GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock > Struct Template Reference

GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle&lt; ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock &gt; Struct Template Reference#

Composable Kernel: ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock > Struct Template Reference
Classes | Public Types | Static Public Member Functions | Static Public Attributes | List of all members
ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock > Struct Template Reference

#include <gridwise_gemm_xdl_waveletmodel_cshuffle.hpp>

Classes

struct  TileLoadThreadGroup
struct  TileMathThreadGroup

Public Types

using CShuffleBlockTransferThreadGroup = ThisThreadBlock<TileMathThreadGroupSize>
using GridwiseGemmLoad = GridwiseGemmLoadWave<TileLoadThreadGroup, NumGemmKPrefetchStage>
using GridwiseGemmMath = GridwiseGemmMathWave<TileMathThreadGroup, NumGemmKPrefetchStage>
using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
using DefaultBlock2ETileMap

Static Public Member Functions

__host__ static __device__ constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1 ()
__host__ static __device__ constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1 ()
__host__ static __device__ constexpr auto GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ()
__host__ static __device__ constexpr index_t GetSharedMemoryNumberOfByte ()
template<InMemoryDataOperationEnum CGlobalMemoryDataOperation_ = InMemoryDataOperationEnum::Set>
static __device__ bool constexpr IsValidCompilationParameter ()
template<typename Block2ETileMap>
__host__ static __device__ constexpr bool CheckValidity (const AGridDesc_M_K &a_grid_desc_m_k, const BGridDesc_N_K &b_grid_desc_n_k, const EGridDesc_M_N &e_grid_desc_m_n, const Block2ETileMap &)
__host__ static __device__ constexpr bool CalculateHasMainKBlockLoop (index_t K)
__host__ static __device__ constexpr auto MakeDefaultBlock2ETileMap (const EGridDesc_M_N &e_grid_desc_m_n)
__host__ static __device__ constexpr index_t CalculateGridSize (const EGridDesc_M_N &e_grid_desc_m_n)
__host__ static __device__ constexpr auto MakeDefaultAGridDescriptor_AK0_M_AK1 (const AGridDesc_M_K &a_grid_desc_m_k)
__host__ static __device__ constexpr auto MakeDefaultBGridDescriptor_BK0_N_BK1 (const BGridDesc_N_K &b_grid_desc_n_k)
template<typename EGridDescriptor_M_N>
__host__ static __device__ constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock (const EGridDescriptor_M_N &e_grid_desc_m_n)
template<bool HasMainKBlockLoop, typename AGridDesc_AK0_M_AK1, typename BGridDesc_BK0_N_BK1, typename Block2ETileMap>
static __device__ void Run (const ABDataType *__restrict__ p_a_grid, const ABDataType *__restrict__ p_b_grid, EDataType *__restrict__ p_e_grid, void *__restrict__ p_shared, const AElementwiseOperation &a_element_op, const BElementwiseOperation &b_element_op, const EElementwiseOperation &e_element_op, const AGridDesc_AK0_M_AK1 &a_grid_desc_ak0_m_ak1, const BGridDesc_BK0_N_BK1 &b_grid_desc_bk0_n_bk1, const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock &e_grid_desc_mblock_mperblock_nblock_nperblock, const Block2ETileMap &block_2_etile_map)

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 AK1 = Number<AK1Value>{}
static constexpr auto BK1 = Number<BK1Value>{}
static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{}
static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{}
static constexpr auto BlockSize = math::max(TileLoadThreadGroupSize, TileMathThreadGroupSize)

Member Typedef Documentation

◆ CShuffleBlockTransferThreadGroup

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
using ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::CShuffleBlockTransferThreadGroup = ThisThreadBlock<TileMathThreadGroupSize>

◆ DefaultBlock2ETileMap

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
using ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::DefaultBlock2ETileMap
Initial value:
remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>
ck::remove_cvref_t
remove_cv_t< remove_reference_t< T > > remove_cvref_t
Definition type.hpp:297
ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle::MakeDefaultBlock2ETileMap
__host__ static __device__ constexpr auto MakeDefaultBlock2ETileMap(const EGridDesc_M_N &e_grid_desc_m_n)
Definition gridwise_gemm_xdl_waveletmodel_cshuffle.hpp:252

◆ EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
using ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
Initial value:
remove_cvref_t<decltype(MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
EGridDesc_M_N{}))>
ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
__host__ static __device__ constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDescriptor_M_N &e_grid_desc_m_n)
Definition gridwise_gemm_xdl_waveletmodel_cshuffle.hpp:331

◆ GridwiseGemmLoad

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
using ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::GridwiseGemmLoad = GridwiseGemmLoadWave<TileLoadThreadGroup, NumGemmKPrefetchStage>

◆ GridwiseGemmMath

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
using ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::GridwiseGemmMath = GridwiseGemmMathWave<TileMathThreadGroup, NumGemmKPrefetchStage>

Member Function Documentation

◆ CalculateGridSize()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
__host__ static __device__ constexpr index_t ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::CalculateGridSize ( const EGridDesc_M_N & e_grid_desc_m_n)
inlinestaticconstexpr

◆ CalculateHasMainKBlockLoop()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
__host__ static __device__ constexpr bool ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::CalculateHasMainKBlockLoop ( index_t K)
inlinestaticconstexpr

◆ CheckValidity()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
template<typename Block2ETileMap>
__host__ static __device__ constexpr bool ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::CheckValidity ( const AGridDesc_M_K & a_grid_desc_m_k,
const BGridDesc_N_K & b_grid_desc_n_k,
const EGridDesc_M_N & e_grid_desc_m_n,
const Block2ETileMap &  )
inlinestaticconstexpr

◆ GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
__host__ static __device__ constexpr auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1 ( )
inlinestaticconstexpr

◆ GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
__host__ static __device__ constexpr auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1 ( )
inlinestaticconstexpr

◆ GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
__host__ static __device__ constexpr auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ( )
inlinestaticconstexpr

◆ GetSharedMemoryNumberOfByte()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
__host__ static __device__ constexpr index_t ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::GetSharedMemoryNumberOfByte ( )
inlinestaticconstexpr

◆ IsValidCompilationParameter()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
template<InMemoryDataOperationEnum CGlobalMemoryDataOperation_ = InMemoryDataOperationEnum::Set>
__device__ bool constexpr ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::IsValidCompilationParameter ( )
inlinestaticconstexpr

◆ MakeDefaultAGridDescriptor_AK0_M_AK1()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
__host__ static __device__ constexpr auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::MakeDefaultAGridDescriptor_AK0_M_AK1 ( const AGridDesc_M_K & a_grid_desc_m_k)
inlinestaticconstexpr

◆ MakeDefaultBGridDescriptor_BK0_N_BK1()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
__host__ static __device__ constexpr auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::MakeDefaultBGridDescriptor_BK0_N_BK1 ( const BGridDesc_N_K & b_grid_desc_n_k)
inlinestaticconstexpr

◆ MakeDefaultBlock2ETileMap()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
__host__ static __device__ constexpr auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::MakeDefaultBlock2ETileMap ( const EGridDesc_M_N & e_grid_desc_m_n)
inlinestaticconstexpr

◆ MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
template<typename EGridDescriptor_M_N>
__host__ static __device__ constexpr auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ( const EGridDescriptor_M_N & e_grid_desc_m_n)
inlinestaticconstexpr

◆ Run()

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
template<bool HasMainKBlockLoop, typename AGridDesc_AK0_M_AK1, typename BGridDesc_BK0_N_BK1, typename Block2ETileMap>
__device__ void ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::Run ( const ABDataType *__restrict__ p_a_grid,
const ABDataType *__restrict__ p_b_grid,
EDataType *__restrict__ p_e_grid,
void *__restrict__ p_shared,
const AElementwiseOperation & a_element_op,
const BElementwiseOperation & b_element_op,
const EElementwiseOperation & e_element_op,
const AGridDesc_AK0_M_AK1 & a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 & b_grid_desc_bk0_n_bk1,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock & e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap & block_2_etile_map )
inlinestatic

Member Data Documentation

◆ AK0PerBlock

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::AK0PerBlock = Number<KPerBlock / AK1Value>{}
staticconstexpr

◆ AK1

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::AK1 = Number<AK1Value>{}
staticconstexpr

◆ BK0PerBlock

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::BK0PerBlock = Number<KPerBlock / BK1Value>{}
staticconstexpr

◆ BK1

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::BK1 = Number<BK1Value>{}
staticconstexpr

◆ BlockSize

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::BlockSize = math::max(TileLoadThreadGroupSize, TileMathThreadGroupSize)
staticconstexpr

◆ I0

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::I0 = Number<0>{}
staticconstexpr

◆ I1

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::I1 = Number<1>{}
staticconstexpr

◆ I2

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::I2 = Number<2>{}
staticconstexpr

◆ I3

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::I3 = Number<3>{}
staticconstexpr

◆ I4

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::I4 = Number<4>{}
staticconstexpr

◆ I5

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::I5 = Number<5>{}
staticconstexpr

◆ I6

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::I6 = Number<6>{}
staticconstexpr

◆ I7

template<typename ABDataType, typename FloatGemmAcc, typename EDataTypeShuffle, typename EDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename EElementwiseOperation, InMemoryDataOperationEnum CGlobalMemoryDataOperation, typename AGridDesc_M_K, typename BGridDesc_N_K, typename EGridDesc_M_N, index_t NumGemmKPrefetchStage, index_t TileLoadThreadGroupSize, index_t TileMathThreadGroupSize, 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, index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
auto ck::GridwiseGemm_k0mk1_k0nk1_mn_xdl_waveletmodel_cshuffle< ABDataType, FloatGemmAcc, EDataTypeShuffle, EDataType, AElementwiseOperation, BElementwiseOperation, EElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc_M_K, BGridDesc_N_K, EGridDesc_M_N, NumGemmKPrefetchStage, TileLoadThreadGroupSize, TileMathThreadGroupSize, 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, CShuffleBlockTransferScalarPerVector_NPerBlock >::I7 = Number<7>{}
staticconstexpr
The documentation for this struct was generated from the following file: