device_gemm_xdl_waveletmodel_cshuffle.hpp Source File

device_gemm_xdl_waveletmodel_cshuffle.hpp Source File#

Composable Kernel: device_gemm_xdl_waveletmodel_cshuffle.hpp Source File
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
 
#pragma once
 
#include <iostream>
#include <sstream>
 
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_gemm.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_waveletmodel_cshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
 
namespace ck {
 
template <typename GridwiseGemm,
          typename ABDataType,
          typename EDataType,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename EElementwiseOperation,
          typename AGridDesc_AK0_M_AK1,
          typename BGridDesc_BK0_N_BK1,
          typename EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
          typename Block2ETileMap,
          bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_WAVELET_MAX_THREAD_PER_BLOCK, CK_WAVELET_MIN_BLOCK_PER_CU)
#endif
    kernel_gemm_xdl_waveletmodel_cshuffle(const ABDataType* __restrict__ p_a_grid,
                                          const ABDataType* __restrict__ p_b_grid,
                                          EDataType* __restrict__ p_e_grid,
                                          const AElementwiseOperation a_element_op,
                                          const BElementwiseOperation b_element_op,
                                          const EElementwiseOperation e_element_op,
                                          const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
                                          const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
                                          const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
                                              e_grid_desc_mblock_mperblock_nblock_nperblock,
                                          const Block2ETileMap block_2_etile_map)
{
#if defined(__gfx9__) || defined(__gfx11__) || defined(__gfx12__)
    if constexpr(GridwiseGemm::template IsValidCompilationParameter<>())
    {
        __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
        GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
                                                      p_b_grid,
                                                      p_e_grid,
                                                      p_shared,
                                                      a_element_op,
                                                      b_element_op,
                                                      e_element_op,
                                                      a_grid_desc_ak0_m_ak1,
                                                      b_grid_desc_bk0_n_bk1,
                                                      e_grid_desc_mblock_mperblock_nblock_nperblock,
                                                      block_2_etile_map);
    }
#else
    ignore = p_a_grid;
    ignore = p_b_grid;
    ignore = p_e_grid;
    ignore = a_element_op;
    ignore = b_element_op;
    ignore = e_element_op;
    ignore = a_grid_desc_ak0_m_ak1;
    ignore = b_grid_desc_bk0_n_bk1;
    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock;
    ignore = block_2_etile_map;
#endif
}
 
} // namespace ck
 
namespace ck {
namespace tensor_operation {
namespace device {
 
template <typename ALayout,
          typename BLayout,
          typename ELayout,
          typename ADataType,
          typename BDataType,
          typename GemmAcEDataType,
          typename CShuffleDataType,
          typename EDataType,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename CDEElementwiseOperation,
          GemmSpecialization GemmSpec,
          index_t NumGemmKPrefetchStage,
          index_t TileLoadThreadGroupSize,
          index_t TileMathThreadGroupSize,
          index_t MPerBlock,
          index_t NPerBlock,
          index_t KPerBlock,
          index_t AK1,
          index_t BK1,
          index_t MPerXDL,
          index_t NPerXDL,
          index_t MXdlPerWave,
          index_t NXdlPerWave,
          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
          typename ABlockTransferThreadClusterArrangeOrder,
          typename ABlockTransferSrcAccessOrder,
          index_t ABlockTransferSrcVectorDim,
          index_t ABlockTransferSrcScalarPerVector,
          index_t ABlockTransferDstScalarPerVector_AK1,
          bool ABlockLdsExtraM,
          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
          typename BBlockTransferThreadClusterArrangeOrder,
          typename BBlockTransferSrcAccessOrder,
          index_t BBlockTransferSrcVectorDim,
          index_t BBlockTransferSrcScalarPerVector,
          index_t BBlockTransferDstScalarPerVector_BK1,
          bool BBlockLdsExtraN,
          index_t CShuffleMXdlPerWavePerShuffle,
          index_t CShuffleNXdlPerWavePerShuffle,
          typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
          index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
struct DeviceGemm_Xdl_WaveletModel_CShuffle : public DeviceGemm<ALayout,
                                                                BLayout,
                                                                ELayout,
                                                                ADataType,
                                                                BDataType,
                                                                EDataType,
                                                                AElementwiseOperation,
                                                                BElementwiseOperation,
                                                                CDEElementwiseOperation>
{
    static constexpr auto BlockSize = math::max(TileLoadThreadGroupSize, TileMathThreadGroupSize);
    GET_NXDL_PER_WAVE_IMPL
    static constexpr auto NXdlPerWave64 = GetNXdlPerWave<true>();
    static constexpr auto NXdlPerWave32 = GetNXdlPerWave<false>();
 
    using DeviceOp = DeviceGemm_Xdl_WaveletModel_CShuffle;
 
    static constexpr auto I0 = Number<0>{};
    static constexpr auto I1 = Number<1>{};
    static constexpr auto I2 = Number<2>{};
 
    static constexpr auto matrix_padder =
        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
 
    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
    {
        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);
    }
 
    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, 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(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 ELay>
    static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
    {
        const auto e_grid_desc_mraw_nraw = [&]() {
            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
            {
                return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
                                                    make_tuple(StrideE, I1));
            }
            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::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);
    }
 
    using AGridDesc_M_K = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
    using BGridDesc_N_K = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
    using EGridDesc_M_N = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
 
    // GridwiseGemm
    template <index_t NXdlPerWave_>
    using GridwiseGemmBase = GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle<
        ADataType, // TODO: distinguish A/B datatype
        GemmAcEDataType,
        CShuffleDataType,
        EDataType,
        AElementwiseOperation,
        BElementwiseOperation,
        CDEElementwiseOperation,
        InMemoryDataOperationEnum::Set,
        AGridDesc_M_K,
        BGridDesc_N_K,
        EGridDesc_M_N,
        NumGemmKPrefetchStage,
        TileLoadThreadGroupSize,
        TileMathThreadGroupSize,
        MPerBlock,
        NPerBlock,
        KPerBlock,
        AK1,
        BK1,
        MPerXDL,
        NPerXDL,
        MXdlPerWave,
        NXdlPerWave_,
        ABlockTransferThreadClusterLengths_AK0_M_AK1,
        ABlockTransferThreadClusterArrangeOrder,
        ABlockTransferSrcAccessOrder,
        ABlockTransferSrcVectorDim,
        ABlockTransferSrcScalarPerVector,
        ABlockTransferDstScalarPerVector_AK1,
        false,
        ABlockLdsExtraM,
        BBlockTransferThreadClusterLengths_BK0_N_BK1,
        BBlockTransferThreadClusterArrangeOrder,
        BBlockTransferSrcAccessOrder,
        BBlockTransferSrcVectorDim,
        BBlockTransferSrcScalarPerVector,
        BBlockTransferDstScalarPerVector_BK1,
        false,
        BBlockLdsExtraN,
        CShuffleMXdlPerWavePerShuffle,
        CShuffleNXdlPerWavePerShuffle,
        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
        CShuffleBlockTransferScalarPerVector_NPerBlock>;
    using GridwiseGemm64 = GridwiseGemmBase<math::max(NXdlPerWave64, 1)>;
    using GridwiseGemm32 = GridwiseGemmBase<NXdlPerWave32>;
 
    using AGridDesc_AK0_M_AK1 =
        remove_cvref_t<decltype(GridwiseGemm64::MakeDefaultAGridDescriptor_AK0_M_AK1(
            AGridDesc_M_K{}))>;
    using BGridDesc_BK0_N_BK1 =
        remove_cvref_t<decltype(GridwiseGemm64::MakeDefaultBGridDescriptor_BK0_N_BK1(
            BGridDesc_N_K{}))>;
 
    using Block2ETileMap = typename GridwiseGemm64::DefaultBlock2ETileMap;
 
    // Argument
    struct Argument : public BaseArgument
    {
        Argument(const ADataType* p_a_grid,
                 const BDataType* p_b_grid,
                 EDataType* p_e_grid,
                 index_t MRaw,
                 index_t NRaw,
                 index_t KRaw,
                 index_t StrideA,
                 index_t StrideB,
                 index_t StrideE,
                 AElementwiseOperation a_element_op,
                 BElementwiseOperation b_element_op,
                 CDEElementwiseOperation cde_element_op)
            : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
              p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
              p_e_grid_{static_cast<EDataType*>(p_e_grid)},
              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K(MRaw, KRaw, StrideA)},
              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K(KRaw, NRaw, StrideB)},
              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(MRaw, NRaw, StrideE)},
              a_grid_desc_ak0_m_ak1_{
                  GridwiseGemm64::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
              b_grid_desc_bk0_n_bk1_{
                  GridwiseGemm64::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
              block_2_etile_map_{GridwiseGemm64::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
              a_element_op_{a_element_op},
              b_element_op_{b_element_op},
              cde_element_op_{cde_element_op}
        {
        }
 
        void Print() const
        {
            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
        }
 
        //  private:
        // pointers
        const ADataType* p_a_grid_;
        const BDataType* p_b_grid_;
        EDataType* p_e_grid_;
 
        // tensor descriptors for problem definiton
        AGridDesc_M_K a_grid_desc_m_k_;
        BGridDesc_N_K b_grid_desc_n_k_;
        EGridDesc_M_N e_grid_desc_m_n_;
 
        // tensor descriptors for block/thread-wise copy
        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
 
        // block-to-e-tile map
        Block2ETileMap block_2_etile_map_;
 
        // element-wise op
        AElementwiseOperation a_element_op_;
        BElementwiseOperation b_element_op_;
        CDEElementwiseOperation cde_element_op_;
    };
 
    // Invoker
    struct Invoker : public BaseInvoker
    {
        using Argument = DeviceOp::Argument;
 
        template <typename GridwiseGemm>
        float RunImp(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
        {
#if 0
            {
                std::cout << "arg.a_grid_desc_ak0_m_ak1_{"
                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) << ", "
                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I1) << ", "
                          << arg.a_grid_desc_ak0_m_ak1_.GetLength(I2) << "}" << std::endl;
 
                std::cout << "arg.b_grid_desc_bk0_n_bk1_{"
                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I0) << ", "
                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I1) << ", "
                          << arg.b_grid_desc_bk0_n_bk1_.GetLength(I2) << "}" << std::endl;
 
                std::cout << "arg.e_grid_desc_m_n_{ " << arg.e_grid_desc_m_n_.GetLength(I0) << ", "
                          << arg.e_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
            }
#endif
 
            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
                                            arg.b_grid_desc_n_k_,
                                            arg.e_grid_desc_m_n_,
                                            arg.block_2_etile_map_))
            {
                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
            }
            auto e_grid_desc_mblock_mperblock_nblock_nperblock =
                GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
                    arg.e_grid_desc_m_n_);
 
            const index_t grid_size = GridwiseGemm::CalculateGridSize(arg.e_grid_desc_m_n_);
            const auto K            = arg.a_grid_desc_m_k_.GetLength(I1);
 
            auto launch_kernel = [&](auto has_main_k_block_loop) {
                constexpr bool has_main_loop = has_main_k_block_loop.value;
 
                const auto kernel = kernel_gemm_xdl_waveletmodel_cshuffle<
                    GridwiseGemm,
                    ADataType, // TODO: distiguish A/B datatype
                    EDataType,
                    AElementwiseOperation,
                    BElementwiseOperation,
                    CDEElementwiseOperation,
                    DeviceOp::AGridDesc_AK0_M_AK1,
                    DeviceOp::BGridDesc_BK0_N_BK1,
                    typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
                    typename GridwiseGemm::DefaultBlock2ETileMap,
                    has_main_loop>;
 
                return launch_and_time_kernel(
                    stream_config,
                    kernel,
                    dim3(grid_size),
                    dim3(TileLoadThreadGroupSize + TileMathThreadGroupSize),
                    0,
                    arg.p_a_grid_,
                    arg.p_b_grid_,
                    arg.p_e_grid_,
                    arg.a_element_op_,
                    arg.b_element_op_,
                    arg.cde_element_op_,
                    arg.a_grid_desc_ak0_m_ak1_,
                    arg.b_grid_desc_bk0_n_bk1_,
                    e_grid_desc_mblock_mperblock_nblock_nperblock,
                    arg.block_2_etile_map_);
            };
 
            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
            {
                return launch_kernel(integral_constant<bool, true>{});
            }
            else
            {
                return launch_kernel(integral_constant<bool, false>{});
            }
        }
 
        INVOKER_RUN_IMPL
 
        // polymorphic
        float Run(const BaseArgument* p_arg,
                  const StreamConfig& stream_config = StreamConfig{}) override
        {
            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
        }
    };
 
    static bool IsSupportedArgument(const Argument& arg)
    {
        if(!ck::is_xdl_wmma_supported<ADataType, BDataType, MPerXDL, NPerXDL>())
        {
            return false;
        }
        if(get_warp_size() == 64)
        {
            if constexpr(NXdlPerWave64 > 0)
            {
                return GridwiseGemm64::CheckValidity(arg.a_grid_desc_m_k_,
                                                     arg.b_grid_desc_n_k_,
                                                     arg.e_grid_desc_m_n_,
                                                     arg.block_2_etile_map_);
            }
        }
        else
        {
            if constexpr(NXdlPerWave32 > 0)
            {
                return GridwiseGemm32::CheckValidity(arg.a_grid_desc_m_k_,
                                                     arg.b_grid_desc_n_k_,
                                                     arg.e_grid_desc_m_n_,
                                                     arg.block_2_etile_map_);
            }
        }
        return false;
    }
 
    // polymorphic
    bool IsSupportedArgument(const BaseArgument* p_arg) override
    {
        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
    }
 
    static auto MakeArgument(const ADataType* p_a,
                             const BDataType* p_b,
                             EDataType* p_e,
                             index_t MRaw,
                             index_t NRaw,
                             index_t KRaw,
                             index_t StrideA,
                             index_t StrideB,
                             index_t StrideE,
                             AElementwiseOperation a_element_op,
                             BElementwiseOperation b_element_op,
                             CDEElementwiseOperation cde_element_op)
    {
        return Argument{p_a,
                        p_b,
                        p_e,
                        MRaw,
                        NRaw,
                        KRaw,
                        StrideA,
                        StrideB,
                        StrideE,
                        a_element_op,
                        b_element_op,
                        cde_element_op};
    }
 
    static auto MakeInvoker() { return Invoker{}; }
 
    // polymorphic
    std::unique_ptr<BaseArgument>
    MakeArgumentPointer(const void* p_a,
                        const void* p_b,
                        void* p_e,
                        index_t MRaw,
                        index_t NRaw,
                        index_t KRaw,
                        index_t StrideA,
                        index_t StrideB,
                        index_t StrideE,
                        AElementwiseOperation a_element_op,
                        BElementwiseOperation b_element_op,
                        CDEElementwiseOperation cde_element_op) override
    {
        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
                                          static_cast<const BDataType*>(p_b),
                                          static_cast<EDataType*>(p_e),
                                          MRaw,
                                          NRaw,
                                          KRaw,
                                          StrideA,
                                          StrideB,
                                          StrideE,
                                          a_element_op,
                                          b_element_op,
                                          cde_element_op);
    }
 
    // polymorphic
    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
    {
        return std::make_unique<Invoker>(Invoker{});
    }
 
    // polymorphic
    std::string GetTypeString() const override
    {
        auto str = std::stringstream();
 
        // clang-format off
        str << "DeviceGemm_Xdl_WaveletModel_CShuffle"
            << "<"
        << TileLoadThreadGroupSize << ", "
            << TileMathThreadGroupSize << ", "
            << MPerBlock << ", "
            << NPerBlock << ", "
            << KPerBlock << ", "
            << AK1 << ", "
            << BK1
            << ">";
        // clang-format on
 
        return str.str();
    }
};
 
} // namespace device
} // namespace tensor_operation
} // namespace ck