merge develop

include/ck/tensor_operation/gpu/block/reduction_functions_blockwise.hpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#ifndef CK_REDUCTION_FUNCTIONS_BLOCKWISE_HPP
+#define CK_REDUCTION_FUNCTIONS_BLOCKWISE_HPP
+
+#include "data_type.hpp"
+
+#include "reduction_common.hpp"
+#include "reduction_operator.hpp"
+#include "reduction_functions_accumulate.hpp"
+
+namespace ck {
+
+template <typename Buffer1dDescType,
+          typename AccDataType,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          bool ReorderThreadClusters,
+          typename OpReduce,
+          bool PropagateNan>
+struct PartitionedBlockwiseReductionOn1dBuffer
+{
+    static constexpr auto buffer_1d_desc = Buffer1dDescType{};
+
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "The product of cluster lengths should be same as BlockSize!");
+    static_assert(KThreadClusterSize > 1, "Parallel reduction need work on at least two elements");
+
+    static_assert(buffer_1d_desc.GetElementSize() == BlockSize,
+                  "The buffer size should be the same as BlockSize!");
+
+    using Accumulation = detail::AccumulateWithNanCheck<PropagateNan, OpReduce, AccDataType>;
+
+    template <typename BufferType>
+    __device__ static void Reduce(BufferType& block_buffer,
+                                  AccDataType& accuData,
+                                  index_t thread_m_cluster_id,
+                                  index_t thread_k_cluster_id)
+    {
+        constexpr auto cluster_len_shift = get_shift<KThreadClusterSize>();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << (cluster_len_shift - 1 - I());
+
+            if(thread_k_cluster_id < indOffset)
+            {
+                // consider the thread clusters order, ensure the contiguous locations are accessed
+                // by contiguous Thread-ID
+                index_t offset1 =
+                    ReorderThreadClusters
+                        ? buffer_1d_desc.CalculateOffset(make_tuple(
+                              thread_k_cluster_id * MThreadClusterSize + thread_m_cluster_id))
+                        : buffer_1d_desc.CalculateOffset(make_tuple(
+                              thread_m_cluster_id * KThreadClusterSize + thread_k_cluster_id));
+                index_t offset2 = ReorderThreadClusters
+                                      ? buffer_1d_desc.CalculateOffset(make_tuple(
+                                            (thread_k_cluster_id + indOffset) * MThreadClusterSize +
+                                            thread_m_cluster_id))
+                                      : buffer_1d_desc.CalculateOffset(
+                                            make_tuple(thread_m_cluster_id * KThreadClusterSize +
+                                                       (thread_k_cluster_id + indOffset)));
+
+                AccDataType opData1 = type_convert<AccDataType>(block_buffer[offset1]);
+                AccDataType opData2 = type_convert<AccDataType>(block_buffer[offset2]);
+                Accumulation::Calculate(opData1, opData2);
+                block_buffer(offset1) = type_convert<AccDataType>(opData1);
+            }
+
+            __syncthreads();
+        });
+
+        index_t offset = ReorderThreadClusters
+                             ? buffer_1d_desc.CalculateOffset(make_tuple(thread_m_cluster_id))
+                             : buffer_1d_desc.CalculateOffset(
+                                   make_tuple(thread_m_cluster_id * KThreadClusterSize));
+
+        accuData = type_convert<AccDataType>(block_buffer[offset]);
+    };
+};
+
+template <typename Buffer1dDescType,
+          typename AccDataType,
+          typename IndexDataType,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          bool ReorderThreadClusters,
