DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize > Struct Template Reference

DeviceNormalizationFwdSplitKImpl&lt; XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize &gt; Struct Template Reference#

Composable Kernel: ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize > Struct Template Reference
Classes | Public Types | Public Member Functions | Static Public Member Functions | Static Public Attributes | List of all members
ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize > Struct Template Reference

#include <device_normalization_fwd_splitk_impl.hpp>

Inheritance diagram for ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >:
ck::tensor_operation::device::DeviceNormalizationFwd< XDataType, GammaDataType, BetaDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim > ck::tensor_operation::device::BaseOperator

Classes

struct  Argument
struct  Invoker

Public Types

using WorkspaceMeanVarDataType = SaveMeanInvStdDataType
using PassThrough = tensor_operation::element_wise::PassThrough
using SrcGridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1))
using Kernel1MeanVarGridDesc_M_KBlock
using Kernel2MeanVarGridDesc_M_KBlock
using Kernel2CountGridDesc_M_KBlock
using SaveMeanInvStdGridDesc_M = decltype(MakeSaveMeanInvStdDescriptor_M({1}, {1}))
using GridwiseWelford
using GridwiseWelfordNormalization

Public Member Functions

size_t GetWorkSpaceSize (const BaseArgument *pArg) const override
void SetWorkSpacePointer (BaseArgument *pArg, void *p_workspace, const StreamConfig &=StreamConfig{}) const override
bool IsSupportedArgument (const BaseArgument *p_arg) override
std::unique_ptr< BaseArgument > MakeArgumentPointer (const std::vector< index_t > lengths, const std::vector< index_t > xStrides, const std::vector< index_t > gammaStrides, const std::vector< index_t > betaStrides, const std::vector< index_t > yStrides, const std::vector< index_t > saveMeanStrides, const std::vector< index_t > saveInvStdStrides, const std::vector< index_t > reduceDims, double epsilon, const void *p_x, const void *p_gamma, const void *p_beta, void *p_y, void *p_saveMean, void *p_saveInvStd, YElementwiseOperation y_elementwise_op) override
std::unique_ptr< BaseInvoker > MakeInvokerPointer () override
std::string GetTypeString () const override
Public Member Functions inherited from ck::tensor_operation::device::BaseOperator
 BaseOperator ()=default
 BaseOperator (const BaseOperator &)=default
BaseOperator & operator= (const BaseOperator &)=default
virtual std::string GetInstanceString () const
virtual std::string GetTypeIdName () const
virtual std::optional< std::string > GetObjectName () const
virtual std::optional< std::string > GetTemplateInfo () const
virtual std::string GetTypeIdHashCode () const
virtual ~BaseOperator ()

Static Public Member Functions

static auto MakeSrc2dDescriptor (const std::vector< index_t > &inLengths, const std::vector< index_t > &inStrides, int kBlockSize, int numBlockTileIteration)
template<typename DoPads, index_t MPerTile, index_t KPerTile>
static auto MakeWorkspaceMeanVarDescriptor_M_K (index_t M, index_t K)
template<typename DoPads, index_t MPerTile, index_t KPerTile>
static auto MakeWorkspaceCountDescriptor_M_K (index_t M, index_t K)
static auto MakeSaveMeanInvStdDescriptor_M (const std::vector< index_t > &lengths, const std::vector< index_t > &strides)

Static Public Attributes

static constexpr auto I0 = Number<0>{}
static constexpr auto I1 = Number<1>{}
static constexpr index_t NumInvariantDim = Rank - NumReduceDim
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize
static constexpr bool reduceAllDim = (NumInvariantDim == 0)

Member Typedef Documentation

◆ GridwiseWelford

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
using ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::GridwiseWelford
Initial value:
GridwiseNormalizationSplitK1st<XDataType,
ComputeDataType,
WorkspaceMeanVarDataType,
SrcGridDesc_M_K,
Kernel1MeanVarGridDesc_M_KBlock,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XYVectorDim,
XSrcVectorSize>
ck::GridwiseNormalizationSplitK1st
Definition gridwise_normalization_splitk_1st.hpp:28
ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl::Kernel1MeanVarGridDesc_M_KBlock
decltype(MakeWorkspaceMeanVarDescriptor_M_K< Sequence< true, false >, 1, 1 >(1, 1)) Kernel1MeanVarGridDesc_M_KBlock
Definition device_normalization_fwd_splitk_impl.hpp:285
ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl::SrcGridDesc_M_K
decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1)) SrcGridDesc_M_K
Definition device_normalization_fwd_splitk_impl.hpp:284
ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl::WorkspaceMeanVarDataType
SaveMeanInvStdDataType WorkspaceMeanVarDataType
Definition device_normalization_fwd_splitk_impl.hpp:146

◆ GridwiseWelfordNormalization

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
using ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::GridwiseWelfordNormalization
Initial value:
GridwiseNormalizationSplitK2nd<WorkspaceMeanVarDataType,
XDataType,
GammaDataType,
BetaDataType,
YDataType,
SaveMeanInvStdDataType,
ComputeDataType,
YElementwiseOperation,
Kernel2MeanVarGridDesc_M_KBlock,
Kernel2CountGridDesc_M_KBlock,
SrcGridDesc_M_K,
SaveMeanInvStdGridDesc_M,
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
XYVectorDim,
XSrcVectorSize,
GammaSrcVectorDim,
GammaSrcVectorSize,
BetaSrcVectorDim,
BetaSrcVectorSize,
XYVectorDim,
YDstVectorSize,
SaveMeanInvStdDstVectorSize>
ck::GridwiseNormalizationSplitK2nd
Definition gridwise_normalization_splitk_2nd.hpp:42
ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl::Kernel2MeanVarGridDesc_M_KBlock
decltype(MakeWorkspaceMeanVarDescriptor_M_K< Sequence< true, true >, 1, 1 >(1, 1)) Kernel2MeanVarGridDesc_M_KBlock
Definition device_normalization_fwd_splitk_impl.hpp:288
ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl::SaveMeanInvStdGridDesc_M
decltype(MakeSaveMeanInvStdDescriptor_M({1}, {1})) SaveMeanInvStdGridDesc_M
Definition device_normalization_fwd_splitk_impl.hpp:294
ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl::Kernel2CountGridDesc_M_KBlock
decltype(MakeWorkspaceCountDescriptor_M_K< Sequence< true, true >, 1, 1 >(1, 1)) Kernel2CountGridDesc_M_KBlock
Definition device_normalization_fwd_splitk_impl.hpp:291

◆ Kernel1MeanVarGridDesc_M_KBlock

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
using ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::Kernel1MeanVarGridDesc_M_KBlock
Initial value:
decltype(MakeWorkspaceMeanVarDescriptor_M_K<Sequence<true, false>, 1, 1>(1, 1))
ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl::MakeWorkspaceMeanVarDescriptor_M_K
static auto MakeWorkspaceMeanVarDescriptor_M_K(index_t M, index_t K)
Definition device_normalization_fwd_splitk_impl.hpp:240

◆ Kernel2CountGridDesc_M_KBlock

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
using ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::Kernel2CountGridDesc_M_KBlock
Initial value:
decltype(MakeWorkspaceCountDescriptor_M_K<Sequence<true, true>, 1, 1>(1, 1))
ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl::MakeWorkspaceCountDescriptor_M_K
static auto MakeWorkspaceCountDescriptor_M_K(index_t M, index_t K)
Definition device_normalization_fwd_splitk_impl.hpp:248

◆ Kernel2MeanVarGridDesc_M_KBlock

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
using ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::Kernel2MeanVarGridDesc_M_KBlock
Initial value:
decltype(MakeWorkspaceMeanVarDescriptor_M_K<Sequence<true, true>, 1, 1>(1, 1))

◆ PassThrough

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
using ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::PassThrough = tensor_operation::element_wise::PassThrough

◆ SaveMeanInvStdGridDesc_M

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
using ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::SaveMeanInvStdGridDesc_M = decltype(MakeSaveMeanInvStdDescriptor_M({1}, {1}))

◆ SrcGridDesc_M_K

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
using ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::SrcGridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1))

◆ WorkspaceMeanVarDataType

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
using ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::WorkspaceMeanVarDataType = SaveMeanInvStdDataType

Member Function Documentation

◆ GetTypeString()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
std::string ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::GetTypeString ( ) const
inlineoverridevirtual

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

◆ GetWorkSpaceSize()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
size_t ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::GetWorkSpaceSize ( const BaseArgument * pArg) const
inlineoverridevirtual

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

◆ IsSupportedArgument()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
bool ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::IsSupportedArgument ( const BaseArgument * p_arg)
inlineoverridevirtual

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

◆ MakeArgumentPointer()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
std::unique_ptr< BaseArgument > ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::MakeArgumentPointer ( const std::vector< index_t > lengths,
const std::vector< index_t > xStrides,
const std::vector< index_t > gammaStrides,
const std::vector< index_t > betaStrides,
const std::vector< index_t > yStrides,
const std::vector< index_t > saveMeanStrides,
const std::vector< index_t > saveInvStdStrides,
const std::vector< index_t > reduceDims,
double epsilon,
const void * p_x,
const void * p_gamma,
const void * p_beta,
void * p_y,
void * p_saveMean,
void * p_saveInvStd,
YElementwiseOperation y_elementwise_op )
inlineoverridevirtual

Implements ck::tensor_operation::device::DeviceNormalizationFwd< XDataType, GammaDataType, BetaDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim >.

◆ MakeInvokerPointer()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
std::unique_ptr< BaseInvoker > ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::MakeInvokerPointer ( )
inlineoverridevirtual

Implements ck::tensor_operation::device::DeviceNormalizationFwd< XDataType, GammaDataType, BetaDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim >.

◆ MakeSaveMeanInvStdDescriptor_M()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
auto ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::MakeSaveMeanInvStdDescriptor_M ( const std::vector< index_t > & lengths,
const std::vector< index_t > & strides )
inlinestatic

◆ MakeSrc2dDescriptor()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
auto ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::MakeSrc2dDescriptor ( const std::vector< index_t > & inLengths,
const std::vector< index_t > & inStrides,
int kBlockSize,
int numBlockTileIteration )
inlinestatic

◆ MakeWorkspaceCountDescriptor_M_K()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
template<typename DoPads, index_t MPerTile, index_t KPerTile>
auto ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::MakeWorkspaceCountDescriptor_M_K ( index_t M,
index_t K )
inlinestatic

◆ MakeWorkspaceMeanVarDescriptor_M_K()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
template<typename DoPads, index_t MPerTile, index_t KPerTile>
auto ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::MakeWorkspaceMeanVarDescriptor_M_K ( index_t M,
index_t K )
inlinestatic

◆ SetWorkSpacePointer()

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
void ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::SetWorkSpacePointer ( BaseArgument * pArg,
void * p_workspace,
const StreamConfig & = StreamConfig{} ) const
inlineoverridevirtual

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

Member Data Documentation

◆ I0

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
auto ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::I0 = Number<0>{}
staticconstexpr

◆ I1

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
auto ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::I1 = Number<1>{}
staticconstexpr

◆ K_BlockTileSize

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
index_t ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::K_BlockTileSize = KThreadClusterSize * KThreadSliceSize
staticconstexpr

◆ M_BlockTileSize

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
index_t ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::M_BlockTileSize = MThreadClusterSize * MThreadSliceSize
staticconstexpr

◆ NumInvariantDim

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
index_t ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::NumInvariantDim = Rank - NumReduceDim
staticconstexpr

◆ reduceAllDim

template<typename XDataType, typename GammaDataType, typename BetaDataType, typename ComputeDataType, typename YDataType, typename SaveMeanInvStdDataType, typename YElementwiseOperation, index_t Rank, index_t NumReduceDim, index_t BlockSize, index_t MThreadClusterSize, index_t KThreadClusterSize, index_t MThreadSliceSize, index_t KThreadSliceSize, index_t XYVectorDim, index_t XSrcVectorSize, index_t GammaSrcVectorDim, index_t GammaSrcVectorSize, index_t BetaSrcVectorDim, index_t BetaSrcVectorSize, index_t YDstVectorSize, index_t SaveMeanInvStdDstVectorSize>
bool ck::tensor_operation::device::DeviceNormalizationFwdSplitKImpl< XDataType, GammaDataType, BetaDataType, ComputeDataType, YDataType, SaveMeanInvStdDataType, YElementwiseOperation, Rank, NumReduceDim, BlockSize, MThreadClusterSize, KThreadClusterSize, MThreadSliceSize, KThreadSliceSize, XYVectorDim, XSrcVectorSize, GammaSrcVectorDim, GammaSrcVectorSize, BetaSrcVectorDim, BetaSrcVectorSize, YDstVectorSize, SaveMeanInvStdDstVectorSize >::reduceAllDim = (NumInvariantDim == 0)
staticconstexpr
The documentation for this struct was generated from the following file: