Argument Struct Reference

Argument Struct Reference#

Composable Kernel: ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument Struct Reference
Public Member Functions | Public Attributes | List of all members
ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument Struct Reference

#include <device_gemm_multiple_d_layernorm_xdl_cshuffle.hpp>

Inheritance diagram for ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument:
ck::tensor_operation::device::BaseArgument

Public Member Functions

 Argument (const void *p_a_grid, const void *p_b_grid, std::array< const void *, NumDTensor > p_ds_grid, const void *p_gamma_grid, const void *p_beta_grid, void *p_h_grid, index_t MRaw, index_t NRaw, index_t KRaw, index_t StrideA, index_t StrideB, std::array< index_t, NumDTensor > StrideDs, index_t StrideH, double epsilon, AElementwiseOperation a_element_op, BElementwiseOperation b_element_op, CDEElementwiseOperation cde_element_op, HElementwiseOperation h_element_op)
void Print () const
Public Member Functions inherited from ck::tensor_operation::device::BaseArgument
 BaseArgument ()=default
 BaseArgument (const BaseArgument &)=default
BaseArgument & operator= (const BaseArgument &)=default
virtual ~BaseArgument ()

Public Attributes

const ADataType * p_a_grid_
const BDataType * p_b_grid_
GridwiseGemmWelford64::DsGridPointer p_ds_grid_
void * p_workspace_e_grid_
void * p_workspace_mean_
void * p_workspace_var_
void * p_workspace_count_
const GammaDataType * p_gamma_grid_
const BetaDataType * p_beta_grid_
HDataType * p_h_grid_
AGridDesc_M_K a_grid_desc_m_k_
BGridDesc_N_K b_grid_desc_n_k_
DsGridDesc_M_N ds_grid_desc_m_n_
EHGridDesc_M_N gemm_e_grid_desc_m_n_
EHGridDesc_M_N layernorm_e_grid_desc_m_n_
GemmMeanVarGridDesc_M_NBlock gemm_mean_var_grid_desc_m_nblock_
GemmCountGridDesc_M_NBlock gemm_count_grid_desc_m_nblock_
LayernormMeanVarGridDesc_M_NBlock layernorm_mean_var_grid_desc_m_nblock_
LayernormCountGridDesc_M_NBlock layernorm_count_grid_desc_m_nblock_
GammaBetaGridDesc_N gamma_grid_desc_n_
GammaBetaGridDesc_N beta_grid_desc_n_
EHGridDesc_M_N h_grid_desc_m_n_
GridwiseGemmWelford64::DefaultAGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_
GridwiseGemmWelford64::DefaultBGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_
GridwiseGemmWelford64::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ds_grid_desc_mblock_mperblock_nblock_nperblock_
GridwiseGemmWelford64::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_
GridwiseGemmWelford64::MeanVarGridDescriptor_MBlock_MPerBlock_NBlock gemm_mean_var_grid_desc_mblock_mperblock_nblock_
GridwiseGemmWelford64::CountGridDescriptor_MBlock_MPerBlock_NBlock gemm_count_grid_desc_mblock_mperblock_nblock_
Block2ETileMap block_2_etile_map_
AElementwiseOperation a_element_op_
BElementwiseOperation b_element_op_
CDEElementwiseOperation cde_element_op_
HElementwiseOperation h_element_op_
index_t MRaw_
index_t NRaw_
index_t KRaw_
index_t gemm_nblock_
AccDataType epsilon_
Public Attributes inherited from ck::tensor_operation::device::BaseArgument
void * p_workspace_ = nullptr

Constructor & Destructor Documentation

◆ Argument()

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::Argument ( const void * p_a_grid,
const void * p_b_grid,
std::array< const void *, NumDTensor > p_ds_grid,
const void * p_gamma_grid,
const void * p_beta_grid,
void * p_h_grid,
index_t MRaw,
index_t NRaw,
index_t KRaw,
index_t StrideA,
index_t StrideB,
std::array< index_t, NumDTensor > StrideDs,
index_t StrideH,
double epsilon,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
HElementwiseOperation h_element_op )
inline

Member Function Documentation

◆ Print()

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
void ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::Print ( ) const
inline

Member Data Documentation

◆ a_element_op_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
AElementwiseOperation ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::a_element_op_

◆ a_grid_desc_ak0_m_ak1_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GridwiseGemmWelford64::DefaultAGridDesc_AK0_M_AK1 ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::a_grid_desc_ak0_m_ak1_

◆ a_grid_desc_m_k_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
AGridDesc_M_K ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::a_grid_desc_m_k_

◆ b_element_op_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
BElementwiseOperation ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::b_element_op_

◆ b_grid_desc_bk0_n_bk1_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GridwiseGemmWelford64::DefaultBGridDesc_BK0_N_BK1 ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::b_grid_desc_bk0_n_bk1_

◆ b_grid_desc_n_k_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
BGridDesc_N_K ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::b_grid_desc_n_k_

◆ beta_grid_desc_n_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GammaBetaGridDesc_N ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::beta_grid_desc_n_

◆ block_2_etile_map_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
Block2ETileMap ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::block_2_etile_map_

◆ cde_element_op_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
CDEElementwiseOperation ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::cde_element_op_

◆ ds_grid_desc_m_n_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
DsGridDesc_M_N ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::ds_grid_desc_m_n_

◆ ds_grid_desc_mblock_mperblock_nblock_nperblock_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GridwiseGemmWelford64::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::ds_grid_desc_mblock_mperblock_nblock_nperblock_

◆ e_grid_desc_mblock_mperblock_nblock_nperblock_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GridwiseGemmWelford64::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::e_grid_desc_mblock_mperblock_nblock_nperblock_

◆ epsilon_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
AccDataType ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::epsilon_

◆ gamma_grid_desc_n_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GammaBetaGridDesc_N ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::gamma_grid_desc_n_

◆ gemm_count_grid_desc_m_nblock_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GemmCountGridDesc_M_NBlock ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::gemm_count_grid_desc_m_nblock_

◆ gemm_count_grid_desc_mblock_mperblock_nblock_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GridwiseGemmWelford64::CountGridDescriptor_MBlock_MPerBlock_NBlock ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::gemm_count_grid_desc_mblock_mperblock_nblock_

◆ gemm_e_grid_desc_m_n_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
EHGridDesc_M_N ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::gemm_e_grid_desc_m_n_

◆ gemm_mean_var_grid_desc_m_nblock_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GemmMeanVarGridDesc_M_NBlock ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::gemm_mean_var_grid_desc_m_nblock_

◆ gemm_mean_var_grid_desc_mblock_mperblock_nblock_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GridwiseGemmWelford64::MeanVarGridDescriptor_MBlock_MPerBlock_NBlock ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::gemm_mean_var_grid_desc_mblock_mperblock_nblock_

◆ gemm_nblock_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
index_t ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::gemm_nblock_

◆ h_element_op_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
HElementwiseOperation ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::h_element_op_

◆ h_grid_desc_m_n_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
EHGridDesc_M_N ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::h_grid_desc_m_n_

◆ KRaw_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
index_t ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::KRaw_

◆ layernorm_count_grid_desc_m_nblock_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
LayernormCountGridDesc_M_NBlock ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::layernorm_count_grid_desc_m_nblock_

◆ layernorm_e_grid_desc_m_n_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
EHGridDesc_M_N ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::layernorm_e_grid_desc_m_n_

◆ layernorm_mean_var_grid_desc_m_nblock_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
LayernormMeanVarGridDesc_M_NBlock ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::layernorm_mean_var_grid_desc_m_nblock_

◆ MRaw_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
index_t ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::MRaw_

◆ NRaw_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
index_t ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::NRaw_

◆ p_a_grid_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
const ADataType* ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_a_grid_

◆ p_b_grid_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
const BDataType* ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_b_grid_

◆ p_beta_grid_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
const BetaDataType* ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_beta_grid_

◆ p_ds_grid_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
GridwiseGemmWelford64::DsGridPointer ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_ds_grid_

◆ p_gamma_grid_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
const GammaDataType* ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_gamma_grid_

◆ p_h_grid_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
HDataType* ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_h_grid_

◆ p_workspace_count_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
void* ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_workspace_count_

◆ p_workspace_e_grid_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
void* ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_workspace_e_grid_

◆ p_workspace_mean_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
void* ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_workspace_mean_

◆ p_workspace_var_

template<typename ALayout, typename BLayout, typename DsLayout, typename HLayout, typename ADataType, typename BDataType, typename AccDataType, typename CShuffleDataType, typename DsDataType, typename EMeanVarDataType, typename GammaDataType, typename BetaDataType, typename HDataType, typename AElementwiseOperation, typename BElementwiseOperation, typename CDEElementwiseOperation, typename HElementwiseOperation, GemmSpecialization GemmSpec, index_t NumGemmKPrefetchStage, index_t BlockSize, index_t MPerBlock, index_t NPerBlock, index_t KPerBlock, index_t AK1, index_t BK1, index_t MPerXDL, index_t NPerXDL, index_t MXdlPerWave, index_t NXdlPerWave, typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, index_t ABlockTransferSrcVectorDim, index_t ABlockTransferSrcScalarPerVector, index_t ABlockTransferDstScalarPerVector_AK1, bool ABlockLdsExtraM, typename BBlockTransferThreadClusterLengths_BK0_N_BK1, typename BBlockTransferThreadClusterArrangeOrder, typename BBlockTransferSrcAccessOrder, index_t BBlockTransferSrcVectorDim, index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferDstScalarPerVector_BK1, bool BBlockLdsExtraN, index_t CShuffleMXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle, typename PostShuffleThreadClusterSize_M_N, index_t PostShuffleScalarPerVector, typename LayernormThreadClusterSize_M_N, index_t LayernormThreadSliceSize_M, LoopScheduler LoopSched = make_default_loop_scheduler(), PipelineVersion PipelineVer = PipelineVersion::v1>
void* ck::tensor_operation::device::DeviceGemmMultipleDLayernorm_Xdl_CShuffle< ALayout, BLayout, DsLayout, HLayout, ADataType, BDataType, AccDataType, CShuffleDataType, DsDataType, EMeanVarDataType, GammaDataType, BetaDataType, HDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation, HElementwiseOperation, GemmSpec, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave, ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1, BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector, BBlockTransferDstScalarPerVector_BK1, BBlockLdsExtraN, CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle, PostShuffleThreadClusterSize_M_N, PostShuffleScalarPerVector, LayernormThreadClusterSize_M_N, LayernormThreadSliceSize_M, LoopSched, PipelineVer >::Argument::p_workspace_var_
The documentation for this struct was generated from the following file: