Invoker
    struct Invoker : public BaseInvoker
    {
        using Argument = DeviceOp::Argument;
 
        template <typename GridwiseGemm>
        float RunImp(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
        {
            float ave_time = 0;
            for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
            {
                if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
                {
                    {
                        std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
                                  << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
                                  << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I1) << ", "
                                  << arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2) << "}"
                                  << std::endl;
 
                        std::cout << "arg.b_grid_desc_k0_n_k1_container_{"
                                  << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I0) << ", "
                                  << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I1) << ", "
                                  << arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I2) << "}"
                                  << std::endl;
 
                        std::cout << "arg.c_grid_desc_m_n_container_{ "
                                  << arg.c_grid_desc_m_n_container_[i].GetLength(I0) << ", "
                                  << arg.c_grid_desc_m_n_container_[i].GetLength(I1) << "}"
                                  << std::endl;
                    }
                }
 
                if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
                                                arg.b_grid_desc_k0_n_k1_container_[i],
                                                arg.c_grid_desc_m_n_container_[i]))
                {
                    throw std::runtime_error(
                        "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
                }
 
                const auto [gdx, gdy, gdz] =
                    GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_container_[i]);
 
                const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
                               arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
 
                if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
                {
                    const auto kernel =
                        kernel_gemm_xdlops_v2r3<GridwiseGemm,
                                                ADataType, // TODO: distiguish A/B datatype
                                                CDataType,
                                                DeviceOp::AGridDesc_K0_M_K1,
                                                DeviceOp::BGridDesc_K0_N_K1,
                                                DeviceOp::CGridDesc_M_N,
                                                true>;
 
                    ave_time += launch_and_time_kernel(stream_config,
                                                       kernel,
                                                       dim3(gdx, gdy, gdz),
                                                       dim3(BlockSize),
                                                       0,
                                                       arg.p_a_grid_,
                                                       arg.p_b_grid_,
                                                       arg.p_c_grid_,
                                                       arg.a_grid_desc_k0_m_k1_container_[i],
                                                       arg.b_grid_desc_k0_n_k1_container_[i],
                                                       arg.c_grid_desc_m_n_container_[i]);
                }
                else
                {
                    const auto kernel =
                        kernel_gemm_xdlops_v2r3<GridwiseGemm,
                                                ADataType, // TODO: distiguish A/B datatype
                                                CDataType,
                                                DeviceOp::AGridDesc_K0_M_K1,
                                                DeviceOp::BGridDesc_K0_N_K1,
                                                DeviceOp::CGridDesc_M_N,
                                                false>;
 
                    ave_time += launch_and_time_kernel(stream_config,
                                                       kernel,
                                                       dim3(gdx, gdy, gdz),
                                                       dim3(BlockSize),
                                                       0,
                                                       arg.p_a_grid_,
                                                       arg.p_b_grid_,
                                                       arg.p_c_grid_,
                                                       arg.a_grid_desc_k0_m_k1_container_[i],
                                                       arg.b_grid_desc_k0_n_k1_container_[i],
                                                       arg.c_grid_desc_m_n_container_[i]);
                }
            }
            return ave_time;
        }
 
        INVOKER_RUN_IMPL
 
        float Run(const BaseArgument* p_arg,
                  const StreamConfig& stream_config = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
        }
    };
 
    static constexpr bool IsValidCompilationParameter()
    {
        // TODO: properly implement this check
        return true;
    }
 
    static bool IsSupportedArgument(const Argument& arg)
    {
        if(!ck::is_xdl_wmma_supported<ADataType, BDataType, MPerXDL, NPerXDL>())
        {
            return false;
        }
        if constexpr(ConvBackwardDataSpecialization ==
                     ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0)
        {
            // check if it's 1x1, stride=1 pad = 0 conv
            if(!(arg.filter_spatial_lengths_[0] == 1 && arg.filter_spatial_lengths_[1] == 1 &&
                 arg.conv_filter_strides_[0] == 1 && arg.conv_filter_strides_[1] == 1 &&
                 arg.input_left_pads_[0] == 0 && arg.input_left_pads_[1] == 0 &&
                 arg.input_right_pads_[0] == 0 && arg.input_right_pads_[1] == 0))
            {
                return false;
            }
        }
 
        // vector load A/B matrix from global memory
        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 1 &&
             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
        {
            return false;
        }
 
        // vector store C matrix into global memory
        if(!(arg.Conv_C_ % CThreadTransferDstScalarPerVector == 0))
        {
            return false;
        }
 
        // Gridwise GEMM size
        bool isWave64 = get_warp_size() == 64;
        for(std::size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
        {
            bool valid = false;
            if(isWave64)
            {
                if constexpr(NXdlPerWave64 > 0)
                {
                    valid = GridwiseGemm64::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
                                                          arg.b_grid_desc_k0_n_k1_container_[i],
                                                          arg.c_grid_desc_m_n_container_[i]);
                }
            }
            else
            {
                if constexpr(NXdlPerWave32 > 0)
                {
                    valid = GridwiseGemm32::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
                                                          arg.b_grid_desc_k0_n_k1_container_[i],
                                                          arg.c_grid_desc_m_n_container_[i]);
                }
            }
            if(!valid)
                return false;
        }
        return true;
    }
 
    bool IsSupportedArgument(const BaseArgument* p_arg) override
    {
        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
    }
 
    static auto MakeArgument(InDataType* p_in_grid,
                             const WeiDataType* p_wei_grid,
                             const OutDataType* p_out_grid,
                             ck::index_t N,
                             ck::index_t K,
                             ck::index_t C,
                             std::vector<ck::index_t> input_spatial_lengths,
                             std::vector<ck::index_t> filter_spatial_lengths,
                             std::vector<ck::index_t> output_spatial_lengths,
                             std::vector<ck::index_t> conv_filter_strides,
                             std::vector<ck::index_t> conv_filter_dilations,
                             std::vector<ck::index_t> input_left_pads,
                             std::vector<ck::index_t> input_right_pads)
    {
        return Argument{p_in_grid,
                        p_wei_grid,
                        p_out_grid,
                        N,
                        K,
                        C,
                        input_spatial_lengths,
                        filter_spatial_lengths,
                        output_spatial_lengths,
                        conv_filter_strides,
                        conv_filter_dilations,
                        input_left_pads,
                        input_right_pads};
    }
 
    static auto MakeInvoker() { return Invoker{}; }
 
    std::unique_ptr<BaseArgument>
    MakeArgumentPointer(void* p_in_grid,
                        const void* p_wei_grid,
                        const void* p_out_grid,
                        ck::index_t N,
                        ck::index_t K,
                        ck::index_t C,
                        std::vector<ck::index_t> input_spatial_lengths,
                        std::vector<ck::index_t> filter_spatial_lengths,
                        std::vector<ck::index_t> output_spatial_lengths,
                        std::vector<ck::index_t> conv_filter_strides,
                        std::vector<ck::index_t> conv_filter_dilations,
                        std::vector<ck::index_t> input_left_pads,
                        std::vector<ck::index_t> input_right_pads,
                        InElementwiseOperation,
                        WeiElementwiseOperation,
                        OutElementwiseOperation) override
    {
        return std::make_unique<Argument>(static_cast<InDataType*>(p_in_grid),
                                          static_cast<const WeiDataType*>(p_wei_grid),
                                          static_cast<const OutDataType*>(p_out_grid),
                                          N,
                                          K,
                                          C,
                                          input_spatial_lengths,
                                          filter_spatial_lengths,
                                          output_spatial_lengths,
                                          conv_filter_strides,
                                          conv_filter_dilations,
                                          input_left_pads,
                                          input_right_pads);
    }
 
    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
    {
        return std::make_unique<Invoker>(Invoker{});
    }
 
    std::string GetTypeString() const override
    {
        auto str = std::stringstream();
 
        // clang-format off
        str << "DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
            << "<"
            << BlockSize << ", "
            << MPerBlock << ", "
            << NPerBlock << ", "
            << K0PerBlock << ", "
            << K1 << ", "
            << MXdlPerWave << ", "
            << NXdlPerWave << ", "
            << ABlockTransferSrcScalarPerVector << ", "
            << ABlockTransferDstScalarPerVector_K1 << ", "
            << BBlockTransferSrcScalarPerVector << ", "
            << BBlockTransferDstScalarPerVector_K1
            << ">";
        // clang-format on
 
        return str.str();
    }
};
 
} // namespace device
} // namespace tensor_operation
} // namespace ck