/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v2.hpp Source File

/home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v2.hpp Source File#

Composable Kernel: /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v2.hpp Source File
Go to the documentation of this 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/grid/gridwise_gemm_xdl_cshuffle_v2.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
  
 namespace ck {
 namespace tensor_operation {
 namespace device {
  
 // Note: inter-wave loop scheduler is rolled out to c-shuffle version first. Becuase non c-shuffle
 // version currently has compiler issues with register spill which further causes validation
 // failures.
 template <typename ALayout,
           typename BLayout,
           typename CLayout,
           typename ADataType,
           typename BDataType,
           typename CDataType,
           typename GemmAccDataType,
           typename CShuffleDataType,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CElementwiseOperation,
           GemmSpecialization GemmSpec,
           index_t NumGemmKPrefetchStage,
           index_t BlockSize,
           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,
           LoopScheduler LoopSched     = make_default_loop_scheduler(),
           PipelineVersion PipelineVer = PipelineVersion::v1,
           typename ComputeTypeA       = CDataType,
           typename ComputeTypeB       = ComputeTypeA>
 struct DeviceGemm_Xdl_CShuffleV2 : public DeviceGemm<ALayout,
                                                      BLayout,
                                                      CLayout,
                                                      ADataType,
                                                      BDataType,
                                                      CDataType,
                                                      AElementwiseOperation,
                                                      BElementwiseOperation,
                                                      CElementwiseOperation>
 {
     using DeviceOp = DeviceGemm_Xdl_CShuffleV2;
     GET_NXDL_PER_WAVE_IMPL
     static constexpr auto NXdlPerWave64 = GetNXdlPerWave<true>();
     static constexpr auto NXdlPerWave32 = GetNXdlPerWave<false>();
  
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
     static constexpr auto I2 = Number<2>{};
  
     // GridwiseGemm
     template <index_t NXdlPerWave_>
     using GridwiseGemmBase = GridwiseGemm_xdl_cshuffle_v2<
         ALayout,
         BLayout,
         CLayout,
         ADataType,
         BDataType,
         GemmAccDataType,
         CShuffleDataType,
         CDataType,
         AElementwiseOperation,
         BElementwiseOperation,
         CElementwiseOperation,
         GemmSpec,
         InMemoryDataOperationEnum::Set,
         NumGemmKPrefetchStage,
         BlockSize,
         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,
         LoopSched,
         PipelineVer,
         ComputeTypeA,
         ComputeTypeB>;
     using GridwiseGemm64 = GridwiseGemmBase<math::max(NXdlPerWave64, 1)>;
     using GridwiseGemm32 = GridwiseGemmBase<NXdlPerWave32>;
  
     using Argument = typename GridwiseGemm64::Argument;
  
     // 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);
  
             float ave_time = 0;
             const auto K   = GridwiseGemm::CalculateAK0(arg.K) * AK1;
  
             if(GridwiseGemm::CalculateKBlockLoopTailNum(K) == 3)
             {
                 const auto kernel = kernel_gemm_xdl_cshuffle_v2<GridwiseGemm, true>;
                 ave_time          = launch_and_time_kernel(
                     stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
             }
             else
             {
                 const auto kernel = kernel_gemm_xdl_cshuffle_v2<GridwiseGemm, true, 2>;
                 ave_time          = launch_and_time_kernel(
                     stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, arg);
             }
  
             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((arg.K % AK1 != 0 || arg.K % BK1 != 0) && !(GemmSpec == GemmSpecialization::MKPadding ||
                                                        GemmSpec == GemmSpecialization::NKPadding ||
                                                        GemmSpec == GemmSpecialization::MNKPadding ||
                                                        GemmSpec == GemmSpecialization::KPadding))
         {
             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 ADataType* p_a,
                              const BDataType* p_b,
                              CDataType* p_c,
                              index_t M,
                              index_t N,
                              index_t K,
                              index_t StrideA,
                              index_t StrideB,
                              index_t StrideC,
                              AElementwiseOperation,
                              BElementwiseOperation,
                              CElementwiseOperation)
     {
         return Argument{p_a, p_b, p_c, M, N, K, StrideA, StrideB, StrideC};
     }
  
     static auto MakeInvoker() { return Invoker{}; }
  
     // polymorphic
     std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
                                                       const void* p_b,
                                                       void* p_c,
                                                       index_t M,
                                                       index_t N,
                                                       index_t K,
                                                       index_t StrideA,
                                                       index_t StrideB,
                                                       index_t StrideC,
                                                       AElementwiseOperation,
                                                       BElementwiseOperation,
                                                       CElementwiseOperation) override
     {
         return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
                                           static_cast<const BDataType*>(p_b),
                                           static_cast<CDataType*>(p_c),
                                           M,
                                           N,
                                           K,
                                           StrideA,
                                           StrideB,
                                           StrideC);
     }
  
     // 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<LoopScheduler, std::string> LoopSchedToString{
             {LoopScheduler::Default, "Default"}, {LoopScheduler::Interwave, "Interwave"}};
  
         std::map<PipelineVersion, std::string> PipelineVersionToString{{PipelineVersion::v1, "v1"},
                                                                        {PipelineVersion::v2, "v2"}};
  
         // clang-format off
         str << "DeviceGemm_Xdl_CShuffleV2"
             << "<"
             << getGemmSpecializationString(GemmSpec) << ", "
             << BlockSize << ", "
             << MPerBlock << ", "
             << NPerBlock << ", "
             << KPerBlock << ", "
             << AK1 << ", "
             << BK1 << ", "
             << MPerXDL << ", "
             << NPerXDL << ", "
             << MXdlPerWave << ", "
             << NXdlPerWave << ", "
             << ABlockTransferSrcScalarPerVector << ", "
             << BBlockTransferSrcScalarPerVector << ", "
             << CShuffleMXdlPerWavePerShuffle << ", "
             << CShuffleNXdlPerWavePerShuffle
             << ">"
             << " LoopScheduler: "
             << LoopSchedToString[LoopSched] << ", "
             << "PipelineVersion: "
             << PipelineVersionToString[PipelineVer];
         // clang-format on
  
         return str.str();
     }
 };
  
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck