Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel...

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel into barkocot/grouped-conv-weight-fp16-c1-k1

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_dpp.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+namespace ck {
+
+template <typename GridwiseGemm, bool HasMainKBlockLoop>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+#if CK_USE_WAVES_PER_EU
+        __attribute__((amdgpu_waves_per_eu(CK_MIN_WAVES_PER_EU, CK_MAX_WAVES_PER_EU)))
+#endif
+        kernel_gemm_dpp(const typename GridwiseGemm::Argument karg)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1030__))
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    const auto a_grid_desc_ak0_m_ak1 = amd_wave_read_first_lane(
+        GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(karg.M, karg.K, karg.AK0, karg.StrideA));
+    const auto b_grid_desc_bk0_n_bk1 = amd_wave_read_first_lane(
+        GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(karg.K, karg.N, karg.BK0, karg.StrideB));
+    const auto c_grid_desc_m_n = amd_wave_read_first_lane(
+        GridwiseGemm::MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC));
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(karg.p_a_grid,
+                                                  karg.p_b_grid,
+                                                  karg.p_c_grid,
+                                                  p_shared,
+                                                  a_grid_desc_ak0_m_ak1,
+                                                  b_grid_desc_bk0_n_bk1,
+                                                  c_grid_desc_m_n);
+#else
+    ignore = karg;
+#endif
+}
+
+template <index_t BlockSize,
+          typename ABDataType,
+          typename AccDataType,
+          typename CDataType,
+          InMemoryDataOperationEnum CGlobalMemoryDataOperation,
+          typename ALayout,
+          typename BLayout,
+          typename CLayout,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CElementwiseOperation,
+          tensor_operation::device::GemmSpecialization GemmSpec,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t MPerDpp,
+          index_t NPerDpp,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MDppPerWave,
+          index_t NDppPerWave,
+          typename ABlockTransferThreadClusterLengths_K0_M_K1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_K1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          bool ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_K0_N_K1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_K1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          bool BBlockLdsExtraN,
+          typename CThreadTransferSrcDstAccessOrder,
+          index_t CThreadTransferSrcDstVectorDim,
+          index_t CThreadTransferDstScalarPerVector,
+          index_t NumGemmKPrefetchStage = 1,
+          PipelineVersion PipelineVer   = PipelineVersion::v1>
+struct GridwiseGemm_ak0mak1_bk0nbk1_mn_dpp
+{
+    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 AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    static constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+    // return block_id to C matrix tile idx (m0, n0) mapping
+    using Block2CTileMap = BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock>;
+
+    __host__ static auto CalculateGridSize(index_t M, index_t N)
+    {
+        return std::make_tuple(Block2CTileMap::CalculateGridSize(M, N), 1, 1);
+    }
+
+    __host__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_divide_ceil(M, MPerBlock) * MPerBlock;
+    }
+
+    __host__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_divide_ceil(N, NPerBlock) * NPerBlock;
+    }
+
+    __host__ static auto CalculateAK0(index_t K) { return math::integer_divide_floor(K, AK1Value); }
+    __host__ static auto CalculateBK0(index_t K) { return math::integer_divide_floor(K, BK1Value); }
+
+    // Argument
+    struct Problem
+    {
+        __host__ Problem(index_t M_,
+                         index_t N_,
+                         index_t K_,
+                         index_t StrideA_,
+                         index_t StrideB_,
+                         index_t StrideC_)
+            : M{M_},
+              N{N_},
+              K{K_},
+              StrideA{StrideA_},
+              StrideB{StrideB_},
+              StrideC{StrideC_},
+              MPadded{CalculateMPadded(M_)},
+              NPadded{CalculateNPadded(N_)},
+              AK0{CalculateAK0(K)},
+              BK0{CalculateBK0(K)}
+        {
+        }
+
+        __host__ void Print() const
+        {
+            std::cout << "problem {"
+                      << "M:" << M << ", "
+                      << "N:" << N << ", "
+                      << "K:" << K << ", "
+                      << "SA:" << StrideA << ", "
+                      << "SB:" << StrideB << ", "
+                      << "SC:" << StrideC << ", "
+                      << "MP:" << MPadded << ", "
+                      << "NP:" << NPadded << ", "
+                      << "AK0:" << AK0 << ", "
+                      << "BK0:" << BK0 << "}" << std::endl;
+        }
+
+        index_t M;
+        index_t N;
+        index_t K;
+        index_t StrideA;
+        index_t StrideB;
+        index_t StrideC;
+        index_t MPadded;
+        index_t NPadded;
+        index_t AK0;
+        index_t BK0;
+    };
+
+    // Argument
+    struct Argument : public Problem, public tensor_operation::device::BaseArgument
+    {
+        __host__ Argument(const ABDataType* p_a_grid_,
+                          const ABDataType* p_b_grid_,
+                          CDataType* p_c_grid_,
+                          index_t M_,
+                          index_t N_,
+                          index_t K_,
+                          index_t StrideA_,
+                          index_t StrideB_,
+                          index_t StrideC_)
+            : Problem{M_, N_, K_, StrideA_, StrideB_, StrideC_},
+              p_a_grid{p_a_grid_},
+              p_b_grid{p_b_grid_},
+              p_c_grid{p_c_grid_}
+        {
+        }
+
+        const ABDataType* p_a_grid;
+        const ABDataType* p_b_grid;
+        CDataType* p_c_grid;
+    };
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = [&]() {
+            if constexpr(ABlockLdsExtraM)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<AK0PerBlock>{}, Number<MPerBlock>{}, AK1),
+                    make_tuple(Number<MPerBlock + 1>{} * AK1, AK1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<AK0PerBlock>{}, Number<MPerBlock>{}, AK1), max_lds_align);
+            }
+        }();
+
+        return a_block_desc_ak0_m_ak1;
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = [&]() {
+            if constexpr(BBlockLdsExtraN)
+            {
+                return make_naive_tensor_descriptor(
+                    make_tuple(Number<BK0PerBlock>{}, Number<NPerBlock>{}, BK1),
+                    make_tuple(Number<NPerBlock + 1>{} * BK1, BK1, I1));
+            }
+            else
+            {
+                return make_naive_tensor_descriptor_aligned(
+                    make_tuple(Number<BK0PerBlock>{}, Number<NPerBlock>{}, BK1), max_lds_align);
+            }
+        }();
+
+        return b_block_desc_bk0_n_bk1;
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        return (a_block_space_size_aligned + b_block_space_size_aligned) * sizeof(ABDataType);
+    }
+
+    __host__ static constexpr bool CheckValidity(const Problem& problem)
+    {
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(AK1)>>::value,
+                      "Wrong! AK1 must be known at the time of compilation.");
+        static_assert(is_known_at_compile_time<remove_cv_t<decltype(BK1)>>::value,
+                      "Wrong! BK1 must be known at the time of compilation.");
+
+        static_assert(
+            MPerBlock % (MPerDpp * MDppPerWave) == 0,
+            "Invalid tuning parameters! MPerBlock must be divisible by MPerDpp * MDppPerWave.");
+        static_assert(
+            NPerBlock % (NPerDpp * NDppPerWave) == 0,
+            "Invalid tuning parameters! NPerBlock must be divisible by NPerDpp * NDppPerWave.");
+
+        static_assert(
+            KPerBlock % AK1Value == 0 && KPerBlock % BK1Value == 0,
+            "Invalid tuning parameters! KPerBlock must be divisible by both AK1 and BK1.");
+
+        static_assert(AK1Value % ABlockTransferDstScalarPerVector_K1 == 0,
+                      "Invalid tuning parameters! AK1Value must be divisible by "
+                      "ABlockTransferDstScalarPerVector_K1");
+
+        static_assert(BK1Value % BBlockTransferDstScalarPerVector_K1 == 0,
+                      "Invalid tuning parameters! BK1Value must be divisible by "
+                      "BBlockTransferDstScalarPerVector_K1");
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(problem.M % MPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                       GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
+        {
+            if(!(problem.N % NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+        {
+            if(problem.K % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.M % ABlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+        {
+            if(problem.N % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+        else
+        {
+            if(problem.K % BBlockTransferSrcScalarPerVector != 0)
+            {
+                return false;
+            }
+        }
+
+        if(problem.K % KPerBlock != 0)
+        {
+            return false;
+        }
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = problem.K / KPerBlock;
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const auto num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    template <typename CGridDesc>
+    __host__ __device__ static constexpr auto
+    MakeCGridDescriptor_M0_N0_M1_N1_M2_N2(const CGridDesc& c_grid_desc_m_n)
+    {
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), DppSelector<ABDataType, MPerDpp, NPerDpp>::selected_dpp.k_per_dpp);
+
+        using BlockwiseGemm =
+            BlockwiseGemmDpp_ak0mak1_bk0nbk1_m0n0m1n1m2n2<BlockSize,
+                                                          ABDataType,
+                                                          AccDataType,
+                                                          decltype(a_block_desc_ak0_m_ak1),
+                                                          decltype(b_block_desc_bk0_n_bk1),
+                                                          MPerDpp,
+                                                          NPerDpp,
+                                                          MDppPerWave,
+                                                          NDppPerWave,
+                                                          KPack>;
+
+        return BlockwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_N2(c_grid_desc_m_n);
+    }
+
+    static constexpr auto matrix_padder =
+        ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+            MPerBlock, NPerBlock, KPerBlock};
+
+    __device__ static auto
+    MakeAGridDescriptor_AK0_M_AK1(index_t M, index_t K, index_t AK0, index_t StrideA)
+    {
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto a_grid_desc_m_k = matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+        return transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1Value)),
+                       make_pass_through_transform(M)),
+            make_tuple(Sequence<1>{}, Sequence<0>{}),
+            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    __device__ static auto
+    MakeBGridDescriptor_BK0_N_BK1(index_t K, index_t N, index_t BK0, index_t StrideB)
+    {
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(N, K), make_tuple(StrideB, I1));
+            }
+        }();
+
+        const auto b_grid_desc_n_k = matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+        return transform_tensor_descriptor(
+            b_grid_desc_n_k,
+            make_tuple(make_pass_through_transform(N),
+                       make_unmerge_transform(make_tuple(BK0, BK1Value))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+    }
+
+    __device__ static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
+    {
+        const auto c_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(c_grid_desc_mraw_nraw);
+    }
+
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename CGridDesc_M_N>
+    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
+                               const ABDataType* __restrict__ p_b_grid,
+                               CDataType* __restrict__ p_c_grid,
+                               void* __restrict__ p_shared,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const CGridDesc_M_N& c_grid_desc_m_n)
+    {
+        const auto c_grid_desc_m0_n0_m1_n1_m2_n2 =
+            MakeCGridDescriptor_M0_N0_M1_N1_M2_N2(c_grid_desc_m_n);
+
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
+        auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_c_grid, c_grid_desc_m0_n0_m1_n1_m2_n2.GetElementSpaceSize());
+
+        const AElementwiseOperation a_element_op{};
+        const BElementwiseOperation b_element_op{};
+        const CElementwiseOperation c_element_op{};
+
+        const auto block_2_ctile_map =
+            Block2CTileMap{c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1)};
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        if(!block_2_ctile_map.ValidCTileIndex(
+               block_work_idx,
+               make_tuple(c_grid_desc_m0_n0_m1_n1_m2_n2.GetLength(I0),
+                          c_grid_desc_m0_n0_m1_n1_m2_n2.GetLength(I1))))
+        {
+            return;
+        }
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_K0_M_K1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(a_grid_desc_ak0_m_ak1),
+                                                decltype(a_block_desc_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                ABlockTransferSrcVectorDim,
+                                                2,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_ak0_m_ak1,
+                make_multi_index(0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_ak0_m_ak1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_K0_N_K1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataType,
+                                                decltype(b_grid_desc_bk0_n_bk1),
+                                                decltype(b_block_desc_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<1, 0, 2>,
+                                                BBlockTransferSrcVectorDim,
+                                                2,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_K1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_bk0_n_bk1,
+                make_multi_index(0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_bk0_n_bk1,
+                make_multi_index(0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[AK0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[BK0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        constexpr index_t KPack = math::max(
+            math::lcm(AK1, BK1), DppSelector<ABDataType, MPerDpp, NPerDpp>::selected_dpp.k_per_dpp);
+        auto blockwise_gemm =
+            BlockwiseGemmDpp_ak0mak1_bk0nbk1_m0n0m1n1m2n2<BlockSize,
+                                                          ABDataType,
+                                                          AccDataType,
+                                                          decltype(a_block_desc_ak0_m_ak1),
+                                                          decltype(b_block_desc_bk0_n_bk1),
+                                                          MPerDpp,
+                                                          NPerDpp,
+                                                          MDppPerWave,
+                                                          NDppPerWave,
+                                                          KPack>();
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(AK0PerBlock, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(BK0PerBlock, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto AK0 = a_grid_desc_ak0_m_ak1.GetLength(I0);
+        // (AK0 / AK0PerBlock) is always equal to (BK0 / BK0PerBlock)
+        const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(AK0 / AK0PerBlock);
+
+        GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
+                                                          a_block_desc_ak0_m_ak1,
+                                                          a_blockwise_copy,
+                                                          a_grid_buf,
+                                                          a_block_buf,
+                                                          a_block_slice_copy_step,
+                                                          b_grid_desc_bk0_n_bk1,
+                                                          b_block_desc_bk0_n_bk1,
+                                                          b_blockwise_copy,
+                                                          b_grid_buf,
+                                                          b_block_buf,
+                                                          b_block_slice_copy_step,
+                                                          blockwise_gemm,
+                                                          c_thread_buf,
+                                                          num_k_block_main_loop);
+
+        // output: register to global memory
+        {
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2();
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2 =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2.GetLength(I4);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2.GetLength(I5);
+
+            constexpr auto MPerThread = c_thread_desc_m0_n0_m1_n1_m2_n2.GetLength(I4);
+            constexpr auto NPerThread = c_thread_desc_m0_n0_m1_n1_m2_n2.GetLength(I5);
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0);
+
+            const index_t m_thread_data_on_grid =
+                m_block_data_idx_on_grid + c_thread_mtx_on_block[I0];
+
+            const index_t n_thread_data_on_grid =
+                n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_grid_to_m0_m1_m2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_grid_idx =
+                m_thread_data_on_grid_to_m0_m1_m2_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_grid));
+
+            const auto n_thread_data_on_grid_to_n0_n1_n2_adaptor = make_single_stage_tensor_adaptor(
+                make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                make_tuple(Sequence<0, 1, 2>{}),
+                make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_grid_idx =
+                n_thread_data_on_grid_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_grid));
+
+            auto c_thread_copy =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_n2),
+                                                   decltype(c_grid_desc_m0_n0_m1_n1_m2_n2),
+                                                   CElementwiseOperation,
+                                                   Sequence<M0, N0, I1, I1, MPerThread, NPerThread>,
+                                                   CThreadTransferSrcDstAccessOrder,
+                                                   CThreadTransferSrcDstVectorDim,
+                                                   CThreadTransferDstScalarPerVector,
+                                                   CGlobalMemoryDataOperation,
+                                                   1,
+                                                   true>{
+                    c_grid_desc_m0_n0_m1_n1_m2_n2,
+                    make_multi_index(m_thread_data_on_grid_idx[I0],
+                                     n_thread_data_on_grid_idx[I0],
+                                     m_thread_data_on_grid_idx[I1],
+                                     n_thread_data_on_grid_idx[I1],
+                                     m_thread_data_on_grid_idx[I2],
+                                     n_thread_data_on_grid_idx[I2]),
+                    c_element_op};
+
+            c_thread_copy.Run(c_thread_desc_m0_n0_m1_n1_m2_n2,
+                              make_tuple(I0, I0, I0, I0, I0, I0),
+                              c_thread_buf,
+                              c_grid_desc_m0_n0_m1_n1_m2_n2,
+                              c_grid_buf);
+        }
+    }
+};
+
+} // namespace ck

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

@@ -15,6 +15,9 @@
 #include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+
 namespace ck {
 
 // GEMM:
@@ -74,6 +77,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
 {
     static constexpr index_t NumDTensor = DsDataType::Size();
 
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
@@ -263,6 +268,8 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
         static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
                           (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
                       "Invalid tuning param!");
+        static_assert(KPerBlock % AK1Value == 0 && KPerBlock % BK1Value == 0,
+                      "KPerBlock must be divisible by AK1Value and BK1Value!");
 
         const auto M  = a_grid_desc_m_k.GetLength(I0);
         const auto N  = b_grid_desc_n_k.GetLength(I0);
@@ -330,6 +337,94 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
 
     using DsGridPointer = decltype(MakeDsGridPointer());
 
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+
+        const auto a_grid_desc_mraw_kraw = [&]() {
+            if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, KRaw),
+                                                    make_tuple(I1, StrideA));
+            }
+        }();
+
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+
+        const auto b_grid_desc_nraw_kraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(I1, StrideB));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(NRaw, KRaw),
+                                                    make_tuple(StrideB, I1));
+            }
+        }();
+
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
+    }
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
     template <bool HasMainKBlockLoop,
               typename AGridDesc_AK0_M_AK1,
               typename BGridDesc_BK0_N_BK1,
@@ -758,6 +853,85 @@ struct GridwiseGemmMultipleD_xdl_cshuffle
             });
         }
     }
+
+    template <bool HasMainKBlockLoop,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ADataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const BDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        // tensor descriptors for problem definiton
+        const auto a_grid_desc_m_k = MakeAGridDescriptor_M_K<ALayout, GemmSpec>(M, K, StrideA);
+        const auto b_grid_desc_n_k = MakeBGridDescriptor_N_K<BLayout, GemmSpec>(K, N, StrideB);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_ak0_m_ak1 = MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+
+        const auto b_grid_desc_bk0_n_bk1 = MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        Run<HasMainKBlockLoop>(p_a_grid,
+                               p_b_grid,
+                               p_ds_grid,
+                               p_e_grid,
+                               p_shared,
+                               a_element_op,
+                               b_element_op,
+                               cde_element_op,
+                               a_grid_desc_ak0_m_ak1,
+                               b_grid_desc_bk0_n_bk1,
+                               ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               block_2_etile_map);
+    }
 };
 
 } // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/multi_index_transform_helper.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
+#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
+#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+
+namespace ck {
+
+// GEMM:
+//   input : A[M, K]
+//   input : B[N, K]
+//   input : D0[M, N], D1[M, N], ...
+//   output : E[M, N]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+// Assume:
+//   D0, D1, ... and E have the same layout
+template <typename ABDataType, // FIXME: don't assume A/B have same datatype
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1Value,
+          index_t BK1Value,
+          index_t MPerXdl,
+          index_t NPerXdl,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          bool AThreadTransferSrcResetCoordinateAfterRun,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          bool BThreadTransferSrcResetCoordinateAfterRun,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+          LoopScheduler LoopSched,
+          PipelineVersion PipelineVer = PipelineVersion::v1>
+struct GridwiseGemmMultipleD_xdl_splitk_cshuffle
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    using GemmSpecialization = ck::tensor_operation::device::GemmSpecialization;
+
+    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 I6 = Number<6>{};
+    static constexpr auto I7 = Number<7>{};
+
+    // K1 should be Number<...>
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
+
+    using ThisThreadBlock = ThisThreadBlock<BlockSize>;
+
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
+
+    // denorm test fix, required to work around fp16 mfma issue
+    // we convert fp16->fp32->bf16 and execute bf16 mfma instruction
+    // when mfma if fixed, remove this section and update
+    // ABDataTypeAdjusted -> ABDataType throughout this file
+#if CK_WORKAROUND_DENORM_FIX
+    using ABDataTypeAdjusted =
+        conditional_t<is_same_v<ABDataType, ck::half_t>, ck::bhalf_t, ABDataType>;
+#else
+    using ABDataTypeAdjusted = ABDataType;
+#endif
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
+                       AK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
+                       BK1,
+                       I1));
+    }
+
+    __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
+    {
+        // A matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
+            make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
+    {
+        // B matrix in LDS memory, dst of blockwise copy
+        return make_naive_tensor_descriptor(
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
+            make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
+    }
+
+    __host__ __device__ static constexpr auto
+    GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
+    {
+        constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+        constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            make_naive_tensor_descriptor_packed(
+                make_tuple(I1,
+                           Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
+                           I1,
+                           Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
+
+        return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // ck::Tuple<const D0DataType*, const D1DataType*, ...>
+    static constexpr auto MakeDsGridPointer()
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                return static_cast<const DDataType*>(nullptr);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
+    {
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
+
+        // LDS allocation for C shuffle in LDS
+        constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+            GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+        constexpr auto c_block_size =
+            c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
+
+        return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    __host__ __device__ static auto CalculateMPadded(index_t M)
+    {
+        return math::integer_least_multiple(M, MPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateNPadded(index_t N)
+    {
+        return math::integer_least_multiple(N, NPerBlock);
+    }
+
+    __host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch)
+    {
+        return math::integer_least_multiple(K, KPerBlock * K_Batch);
+    }
+
+    template <typename ALayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeAGridDescriptor_KBatch_AK0_M_AK1(index_t M, index_t K, index_t StrideA, index_t KBatch)
+    {
+        const auto a_grid_desc_m_k = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
+            }
+        }();
+
+        const auto MPad = CalculateMPadded(M);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+            a_grid_desc_m_k,
+            make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto AK0 = KPad / (KBatch * AK1);
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, AK0, AK1)),
+                           make_pass_through_transform(M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    template <typename BLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeBGridDescriptor_KBatch_BK0_N_BK1(index_t K, index_t N, index_t StrideB, index_t KBatch)
+    {
+        const auto b_grid_desc_k_n = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
+            }
+        }();
+
+        const auto NPad = CalculateNPadded(N);
+        const auto KPad = CalculateKPadded(K, KBatch);
+
+        const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+            b_grid_desc_k_n,
+            make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto BK0 = KPad / (KBatch * BK1);
+
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
+                     GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            // const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+        else
+        {
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(KBatch, BK0, BK1)),
+                           make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+        }
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const DsGridDesc_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    template <typename EGridDesc_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
+    }
+
+    template <typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              GemmSpecialization GemmSpec>
+    __host__ __device__ static constexpr bool
+    CheckValidity(const index_t M,
+                  const index_t N,
+                  const index_t K,
+                  const index_t StrideA,
+                  const index_t StrideB,
+                  const std::array<index_t, NumDTensor> StrideDs,
+                  const index_t StrideE,
+                  const index_t KBatch)
+    {
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        ignore = StrideDs;
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+#if 0
+        // check tile size
+        if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
+            return false;
+        }
+#endif
+
+        // check gridwise gemm pipeline
+        const auto num_k_loop = K / KPerBlock;
+
+        if(!GridwiseGemmPipe::IsSupported(num_k_loop))
+        {
+            return false;
+        }
+
+        // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
+        return true;
+    }
+
+    __host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
+    {
+        const index_t num_loop = K / KPerBlock;
+
+        return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
+    }
+
+    using DsGridPointer = decltype(MakeDsGridPointer());
+
+    template <typename ELayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
+    {
+        constexpr auto matrix_padder =
+            ck::tensor_operation::device::MatrixPadder<GemmSpec, index_t, index_t, index_t>{
+                MPerBlock, NPerBlock, KPerBlock};
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(StrideE, I1));
+            }
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELayout>::value)
+            {
+                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
+                                                    make_tuple(I1, StrideE));
+            }
+        }();
+
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
+
+    template <typename DsLayout, GemmSpecialization GemmSpec>
+    __host__ __device__ static auto
+    MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                             const std::array<index_t, NumDTensor>& NRaws,
+                             const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return MakeEGridDescriptor_M_N<DLayout, GemmSpec>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    __device__ __host__ static constexpr auto GetMPerBlock() { return MPerBlock; }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              index_t NumDTensor_,
+              typename DsDataType_,
+              typename AGridDesc_KBatch_AK0_M_AK1,
+              typename BGridDesc_KBatch_BK0_N_BK1,
+              typename DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+              typename CDEElementwiseOperation_,
+              typename Block2ETileMap>
+    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
+                               const ABDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const index_t KBatch,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation_& cde_element_op,
+                               const AGridDesc_KBatch_AK0_M_AK1& a_grid_desc_kbatch_ak0_m_ak1,
+                               const BGridDesc_KBatch_BK0_N_BK1& b_grid_desc_kbatch_bk0_n_bk1,
+                               const DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_a_grid, a_grid_desc_kbatch_ak0_m_ak1.GetElementSpaceSize());
+
+        const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_b_grid, b_grid_desc_kbatch_bk0_n_bk1.GetElementSpaceSize());
+
+        const auto ds_grid_buf = generate_tuple(
+            [&](auto i) {
+                return make_dynamic_buffer<AddressSpaceEnum::Global>(
+                    p_ds_grid[i],
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
+            },
+            Number<NumDTensor_>{});
+
+        auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
+            p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+        // divide block work by [M, N]
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        // HACK: this force m/n_block_data_idx_on_grid into SGPR
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        const index_t m_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
+
+        const index_t n_block_data_idx_on_grid =
+            __builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
+
+        // lds max alignment
+        constexpr auto max_lds_align = math::lcm(AK1, BK1);
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_kbatch_ak0_m_ak1 =
+            GetABlockDescriptor_KBatch_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_kbatch_bk0_n_bk1 =
+            GetBBlockDescriptor_KBatch_BK0PerBlock_NPerBlock_BK1();
+
+        // A matrix blockwise copy
+        auto a_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                AElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, AK0PerBlock, MPerBlock, AK1>,
+                                                ABlockTransferThreadClusterLengths_KBatch_AK0_M_AK1,
+                                                ABlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataTypeAdjusted,
+                                                decltype(a_grid_desc_kbatch_ak0_m_ak1),
+                                                decltype(a_block_desc_kbatch_ak0_m_ak1),
+                                                ABlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                ABlockTransferSrcVectorDim,
+                                                3,
+                                                ABlockTransferSrcScalarPerVector,
+                                                ABlockTransferDstScalarPerVector_AK1,
+                                                1,
+                                                1,
+                                                AThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                a_grid_desc_kbatch_ak0_m_ak1,
+                make_multi_index(kbatch_id, 0, m_block_data_idx_on_grid, 0),
+                a_element_op,
+                a_block_desc_kbatch_ak0_m_ak1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // B matrix blockwise copy
+        auto b_blockwise_copy =
+            ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
+                                                BElementwiseOperation,
+                                                ck::tensor_operation::element_wise::PassThrough,
+                                                InMemoryDataOperationEnum::Set,
+                                                Sequence<1, BK0PerBlock, NPerBlock, BK1>,
+                                                BBlockTransferThreadClusterLengths_KBatch_BK0_N_BK1,
+                                                BBlockTransferThreadClusterArrangeOrder,
+                                                ABDataType,
+                                                ABDataTypeAdjusted,
+                                                decltype(b_grid_desc_kbatch_bk0_n_bk1),
+                                                decltype(b_block_desc_kbatch_bk0_n_bk1),
+                                                BBlockTransferSrcAccessOrder,
+                                                Sequence<2, 0, 1, 3>,
+                                                BBlockTransferSrcVectorDim,
+                                                3,
+                                                BBlockTransferSrcScalarPerVector,
+                                                BBlockTransferDstScalarPerVector_BK1,
+                                                1,
+                                                1,
+                                                BThreadTransferSrcResetCoordinateAfterRun,
+                                                true,
+                                                NumGemmKPrefetchStage>(
+                b_grid_desc_kbatch_bk0_n_bk1,
+                make_multi_index(kbatch_id, 0, n_block_data_idx_on_grid, 0),
+                b_element_op,
+                b_block_desc_kbatch_bk0_n_bk1,
+                make_multi_index(0, 0, 0, 0),
+                ck::tensor_operation::element_wise::PassThrough{});
+
+        // A matrix in LDS memory, dst of blockwise copy
+        constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
+
+        // B matrix in LDS memory, dst of blockwise copy
+        constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
+
+        // GEMM definition
+        //   c_mtx += transpose(a_mtx) * b_mtx
+        //     a_mtx[K0PerBlock, MPerBlock] is in LDS
+        //     b_mtx[K0PerBlock, NPerBlock] is in LDS
+        //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
+        //       register
+        // sanity check
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ABDataTypeAdjusted, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+
+        auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
+            BlockSize,
+            ABDataTypeAdjusted,
+            AccDataType,
+            decltype(a_block_desc_ak0_m_ak1),
+            decltype(b_block_desc_bk0_n_bk1),
+            MPerXdl,
+            NPerXdl,
+            MXdlPerWave,
+            NXdlPerWave,
+            KPack,
+            LoopSched>();
+
+#if 1
+        if(block_work_idx[I0] == 0)
+        {
+            const index_t nThreadSize = CDEShuffleBlockTransferScalarPerVector_NPerBlock;
+            const index_t numNThreads = NPerBlock / nThreadSize;
+            const index_t numMThreads = BlockSize / numNThreads;
+            const index_t mThreadSize = MPerBlock / numMThreads;
+
+            const index_t m_tid = get_thread_local_1d_id() / numNThreads;
+            const index_t n_tid = get_thread_local_1d_id() % numNThreads;
+
+            auto c_thread_desc_mblock_mperblock_nblock_nperblock =
+                make_naive_tensor_descriptor_packed(
+                    make_tuple(I1, Number<mThreadSize>{}, I1, Number<nThreadSize>{}));
+
+            StaticBuffer<AddressSpaceEnum::Vgpr,
+                         EDataType,
+                         c_thread_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize(),
+                         true>
+                e_thread_zero_buf;
+
+            auto c_thread_copy = ThreadwiseTensorSliceTransfer_v1r3<
+                EDataType,
+                EDataType,
+                decltype(c_thread_desc_mblock_mperblock_nblock_nperblock),
+                decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                ck::tensor_operation::element_wise::PassThrough,
+                Sequence<1, mThreadSize, 1, nThreadSize>,
+                Sequence<0, 1, 2, 3>,
+                3,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                InMemoryDataOperationEnum::Set,
+                1,
+                true>{e_grid_desc_mblock_mperblock_nblock_nperblock,
+                      make_multi_index(block_work_idx[I1],
+                                       m_tid * mThreadSize,
+                                       block_work_idx[I2],
+                                       n_tid * nThreadSize),
+                      ck::tensor_operation::element_wise::PassThrough{}};
+
+            c_thread_copy.Run(c_thread_desc_mblock_mperblock_nblock_nperblock,
+                              make_tuple(I0, I0, I0, I0),
+                              e_thread_zero_buf,
+                              e_grid_desc_mblock_mperblock_nblock_nperblock,
+                              e_grid_buf);
+
+            __syncthreads();
+
+            if(threadIdx.x == 0)
+            {
+                atomicAdd(barrier_count_finished, 1);
+            }
+        }
+#endif
+
+        auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
+
+        // LDS allocation for A and B: be careful of alignment
+        constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
+
+        auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataTypeAdjusted*>(p_shared),
+            a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+
+        auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+            static_cast<ABDataTypeAdjusted*>(p_shared) + a_block_space_size_aligned,
+            b_block_desc_bk0_n_bk1.GetElementSpaceSize());
+
+        constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
+        constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
+
+        // gridwise GEMM pipeline
+        const auto gridwise_gemm_pipeline =
+            GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>();
+
+        const index_t num_k_block_main_loop =
+            __builtin_amdgcn_readfirstlane((a_grid_desc_kbatch_ak0_m_ak1.GetLength(I1) *
+                                            a_grid_desc_kbatch_ak0_m_ak1.GetLength(I3)) /
+                                           KPerBlock);
+
+        gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_kbatch_ak0_m_ak1,
+                                                               a_block_desc_kbatch_ak0_m_ak1,
+                                                               a_blockwise_copy,
+                                                               a_grid_buf,
+                                                               a_block_buf,
+                                                               a_block_slice_copy_step,
+                                                               b_grid_desc_kbatch_bk0_n_bk1,
+                                                               b_block_desc_kbatch_bk0_n_bk1,
+                                                               b_blockwise_copy,
+                                                               b_grid_buf,
+                                                               b_block_buf,
+                                                               b_block_slice_copy_step,
+                                                               blockwise_gemm,
+                                                               c_thread_buf,
+                                                               num_k_block_main_loop);
+
+        // shuffle C and write out
+        {
+            if(threadIdx.x == 0)
+            {
+                while(__atomic_load_n(barrier_count_finished, __ATOMIC_RELAXED) == 0) {}
+            }
+
+            __syncthreads();
+
+            static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
+                              NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
+                          "wrong!");
+
+            constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
+            constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
+
+            // TODO: hacky, fix it!
+            constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
+                blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            // TODO: hacky, fix it!
+            // c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
+                blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
+
+            constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
+            constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
+            constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
+            constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
+            constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
+            constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
+            constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
+            constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
+
+            constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
+                GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
+
+            auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
+                static_cast<CShuffleDataType*>(p_shared),
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
+
+            constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
+                c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
+                make_tuple(
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
+                        M1,                                      // M1 = MWave
+                        M2,                                      // M2 * M3 * M4 = MPerXdl
+                        M3,
+                        M4)),
+                    make_freeze_transform(I0),
+                    make_unmerge_transform(make_tuple(
+                        Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
+                        N1,                                      // N1 = NWave
+                        N2))),                                   // N2 = NPerXdl
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(
+                    Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
+
+            // calculate origin of thread output tensor on global memory
+            //     blockwise GEMM c matrix starting index
+            const auto c_thread_mtx_on_block =
+                blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
+
+            const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
+            const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
+
+            const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
+                    make_tuple(Sequence<0, 1, 2, 3, 4>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto m_thread_data_on_block_idx =
+                m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
+                    make_multi_index(m_thread_data_on_block));
+
+            const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
+                make_single_stage_tensor_adaptor(
+                    make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
+                    make_tuple(Sequence<0, 1, 2>{}),
+                    make_tuple(Sequence<0>{}));
+
+            const auto n_thread_data_on_block_idx =
+                n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
+                    make_multi_index(n_thread_data_on_block));
+
+            // shuffle: threadwise copy C from VGPR to LDS
+            auto c_thread_copy_vgpr_to_lds =
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
+                                                   decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
+                                                   ck::tensor_operation::element_wise::PassThrough,
+                                                   Sequence<CShuffleMXdlPerWavePerShuffle,
+                                                            CShuffleNXdlPerWavePerShuffle,
+                                                            I1,
+                                                            I1,
+                                                            M2,
+                                                            I1,
+                                                            M4,
+                                                            I1>,
+                                                   Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                                   7,
+                                                   1,
+                                                   InMemoryDataOperationEnum::Set,
+                                                   1,
+                                                   true>{
+                    c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                    make_multi_index(0,
+                                     0,
+                                     m_thread_data_on_block_idx[I1],
+                                     n_thread_data_on_block_idx[I1],
+                                     m_thread_data_on_block_idx[I2],
+                                     m_thread_data_on_block_idx[I3],
+                                     m_thread_data_on_block_idx[I4],
+                                     n_thread_data_on_block_idx[I2]),
+                    ck::tensor_operation::element_wise::PassThrough{}};
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_desc_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of reference to C/Ds tensor descriptors
+            const auto c_ds_buf_refs = concat_tuple_of_reference(
+                tie(c_shuffle_block_buf),
+                generate_tie(
+                    [&](auto i) -> const auto& // return type should be reference
+                    { return ds_grid_buf[i]; },
+                    Number<NumDTensor_>{}));
+
+            // tuple of starting index of C/Ds blockwise copy
+            const auto idx_c_ds_block_begin = container_concat(
+                make_tuple(make_multi_index(0, 0, 0, 0)),
+                generate_tuple(
+                    [&](auto) {
+                        return make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0);
+                    },
+                    Number<NumDTensor_>{}));
+
+            // space filling curve for threadwise C in VGPR before shuffle
+            constexpr auto sfc_c_vgpr =
+                SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
+                                  Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
+                                  Sequence<CShuffleMXdlPerWavePerShuffle,
+                                           CShuffleNXdlPerWavePerShuffle,
+                                           1,
+                                           1,
+                                           M2,
+                                           1,
+                                           M4,
+                                           1>>{};
+
+            // space filling curve for shuffled blockwise C/D/E
+            constexpr auto sfc_cde_block =
+                SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
+                                  Sequence<0, 2, 1, 3>,
+                                  Sequence<1,
+                                           CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                                           1,
+                                           CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
+
+            constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
+
+            static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
+
+            // blockwise copy C/D/E between LDS and global
+            auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
+                ThisThreadBlock,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType_{})),
+                Tuple<EDataType>,
+                decltype(c_ds_desc_refs),
+                decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
+                CDEElementwiseOperation_,
+                Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
+                                                                            // Sequence support
+                                                                            // arbitray type
+                Sequence<1,
+                         CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
+                         1,
+                         CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
+                CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+                Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
+                Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
+                3,                    // index_t VectorDim,
+                CDEShuffleBlockTransferScalarPerVector_NPerBlock,
+                sequence_merge_t<
+                    Sequence<true>,
+                    uniform_sequence_gen_t<NumDTensor_,
+                                           false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
+                Sequence<false>>                    // ThreadTransferDstResetCoordinateAfterRunFlags
+                {c_ds_desc_refs,
+                 idx_c_ds_block_begin,
+                 tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                 make_tuple(make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0)),
+                 cde_element_op};
+
+            static_for<0, num_access, 1>{}([&](auto access_id) {
+                // make sure it's safe to write to LDS
+                block_sync_lds();
+
+                // each thread write its data from VGPR to LDS
+                c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
+                                              c_thread_buf,
+                                              c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
+                                              c_shuffle_block_buf);
+
+                // make sure it's safe to read from LDS
+                block_sync_lds();
+
+                // each block copy its data from LDS to global
+                cde_block_copy_lds_and_global.Run(
+                    c_ds_desc_refs,
+                    c_ds_buf_refs,
+                    tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                    tie(e_grid_buf));
+
+                if constexpr(access_id < num_access - 1)
+                {
+                    constexpr auto cde_lds_and_global_step =
+                        sfc_cde_block.GetForwardStep(access_id);
+
+                    // move on Ds
+                    static_for<0, NumDTensor_, 1>{}([&](auto i) {
+                        cde_block_copy_lds_and_global.MoveSrcSliceWindow(
+                            c_ds_desc_refs, i + I1, cde_lds_and_global_step);
+                    });
+
+                    // move on E
+                    cde_block_copy_lds_and_global.MoveDstSliceWindow(
+                        tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
+                        I0,
+                        cde_lds_and_global_step);
+                }
+            });
+
+            if(threadIdx.x == 0)
+            {
+                index_t k_id_finished_t = atomicAdd(barrier_count_finished, 1);
+
+                if(k_id_finished_t == KBatch)
+                {
+                    *barrier_count_finished = 0;
+                }
+            }
+        }
+    }
+
+    template <bool HasMainKBlockLoop,
+              InMemoryDataOperationEnum EGlobalMemoryDataOperation,
+              GemmSpecialization GemmSpec,
+              typename ALayout,
+              typename BLayout,
+              typename DsLayout,
+              typename ELayout,
+              typename Block2ETileMap>
+    __device__ static void Run(const void* __restrict__ p_a_grid_,
+                               const void* __restrict__ p_b_grid_,
+                               DsGridPointer p_ds_grid,
+                               void* __restrict__ p_e_grid_,
+                               void* __restrict__ p_shared,
+                               uint32_t* barrier_count_finished,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const index_t M,
+                               const index_t N,
+                               const index_t K,
+                               const index_t StrideA,
+                               const index_t StrideB,
+                               const std::array<index_t, NumDTensor> StrideDs,
+                               const index_t StrideE,
+                               const index_t KBatch,
+                               const Block2ETileMap& block_2_etile_map)
+    {
+        const auto p_a_grid = reinterpret_cast<const ABDataType*>(p_a_grid_);
+        const auto p_b_grid = reinterpret_cast<const ABDataType*>(p_b_grid_);
+        const auto p_e_grid = reinterpret_cast<EDataType*>(p_e_grid_);
+
+        using DsGridDesc_M_N =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_M_N<DsLayout, GemmSpec>({}, {}, {}))>;
+
+        DsGridDesc_M_N ds_grid_desc_m_n;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+            ds_grid_desc_m_n(j) = MakeEGridDescriptor_M_N<DLayout, GemmSpec>(M, N, StrideDs[j]);
+        });
+
+        const auto e_grid_desc_m_n = MakeEGridDescriptor_M_N<ELayout, GemmSpec>(M, N, StrideE);
+
+        // tensor descriptors for block/thread-wise copy
+        const auto a_grid_desc_kbatch_ak0_m_ak1 =
+            MakeAGridDescriptor_KBatch_AK0_M_AK1<ALayout, GemmSpec>(M, K, StrideA, KBatch);
+
+        const auto b_grid_desc_kbatch_bk0_n_bk1 =
+            MakeBGridDescriptor_KBatch_BK0_N_BK1<BLayout, GemmSpec>(K, N, StrideB, KBatch);
+
+        using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+            remove_cvref_t<decltype(MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                DsGridDesc_M_N{}))>;
+
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock;
+
+        static_for<0, NumDTensor, 1>{}([&](auto j) {
+            ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
+                MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[j]);
+        });
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+            MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(e_grid_desc_m_n);
+
+        const auto block_work_idx =
+            block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
+
+        const index_t kbatch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
+
+        if(kbatch_id == KBatch - 1)
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, NumDTensor, DsDataType>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                cde_element_op,
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+        else
+        {
+            Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, 0, Tuple<>>(
+                p_a_grid,
+                p_b_grid,
+                p_ds_grid,
+                p_e_grid,
+                p_shared,
+                barrier_count_finished,
+                KBatch,
+                a_element_op,
+                b_element_op,
+                ck::tensor_operation::element_wise::PassThrough{},
+                a_grid_desc_kbatch_ak0_m_ak1,
+                b_grid_desc_kbatch_bk0_n_bk1,
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                e_grid_desc_mblock_mperblock_nblock_nperblock,
+                block_2_etile_map);
+        }
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp

@@ -4,7 +4,8 @@
 #pragma once
 
 #include "ck/utility/common_header.hpp"
-#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/utility/loop_scheduler.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 
 namespace ck {
 

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

@@ -37,7 +37,8 @@ __global__ void
                                    index_t StrideC,
                                    typename GridwiseGemm::Block2CTileMap block_mapping)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
 
     __shared__ uint8_t p_shared[shared_size];

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp

@@ -194,7 +194,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
               StrideC{StrideC_},
               MPadded{CalculateMPadded(M_)},
               NPadded{CalculateNPadded(N_)},
-              K0{CalculateK0(K)}
+              K0{CalculateK0(K_)}
         {
         }
 
@@ -383,7 +383,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
 
     __host__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
     {
-        const index_t num_loop = K / (K0PerBlock * K1);
+        const index_t num_loop = math::integer_divide_ceil(K, K0PerBlock * K1);
 
         return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
     }
@@ -840,7 +840,25 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext
             }
         }();
 
-        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            const auto K0Pad = math::integer_divide_ceil(K0, K0PerBlock) * K0PerBlock;
+            const auto KPad  = K0Pad * K1Value;
+
+            const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
+                a_grid_desc_m_k,
+                make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return transform_tensor_descriptor(
+                a_grid_desc_m_kpad,
+                make_tuple(make_unmerge_transform(make_tuple(K0Pad, K1Value)),
+                           make_right_pad_transform(M, MPad - M)),
+                make_tuple(Sequence<1>{}, Sequence<0>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
         {
             return transform_tensor_descriptor(
                 a_grid_desc_m_k,
@@ -874,7 +892,26 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext
             }
         }();
 
-        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
+        if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
+        {
+            const auto K0Pad = math::integer_divide_ceil(K0, K0PerBlock) * K0PerBlock;
+            const auto KPad  = K0Pad * K1Value;
+
+            const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
+                b_grid_desc_k_n,
+                make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return transform_tensor_descriptor(
+                b_grid_desc_kpad_n,
+                make_tuple(make_unmerge_transform(make_tuple(K0Pad, K1Value)),
+                           make_right_pad_transform(N, NPad - N)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+        }
+
+        else if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
         {
             return transform_tensor_descriptor(
                 b_grid_desc_k_n,

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

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

include/ck/tensor_operation/gpu/thread/threadwise_contraction_dl_dpp8.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include "ck/utility/amd_gemm_dpp.hpp"
-#include "ck/utility/common_header.hpp"
-#include "ck/utility/inner_product_dpp8.hpp"
-#include "ck/utility/math.hpp"
-
-namespace ck {
-
-/**
- * Threadwise contraction using dot instructions with DPP8 modifier.
- *
- * Assumptions:
- *   1. `AThreadDesc_TK0_TM0_TM1_TK1`, `BThreadDesc_TK0_TN0_TN1_TK1`, `CThreadDesc_TM0_TM1_TN0_TN1`
- *      are known at compile-time;
- *   2. `AOriginIdx`, `BOriginIdx`, `COriginIdx` are known at compile-time;
- *   3. `TM0` is equal to 1 and `TN0` is equal to 1;
- *   4. When `ShareA` is set (unset, respectively), `TM1` (`TN1`, respectively) is divisible by
- *      the size of the lane group (`dpp8::lane_group_size`).
- */
-template <typename FloatA,
-          typename FloatB,
-          typename FloatC,
-          typename AThreadDesc_TK0_TM0_TM1_TK1,
-          typename BThreadDesc_TK0_TN0_TN1_TK1,
-          typename CThreadDesc_TM0_TM1_TN0_TN1,
-          typename TKLengths,
-          typename TMLengths,
-          typename TNLengths,
-          bool ShareA,
-          typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
-                                 BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
-                                 CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
-                             bool>::type = false>
-struct ThreadwiseContractionDlDpp8_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1
-{
-
-    static constexpr auto I0 = Number<0>{};
-    static constexpr auto I1 = Number<1>{};
-
-    static constexpr index_t TK0 = TKLengths{}[I0];
-    static constexpr index_t TK1 = TKLengths{}[I1];
-    static constexpr index_t TM0 = TMLengths{}[I0];
-    static constexpr index_t TM1 = TMLengths{}[I1];
-    static constexpr index_t TN0 = TNLengths{}[I0];
-    static constexpr index_t TN1 = TNLengths{}[I1];
-
-    static_assert(TM0 == 1 && TN0 == 1);
-
-    static_assert((ShareA && TM1 % dpp8::lane_group_size == 0) ||
-                  (!ShareA && TN1 % dpp8::lane_group_size == 0));
-    static constexpr index_t shared_elems_per_lane =
-        ShareA ? TM1 / dpp8::lane_group_size : TN1 / dpp8::lane_group_size;
-
-    __device__ constexpr ThreadwiseContractionDlDpp8_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1()
-    {
-        static_assert(AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
-                          BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
-                          CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
-                      "wrong! Desc should be known at compile-time");
-
-        static_assert(TKLengths::Size() == 2 && TMLengths::Size() == 2 && TNLengths::Size() == 2,
-                      "wrong!");
-    }
-
-    template <typename ABuffer,
-              typename AOriginIdx,
-              typename BBuffer,
-              typename BOriginIdx,
-              typename CBuffer,
-              typename COriginIdx>
-    __device__ static void Run(const ABuffer& a_buf,
-                               AOriginIdx,
-                               const BBuffer& b_buf,
-                               BOriginIdx,
-                               CBuffer& c_buf,
-                               COriginIdx)
-    {
-        static_assert(is_known_at_compile_time<remove_cvref_t<AOriginIdx>>::value &&
-                          is_known_at_compile_time<remove_cvref_t<BOriginIdx>>::value &&
-                          is_known_at_compile_time<remove_cvref_t<COriginIdx>>::value,
-                      "wrong! AOriginIdx, BOriginIdx, COringinIdx should be known at compile-time");
-
-        static_assert(
-            is_same<remove_cvref_t<typename ABuffer::type>, remove_cvref_t<FloatA>>::value &&
-            is_same<remove_cvref_t<typename BBuffer::type>, remove_cvref_t<FloatB>>::value &&
-            is_same<remove_cvref_t<typename CBuffer::type>, remove_cvref_t<FloatC>>::value &&
-            "wrong! inconsistent type");
-
-        constexpr auto a_origin_idx = to_multi_index(AOriginIdx{});
-        constexpr auto b_origin_idx = to_multi_index(BOriginIdx{});
-        constexpr auto c_origin_idx = to_multi_index(COriginIdx{});
-
-        static_for<0, TK0, 1>{}([&](auto tk0) {
-            static_for<0, TM1, 1>{}([&](auto tm1) {
-                static_for<0, TN1, 1>{}([&](auto tn1) {
-                    vector_type<FloatA, TK1> a_vec;
-                    vector_type<FloatB, TK1> b_vec;
-
-                    static_for<0, TK1, 1>{}([&](auto tk1) {
-                        constexpr index_t local_tm1 = ShareA ? tm1 % shared_elems_per_lane : tm1;
-                        constexpr index_t a_offset  = AThreadDesc_TK0_TM0_TM1_TK1{}.CalculateOffset(
-                            a_origin_idx + make_multi_index(tk0, 0, local_tm1, tk1));
-
-                        constexpr index_t local_tn1 = ShareA ? tn1 : tn1 % shared_elems_per_lane;
-                        constexpr index_t b_offset  = BThreadDesc_TK0_TN0_TN1_TK1{}.CalculateOffset(
-                            b_origin_idx + make_multi_index(tk0, 0, local_tn1, tk1));
-
-                        a_vec.template AsType<FloatA>()(tk1) = a_buf[Number<a_offset>{}];
-                        b_vec.template AsType<FloatB>()(tk1) = b_buf[Number<b_offset>{}];
-                    });
-
-                    using a_vector_t = typename vector_type<FloatA, TK1>::type;
-                    using b_vector_t = typename vector_type<FloatB, TK1>::type;
-
-                    constexpr index_t c_offset = CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
-                        c_origin_idx + make_multi_index(0, tm1, 0, tn1));
-
-                    constexpr int src_lane =
-                        ShareA ? (tm1 / shared_elems_per_lane) % dpp8::lane_group_size
-                               : (tn1 / shared_elems_per_lane) % dpp8::lane_group_size;
-
-                    dpp8::inner_product_dpp<a_vector_t, b_vector_t, FloatC, src_lane, ShareA>(
-                        a_vec.template AsType<a_vector_t>()[I0],
-                        b_vec.template AsType<b_vector_t>()[I0],
-                        c_buf(Number<c_offset>{}));
-                });
-            });
-        });
-    }
-};
-
-} // namespace ck

include/ck/tensor_operation/gpu/warp/dpp_gemm.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/amd_gemm_dpp.hpp"
+#include "ck/utility/common_header.hpp"
+#include "ck/utility/math.hpp"
+
+namespace ck {
+
+enum struct DppInstr
+{
+    dpp8_f16_16x16x2 = 0,
+    dpp8_f16_8x32x2,
+    dpp8_f16_32x8x2
+};
+
+/**
+ * Structure representing DPP GEMM executed by a single wavefront.
+ *
+ * Each structure instantiation must contain the following fields:
+ * - wave_size - number of threads that execute a single DPP GEMM operation, usually equal to the
+ *               number of threads in a wavefront;
+ * - lanegroup_size - number of threads (lanes) that share data using DPP instruction modifier,
+ *                    it's 8 in case of DPP8;
+ * - m_per_wave - size along M dimension of matrix C that is processed in a single DPP GEMM
+ *                operation;
+ * - n_per_wave - size along N dimension of matrix C that is processed in a single DPP GEMM
+ *                operation;
+ * - m_per_lanegroup - size along M dimension that is processed by a single lanegroup;
+ * - n_per_lanegroup - size along N dimension that is processed by a single lanegroup;
+ * - m_per_thread - size along M dimension of the tile calculated by a single thread;
+ * - n_per_thread - size along N dimension of the tile calculated by a single thread;
+ * - k_per_dpp - size along K dimension that is reduced in a single DPP GEMM operation;
+ * - share_a - indicates whether we share matrix A or matrix B between lanes using DPP modifiers.
+ *
+ * Not all the combinarions are supported now, for current restrictions see the static asserts
+ * in the DppSelector's contructor.
+ */
+template <DppInstr instr>
+struct dpp_type;
+
+template <>
+struct dpp_type<DppInstr::dpp8_f16_32x8x2>
+{
+    static constexpr index_t wave_size       = 32;
+    static constexpr index_t lanegroup_size  = 8;
+    static constexpr index_t m_per_wave      = 32;
+    static constexpr index_t n_per_wave      = 8;
+    static constexpr index_t m_per_lanegroup = 8;
+    static constexpr index_t n_per_lanegroup = 8;
+    static constexpr index_t m_per_thread    = 8;
+    static constexpr index_t n_per_thread    = 1;
+    static constexpr index_t k_per_dpp       = 2;
+    static constexpr bool share_a            = true;
+    using BaseType                           = half_t;
+
+    template <index_t MPerDpp, index_t NPerDpp, class ADataType, class BDataType, class CDataType>
+    __device__ void run(const ADataType& a, const BDataType& b, CDataType& reg_c) const
+    {
+        dpp8::DppLanegroupGemm<m_per_thread,
+                               n_per_thread,
+                               k_per_dpp,
+                               BaseType,
+                               ADataType,
+                               BDataType,
+                               CDataType,
+                               share_a>{}
+            .Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct dpp_type<DppInstr::dpp8_f16_8x32x2>
+{
+    static constexpr index_t wave_size       = 32;
+    static constexpr index_t lanegroup_size  = 8;
+    static constexpr index_t m_per_wave      = 8;
+    static constexpr index_t n_per_wave      = 32;
+    static constexpr index_t m_per_lanegroup = 8;
+    static constexpr index_t n_per_lanegroup = 8;
+    static constexpr index_t m_per_thread    = 8;
+    static constexpr index_t n_per_thread    = 1;
+    static constexpr index_t k_per_dpp       = 2;
+    static constexpr bool share_a            = true;
+    using BaseType                           = half_t;
+
+    template <index_t MPerDpp, index_t NPerDpp, class ADataType, class BDataType, class CDataType>
+    __device__ void run(const ADataType& a, const BDataType& b, CDataType& reg_c) const
+    {
+        dpp8::DppLanegroupGemm<m_per_thread,
+                               n_per_thread,
+                               k_per_dpp,
+                               BaseType,
+                               ADataType,
+                               BDataType,
+                               CDataType,
+                               share_a>{}
+            .Run(a, b, reg_c);
+    }
+};
+
+template <>
+struct dpp_type<DppInstr::dpp8_f16_16x16x2>
+{
+    static constexpr index_t wave_size       = 32;
+    static constexpr index_t lanegroup_size  = 8;
+    static constexpr index_t m_per_wave      = 16;
+    static constexpr index_t n_per_wave      = 16;
+    static constexpr index_t m_per_lanegroup = 8;
+    static constexpr index_t n_per_lanegroup = 8;
+    static constexpr index_t m_per_thread    = 8;
+    static constexpr index_t n_per_thread    = 1;
+    static constexpr index_t k_per_dpp       = 2;
+    static constexpr bool share_a            = true;
+    using BaseType                           = half_t;
+
+    template <index_t MPerDpp, index_t NPerDpp, class ADataType, class BDataType, class CDataType>
+    __device__ void run(const ADataType& a, const BDataType& b, CDataType& reg_c) const
+    {
+        dpp8::DppLanegroupGemm<m_per_thread,
+                               n_per_thread,
+                               k_per_dpp,
+                               BaseType,
+                               ADataType,
+                               BDataType,
+                               CDataType,
+                               share_a>{}
+            .Run(a, b, reg_c);
+    }
+};
+
+template <typename BaseType, index_t MPerDpp, index_t NPerDpp>
+struct DppSelector
+{
+    template <typename BaseType_, index_t MPerDpp_, index_t NPerDpp_>
+    static constexpr auto GetDpp();
+
+    template <>
+    static constexpr auto GetDpp<half_t, 8, 32>()
+    {
+        return DppInstr::dpp8_f16_8x32x2;
+    }
+
+    template <>
+    static constexpr auto GetDpp<half_t, 16, 16>()
+    {
+        return DppInstr::dpp8_f16_16x16x2;
+    }
+
+    template <>
+    static constexpr auto GetDpp<half_t, 32, 8>()
+    {
+        return DppInstr::dpp8_f16_32x8x2;
+    }
+
+    static constexpr auto selected_dpp = dpp_type<GetDpp<BaseType, MPerDpp, NPerDpp>()>{};
+
+    __host__ __device__ constexpr DppSelector()
+    {
+        static_assert(selected_dpp.m_per_wave % selected_dpp.m_per_lanegroup == 0);
+        static_assert(selected_dpp.n_per_wave % selected_dpp.n_per_lanegroup == 0);
+
+        static_assert(selected_dpp.k_per_dpp % 2 == 0);
+
+        static_assert(selected_dpp.wave_size % selected_dpp.lanegroup_size == 0);
+        constexpr index_t num_dpp_per_wave = selected_dpp.wave_size / selected_dpp.lanegroup_size;
+        constexpr index_t num_wave_c_elems = selected_dpp.m_per_wave * selected_dpp.n_per_wave;
+        constexpr index_t num_dpp_c_elems =
+            selected_dpp.m_per_lanegroup * selected_dpp.n_per_lanegroup;
+        static_assert(num_wave_c_elems % num_dpp_c_elems == 0);
+        static_assert(num_dpp_per_wave == num_wave_c_elems / num_dpp_c_elems);
+
+        if constexpr(selected_dpp.share_a)
+        {
+            static_assert(selected_dpp.m_per_lanegroup == selected_dpp.m_per_thread);
+            static_assert(selected_dpp.n_per_lanegroup % selected_dpp.n_per_thread == 0);
+            static_assert(selected_dpp.n_per_lanegroup / selected_dpp.n_per_thread ==
+                          selected_dpp.lanegroup_size);
+        }
+        else
+        {
+            static_assert(selected_dpp.m_per_lanegroup % selected_dpp.n_per_thread == 0);
+            static_assert(selected_dpp.m_per_lanegroup / selected_dpp.n_per_thread ==
+                          selected_dpp.lanegroup_size);
+            static_assert(selected_dpp.n_per_lanegroup == selected_dpp.n_per_thread);
+        }
+
+        // Below checks come from the restrictions of the current implementation, could be removed
+        // in the future when the implementation is more generalized.
+        static_assert(selected_dpp.share_a);
+        static_assert(selected_dpp.n_per_thread == 1);
+        static_assert(selected_dpp.m_per_thread == selected_dpp.lanegroup_size);
+        static_assert(selected_dpp.m_per_lanegroup == selected_dpp.m_per_thread);
+        static_assert(selected_dpp.n_per_lanegroup ==
+                      selected_dpp.n_per_thread * selected_dpp.lanegroup_size);
+    }
+
+    static constexpr index_t GetK1PerDpp() { return selected_dpp.k_per_dpp; }
+};
+
+template <typename BaseType, index_t MPerDpp, index_t NPerDpp, index_t KPack>
+struct DppGemm
+{
+    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>{};
+
+    using CIndex   = MultiIndex<2>;
+    using CIndex4D = MultiIndex<4>;
+
+    __host__ __device__ constexpr DppGemm()
+    {
+        static_assert(MPerDpp == 8 || MPerDpp == 16 || MPerDpp == 32,
+                      "MPerDpp must be either 8, 16 or 32.");
+        static_assert(NPerDpp == 8 || NPerDpp == 16 || NPerDpp == 32,
+                      "NPerDpp must be either 8, 16 or 32.");
+
+        static_assert(KPack % dpp_instr.k_per_dpp == 0, "KPack must be divisible by k_per_dpp.");
+    }
+
+    __device__ static constexpr index_t GetRegSizePerDpp()
+    {
+        return MPerDpp * NPerDpp / dpp_instr.wave_size;
+    }
+
+    template <class ADataType, class BDataType, class CDataType>
+    __device__ void
+    Run(const ADataType& p_a_wave, const BDataType& p_b_wave, CDataType& p_c_thread) const
+    {
+        static_assert(is_same<BaseType, double>::value || is_same<BaseType, float>::value ||
+                          is_same<BaseType, half_t>::value || is_same<BaseType, bhalf_t>::value ||
+                          is_same<BaseType, int8_t>::value || is_same<BaseType, f8_t>::value,
+                      "base BaseType must be double, float, half, bfloat16, and int8_t!");
+
+        static_for<0, KPack / dpp_instr.k_per_dpp, 1>{}([&](auto k) {
+            dpp_instr.template run<MPerDpp, NPerDpp>(p_a_wave[k], p_b_wave[k], p_c_thread);
+        });
+    }
+
+    __device__ static auto GetLaneIdInWave()
+    {
+        return get_thread_local_1d_id() % dpp_instr.wave_size;
+    }
+
+    __device__ static auto GetWaveId() { return get_thread_local_1d_id() / dpp_instr.wave_size; }
+
+    __device__ static auto GetLaneIdInLaneGroup()
+    {
+        return get_thread_local_1d_id() % dpp_instr.lanegroup_size;
+    }
+
+    __device__ static auto GetLaneGroupIdInWave()
+    {
+        return GetLaneIdInWave() / dpp_instr.lanegroup_size;
+    }
+
+    __device__ static auto GetDppOpIdx()
+    {
+        const auto lanegroupId = GetLaneGroupIdInWave();
+
+        constexpr auto lanegroup_idx_1d_to_dpp_idx_2d_adaptor = make_single_stage_tensor_adaptor(
+            make_tuple(
+                make_merge_transform(make_tuple(dpp_instr.m_per_wave / dpp_instr.m_per_lanegroup,
+                                                dpp_instr.n_per_wave / dpp_instr.n_per_lanegroup))),
+            make_tuple(Sequence<0, 1>{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto dpp_idx = lanegroup_idx_1d_to_dpp_idx_2d_adaptor.CalculateBottomIndex(
+            make_multi_index(lanegroupId));
+
+        const auto m_dpp_idx = dpp_idx[I0];
+        const auto n_dpp_idx = dpp_idx[I1];
+
+        return make_tuple(m_dpp_idx, n_dpp_idx);
+    }
+
+    __host__ __device__ static auto CalculateAThreadOriginDataIndex_K_M()
+    {
+        const auto laneId   = get_thread_local_1d_id();
+        const auto wave_row = laneId / dpp_instr.n_per_wave;
+        auto m_idx          = dpp_instr.m_per_thread * wave_row + GetLaneIdInLaneGroup();
+        return make_tuple(0, m_idx % dpp_instr.m_per_wave);
+    }
+
+    __host__ __device__ static auto CalculateBThreadOriginDataIndex_K_N()
+    {
+        const auto laneId = get_thread_local_1d_id();
+        return make_tuple(0, laneId % dpp_instr.n_per_wave);
+    }
+
+    __device__ static CIndex GetBeginOfThreadBlk()
+    {
+        const auto dpp_op_idx = GetDppOpIdx();
+
+        const auto m_dpp_op_idx = dpp_op_idx[I0];
+        const auto n_dpp_op_idx = dpp_op_idx[I1];
+
+        index_t n_offset = n_dpp_op_idx * dpp_instr.n_per_lanegroup + GetLaneIdInLaneGroup();
+        index_t m_offset = m_dpp_op_idx * dpp_instr.m_per_lanegroup;
+
+        return CIndex{m_offset, n_offset};
+    }
+
+    static constexpr auto dpp = DppSelector<BaseType, MPerDpp, NPerDpp>{};
+
+    static constexpr auto dpp_instr = dpp.selected_dpp;
+
+    static constexpr auto K0PerDpp = 1;
+    static constexpr auto K1PerDpp = dpp.GetK1PerDpp();
+
+    __host__ __device__ static constexpr auto GetCMNThreadBlkLengths()
+    {
+        return make_tuple(Number<dpp_instr.m_per_thread>{}, Number<dpp_instr.n_per_thread>{});
+    }
+};
+
+} // namespace ck

include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp

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

include/ck/utility/amd_gemm_dpp.hpp

@@ -5,17 +5,63 @@
 
 #include "ck/utility/common_header.hpp"
 #include "ck/utility/math.hpp"
-#include "ck/utility/amd_gemm_dpp.hpp"
+#include "ck/utility/inner_product_dpp8.hpp"
 
 namespace ck {
 
 namespace dpp8 {
 
-/// Number of lanes that can share data using DPP8 modifiers.
-constexpr index_t lane_group_size = 8;
+template <class ABDataType>
+struct dpp_datatypes;
 
-__device__ index_t get_lane_group_local_idx() { return threadIdx.x / lane_group_size; }
-__device__ index_t get_thread_idx_in_lane_group() { return threadIdx.x % lane_group_size; }
+template <>
+struct dpp_datatypes<half_t>
+{
+    // Dot product of `half2_t` and `half2_t` to get `float`. Reducing 2 elements from K in a
+    // single instruction.
+    using a_dtype                        = half_t;
+    using b_dtype                        = half_t;
+    using c_dtype                        = float;
+    static constexpr index_t k_per_instr = 2;
+};
+
+template <index_t MPerThread,
+          index_t NPerThread,
+          index_t KPerThread,
+          class BaseInputType,
+          class AVecDataType,
+          class BVecDataType,
+          class CVecDataType,
+          bool ShareA>
+struct DppLanegroupGemm
+{
+    using datatypes_conf = dpp_datatypes<BaseInputType>;
+    using ADataType      = typename datatypes_conf::a_dtype;
+    using BDataType      = typename datatypes_conf::b_dtype;
+    using CDataType      = typename datatypes_conf::c_dtype;
+
+    __device__ void Run(const AVecDataType& a_vec, const BVecDataType& b_vec, CVecDataType& c_vec)
+    {
+        constexpr index_t num_c_elems_per_thread = ShareA ? MPerThread : NPerThread;
+
+        const vector_type<ADataType, KPerThread> a_vector{a_vec};
+        const vector_type<BDataType, KPerThread> b_vector{b_vec};
+
+        static_for<0, num_c_elems_per_thread, 1>{}([&](auto c_idx) {
+            float c = c_vec.template AsType<CDataType>()(c_idx);
+            // Next `c_idx` implies that we need to pull data from the next lane.
+            constexpr index_t source_lane = c_idx;
+            static_for<0, KPerThread / datatypes_conf::k_per_instr, 1>{}([&](auto k_chunk) {
+                const auto a_k_vec = a_vector.template AsType<AVecDataType>()[k_chunk];
+                const auto b_k_vec = b_vector.template AsType<BVecDataType>()[k_chunk];
+                ck::dpp8::
+                    inner_product_dpp<AVecDataType, BVecDataType, CDataType, source_lane, ShareA>(
+                        a_k_vec, b_k_vec, c);
+            });
+            c_vec.template AsType<CDataType>()(c_idx) = c;
+        });
+    }
+};
 
 } // namespace dpp8
 

include/ck/utility/inner_product_dpp8.hpp

@@ -2,6 +2,7 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
+
 #include "amd_gemm_dpp.hpp"
 #include "data_type.hpp"
 #include "type_convert.hpp"
@@ -10,6 +11,9 @@ namespace ck {
 
 namespace dpp8 {
 
+/// Number of lanes that can share data using DPP8 modifiers.
+constexpr index_t lane_group_size = 8;
+
 template <int SrcLaneIdx>
 __device__ void inline_v_dot2c_dpp8_instr(const half2_t& a, const half2_t& b, float& c);
 

include/ck/utility/loop_scheduler.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_adaptor.hpp"
+
+namespace ck {
+
+enum struct LoopScheduler
+{
+    Default,
+    Interwave,
+};
+
+constexpr LoopScheduler make_default_loop_scheduler()
+{
+#if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
+    return LoopScheduler::Interwave;
+#else
+    return LoopScheduler::Default;
+#endif // if CK_EXPERIMENTAL_DEFAULT_TO_INTER_WAVE_SCHEDULING
+}
+
+} // namespace ck

include/ck/utility/reduction_operator.hpp

@@ -116,7 +116,15 @@ struct Max
     template <typename T>
     __host__ __device__ static constexpr T GetIdentityValue()
     {
-        return NumericLimits<T>::Lowest();
+        if constexpr(is_same_v<T, bhalf_t>)
+        {
+            float val = NumericLimits<float>::Lowest();
+            return type_convert<bhalf_t>(val);
+        }
+        else
+        {
+            return NumericLimits<T>::Lowest();
+        }
     };
 
     __host__ __device__ static constexpr bool
@@ -138,6 +146,15 @@ struct Max
             a = b;
     }
 
+    __host__ __device__ inline constexpr void operator()(bhalf_t& a, bhalf_t b) const
+    {
+        float a_ = type_convert<float>(a);
+        float b_ = type_convert<float>(b);
+
+        if(a_ < b_)
+            a = b;
+    }
+
     template <typename T>
     __host__ __device__ inline constexpr void operator()(T& a, T b, bool& changed) const
     {
@@ -152,6 +169,18 @@ struct Max
             changed = true;
         }
     }
+
+    __host__ __device__ inline constexpr void operator()(bhalf_t& a, bhalf_t b, bool& changed) const
+    {
+        float a_ = type_convert<float>(a);
+        float b_ = type_convert<float>(b);
+
+        if(a_ < b_)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
 };
 
 struct Min
@@ -159,6 +188,15 @@ struct Min
     template <typename T>
     __host__ __device__ static constexpr T GetIdentityValue()
     {
+        if constexpr(is_same_v<T, bhalf_t>)
+        {
+            float val = NumericLimits<float>::Max();
+            return type_convert<bhalf_t>(val);
+        }
+        else
+        {
+            return NumericLimits<T>::Max();
+        }
         return NumericLimits<T>::Max();
     };
 
@@ -181,6 +219,15 @@ struct Min
             a = b;
     }
 
+    __host__ __device__ inline constexpr void operator()(bhalf_t& a, bhalf_t b) const
+    {
+        float a_ = type_convert<float>(a);
+        float b_ = type_convert<float>(b);
+
+        if(a_ > b_)
+            a = b;
+    }
+
     template <typename T>
     __host__ __device__ inline constexpr void operator()(T& a, T b, bool& changed) const
     {
@@ -195,6 +242,18 @@ struct Min
             changed = true;
         }
     }
+
+    __host__ __device__ inline constexpr void operator()(bhalf_t& a, bhalf_t b, bool& changed) const
+    {
+        float a_ = type_convert<float>(a);
+        float b_ = type_convert<float>(b);
+
+        if(a_ > b_)
+        {
+            a       = b;
+            changed = true;
+        }
+    }
 };
 
 struct AMax

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

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

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

@@ -53,7 +53,16 @@ struct ReferenceMaxPoolBwd : public device::BaseOperator
             {
                 int index = arg.indices_.mData[i];
                 if(index >= 0 && index < din_length)
-                    buf[index] += ck::type_convert<ConputeDataType>(arg.dout_.mData[i]);
+                {
+                    if constexpr(is_same_v<ConputeDataType, bhalf_t>)
+                    {
+                        float buf_val = ck::type_convert<float>(buf[index]);
+                        buf_val += ck::type_convert<float>(arg.dout_.mData[i]);
+                        buf[index] = ck::type_convert<ConputeDataType>(buf_val);
+                    }
+                    else
+                        buf[index] += ck::type_convert<ConputeDataType>(arg.dout_.mData[i]);
+                }
             }
 
             for(int i = 0; i < din_length; ++i)

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

@@ -256,10 +256,12 @@ struct ReferencePoolingFwd : public device::BaseOperator
 
                     for(ck::index_t y = 0; y < arg.window_spatial_lengths_[0]; ++y)
                     {
-                        ck::index_t hi = ho * arg.window_strides_[0] + y - arg.in_left_pads_[0];
+                        ck::index_t hi = ho * arg.window_strides_[0] +
+                                         y * arg.window_dilations_[0] - arg.in_left_pads_[0];
                         for(ck::index_t x = 0; x < arg.window_spatial_lengths_[1]; ++x)
                         {
-                            ck::index_t wi = wo * arg.window_strides_[1] + x - arg.in_left_pads_[1];
+                            ck::index_t wi = wo * arg.window_strides_[1] +
+                                             x * arg.window_dilations_[1] - arg.in_left_pads_[1];
                             if(hi >= 0 &&
                                hi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[2]) &&
                                wi >= 0 &&

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -101,6 +101,7 @@ using MultiplyAdd    = ck::tensor_operation::element_wise::MultiplyAdd;
 using ScaleAdd       = ck::tensor_operation::element_wise::ScaleAdd;
 using Gelu           = ck::tensor_operation::element_wise::Gelu;
 using Swish          = ck::tensor_operation::element_wise::Swish;
+using Add            = ck::tensor_operation::element_wise::Add;
 
 template <typename Activation>
 using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;

library/include/ck/library/tensor_operation_instance/gpu/avg_pool3d_bwd.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/tensor_operation/gpu/device/device_avgpool_bwd.hpp"
+#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+#ifdef CK_ENABLE_FP16
+void add_device_avgpool_bwd_ndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F16, F16, NDHWC, NDHWC>>>&);
+#endif
+#ifdef CK_ENABLE_BF16
+void add_device_avgpool_bwd_ndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, BF16, BF16, NDHWC, NDHWC>>>&);
+#endif
+#ifdef CK_ENABLE_FP32
+void add_device_avgpool_bwd_ndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceAvgPoolBwd<3, F32, F32, NDHWC, NDHWC>>>&);
+#endif
+template <typename DOutDataType, typename DInDataType, typename InLayout, typename OutLayout>
+struct DeviceOperationInstanceFactory<
+    ck::tensor_operation::device::
+        DeviceAvgPoolBwd<3, DOutDataType, DInDataType, InLayout, OutLayout>>
+{
+    using DeviceOp = DeviceAvgPoolBwd<3, DOutDataType, DInDataType, InLayout, OutLayout>;
+
+    static auto GetInstances()
+    {
+        std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
+        if constexpr(is_same_v<InLayout, NDHWC> && is_same_v<OutLayout, NDHWC>)
+        {
+#ifdef CK_ENABLE_FP16
+            if constexpr(is_same_v<DOutDataType, F16> && is_same_v<DInDataType, F16>)
+                add_device_avgpool_bwd_ndhwc_f16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_BF16
+            else if constexpr(is_same_v<DOutDataType, BF16> && is_same_v<DInDataType, BF16>)
+                add_device_avgpool_bwd_ndhwc_bf16_instances(op_ptrs);
+#endif
+#ifdef CK_ENABLE_FP32
+            else if constexpr(is_same_v<DOutDataType, F32> && is_same_v<DInDataType, F32>)
+                add_device_avgpool_bwd_ndhwc_f32_instances(op_ptrs);
+#endif
+        }
+
+        return op_ptrs;
+    }
+};
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -23,7 +23,7 @@ void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
         DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances(
+void add_device_gemm_dpp_f16_f16_f16_km_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
@@ -38,7 +38,7 @@ void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(
         DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances(
+void add_device_gemm_dpp_f16_f16_f16_km_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
@@ -53,7 +53,7 @@ void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(
         DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances(
+void add_device_gemm_dpp_f16_f16_f16_mk_kn_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
@@ -68,7 +68,7 @@ void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(
         DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 
-void add_device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances(
+void add_device_gemm_dpp_f16_f16_f16_mk_nk_mn_instances(
     std::vector<std::unique_ptr<
         DeviceGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
@@ -374,7 +374,7 @@ struct DeviceOperationInstanceFactory<
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_mk_kn_mn_irregular_instances(op_ptrs);
-                add_device_gemm_dl_dpp8_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
+                add_device_gemm_dpp_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
 #endif
                 add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
             }
@@ -385,7 +385,7 @@ struct DeviceOperationInstanceFactory<
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_mk_nk_mn_irregular_instances(op_ptrs);
-                add_device_gemm_dl_dpp8_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
+                add_device_gemm_dpp_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
 #endif
                 add_device_gemm_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
                 add_device_gemm_xdl_c_shuffle_2_stage_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
@@ -397,7 +397,7 @@ struct DeviceOperationInstanceFactory<
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_km_kn_mn_irregular_instances(op_ptrs);
-                add_device_gemm_dl_dpp8_f16_f16_f16_km_kn_mn_instances(op_ptrs);
+                add_device_gemm_dpp_f16_f16_f16_km_kn_mn_instances(op_ptrs);
 #endif
                 add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs);
             }
@@ -408,7 +408,7 @@ struct DeviceOperationInstanceFactory<
 #ifdef DL_KERNELS
                 add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(op_ptrs);
                 add_device_gemm_dl_f16_f16_f16_km_nk_mn_irregular_instances(op_ptrs);
-                add_device_gemm_dl_dpp8_f16_f16_f16_km_nk_mn_instances(op_ptrs);
+                add_device_gemm_dpp_f16_f16_f16_km_nk_mn_instances(op_ptrs);
 #endif
                 add_device_gemm_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs);
             }