/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp Source File
Go to the documentation of this file.
 // 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(__gfx103__) || defined(__gfx11__))
     __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