/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma_cshuffle_v3.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_wmma_cshuffle_v3.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_wmma_cshuffle_v3.hpp Source File
Go to the documentation of this file.
 // SPDX-License-Identifier: MIT
 // Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
  
 #pragma once
  
 #include "ck/utility/env.hpp"
 #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/block/blockwise_gemm_pipeline_wmma_selector.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_v7r3.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/grid/gridwise_gemm_wmma_cshuffle_v3_common.hpp"
  
 namespace ck {
  
 //          operations that could be applied on each tensor respectively. The CDE_op is an
 //          elementwise operation applied to the C and all D tensors.
 template <typename ALayout,
           typename BLayout,
           typename DsLayout,
           typename ELayout,
           typename ADataType,
           typename BDataType,
           typename AccDataType,
           typename CShuffleDataType,
           typename DsDataType,
           typename EDataType,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CDEElementwiseOperation,
           tensor_operation::device::GemmSpecialization GemmSpec,
           index_t BlockSize,
           index_t MPerBlock,
           index_t NPerBlock,
           index_t KPerBlock,
           index_t AK1Value,
           index_t BK1Value,
           index_t MPerWmma,
           index_t NPerWmma,
           index_t MRepeat,
           index_t NRepeat,
           typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
           typename ABlockTransferThreadClusterArrangeOrder,
           typename ABlockTransferSrcAccessOrder,
           index_t ABlockTransferSrcVectorDim,
           index_t ABlockTransferSrcScalarPerVector,
           index_t ABlockTransferDstScalarPerVector_AK1,
           bool AThreadTransferSrcResetCoordinateAfterRun,
           index_t ABlockLdsExtraM,
           typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
           typename BBlockTransferThreadClusterArrangeOrder,
           typename BBlockTransferSrcAccessOrder,
           index_t BBlockTransferSrcVectorDim,
           index_t BBlockTransferSrcScalarPerVector,
           index_t BBlockTransferDstScalarPerVector_BK1,
           bool BThreadTransferSrcResetCoordinateAfterRun,
           index_t BBlockLdsExtraN,
           index_t CShuffleMRepeatPerShuffle,
           index_t CShuffleNRepeatPerShuffle,
           typename CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           typename CDEShuffleBlockTransferScalarPerVectors,
           BlockGemmPipelineScheduler BlkGemmPipeSched,
           BlockGemmPipelineVersion BlkGemmPipelineVer,
           typename ComputeTypeA,
           typename ComputeTypeB,
           bool PermuteA,
           bool PermuteB>
 struct GridwiseGemm_wmma_cshuffle_v3
     : GridwiseGemm_wmma_cshuffle_v3_base<
           ALayout,
           BLayout,
           DsLayout,
           ELayout,
           ADataType,
           BDataType,
           AccDataType,
           CShuffleDataType,
           DsDataType,
           EDataType,
           AElementwiseOperation,
           BElementwiseOperation,
           CDEElementwiseOperation,
           GemmSpec,
           BlockSize,
           MPerBlock,
           NPerBlock,
           KPerBlock,
           AK1Value,
           BK1Value,
           MPerWmma,
           NPerWmma,
           MRepeat,
           NRepeat,
           ABlockTransferThreadClusterLengths_AK0_M_AK1,
           ABlockTransferThreadClusterArrangeOrder,
           ABlockTransferSrcAccessOrder,
           ABlockTransferSrcVectorDim,
           ABlockTransferSrcScalarPerVector,
           ABlockTransferDstScalarPerVector_AK1,
           AThreadTransferSrcResetCoordinateAfterRun,
           ABlockLdsExtraM,
           BBlockTransferThreadClusterLengths_BK0_N_BK1,
           BBlockTransferThreadClusterArrangeOrder,
           BBlockTransferSrcAccessOrder,
           BBlockTransferSrcVectorDim,
           BBlockTransferSrcScalarPerVector,
           BBlockTransferDstScalarPerVector_BK1,
           BThreadTransferSrcResetCoordinateAfterRun,
           BBlockLdsExtraN,
           CShuffleMRepeatPerShuffle,
           CShuffleNRepeatPerShuffle,
           CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           CDEShuffleBlockTransferScalarPerVectors,
           BlkGemmPipeSched,
           BlkGemmPipelineVer,
           ComputeTypeA,
           ComputeTypeB,
           PermuteA,
           PermuteB>
 {
     using Base = GridwiseGemm_wmma_cshuffle_v3_base<
         ALayout,
         BLayout,
         DsLayout,
         ELayout,
         ADataType,
         BDataType,
         AccDataType,
         CShuffleDataType,
         DsDataType,
         EDataType,
         AElementwiseOperation,
         BElementwiseOperation,
         CDEElementwiseOperation,
         GemmSpec,
         BlockSize,
         MPerBlock,
         NPerBlock,
         KPerBlock,
         AK1Value,
         BK1Value,
         MPerWmma,
         NPerWmma,
         MRepeat,
         NRepeat,
         ABlockTransferThreadClusterLengths_AK0_M_AK1,
         ABlockTransferThreadClusterArrangeOrder,
         ABlockTransferSrcAccessOrder,
         ABlockTransferSrcVectorDim,
         ABlockTransferSrcScalarPerVector,
         ABlockTransferDstScalarPerVector_AK1,
         AThreadTransferSrcResetCoordinateAfterRun,
         ABlockLdsExtraM,
         BBlockTransferThreadClusterLengths_BK0_N_BK1,
         BBlockTransferThreadClusterArrangeOrder,
         BBlockTransferSrcAccessOrder,
         BBlockTransferSrcVectorDim,
         BBlockTransferSrcScalarPerVector,
         BBlockTransferDstScalarPerVector_BK1,
         BThreadTransferSrcResetCoordinateAfterRun,
         BBlockLdsExtraN,
         CShuffleMRepeatPerShuffle,
         CShuffleNRepeatPerShuffle,
         CDEShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
         CDEShuffleBlockTransferScalarPerVectors,
         BlkGemmPipeSched,
         BlkGemmPipelineVer,
         ComputeTypeA,
         ComputeTypeB,
         PermuteA,
         PermuteB>;
  
     using Base::I0;
     using Base::I1;
     using Base::I2;
     using Base::I3;
     using Base::I4;
     using Base::I5;
     using Base::I6;
     using Base::I7;
  
     using Base::AK0Number;
     using Base::AK1Number;
     using Base::BK0Number;
     using Base::BK1Number;
  
     using Base::APackedSize;
     using Base::BPackedSize;
  
     using Base::CalculateAK0Padded;
     using Base::CalculateBK0Padded;
     using Base::CalculateKPadded;
     using Base::CalculateKRead;
     using Base::CalculateMBlock;
     using Base::CalculateMPadded;
     using Base::CalculateNBlock;
     using Base::CalculateNPadded;
     using Base::MakeAGridDescriptor_AK0_M_AK1;
     using Base::MakeBGridDescriptor_BK0_N_BK1;
     using Base::MakeDEGridDescriptor_M_N;
     using Base::MakeDsGridDescriptor_M_N;
     using Base::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock;
  
     using Base::GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat;
  
     using Base::MakeDEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock;
  
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
  
     using Base::GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1;
     using Base::GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1;
  
     using Base::NumDTensor;
     using typename Base::DsGridPointer;
  
     struct Problem
     {
         __host__ Problem(index_t M_,
                          index_t N_,
                          index_t K_,
                          index_t StrideA_,
                          index_t StrideB_,
                          std::array<index_t, NumDTensor> StrideDs_,
                          index_t StrideE_,
                          index_t KBatch_)
             : M{M_},
               N{N_},
               K{K_},
               StrideA{StrideA_},
               StrideB{StrideB_},
               StrideDs{StrideDs_},
               StrideE{StrideE_},
               KBatch{KBatch_},
               MPadded{CalculateMPadded(M_)},
               NPadded{CalculateNPadded(N_)},
               KRead{CalculateKRead(K_, KBatch_)},
               KPadded{CalculateKPadded(K_, KBatch_)},
               AK0{CalculateAK0Padded(K_, KBatch_)},
               BK0{CalculateBK0Padded(K_, KBatch_)},
               MBlock{CalculateMBlock(M_)},
               NBlock{CalculateNBlock(N_)}
         {
         }
  
         __host__ void Print() const
         {
             std::cout << "problem {" << "M:" << M << ", " << "N:" << N << ", " << "K:" << K << ", "
                       << "SA:" << StrideA << ", " << "SB:" << StrideB << ", ";
             if constexpr(NumDTensor > 0)
             {
                 std::cout << "SDs: { ";
                 static_for<0, NumDTensor, 1>{}([&](auto i) {
                     std::cout << StrideDs[i] << (i.value < NumDTensor - 1 ? ", " : "");
                 });
                 std::cout << " }, ";
             }
             std::cout << "SE:" << StrideE << ", " << "MP:" << MPadded << ", " << "NP:" << NPadded
                       << ", " << "KRead:" << KRead << ", " << "KP:" << KPadded << ", "
                       << "AK0:" << AK0 << ", " << "BK0:" << BK0 << ", " << "MBlock: " << MBlock
                       << ", " << "NBlock: " << NBlock << "}" << std::endl;
         }
  
         index_t M;
         index_t N;
         index_t K;
         index_t StrideA;
         index_t StrideB;
         std::array<index_t, NumDTensor> StrideDs;
         index_t StrideE;
         index_t KBatch;
         index_t MPadded;
         index_t NPadded;
         index_t KRead;
         index_t KPadded;
         index_t AK0;
         index_t BK0;
         index_t MBlock;
         index_t NBlock;
     };
  
     // Argument
     struct Argument : public tensor_operation::device::BaseArgument, public Problem
     {
         __host__ Argument(const ADataType* p_a_grid_,
                           const BDataType* p_b_grid_,
                           std::array<const void*, NumDTensor> p_ds_grid_,
                           EDataType* p_e_grid_,
                           index_t M_,
                           index_t N_,
                           index_t K_,
                           index_t StrideA_,
                           index_t StrideB_,
                           std::array<index_t, NumDTensor> StrideDs_,
                           index_t StrideE_,
                           index_t k_batch_,
                           AElementwiseOperation a_element_op_,
                           BElementwiseOperation b_element_op_,
                           CDEElementwiseOperation cde_element_op_,
                           bool is_reduce_ = false)
             : Problem{M_, N_, K_, StrideA_, StrideB_, StrideDs_, StrideE_, k_batch_},
               p_a_grid{p_a_grid_},
               p_b_grid{p_b_grid_},
               p_ds_grid{},
               p_e_grid{p_e_grid_},
               a_element_op{a_element_op_},
               b_element_op{b_element_op_},
               cde_element_op{cde_element_op_},
               is_reduce(is_reduce_)
         {
             static_for<0, NumDTensor, 1>{}([&](auto i) {
                 using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
  
                 p_ds_grid(i) = static_cast<const DDataType*>(p_ds_grid_[i]);
             });
         }
  
         __host__ __device__ inline bool IsReduceAdd() const
         {
             return (Problem::KBatch > 1) && is_reduce;
         }
  
         __host__ __device__ inline bool IsAtomicAdd() const
         {
             return (Problem::KBatch > 1) && (!is_reduce);
         }
  
         const ADataType* p_a_grid;
         const BDataType* p_b_grid;
         DsGridPointer p_ds_grid;
         EDataType* p_e_grid;
  
         const AElementwiseOperation a_element_op;
         const BElementwiseOperation b_element_op;
         const CDEElementwiseOperation cde_element_op;
  
         // TODO: it can be used with SplitK+reduction but currently only used with SplitK+atomicAdd
         bool is_reduce;
     };
  
     struct SplitKBatchOffset
     {
  
         __device__ SplitKBatchOffset(Argument& karg, index_t k_id)
         {
             if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
             {
                 a_k_split_offset = k_id * karg.KRead / APackedSize;
             }
             else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, ALayout>)
             {
                 a_k_split_offset = k_id * karg.KRead * karg.StrideA;
             }
  
             if constexpr(is_same_v<tensor_layout::gemm::RowMajor, BLayout>)
             {
                 b_k_split_offset = k_id * karg.KRead * karg.StrideB;
             }
             else if constexpr(is_same_v<tensor_layout::gemm::ColumnMajor, BLayout>)
             {
                 if constexpr(!PermuteB)
                 {
                     b_k_split_offset = k_id * karg.KRead / BPackedSize;
                 }
                 else
                 {
                     const int k0_offset = karg.KRead * karg.N;
                     b_k_split_offset    = k_id * k0_offset / BPackedSize;
                 }
             }
  
             if(k_id < karg.KBatch - 1)
             {
                 karg.K = karg.KRead;
             }
             else
             {
                 karg.K = karg.K - karg.KRead * (karg.KBatch - 1);
             }
  
             if(karg.IsReduceAdd())
             {
                 c_reduce_offset = k_id * karg.M * karg.N;
             }
             else
             {
                 c_reduce_offset = 0;
             }
         }
  
         index_t a_k_split_offset;
         index_t b_k_split_offset;
         index_t c_reduce_offset;
     };
  
     using BlockwiseGemmPipe = typename Base::BlockwiseGemmPipe;
  
     // return block_id to C matrix tile idx (m0, n0) mapping
     // if arch = gfx942
     using Block2CTileMap = BlockToCTileMap_Grouped_M00_N0_M01Adapt<8, MPerBlock, NPerBlock>;
     // using Block2CTileMap = BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>;
  
     __device__ static index_t GetKBlockPerScale() { return 1; }
  
     template <bool HasMainKBlockLoop,
               InMemoryDataOperationEnum EGlobalMemoryDataOperation,
               TailNumber TailNum>
     __device__ static void Run(const ADataType* p_a_grid,
                                const BDataType* p_b_grid,
                                DsGridPointer& p_ds_grid,
                                EDataType* p_e_grid,
                                void* p_shared,
                                const Problem& problem,
                                AElementwiseOperation a_element_op,
                                BElementwiseOperation b_element_op,
                                CDEElementwiseOperation cde_element_op)
     {
         const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
             problem.M, problem.MPadded, problem.K, problem.KPadded, problem.StrideA, problem.AK0);
         const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
             problem.K, problem.KPadded, problem.N, problem.NPadded, problem.StrideB, problem.BK0);
         const auto ds_grid_desc_m_n = MakeDsGridDescriptor_M_N(
             problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideDs);
         const auto e_grid_desc_m_n = Base::template MakeDEGridDescriptor_M_N<ELayout>(
             problem.M, problem.MPadded, problem.N, problem.NPadded, problem.StrideE);
         const auto ds_grid_desc_mblock_mperblock_nblock_nperblock =
             MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
                 ds_grid_desc_m_n, problem.MBlock, problem.NBlock);
         const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
             MakeDEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
                 e_grid_desc_m_n, problem.MBlock, problem.NBlock);
  
         // divide block work by [M, N]
         const auto block_2_ctile_map = Block2CTileMap{problem.M, problem.N, 4};
  
         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(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
                           e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
         {
             return;
         }
  
         const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
         const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
  
         // BScale struct (Empty)
         using BScale        = typename BlockwiseGemmPipe::Empty;
         auto b_scale_struct = BScale{};
  
         const index_t num_k_block_per_scale = GetKBlockPerScale();
  
         Base::template Run<decltype(a_grid_desc_ak0_m_ak1),
                            decltype(b_grid_desc_bk0_n_bk1),
                            decltype(ds_grid_desc_mblock_mperblock_nblock_nperblock),
                            decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
                            decltype(b_scale_struct),
                            HasMainKBlockLoop,
                            EGlobalMemoryDataOperation,
                            TailNum>(p_a_grid,
                                     p_b_grid,
                                     p_ds_grid,
                                     p_e_grid,
                                     p_shared,
                                     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,
                                     a_element_op,
                                     b_element_op,
                                     cde_element_op,
                                     block_m_id,
                                     block_n_id,
                                     num_k_block_per_scale,
                                     b_scale_struct);
     }
  
     // Wrapper function to have __global__ function in common
     // between gemm_universal, b_scale, ab_scale, etc.
     template <bool HasMainKBlockLoop,
               InMemoryDataOperationEnum EGlobalMemoryDataOperation,
               TailNumber TailNum>
     __device__ static void
     Run(void* p_shared, const SplitKBatchOffset& splitk_batch_offset, Argument& karg)
     {
         Run<HasMainKBlockLoop, EGlobalMemoryDataOperation, TailNum>(
             karg.p_a_grid + splitk_batch_offset.a_k_split_offset,
             karg.p_b_grid + splitk_batch_offset.b_k_split_offset,
             karg.p_ds_grid, //; + splitk_batch_offset.c_reduce_offset,
             karg.p_e_grid + splitk_batch_offset.c_reduce_offset,
             p_shared,
             karg,
             karg.a_element_op,
             karg.b_element_op,
             karg.cde_element_op);
     }
 };
  
 } // namespace ck