device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp Source File

device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp Source File#

Composable Kernel: device_gemm_multiple_d_xdl_cshuffle_v3_blockscale_bpreshuffle.hpp Source File
// SPDX-License-Identifier: MIT
// Copyright (c) 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_multiple_d_ab_scale.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/host_utility/flush_cache.hpp"
 
namespace ck {
namespace tensor_operation {
namespace device {
 
template <typename ALayout,
          typename BLayout,
          typename DsLayout,
          typename CLayout,
          typename ADataType,
          typename AScaleDataType,
          typename BDataType,
          typename BScaleDataType,
          typename DsDataType,
          typename CDataType,
          typename GemmAccDataType,
          typename CShuffleDataType,
          typename AElementwiseOperation,
          typename BElementwiseOperation,
          typename CElementwiseOperation,
          GemmSpecialization GemmSpec,
          index_t BlockSize,
          index_t ScaleBlockM,
          index_t ScaleBlockN,
          index_t ScaleBlockK,
          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,
          typename CDEShuffleBlockTransferScalarPerVectors,
          BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave,
          BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1,
          typename ComputeTypeA                       = CDataType,
          typename ComputeTypeB                       = ComputeTypeA,
          typename LDSTypeA                           = ComputeTypeA,
          typename LDSTypeB                           = ComputeTypeB>
struct DeviceGemmMultiD_BlockScale_Xdl_CShuffle_V3_BPreshuffle
    : public DeviceGemmMultipleD_BlockScale_BPreshuffle<ALayout,
                                                        BLayout,
                                                        DsLayout,
                                                        CLayout,
                                                        ADataType,
                                                        AScaleDataType,
                                                        BDataType,
                                                        BScaleDataType,
                                                        DsDataType,
                                                        CDataType,
                                                        ScaleBlockM,
                                                        ScaleBlockN,
                                                        ScaleBlockK,
                                                        AElementwiseOperation,
                                                        BElementwiseOperation,
                                                        CElementwiseOperation>
{
    static constexpr index_t NumDTensor = DsDataType::Size();
    GET_NXDL_PER_WAVE_IMPL
    static constexpr auto NXdlPerWave64 = GetNXdlPerWave<true>();
    static constexpr auto NXdlPerWave32 = GetNXdlPerWave<false>();
 
    // GridwiseGemm
    template <index_t NXdlPerWave_>
    using GridwiseGemmBase = GridwiseGemmMultiD_blockscale_xdl_cshuffle_v3_b_preshuffle<
        ALayout,
        BLayout,
        DsLayout,
        CLayout,
        ADataType,
        BDataType,
        GemmAccDataType,
        CShuffleDataType,
        DsDataType,
        CDataType,
        AElementwiseOperation,
        BElementwiseOperation,
        CElementwiseOperation,
        GemmSpec,
        BlockSize,
        ScaleBlockM,
        ScaleBlockN,
        ScaleBlockK,
        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,
        math::min(CShuffleNXdlPerWavePerShuffle, NXdlPerWave_),
        CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
        CDEShuffleBlockTransferScalarPerVectors,
        BlkGemmPipeSched,
        BlkGemmPipelineVer,
        ComputeTypeA,
        ComputeTypeB,
        LDSTypeA,
        LDSTypeB>;
    using GridwiseGemm64 = GridwiseGemmBase<math::max(NXdlPerWave64, 1)>;
    using GridwiseGemm32 = GridwiseGemmBase<math::max(NXdlPerWave32, 1)>;
 
    using Argument = typename GridwiseGemm64::Argument;
 
    int GetPreShuffleParameters() override { return NPerXDL; }
 
    // Invoker
    struct Invoker : public BaseInvoker
    {
        template <typename GridwiseGemm>
        float RunImp(const typename GridwiseGemm::Argument& arg,
                     const StreamConfig& stream_config = StreamConfig{})
        {
            if(stream_config.log_level_ > 0)
            {
                arg.Print();
            }
 
            if(!GridwiseGemm::CheckValidity(arg))
            {
                throw std::runtime_error("wrong! GridwiseGemm has invalid setting");
            }
 
            index_t gdx, gdy, gdz;
            std::tie(gdx, gdy, gdz) = GridwiseGemm::CalculateGridSize(arg.M, arg.N, arg.KBatch);
 
            float ave_time = 0;
 
            index_t k_grain = arg.KBatch * KPerBlock;
            index_t K_split = (arg.K + k_grain - 1) / k_grain * KPerBlock;
 
            const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K_split);
 
            const auto Run = [&](const auto& kernel) {
                if(stream_config.flush_cache)
                {
                    auto arg_ = arg;
 
                    const auto a_grid_desc_ak0_m_ak1 = GridwiseGemm::MakeAGridDescriptor_AK0_M_AK1(
                        arg_.M, arg_.MPadded, arg_.K, arg_.KPadded, arg_.StrideA, arg_.AK0);
                    const auto b_grid_desc_bk0_n_bk1 = GridwiseGemm::MakeBGridDescriptor_BK0_N_BK1(
                        arg_.K, arg_.KPadded, arg_.N, arg_.NPadded, arg_.StrideB, arg_.BK0);
 
                    auto size_a_buffer =
                        a_grid_desc_ak0_m_ak1.GetElementSpaceSize() * sizeof(ADataType);
                    auto size_b_buffer =
                        b_grid_desc_bk0_n_bk1.GetElementSpaceSize() * sizeof(BDataType);
 
                    ck::utility::RotatingMemWrapper<typename GridwiseGemm::Argument> rotating_mem(
                        arg_, stream_config.rotating_count, size_a_buffer, size_b_buffer);
                    rotating_mem.Print();
 
                    auto run_flush_cache = [&]() {
                        // flush icache
                        ck::utility::flush_icache();
                        // rotating mem
                        rotating_mem.Next();
                        // clear c mem
                        if(arg_.KBatch > 1)
                            hipGetErrorString(hipMemsetAsync(arg_.p_c_grid,
                                                             0,
                                                             arg_.M * arg_.N * sizeof(CDataType),
                                                             stream_config.stream_id_));
                    };
 
                    ave_time = ck::utility::launch_and_time_kernel_with_preprocess<false>(
                        stream_config,
                        run_flush_cache,
                        kernel,
                        dim3(gdx, gdy, gdz),
                        dim3(BlockSize),
                        0,
                        arg_);
                }
                else
                {
                    if(arg.KBatch > 1)
                        hipGetErrorString(hipMemsetAsync(arg.p_c_grid,
                                                         0,
                                                         arg.M * arg.N * sizeof(CDataType),
                                                         stream_config.stream_id_));
 
                    ave_time = launch_and_time_kernel(
                        stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
                }
            };
 
            // unconditional 2 to remove agpr usage
            constexpr index_t minimum_occupancy = 2;
 
            if(has_main_k_block_loop)
            {
                // Tail number always full
                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
                {
                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle<
                                GridwiseGemm,
                                true,
                                InMemoryDataOperationEnum::Set,
                                minimum_occupancy,
                                TailNumber::Odd>;
                        Run(kernel);
                    }
                    else
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle<
                                GridwiseGemm,
                                true,
                                InMemoryDataOperationEnum::Set,
                                minimum_occupancy,
                                TailNumber::Even>;
                        Run(kernel);
                    }
                }
                else if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v3)
                {
                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle_2lds<
                                GridwiseGemm,
                                true,
                                InMemoryDataOperationEnum::Set,
                                minimum_occupancy,
                                TailNumber::Odd>;
                        Run(kernel);
                    }
                    else
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle_2lds<
                                GridwiseGemm,
                                true,
                                InMemoryDataOperationEnum::Set,
                                minimum_occupancy,
                                TailNumber::Even>;
                        Run(kernel);
                    }
                }
            }
            else
            {
                // Tail number always 1
                if constexpr(BlkGemmPipelineVer == BlockGemmPipelineVersion::v1)
                {
                    if(GridwiseGemm::CalculateKBlockLoopTailNum(K_split) == TailNumber::Odd)
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle<
                                GridwiseGemm,
                                false,
                                InMemoryDataOperationEnum::Set,
                                minimum_occupancy,
                                TailNumber::Odd>;
                        Run(kernel);
                    }
                    else
                    {
                        const auto kernel =
                            kernel_gemm_xdl_cshuffle_v3_multi_d_blockscale_b_preshuffle<
                                GridwiseGemm,
                                false,
                                InMemoryDataOperationEnum::Set,
                                minimum_occupancy,
                                TailNumber::Even>;
                        Run(kernel);
                    }
                }
            }
            return ave_time;
        }
 
        INVOKER_RUN3_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 constexpr bool IsValidCompilationParameter()
    {
        // TODO: properly implement this check
        return true;
    }
 
    static bool IsSupportedArgument(const Argument& arg)
    {
        if(!ck::is_xdl_wmma_supported<ComputeTypeA, ComputeTypeB, MPerXDL, NPerXDL>())
        {
            return false;
        }
        // if(ScaleBlockM % MPerBlock != 0 || ScaleBlockN % NPerBlock != 0 || ScaleBlockK !=
        // KPerBlock)
        // {
        //     return false;
        // }
        if(is_gfx11_supported() && arg.KBatch > 1)
        {
            return false;
        }
        if(!is_bf16_atomic_supported() && std::is_same_v<CDataType, ck::bhalf_t> && arg.KBatch > 1)
        {
            return false;
        }
 
        if((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
                                                       GemmSpec == GemmSpecialization::NKPadding ||
                                                       GemmSpec == GemmSpecialization::MNKPadding ||
                                                       GemmSpec == GemmSpecialization::KPadding))
        {
            return false;
        }
 
        // Padding to release this restriction
        if(arg.N % NPerBlock != 0 || arg.K % KPerBlock != 0)
        {
            return false;
        }
 
        if(get_warp_size() == 64)
        {
            if constexpr(NXdlPerWave64 > 0)
            {
                return GridwiseGemm64::CheckValidity(arg);
            }
        }
        else
        {
            if constexpr(NXdlPerWave32 > 0)
            {
                return GridwiseGemm32::CheckValidity(
                    reinterpret_cast<const typename GridwiseGemm32::Argument&>(arg));
            }
        }
        return false;
    }
 
    // polymorphic
    bool IsSupportedArgument(const BaseArgument* p_arg) override
    {
        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
    }
 
    static auto MakeArgument(const void* p_a,
                             const void* p_b,
                             std::array<const void*, NumDTensor> p_ds,
                             void* p_c,
                             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 StrideC,
                             const void* p_a_scale,
                             const void* p_b_scale,
                             AElementwiseOperation a_element_op,
                             BElementwiseOperation b_element_op,
                             CElementwiseOperation c_element_op)
    {
        return Argument{static_cast<const ADataType*>(p_a),
                        static_cast<const BDataType*>(p_b),
                        p_ds,
                        static_cast<CDataType*>(p_c),
                        M,
                        N,
                        K,
                        StrideA,
                        StrideB,
                        StrideDs,
                        StrideC,
                        static_cast<const AScaleDataType*>(p_a_scale),
                        static_cast<const BScaleDataType*>(p_b_scale),
                        1,
                        a_element_op,
                        b_element_op,
                        c_element_op};
    }
 
    static auto MakeInvoker() { return Invoker{}; }
 
    // polymorphic
    std::unique_ptr<BaseArgument>
    MakeArgumentPointer(const void* p_a,
                        const void* p_b,
                        std::array<const void*, NumDTensor> p_ds,
                        void* p_c,
                        const index_t M,
                        const index_t N,
                        const index_t K,
                        const index_t StrideA,
                        const index_t StrideB,
                        const std::array<ck::index_t, NumDTensor> StrideDs,
                        const index_t StrideC,
                        const void* p_a_scale,
                        const void* p_b_scale,
                        AElementwiseOperation a_element_op,
                        BElementwiseOperation b_element_op,
                        CElementwiseOperation c_element_op) override
    {
        return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
                                          static_cast<const BDataType*>(p_b),
                                          p_ds,
                                          static_cast<CDataType*>(p_c),
                                          M,
                                          N,
                                          K,
                                          StrideA,
                                          StrideB,
                                          StrideDs,
                                          StrideC,
                                          static_cast<const AScaleDataType*>(p_a_scale),
                                          static_cast<const BScaleDataType*>(p_b_scale),
                                          1,
                                          a_element_op,
                                          b_element_op,
                                          c_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();
 
        std::map<BlockGemmPipelineScheduler, std::string> BlkGemmPipelineSchedulerToString{
            {BlockGemmPipelineScheduler::Intrawave, "Intrawave"},
            {BlockGemmPipelineScheduler::Interwave, "Interwave"}};
 
        std::map<BlockGemmPipelineVersion, std::string> BlkGemmPipelineVersionToString{
            {BlockGemmPipelineVersion::v1, "v1"},
            {BlockGemmPipelineVersion::v2, "v2"},
            {BlockGemmPipelineVersion::v3, "v3"}};
 
        // clang-format off
        str << "DeviceGemmXdlUniversal"
            << "<"
            << getGemmSpecializationString(GemmSpec) << ", "
            << std::string(ALayout::name)[0]
            << std::string(BLayout::name)[0]
            << std::string(CLayout::name)[0]
            << ">"
            << " BlkSize: "
            << BlockSize << ", "
            << "BlkTile: "
            << MPerBlock<<"x"<<NPerBlock<<"x"<<KPerBlock << ", "
            << "WaveTile: "
            << MPerXDL<<"x"<<NPerXDL << ", "
            << "WaveMap: "
            << MXdlPerWave<<"x" << NXdlPerWave<<", "
            << "VmemReadVec: "
            << ABlockTransferSrcScalarPerVector<<"x"<<BBlockTransferSrcScalarPerVector<<", "
            << "BlkGemmPipelineScheduler: "
            << BlkGemmPipelineSchedulerToString[BlkGemmPipeSched] << ", "
            << "BlkGemmPipelineVersion: "
            << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
            << "BlkGemmPipelinePrefetchStages: "
            << GridwiseGemm64::BlockwiseGemmPipe::PrefetchStages;
        // clang-format on
 
        return str.str();
    }
};
 
} // namespace device
} // namespace tensor_operation
} // namespace ck