DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType > Struct Template Reference

DeviceGemm_Xdl_CShuffle_LdsDirectLoad&lt; ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType &gt; Struct Template Reference#

Composable Kernel: ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType > Struct Template Reference
Classes | Public Types | Public Member Functions | Static Public Member Functions | Static Public Attributes | List of all members
ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType > Struct Template Reference

#include <device_gemm_xdl_cshuffle_lds_direct_load.hpp>

Inheritance diagram for ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >:
ck::tensor_operation::device::DeviceGemm< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation > ck::tensor_operation::device::BaseOperator

Classes

struct  Invoker

Public Types

template<index_t NXdlPerWave_>
using GridwiseGemmBase
using GridwiseGemm64 = GridwiseGemmBase<math::max(NXdlPerWave64, 1)>
using GridwiseGemm32 = GridwiseGemmBase<NXdlPerWave32>
using Argument = typename GridwiseGemm64::Argument

Public Member Functions

bool IsSupportedArgument (const BaseArgument *p_arg) override
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
std::unique_ptr< BaseInvoker > MakeInvokerPointer () override
std::string GetTypeString () const override
Public Member Functions inherited from ck::tensor_operation::device::BaseOperator
 BaseOperator ()=default
 BaseOperator (const BaseOperator &)=default
BaseOperator & operator= (const BaseOperator &)=default
virtual std::string GetInstanceString () const
virtual std::string GetTypeIdName () const
virtual std::optional< std::string > GetObjectName () const
virtual std::optional< std::string > GetTemplateInfo () const
virtual std::string GetTypeIdHashCode () const
virtual size_t GetWorkSpaceSize (const BaseArgument *) const
virtual void SetWorkSpacePointer (BaseArgument *p_arg, void *p_workspace, const StreamConfig &=StreamConfig{}) const
virtual ~BaseOperator ()

Static Public Member Functions

static bool IsSupportedArgument (const Argument &arg)
static auto MakeArgument (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)
static auto MakeInvoker ()

Static Public Attributes

static GET_NXDL_PER_WAVE_IMPL constexpr auto NXdlPerWave64 = GetNXdlPerWave<true>()
static constexpr auto NXdlPerWave32 = GetNXdlPerWave<false>()
static constexpr auto I1 = Number<1>{}

Member Typedef Documentation

◆ Argument

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
using ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::Argument = typename GridwiseGemm64::Argument

◆ GridwiseGemm32

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
using ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::GridwiseGemm32 = GridwiseGemmBase<NXdlPerWave32>

◆ GridwiseGemm64

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
using ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::GridwiseGemm64 = GridwiseGemmBase<math::max(NXdlPerWave64, 1)>

◆ GridwiseGemmBase

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
template<index_t NXdlPerWave_>
using ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::GridwiseGemmBase

Member Function Documentation

◆ GetTypeString()

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
std::string ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::GetTypeString ( ) const
inlineoverridevirtual

Reimplemented from ck::tensor_operation::device::BaseOperator.

◆ IsSupportedArgument() [1/2]

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
bool ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::IsSupportedArgument ( const Argument & arg)
inlinestatic

◆ IsSupportedArgument() [2/2]

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
bool ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::IsSupportedArgument ( const BaseArgument * p_arg)
inlineoverridevirtual

Reimplemented from ck::tensor_operation::device::BaseOperator.

◆ MakeArgument()

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
auto ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::MakeArgument ( 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 )
inlinestatic

◆ MakeArgumentPointer()

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
std::unique_ptr< BaseArgument > ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::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 )
inlineoverridevirtual

Implements ck::tensor_operation::device::DeviceGemm< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation >.

◆ MakeInvoker()

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
auto ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::MakeInvoker ( )
inlinestatic

◆ MakeInvokerPointer()

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
std::unique_ptr< BaseInvoker > ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::MakeInvokerPointer ( )
inlineoverridevirtual

Implements ck::tensor_operation::device::DeviceGemm< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation >.

Member Data Documentation

◆ I1

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
auto ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::I1 = Number<1>{}
staticconstexpr

◆ NXdlPerWave32

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
auto ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::NXdlPerWave32 = GetNXdlPerWave<false>()
staticconstexpr

◆ NXdlPerWave64

template<typename ALayout, typename BLayout, typename ELayout, typename ADataType, typename BDataType, typename EDataType, typename AccDataType, typename CShuffleDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, 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 ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferScalarPerVector, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferScalarPerVector, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v4, typename ComputeDataType = EDataType>
GET_NXDL_PER_WAVE_IMPL constexpr auto ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle_LdsDirectLoad< ALayout, BLayout, ELayout, ADataType, BDataType, EDataType, AccDataType, CShuffleDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferScalarPerVector, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferScalarPerVector, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEBlockTransferScalarPerVector_NPerBlock, LoopSched, PipelineVer, ComputeDataType >::NXdlPerWave64 = GetNXdlPerWave<true>()
staticconstexpr
The documentation for this struct was generated from the following file: