GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB > Struct Template Reference

• ck
• GridwiseMoeGemm

Classes | Public Types | Static Public Member Functions | Static Public Attributes | List of all members

ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB > Struct Template Reference

#include <gridwise_moe_gemm.hpp>

Classes
struct	Argument
struct	Problem
struct	SplitKBatchOffset

Public Types
using	mfma_selector = MfmaSelector< ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB >
using	DsGridPointer = decltype(MakeDsGridPointer())
using	ThisThreadBlock = ThisThreadBlock< BlockSize >
using	BlockwiseGemmPipe = remove_cvref_t< decltype(BlockGemmBPreshufflePipeline_Selector< BlkGemmPipelineVer, BlkGemmPipeSched, BlockSize, ADataType, BDataType, ComputeTypeA, AccDataType, decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()), decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()), decltype(MakeAMmaTileDescriptor_M0_M1_M2_K(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())), decltype(MakeBMmaTileDescriptor_N0_N1_N2_K(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())), ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, KPack, IsInputGemm >())>

Static Public Member Functions
static constexpr auto	MakeDsGridPointer ()
static __host__ auto	CalculateGridSize (index_t M, index_t N)
__host__ static __device__ auto	CalculateMPadded (index_t M)
__host__ static __device__ auto	CalculateNPadded (index_t N)
__host__ static __device__ auto	CalculateBN0Shuffled (index_t N)
__host__ static __device__ auto	CalculateBK0Shuffled (index_t K)
__host__ static __device__ auto	CalculateKPadded (index_t K)
__host__ static __device__ auto	CalculateAK0Padded (index_t K, index_t K_Batch=1)
__host__ static __device__ auto	CalculateBK0Padded (index_t K, index_t K_Batch=1)
__host__ static __device__ auto	CalculateKPadded (index_t K, index_t K_Batch=1)
__host__ static __device__ auto	CalculateKRead (index_t K, index_t K_Batch=1)
__host__ static __device__ auto	CalculateMBlock (index_t M)
__host__ static __device__ auto	CalculateNBlock (index_t N)
template<index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1 >
__host__ static constexpr __device__ auto	MakeGemmMmaTileDescriptor (const TileDesc_K0_MN_K1 &)
__host__ static __device__ auto	MakeAGridDescriptor_AK0_M_AK1 (IndexType M, IndexType MPad, IndexType K, IndexType KPad, IndexType StrideA, IndexType AK0)
__host__ static __device__ auto	MakeBGridDescriptor_Preshuffled (index_t N0, index_t K0)
__host__ static __device__ auto	MakeBGridDescriptor_BK0_N_BK1 (index_t K, index_t KPad, index_t N, index_t NPad, index_t StrideB, index_t BK0)
template<typename ABlockDesc_AK0_M_AK1 >
__host__ static constexpr __device__ auto	MakeAMmaTileDescriptor_M0_M1_M2_K (const ABlockDesc_AK0_M_AK1 &)
template<typename BBlockDesc_BK0_N_BK1 >
__host__ static constexpr __device__ auto	MakeBMmaTileDescriptor_N0_N1_N2_K (const BBlockDesc_BK0_N_BK1 &)
template<typename ELayout >
__host__ static __device__ auto	MakeCGridDescriptor_M_N (IndexType M, IndexType MPad, IndexType N, IndexType NPad, IndexType StrideC)
template<typename DLayout >
__host__ static __device__ auto	MakeDGridDescriptor_M_N (index_t M, index_t MPad, index_t N, index_t NPad, index_t StrideC)
__host__ static __device__ auto	MakeDsGridDescriptor_M_N (index_t M, index_t MPad, index_t N, index_t NPad, std::array< index_t, NumDTensor > StrideDs)
template<typename DsGridDesc >
static constexpr __device__ auto	MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock (const DsGridDesc &ds_grid_desc_m_n, index_t MBlock, index_t NBlock)
static constexpr __device__ auto	GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1 ()
static constexpr __device__ auto	GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1 ()
static constexpr __device__ auto	GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ()
static constexpr __device__ index_t	GetSharedMemoryNumberOfByte ()
static constexpr __host__ bool	CheckValidity (const Argument &karg)
__host__ static constexpr __device__ bool	CalculateHasMainKBlockLoop (index_t K)
__host__ static constexpr __device__ TailNumber	CalculateKBlockLoopTailNum (index_t K)
template<typename CGridDesc >
static constexpr __device__ auto	MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock (const CGridDesc &c_grid_desc_m_n, index_t MBlock, index_t NBlock)
template<bool HasMainKBlockLoop, InMemoryDataOperationEnum CGlobalMemoryDataOperation, TailNumber TailNum = TailNumber::Odd>
static __device__ void	Run (const index_t *p_sorted_token_ids, const index_t *p_sorted_expert_ids, const index_t *p_max_token_id, const ADataType *p_a_grid, const BDataType *p_b_grid, DsGridPointer &p_ds_grid, CDataType *p_c_grid, void *p_shared, const Problem &problem, AElementwiseOperation a_element_op, BElementwiseOperation b_element_op, CElementwiseOperation c_element_op)
template<bool HasMainKBlockLoop, InMemoryDataOperationEnum CGlobalMemoryDataOperation, TailNumber TailNum = TailNumber::Odd>
static __device__ void	Run_2Lds (const index_t *p_sorted_token_ids, const index_t *p_sorted_expert_ids, const index_t *p_max_token_id, const ADataType *p_a_grid, const BDataType *p_b_grid, DsGridPointer &p_ds_grid, CDataType *p_c_grid, void *p_shared, void *p_shared1, const Problem &problem, AElementwiseOperation a_element_op, BElementwiseOperation b_element_op, CElementwiseOperation c_element_op)

Static Public Attributes
static constexpr auto	I0 = Number<0>{}
static constexpr auto	I1 = Number<1>{}
static constexpr auto	I2 = Number<2>{}
static constexpr auto	I3 = Number<3>{}
static constexpr auto	I4 = Number<4>{}
static constexpr auto	I5 = Number<5>{}
static constexpr auto	I6 = Number<6>{}
static constexpr auto	I7 = Number<7>{}
static constexpr auto	CShuffleBlockTransferScalarPerVector_NPerBlock
static constexpr auto	AK0Number = Number<KPerBlock / AK1Value>{}
static constexpr auto	BK0Number = Number<KPerBlock / BK1Value>{}
static constexpr auto	AK1Number = Number<AK1Value>{}
static constexpr auto	BK1Number = Number<BK1Value>{}
static constexpr auto	BlockSizeNumber = Number<BlockSize>{}
static constexpr index_t	NumDTensor = DsDataType::Size()
static constexpr index_t	KPack
static constexpr index_t	KLane
static constexpr index_t	KGroup
static constexpr index_t	KRepeat = KPerBlock / KLane / (KPack / KGroup)
static constexpr index_t	NLane = NPerXdl
static constexpr index_t	NWave = NPerBlock / NPerXdl / NXdlPerWave
static constexpr index_t	SortedTileSize = MPerBlock
static constexpr index_t	APackedSize
static constexpr index_t	BPackedSize

Member Typedef Documentation

◆ BlockwiseGemmPipe

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

using ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::BlockwiseGemmPipe = remove_cvref_t<decltype(BlockGemmBPreshufflePipeline_Selector< BlkGemmPipelineVer, BlkGemmPipeSched, BlockSize, ADataType, BDataType, ComputeTypeA, AccDataType, decltype(GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()), decltype(GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()), decltype(MakeAMmaTileDescriptor_M0_M1_M2_K( GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1())), decltype(MakeBMmaTileDescriptor_N0_N1_N2_K( GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1())), ABlockTransferSrcScalarPerVector, BBlockTransferSrcScalarPerVector, MPerBlock, NPerBlock, KPerBlock, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, KPack, IsInputGemm>())>

◆ DsGridPointer

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

using ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::DsGridPointer = decltype(MakeDsGridPointer())

◆ mfma_selector

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

using ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::mfma_selector = MfmaSelector<ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB>

◆ ThisThreadBlock

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

using ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::ThisThreadBlock = ThisThreadBlock<BlockSize>

Member Function Documentation

◆ CalculateAK0Padded()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateAK0Padded ( index_t  K, index_t  K_Batch = 1  )

inlinestatic

◆ CalculateBK0Padded()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateBK0Padded ( index_t  K, index_t  K_Batch = 1  )

inlinestatic

◆ CalculateBK0Shuffled()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateBK0Shuffled ( index_t  K )

inlinestatic

◆ CalculateBN0Shuffled()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateBN0Shuffled ( index_t  N )

inlinestatic

◆ CalculateGridSize()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

static __host__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateGridSize ( index_t  M, index_t  N  )

inlinestatic

◆ CalculateHasMainKBlockLoop()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static constexpr __device__ bool ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateHasMainKBlockLoop ( index_t  K )

inlinestaticconstexpr

◆ CalculateKBlockLoopTailNum()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static constexpr __device__ TailNumber ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateKBlockLoopTailNum ( index_t  K )

inlinestaticconstexpr

◆ CalculateKPadded() [1/2]

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateKPadded ( index_t  K )

inlinestatic

◆ CalculateKPadded() [2/2]

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateKPadded ( index_t  K, index_t  K_Batch = 1  )

inlinestatic

◆ CalculateKRead()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateKRead ( index_t  K, index_t  K_Batch = 1  )

inlinestatic

◆ CalculateMBlock()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateMBlock ( index_t  M )

inlinestatic

◆ CalculateMPadded()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateMPadded ( index_t  M )

inlinestatic

◆ CalculateNBlock()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateNBlock ( index_t  N )

inlinestatic

◆ CalculateNPadded()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CalculateNPadded ( index_t  N )

inlinestatic

◆ CheckValidity()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

static constexpr __host__ bool ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CheckValidity ( const Argument &  karg )

inlinestaticconstexpr

◆ GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

static constexpr __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1 ( )

inlinestaticconstexpr

◆ GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

static constexpr __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1 ( )

inlinestaticconstexpr

◆ GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

static constexpr __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ( )

inlinestaticconstexpr

◆ GetSharedMemoryNumberOfByte()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

static constexpr __device__ index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::GetSharedMemoryNumberOfByte ( )

inlinestaticconstexpr

◆ MakeAGridDescriptor_AK0_M_AK1()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeAGridDescriptor_AK0_M_AK1 ( IndexType  M, IndexType  MPad, IndexType  K, IndexType  KPad, IndexType  StrideA, IndexType  AK0  )

inlinestatic

◆ MakeAMmaTileDescriptor_M0_M1_M2_K()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

template<typename ABlockDesc_AK0_M_AK1 >

__host__ static constexpr __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeAMmaTileDescriptor_M0_M1_M2_K ( const ABlockDesc_AK0_M_AK1 &  )

inlinestaticconstexpr

◆ MakeBGridDescriptor_BK0_N_BK1()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeBGridDescriptor_BK0_N_BK1 ( index_t  K, index_t  KPad, index_t  N, index_t  NPad, index_t  StrideB, index_t  BK0  )

inlinestatic

◆ MakeBGridDescriptor_Preshuffled()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeBGridDescriptor_Preshuffled ( index_t  N0, index_t  K0  )

inlinestatic

◆ MakeBMmaTileDescriptor_N0_N1_N2_K()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

template<typename BBlockDesc_BK0_N_BK1 >

__host__ static constexpr __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeBMmaTileDescriptor_N0_N1_N2_K ( const BBlockDesc_BK0_N_BK1 &  )

inlinestaticconstexpr

◆ MakeCGridDescriptor_M_N()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

template<typename ELayout >

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeCGridDescriptor_M_N ( IndexType  M, IndexType  MPad, IndexType  N, IndexType  NPad, IndexType  StrideC  )

inlinestatic

◆ MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

template<typename CGridDesc >

static constexpr __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ( const CGridDesc &  c_grid_desc_m_n, index_t  MBlock, index_t  NBlock  )

inlinestaticconstexpr

◆ MakeDGridDescriptor_M_N()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

template<typename DLayout >

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeDGridDescriptor_M_N ( index_t  M, index_t  MPad, index_t  N, index_t  NPad, index_t  StrideC  )

inlinestatic

◆ MakeDsGridDescriptor_M_N()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

__host__ static __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeDsGridDescriptor_M_N ( index_t  M, index_t  MPad, index_t  N, index_t  NPad, std::array< index_t, NumDTensor >  StrideDs  )

inlinestatic

◆ MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

template<typename DsGridDesc >

static constexpr __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ( const DsGridDesc &  ds_grid_desc_m_n, index_t  MBlock, index_t  NBlock  )

inlinestaticconstexpr

◆ MakeDsGridPointer()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

static constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeDsGridPointer ( )

inlinestaticconstexpr

◆ MakeGemmMmaTileDescriptor()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

template<index_t MNXdlPerWave, index_t MNWaves, index_t MNPerXdl, typename TileDesc_K0_MN_K1 >

__host__ static constexpr __device__ auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::MakeGemmMmaTileDescriptor ( const TileDesc_K0_MN_K1 &  )

inlinestaticconstexpr

◆ Run()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

template<bool HasMainKBlockLoop, InMemoryDataOperationEnum CGlobalMemoryDataOperation, TailNumber TailNum = TailNumber::Odd>

static __device__ void ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::Run ( const index_t *  p_sorted_token_ids, const index_t *  p_sorted_expert_ids, const index_t *  p_max_token_id, const ADataType *  p_a_grid, const BDataType *  p_b_grid, DsGridPointer &  p_ds_grid, CDataType *  p_c_grid, void *  p_shared, const Problem &  problem, AElementwiseOperation  a_element_op, BElementwiseOperation  b_element_op, CElementwiseOperation  c_element_op  )

inlinestatic

◆ Run_2Lds()

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

template<bool HasMainKBlockLoop, InMemoryDataOperationEnum CGlobalMemoryDataOperation, TailNumber TailNum = TailNumber::Odd>

static __device__ void ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::Run_2Lds ( const index_t *  p_sorted_token_ids, const index_t *  p_sorted_expert_ids, const index_t *  p_max_token_id, const ADataType *  p_a_grid, const BDataType *  p_b_grid, DsGridPointer &  p_ds_grid, CDataType *  p_c_grid, void *  p_shared, void *  p_shared1, const Problem &  problem, AElementwiseOperation  a_element_op, BElementwiseOperation  b_element_op, CElementwiseOperation  c_element_op  )

inlinestatic

Member Data Documentation

◆ AK0Number

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::AK0Number = Number<KPerBlock / AK1Value>{}

staticconstexpr

◆ AK1Number

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::AK1Number = Number<AK1Value>{}

staticconstexpr

◆ APackedSize

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::APackedSize

staticconstexpr

Initial value:

= []() {

if constexpr(is_same_v<remove_cvref_t<ADataType>, pk_i4_t>)

return 2;

else

return 1;

}()

ck::is_same_v

constexpr bool is_same_v

Definition: type.hpp:283

◆ BK0Number

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::BK0Number = Number<KPerBlock / BK1Value>{}

staticconstexpr

◆ BK1Number

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::BK1Number = Number<BK1Value>{}

staticconstexpr

◆ BlockSizeNumber

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::BlockSizeNumber = Number<BlockSize>{}

staticconstexpr

◆ BPackedSize

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::BPackedSize

staticconstexpr

Initial value:

= []() {

if constexpr(is_same_v<remove_cvref_t<BDataType>, pk_i4_t>)

return 2;

else

return 1;

}()

◆ CShuffleBlockTransferScalarPerVector_NPerBlock

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::CShuffleBlockTransferScalarPerVector_NPerBlock

staticconstexpr

Initial value:

=

CDEShuffleBlockTransferScalarPerVectors{}[I0]

ck::GridwiseMoeGemm::I0

static constexpr auto I0

Definition: gridwise_moe_gemm.hpp:166

◆ I0

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::I0 = Number<0>{}

staticconstexpr

◆ I1

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::I1 = Number<1>{}

staticconstexpr

◆ I2

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::I2 = Number<2>{}

staticconstexpr

◆ I3

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::I3 = Number<3>{}

staticconstexpr

◆ I4

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::I4 = Number<4>{}

staticconstexpr

◆ I5

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::I5 = Number<5>{}

staticconstexpr

◆ I6

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::I6 = Number<6>{}

staticconstexpr

◆ I7

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr auto ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::I7 = Number<7>{}

staticconstexpr

◆ KGroup

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::KGroup

staticconstexpr

Initial value:

= []() {

if constexpr(is_same_v<remove_cvref_t<BDataType>, f8_t>)

return mfma_selector::selected_mfma.k_per_blk == 32 ? 2 : 1;

else

return 1;

}()

ck::f8_t

f8_fnuz_t f8_t

Definition: amd_ck_fp8.hpp:1737

ck::GridwiseMoeGemm::mfma_selector

MfmaSelector< ComputeTypeA, MPerXdl, NPerXdl, ComputeTypeB > mfma_selector

Definition: gridwise_moe_gemm.hpp:186

◆ KLane

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::KLane

staticconstexpr

Initial value:

=

mfma_selector::GetKPerXdlops() / mfma_selector::GetK1PerXdlops()

ck::MfmaSelector::GetK1PerXdlops

static constexpr index_t GetK1PerXdlops()

Definition: xdlops_gemm.hpp:1700

ck::MfmaSelector::GetKPerXdlops

static constexpr index_t GetKPerXdlops()

Definition: xdlops_gemm.hpp:1694

◆ KPack

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::KPack

staticconstexpr

Initial value:

=

math::max(math::lcm(AK1Number, BK1Number), mfma_selector::selected_mfma.k_per_blk)

ck::math::lcm

Y __host__ constexpr __device__ auto lcm(X x, Y y)

Definition: math.hpp:198

ck::math::max

__host__ constexpr __device__ T max(T x)

Definition: math.hpp:84

ck::GridwiseMoeGemm::AK1Number

static constexpr auto AK1Number

Definition: gridwise_moe_gemm.hpp:180

ck::GridwiseMoeGemm::BK1Number

static constexpr auto BK1Number

Definition: gridwise_moe_gemm.hpp:181

ck::MfmaSelector::selected_mfma

static constexpr auto selected_mfma

Definition: xdlops_gemm.hpp:1647

◆ KRepeat

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::KRepeat = KPerBlock / KLane / (KPack / KGroup)

staticconstexpr

◆ NLane

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::NLane = NPerXdl

staticconstexpr

◆ NumDTensor

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::NumDTensor = DsDataType::Size()

staticconstexpr

◆ NWave

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::NWave = NPerBlock / NPerXdl / NXdlPerWave

staticconstexpr

◆ SortedTileSize

template<typename ALayout , typename BLayout , typename DsLayout , typename CLayout , typename ADataType , typename BDataType , typename AccDataType , typename CShuffleDataType , typename DsDataType , typename CDataType , typename AElementwiseOperation , typename BElementwiseOperation , typename CElementwiseOperation , tensor_operation::device::GemmSpecialization GemmSpec, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1Value, index_t BK1Value, index_t 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 AThreadTransferSrcResetCoordinateAfterRun, index_t ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1 , typename BBlockTransferThreadClusterArrangeOrder , typename BBlockTransferSrcAccessOrder , index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BThreadTransferSrcResetCoordinateAfterRun, index_t BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock , typename CDEShuffleBlockTransferScalarPerVectors , BlockGemmPipelineScheduler BlkGemmPipeSched = BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion BlkGemmPipelineVer = BlockGemmPipelineVersion::v1, index_t ActivationOperation = 0, bool NSwizzle = false, bool IsInputGemm = true, bool MulRoutedWeight = true, bool PerTokenQuant = false, typename IndexType = index_t, typename ComputeTypeA = CDataType, typename ComputeTypeB = ComputeTypeA, typename LDSTypeA = ADataType, typename LDSTypeB = BDataType>

constexpr index_t ck::GridwiseMoeGemm< ALayout, BLayout, DsLayout, CLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, CDataType, AElementwiseOperation, BElementwiseOperation, CElementwiseOperation, GemmSpec, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1Value, BK1Value, MPerXdl, NPerXdl, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, AThreadTransferSrcResetCoordinateAfterRun, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BThreadTransferSrcResetCoordinateAfterRun, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CDEShuffleBlockTransferScalarPerVectors, BlkGemmPipeSched, BlkGemmPipelineVer, ActivationOperation, NSwizzle, IsInputGemm, MulRoutedWeight, PerTokenQuant, IndexType, ComputeTypeA, ComputeTypeB, LDSTypeA, LDSTypeB >::SortedTileSize = MPerBlock

staticconstexpr

---

The documentation for this struct was generated from the following file:

• /home/docs/checkouts/readthedocs.org/user_builds/advanced-micro-devices-composable-kernel/checkouts/develop/include/ck/tensor_operation/gpu/grid/gridwise_moe_gemm.hpp