+          typename OpReduce,
+          bool PropagateNan>
+struct PartitionedBlockwiseReductionWithIndexOn1dBuffer
+{
+    static constexpr auto buffer_1d_desc = Buffer1dDescType{};
+
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
+                  "The product of cluster lengths should be same as BlockSize!");
+    static_assert(KThreadClusterSize > 1, "Parallel reduction need work on at least two elements");
+
+    static_assert(buffer_1d_desc.GetElementSize() == BlockSize,
+                  "The buffer size should be the same as BlockSize!");
+
+    using Accumulation =
+        detail::AccumulateWithIndexAndNanCheck<PropagateNan, OpReduce, AccDataType, IndexDataType>;
+
+    // This interface accumulates on both data values and indices
+    template <typename BufferType, typename IdxBufferType>
+    __device__ static void Reduce(BufferType& block_val_buffer,
+                                  IdxBufferType& block_idx_buffer,
+                                  AccDataType& accuData,
+                                  IndexDataType& accuIndex,
+                                  index_t thread_m_cluster_id,
+                                  index_t thread_k_cluster_id)
+    {
+        constexpr auto cluster_len_shift = get_shift<KThreadClusterSize>();
+
+        static_for<0, cluster_len_shift, 1>{}([&](auto I) {
+            constexpr index_t indOffset = 1 << I();
+
+            if(thread_k_cluster_id % (indOffset * 2) == 0)
+            {
+                // consider the thread clusters order, ensure the contiguous locations are accessed
+                // by contiguous Thread-ID
+                index_t offset1 =
+                    ReorderThreadClusters
+                        ? buffer_1d_desc.CalculateOffset(make_tuple(
+                              thread_k_cluster_id * MThreadClusterSize + thread_m_cluster_id))
+                        : buffer_1d_desc.CalculateOffset(make_tuple(
+                              thread_m_cluster_id * KThreadClusterSize + thread_k_cluster_id));
+                index_t offset2 = ReorderThreadClusters
+                                      ? buffer_1d_desc.CalculateOffset(make_tuple(
+                                            (thread_k_cluster_id + indOffset) * MThreadClusterSize +
+                                            thread_m_cluster_id))
+                                      : buffer_1d_desc.CalculateOffset(
+                                            make_tuple(thread_m_cluster_id * KThreadClusterSize +
+                                                       (thread_k_cluster_id + indOffset)));
+
+                AccDataType opData1      = type_convert<AccDataType>(block_val_buffer[offset1]);
+                AccDataType opData2      = type_convert<AccDataType>(block_val_buffer[offset2]);
+                IndexDataType currIndex1 = block_idx_buffer[offset1];
+                IndexDataType currIndex2 = block_idx_buffer[offset2];
+
+                Accumulation::Calculate(opData1, opData2, currIndex1, currIndex2);
+                block_val_buffer(offset1) = type_convert<AccDataType>(opData1);
+                block_idx_buffer(offset1) = currIndex1;
+            }
+
+            __syncthreads();
+        });
+
+        index_t offset = ReorderThreadClusters
+                             ? buffer_1d_desc.CalculateOffset(make_tuple(thread_m_cluster_id))
+                             : buffer_1d_desc.CalculateOffset(
+                                   make_tuple(thread_m_cluster_id * KThreadClusterSize));
+
+        accuData  = type_convert<AccDataType>(block_val_buffer[offset]);
+        accuIndex = block_idx_buffer[offset];
+    }
+};
+
+}; // end of namespace ck
+
+#endif

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

+#ifndef CONVOLUTION_BACKWARD_DATA_SPECIALIZATION
+#define CONVOLUTION_BACKWARD_DATA_SPECIALIZATION
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+enum ConvolutionBackwardDataSpecialization_t
+{
+    Default,
+    Filter1x1Stride1Pad0,
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

include/ck/tensor_operation/gpu/device/device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp

+#ifndef DEVICE_CONV2D_WRW_XDL_C_SHUFFLE_NHWC_KYXC_NHWK_HPP
+#define DEVICE_CONV2D_WRW_XDL_C_SHUFFLE_NHWC_KYXC_NHWK_HPP
+
+#include <iostream>
+#include <sstream>
+#include "device.hpp"
+#include "device_base.hpp"
+#include "device_conv_backward_weight.hpp"
+#include "convolution_forward_specialization.hpp"
+#include "common_header.hpp"
+#include "tensor_layout.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "gridwise_gemm_xdlops_v2r4r2.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,
+          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,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+struct DeviceConv2dWrWXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
+    : public DeviceConvWrw<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConv2dWrWXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    using ADataType = OutDataType;
+    using BDataType = InDataType;
+    using CDataType = WeiDataType;
+
+    using AElementwiseOperation = OutElementwiseOperation;
+    using BElementwiseOperation = InElementwiseOperation;
+    using CElementwiseOperation = WeiElementwiseOperation;
+
+    // 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 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,
+                                                    ck::index_t batch_k)
+    {
+        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 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 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 GemmKTotal = N * Ho * Wo;
+        const index_t GemmM      = K;
+        const index_t GemmN      = C * X * Y;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        const auto out_gemmktotal_gemmm_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+        const auto in_n_hi_wi_c_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+        // A: output tensor
+        const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+            out_gemmktotal_gemmm_grid_desc,
+            make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                       make_pass_through_transform(GemmM)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+            out_gemmkpad_gemmm_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                       make_pass_through_transform(GemmM)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+        // B: 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_y_ho_x_wo_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(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                make_embed_transform(make_tuple(X, Wo), 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_gemmktotal_gemmn_grid_desc =
+            transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                        make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                                   make_merge_transform(make_tuple(N, Ho, Wo))),
+                                        make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+        const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+            in_gemmktotal_gemmn_grid_desc,
+            make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                       make_pass_through_transform(GemmN)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+            in_gemmkpad_gemmn_grid_desc,
+            make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1Number)),
+                       make_pass_through_transform(GemmN)),
+            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);
+    }
+
+    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}, 1));
+
+    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
+    using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum_t::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        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,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXdl,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
+
+    using GridwiseGemmAtomicAdd = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2<
+        BlockSize,
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum_t::AtomicAdd,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CElementwiseOperation,
+        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,
+        CShuffleMXdlPerWavePerShuffle,
+        CShuffleNXdlPerWavePerShuffle,
+        CBlockTransferScalarPerVector_NWaveNPerXdl,
+        CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock>;
+    // Argument
+    using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        decltype(GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}));
+
+    using Block2CTileMap =
+        decltype(GridwiseGemm::MakeCBlockClusterAdaptor(CGridDesc_M_N{}, 1, 1, 1));
+    struct Argument : public BaseArgument
+    {
+        Argument(const InDataType* p_in_grid,
+                 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,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op,
+                 ck::index_t split_k)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_in_grid},
+              p_c_grid_{p_wei_grid},
+              a_grid_desc_kbatch_k0_m_k1_{},
+              b_grid_desc_kbatch_k0_n_k1_{},
+              c_grid_desc_m_n_{},
+              c_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_ctile_map_{},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{out_element_op},
+              b_element_op_{in_element_op},
+              c_element_op_{wei_element_op},
+              Conv_N_{N},
+              Conv_K_{K},
+              Conv_C_{C},
+              output_spatial_lengths_{output_spatial_lengths},
+              filter_spatial_lengths_{filter_spatial_lengths},
+              conv_filter_strides_{conv_filter_strides},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads},
+              k_batch_{split_k}
+        {
+            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,
+                                                                          k_batch_);
+
+            a_grid_desc_kbatch_k0_m_k1_ = descs[I0];
+            b_grid_desc_kbatch_k0_n_k1_ = descs[I1];
+            c_grid_desc_m_n_            = descs[I2];
+
+            if(GridwiseGemm::CheckValidity(a_grid_desc_kbatch_k0_m_k1_,
+                                           b_grid_desc_kbatch_k0_n_k1_,
+                                           c_grid_desc_m_n_,
+                                           M01_,
+                                           N01_))
+            {
+                c_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                    GridwiseGemm::MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n_);
+
+                block_2_ctile_map_ =
+                    GridwiseGemm::MakeCBlockClusterAdaptor(c_grid_desc_m_n_, M01, N01, k_batch_);
+            }
+        }
+
+        const ADataType* p_a_grid_;
+        const BDataType* p_b_grid_;
+        CDataType* p_c_grid_;
+        AGridDesc_K0_M_K1 a_grid_desc_kbatch_k0_m_k1_;
+        BGridDesc_K0_N_K1 b_grid_desc_kbatch_k0_n_k1_;
+        CGridDesc_M_N c_grid_desc_m_n_;
+        CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock c_grid_desc_mblock_mperblock_nblock_nperblock_;
+        Block2CTileMap block_2_ctile_map_;
+        index_t M01_;
+        index_t N01_;
+        InElementwiseOperation a_element_op_;
+        OutElementwiseOperation b_element_op_;
+        WeiElementwiseOperation c_element_op_;
+        // for checking IsSupportedArgument()
+        index_t Conv_N_;
+        index_t Conv_K_;
+        index_t Conv_C_;
+        std::vector<index_t> output_spatial_lengths_;
+        std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> input_left_pads_;
+        std::vector<index_t> input_right_pads_;
+        index_t k_batch_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        void ShowInfo(const Argument& arg)
+        {
+            std::cout << "arg.a_grid_desc_kbatch_k0_m_k1_{"
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I2) << ", "
+                      << arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.b_grid_desc_kbatch_k0_n_k1_{"
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I0) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I1) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I2) << ", "
+                      << arg.b_grid_desc_kbatch_k0_n_k1_.GetLength(I3) << "}" << std::endl;
+
+            std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
+                      << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
+        }
+
+        float Run(const Argument& arg, int nrepeat = 1)
+        {
+            ShowInfo(arg);
+            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                            arg.b_grid_desc_kbatch_k0_n_k1_,
+                                            arg.c_grid_desc_m_n_,
+                                            arg.M01_,
+                                            arg.N01_))
+            {
+                throw std::runtime_error(
+                    "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+            }
+            const auto kbatch       = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I0);
+            const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_, kbatch);
+
+            const auto K0 = arg.a_grid_desc_kbatch_k0_m_k1_.GetLength(I1);
+
+            const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+            float ave_time = 0;
+
+            const auto Run = [&](const auto& kernel) {
+                if(nrepeat > 0)
+                {
+                    ave_time =
+                        launch_and_time_kernel(kernel,
+                                               nrepeat,
+                                               dim3(grid_size),
+                                               dim3(BlockSize),
+                                               0,
+                                               arg.p_a_grid_,
+                                               arg.p_b_grid_,
+                                               arg.p_c_grid_,
+                                               arg.a_grid_desc_kbatch_k0_m_k1_,
+                                               arg.b_grid_desc_kbatch_k0_n_k1_,
+                                               arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                               arg.a_element_op_,
+                                               arg.b_element_op_,
+                                               arg.c_element_op_,
+                                               arg.block_2_ctile_map_);
+                }
+
+                if(kbatch > 1 || nrepeat <= 0)
+                {
+                    hipGetErrorString(hipMemset(
+                        arg.p_c_grid_,
+                        0,
+                        arg.c_grid_desc_mblock_mperblock_nblock_nperblock_.GetElementSpaceSize() *
+                            sizeof(CDataType)));
+
+                    launch_kernel(kernel,
+                                  dim3(grid_size),
+                                  dim3(BlockSize),
+                                  0,
+                                  arg.p_a_grid_,
+                                  arg.p_b_grid_,
+                                  arg.p_c_grid_,
+                                  arg.a_grid_desc_kbatch_k0_m_k1_,
+                                  arg.b_grid_desc_kbatch_k0_n_k1_,
+                                  arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
+                                  arg.a_element_op_,
+                                  arg.b_element_op_,
+                                  arg.c_element_op_,
+                                  arg.block_2_ctile_map_);
+                }
+            };
+
+            if(has_main_k0_block_loop)
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r4r2<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        true>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r4r2<
+                        GridwiseGemmAtomicAdd,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        true>;
+
+                    Run(kernel);
+                }
+            }
+            else
+            {
+                if(kbatch == 1)
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r4r2<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        false>;
+
+                    Run(kernel);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r4r2<
+                        GridwiseGemmAtomicAdd,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
+                        OutElementwiseOperation,
+                        InElementwiseOperation,
+                        WeiElementwiseOperation,
+                        remove_reference_t<DeviceOp::Block2CTileMap>,
+                        false>;
+
+                    Run(kernel);
+                }
+            }
+
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg, int nrepeat = 1) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        // vector load A/B matrix from global memory
+        if(!(ABlockTransferSrcVectorDim == 2 && BBlockTransferSrcVectorDim == 2 &&
+             arg.Conv_K_ % ABlockTransferSrcScalarPerVector == 0 &&
+             arg.Conv_C_ % BBlockTransferSrcScalarPerVector == 0))
+        {
+            return false;
+        }
+
+        // vector store C matrix into global memory
+        if(!(arg.Conv_C_ % CBlockTransferScalarPerVector_NWaveNPerXdl == 0))
+        {
+            return false;
+        }
+
+        // Gridwise GEMM size
+        return GridwiseGemm::CheckValidity(arg.a_grid_desc_kbatch_k0_m_k1_,
+                                           arg.b_grid_desc_kbatch_k0_n_k1_,
+                                           arg.c_grid_desc_m_n_,
+                                           arg.M01_,
+                                           arg.N01_);
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        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,
+                             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 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,
+                        N,
+                        K,
+                        C,
+                        input_spatial_lengths,
+                        filter_spatial_lengths,
+                        output_spatial_lengths,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op,
+                        split_k};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in_grid,
+                        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 in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op,
+                        ck::index_t split_k) override
+    {
+        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),
+                                          N,
+                                          K,
+                                          C,
+                                          input_spatial_lengths,
+                                          filter_spatial_lengths,
+                                          output_spatial_lengths,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op,
+                                          split_k);
+    }
+
+    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 << "DeviceConv2dWrWXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << K0PerBlock
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

include/ck/tensor_operation/gpu/device/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp

+#ifndef DEVICE_CONV2D_BWD_DATA_XDL_NHWC_KYXC_NHWK_HPP
+#define DEVICE_CONV2D_BWD_DATA_XDL_NHWC_KYXC_NHWK_HPP
+
+#include <iostream>
+#include <sstream>
+#include "device.hpp"
+#include "device_base.hpp"
+#include "device_conv_bwd_data.hpp"
+#include "convolution_backward_data_specialization.hpp"
+#include "common_header.hpp"
+#include "tensor_layout.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+#include "gridwise_gemm_xdlops_v2r3.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_t 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<InElementwiseOperation,
+                               WeiElementwiseOperation,
+                               OutElementwiseOperation>
+{
+    using DeviceOp = DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K;
+
+    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_ytilda,
+                                                    index_t i_xtilda)
+    {
+        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_t::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 YTilda = ConvStrideH / GcdStrideDilationH;
+            const auto XTilda = ConvStrideW / GcdStrideDilationW;
+
+            const auto YDot = math::integer_divide_ceil(Y, YTilda);
+            const auto XDot = math::integer_divide_ceil(X, XTilda);
+
+            const auto HTilda =
+                Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
+            const auto WTilda =
+                Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
+
+            // only work on HTilda and WTilda that contribute to non-padding area of input tensor
+            const auto IHTildaSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadH - ConvDilationH * (YTilda - I1)), ConvStrideH);
+            const auto IWTildaSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadW - ConvDilationW * (XTilda - I1)), ConvStrideW);
+
+            const auto IHTildaSliceEnd = math::min(
+                HTilda, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
+            const auto IWTildaSliceEnd = math::min(
+                WTilda, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
+
+            const auto HTildaSlice = IHTildaSliceEnd - IHTildaSliceBegin;
+            const auto WTildaSlice = IWTildaSliceEnd - IWTildaSliceBegin;
+
+            // GemmK is different for each GEMM
+            const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilda, YTilda);
+            const auto XDotSlice = math::integer_divide_ceil(X - i_xtilda, XTilda);
+
+            // 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_htilda_xdot_wtilda_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, HTilda),
+                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                    make_embed_transform(make_tuple(XDot, WTilda),
+                                         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_htildaslice_xdotslice_wtildaslice_k0_k1_grid_desc =
+                transform_tensor_descriptor(
+                    out_n_ydot_htilda_xdot_wtilda_k_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(HTilda, IHTildaSliceBegin, HTildaSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_slice_transform(WTilda, IWTildaSliceBegin, WTildaSlice),
+                               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_htildaslice_xdotslice_wtildaslice_k0_k1_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K0)),
+                           make_merge_transform(make_tuple(N, HTildaSlice, WTildaSlice)),
+                           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_ytilda_xdot_xtilda_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, YTilda),
+                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                           make_embed_transform(make_tuple(XDot, XTilda),
+                                                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_ytilda_xdot_xtilda_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_ytilda),
+                                                       make_freeze_transform(i_xtilda),
+                                                       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_ytilda_htilda_xtilda_wtilda_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(YTilda, HTilda),
+                                                make_tuple(ConvDilationH, ConvStrideH)),
+                           make_embed_transform(make_tuple(XTilda, WTilda),
+                                                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_htildaslice_wtildaslice_c_grid_desc = transform_tensor_descriptor(
+                in_n_ytilda_htilda_xtilda_wtilda_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_freeze_transform(i_ytilda),
+                           make_slice_transform(HTilda, IHTildaSliceBegin, HTildaSlice),
+                           make_freeze_transform(i_xtilda),
+                           make_slice_transform(WTilda, IWTildaSliceBegin, WTildaSlice),
+                           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_htildaslice_wtildaslice_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, HTildaSlice, WTildaSlice)),
+                           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
+    using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
+        BlockSize,
+        ABDataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CDataType,
+        InMemoryDataOperationEnum_t::Set,
+        AGridDesc_K0_M_K1,
+        BGridDesc_K0_N_K1,
+        CGridDesc_M_N,
+        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>;
+
+    // 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,
+                 ck::index_t M01,
+                 ck::index_t N01,
+                 InElementwiseOperation in_element_op,
+                 WeiElementwiseOperation wei_element_op,
+                 OutElementwiseOperation out_element_op)
+            : p_a_grid_{p_out_grid},
+              p_b_grid_{p_wei_grid},
+              p_c_grid_{p_in_grid},
+              M01_{M01},
+              N01_{N01},
+              a_element_op_{out_element_op},
+              b_element_op_{wei_element_op},
+              c_element_op_{in_element_op},
+              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 YTilda = ConvStrideH / GcdStrideDilationH;
+            const auto XTilda = ConvStrideW / GcdStrideDilationW;
+
+            for(index_t i_ytilda = 0; i_ytilda < YTilda; ++i_ytilda)
+            {
+                for(index_t i_xtilda = 0; i_xtilda < XTilda; ++i_xtilda)
+                {
+                    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_ytilda,
+                        i_xtilda);
+                    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]);
+
+                    if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2], M01_, N01_))
+                    {
+                        c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
+                            GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(descs[I2]));
+
+                        block_2_ctile_map_container_.push_back(
+                            GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01, N01));
+                    }
+                }
+            }
+        }
+
+        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_;
+        std::vector<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>
+            c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_;
+        std::vector<typename GridwiseGemm::DefaultBlock2CTileMap> block_2_ctile_map_container_;
+        index_t M01_;
+        index_t N01_;
+        OutElementwiseOperation a_element_op_;
+        WeiElementwiseOperation b_element_op_;
+        InElementwiseOperation c_element_op_;
+        // 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;
+
+        float Run(const Argument& arg, int nrepeat = 1)
+        {
+            nrepeat        = 1;
+            float ave_time = 0;
+            for(size_t i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+            {
+                {
+                    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;
+
+                    std::cout << "arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I0)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I1)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I2)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I3)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I4)
+                              << ", "
+                              << arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I5)
+                              << " ) " << 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],
+                                                arg.M01_,
+                                                arg.N01_))
+                {
+                    throw std::runtime_error(
+                        "wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
+                }
+
+                const index_t grid_size =
+                    GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_container_[i]);
+
+                const auto K0 = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0);
+
+                const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
+
+                if(has_main_k0_block_loop)
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r3<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<
+                            typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                        OutElementwiseOperation,
+                        WeiElementwiseOperation,
+                        InElementwiseOperation,
+                        remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                        true>;
+
+                    ave_time += launch_and_time_kernel(
+                        kernel,
+                        nrepeat,
+                        dim3(grid_size),
+                        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_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.c_element_op_,
+                        arg.block_2_ctile_map_container_[i]);
+                }
+                else
+                {
+                    const auto kernel = kernel_gemm_xdlops_v2r3<
+                        GridwiseGemm,
+                        ADataType, // TODO: distiguish A/B datatype
+                        CDataType,
+                        remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
+                        remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
+                        remove_reference_t<
+                            typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
+                        OutElementwiseOperation,
+                        WeiElementwiseOperation,
+                        InElementwiseOperation,
+                        remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
+                        false>;
+
+                    ave_time += launch_and_time_kernel(
+                        kernel,
+                        nrepeat,
+                        dim3(grid_size),
+                        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_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.c_element_op_,
+                        arg.block_2_ctile_map_container_[i]);
+                }
+            }
+            return ave_time;
+        }
+
+        float Run(const BaseArgument* p_arg, int nrepeat = 1) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), nrepeat);
+        }
+    };
+
+    static constexpr bool IsValidCompilationParameter()
+    {
+        // TODO: properly implement this check
+        return true;
+    }
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        if constexpr(ConvBackwardDataSpecialization ==
+                     ConvolutionBackwardDataSpecialization_t::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
+        for(int i = 0; i < arg.a_grid_desc_k0_m_k1_container_.size(); i++)
+        {
+            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],
+                                            arg.M01_,
+                                            arg.N01_))
+            {
+                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,
+                             InElementwiseOperation in_element_op,
+                             WeiElementwiseOperation wei_element_op,
+                             OutElementwiseOperation out_element_op)
+    {
+        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,
+                        1,
+                        1,
+                        in_element_op,
+                        wei_element_op,
+                        out_element_op};
+    }
+
+    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 in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) 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,
+                                          1,
+                                          1,
+                                          in_element_op,
+                                          wei_element_op,
+                                          out_element_op);
+    }
+
+    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
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

device_operation/include/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp → include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_bias_activation_add_nhwc_kyxc_nhwk.hpp

@@ -549,8 +549,11 @@ struct
               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}
         {
@@ -625,8 +628,11 @@ struct
         index_t Conv_N_;
         index_t Conv_K_;
         index_t Conv_C_;
+        std::vector<index_t> input_spatial_lengths_;
         std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
         std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> conv_filter_dilations_;
         std::vector<index_t> input_left_pads_;
         std::vector<index_t> input_right_pads_;
     };
@@ -638,6 +644,28 @@ struct
 
         float Run(const Argument& arg, int nrepeat = 1)
         {
+#if 0
+            {
+                std::cout << DeviceOp{}.GetTypeString() << std::endl;
+                std::cout << "N " << arg.Conv_N_ << ", "
+                          << "K " << arg.Conv_K_ << ", "
+                          << "C " << arg.Conv_C_ << ", " << std::endl;
+                std::cout << "Y X " << arg.filter_spatial_lengths_[0] << ", "
+                          << arg.filter_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Hi Wi " << arg.input_spatial_lengths_[0] << ", "
+                          << arg.input_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Ho Wo " << arg.output_spatial_lengths_[0] << ", "
+                          << arg.output_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Strides " << arg.conv_filter_strides_[0] << ", "
+                          << arg.conv_filter_strides_[1] << ", " << std::endl;
+                std::cout << "Dilations " << arg.conv_filter_dilations_[0] << ", "
+                          << arg.conv_filter_dilations_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_left_pads_[0] << ", "
+                          << arg.input_left_pads_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_right_pads_[0] << ", "
+                          << arg.input_right_pads_[1] << ", " << std::endl;
+            }
+
             {
                 std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
                           << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
@@ -656,6 +684,7 @@ struct
                 std::cout << "arg.c1_grid_desc_m_n_{ " << arg.c1_grid_desc_m_n_.GetLength(I0)
                           << ", " << arg.c1_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
             }
+#endif
 
             if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
                                             arg.b_grid_desc_k0_n_k1_,

device_operation/include/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp → include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp

@@ -526,8 +526,11 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
               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}
         {
@@ -590,8 +593,11 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
         index_t Conv_N_;
         index_t Conv_K_;
         index_t Conv_C_;
+        std::vector<index_t> input_spatial_lengths_;
         std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
         std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> conv_filter_dilations_;
         std::vector<index_t> input_left_pads_;
         std::vector<index_t> input_right_pads_;
     };
@@ -603,6 +609,28 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
 
         float Run(const Argument& arg, int nrepeat = 1)
         {
+#if 0
+            {
+                std::cout << DeviceOp{}.GetTypeString() << std::endl;
+                std::cout << "N " << arg.Conv_N_ << ", "
+                          << "K " << arg.Conv_K_ << ", "
+                          << "C " << arg.Conv_C_ << ", " << std::endl;
+                std::cout << "Y X " << arg.filter_spatial_lengths_[0] << ", "
+                          << arg.filter_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Hi Wi " << arg.input_spatial_lengths_[0] << ", "
+                          << arg.input_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Ho Wo " << arg.output_spatial_lengths_[0] << ", "
+                          << arg.output_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Strides " << arg.conv_filter_strides_[0] << ", "
+                          << arg.conv_filter_strides_[1] << ", " << std::endl;
+                std::cout << "Dilations " << arg.conv_filter_dilations_[0] << ", "
+                          << arg.conv_filter_dilations_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_left_pads_[0] << ", "
+                          << arg.input_left_pads_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_right_pads_[0] << ", "
+                          << arg.input_right_pads_[1] << ", " << std::endl;
+            }
+
             {
                 std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
                           << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
@@ -618,6 +646,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
                 std::cout << "arg.c0_grid_desc_m_n_{ " << arg.c0_grid_desc_m_n_.GetLength(I0)
                           << ", " << arg.c0_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
             }
+#endif
 
             if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
                                             arg.b_grid_desc_k0_n_k1_,

device_operation/include/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp → include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp

@@ -422,6 +422,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
         BlockSize,
         ABDataType, // TODO: distinguish A/B datatype
         AccDataType,
+        CDataType, // TODO: Add ShuffleType for DeviceConv2d
         CDataType,
         InMemoryDataOperationEnum_t::Set,
         AGridDesc_K0_M_K1,
@@ -497,8 +498,11 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
               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}
         {
@@ -550,8 +554,11 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
         index_t Conv_N_;
         index_t Conv_K_;
         index_t Conv_C_;
+        std::vector<index_t> input_spatial_lengths_;
         std::vector<index_t> filter_spatial_lengths_;
+        std::vector<index_t> output_spatial_lengths_;
         std::vector<index_t> conv_filter_strides_;
+        std::vector<index_t> conv_filter_dilations_;
         std::vector<index_t> input_left_pads_;
         std::vector<index_t> input_right_pads_;
     };
@@ -563,6 +570,28 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
 
         float Run(const Argument& arg, int nrepeat = 1)
         {
+#if 0
+            {
+                std::cout << DeviceOp{}.GetTypeString() << std::endl;
+                std::cout << "N " << arg.Conv_N_ << ", "
+                          << "K " << arg.Conv_K_ << ", "
+                          << "C " << arg.Conv_C_ << ", " << std::endl;
+                std::cout << "Y X " << arg.filter_spatial_lengths_[0] << ", "
+                          << arg.filter_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Hi Wi " << arg.input_spatial_lengths_[0] << ", "
+                          << arg.input_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Ho Wo " << arg.output_spatial_lengths_[0] << ", "
+                          << arg.output_spatial_lengths_[1] << ", " << std::endl;
+                std::cout << "Strides " << arg.conv_filter_strides_[0] << ", "
+                          << arg.conv_filter_strides_[1] << ", " << std::endl;
+                std::cout << "Dilations " << arg.conv_filter_dilations_[0] << ", "
+                          << arg.conv_filter_dilations_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_left_pads_[0] << ", "
+                          << arg.input_left_pads_[1] << ", " << std::endl;
+                std::cout << "InLeftPads " << arg.input_right_pads_[0] << ", "
+                          << arg.input_right_pads_[1] << ", " << std::endl;
+            }
+
             {
                 std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
                           << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
@@ -597,6 +626,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
                            .GetLength(I5)
                     << "}" << std::endl;
             }
+#endif
 
             if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
                                             arg.b_grid_desc_k0_n_k1_,

device_operation/include/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp → include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp

@@ -452,6 +452,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
 
         float Run(const Argument& arg, int nrepeat = 1)
         {
+#if 0
             {
                 std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
                           << ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
@@ -464,6 +465,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
                 std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
                           << arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
             }
+#endif
 
             if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
                                             arg.b_grid_desc_k0_n_k1_,

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

+#ifndef DEVICE_CONV_WRW_HPP
+#define DEVICE_CONV_WRW_HPP
+
+#include <iostream>
+#include "device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvWrw : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(const void* p_in,
+                        void* p_wei,
+                        const void* p_out,
+                        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 in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op,
+                        ck::index_t split_k) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+using DeviceConvWrwPtr = std::unique_ptr<
+    DeviceConvWrw<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif

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

+#ifndef DEVICE_CONV_BWD_DATA_HPP
+#define DEVICE_CONV_BWD_DATA_HPP
+
+#include <iostream>
+#include "device_base.hpp"
+#include "element_wise_operation.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+struct DeviceConvBwdData : public BaseOperator
+{
+    virtual std::unique_ptr<BaseArgument>
+    MakeArgumentPointer(void* p_in,
+                        const void* p_wei,
+                        const void* p_out,
+                        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 in_element_op,
+                        WeiElementwiseOperation wei_element_op,
+                        OutElementwiseOperation out_element_op) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+template <typename InElementwiseOperation,
+          typename WeiElementwiseOperation,
+          typename OutElementwiseOperation>
+using DeviceConvBwdDataPtr = std::unique_ptr<
+    DeviceConvBwdData<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>>;
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck
+#endif