Composable Kernel: ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg Struct Reference

SelfAttnArg Struct Reference

SelfAttnArg Struct Reference#

Composable Kernel: ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg Struct Reference

Public Member Functions | Public Attributes | List of all members

ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg Struct Reference

#include <device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp>

Inheritance diagram for ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg:

Public Member Functions
	SelfAttnArg (const ADataType p_qkv_grid, CDataType p_out_grid, index_t batch_size, index_t sequence_length, index_t head_count, index_t head_size, float alpha)
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_qkv_grid_
CDataType *	p_out_grid_
index_t	batch_size_
index_t	sequence_length_
index_t	head_count_
index_t	head_size_
float	alpha_
Public Attributes inherited from ck::tensor_operation::device::BaseArgument
void *	p_workspace_ = nullptr

Constructor & Destructor Documentation

◆ SelfAttnArg()

template<index_t NumDimG, index_t NumDimM, index_t NumDimL, index_t NumDimK, index_t NumDimN, typename ADataType, typename B0DataType, typename B1DataType, typename CDataType, typename Acc0BiasDataType, typename Acc0DataType, typename Acc1BiasDataType, typename Acc1DataType, typename CShuffleDataType, typename AElementwiseOperation, typename B0ElementwiseOperation, typename AccElementwiseOperation, typename B1ElementwiseOperation, typename CElementwiseOperation, GemmSpecialization GemmSpec, TensorSpecialization ASpec, TensorSpecialization B0Spec, TensorSpecialization B1Spec, TensorSpecialization CSpec, ck::index_t NumPrefetch, ck::index_t BlockSize, ck::index_t MPerBlock, ck::index_t LPerBlock, ck::index_t KPerBlock, ck::index_t AK1, ck::index_t BK1, ck::index_t NPerBlock, ck::index_t LTilePerBlock, ck::index_t L1, ck::index_t MPerWmma, ck::index_t LPerWmma, ck::index_t NPerWmma, ck::index_t MRepeat, ck::index_t LRepeat, ck::index_t NRepeat, typename ABlockTransferThreadClusterLengths_K0_M_K1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, ck::index_t ABlockTransferSrcVectorDim, ck::index_t ABlockTransferSrcScalarPerVector, ck::index_t ABlockTransferDstScalarPerVector_K1, bool ABlockLdsAddExtraM, typename B0BlockTransferThreadClusterLengths_K0_L_K1, typename B0BlockTransferThreadClusterArrangeOrder, typename B0BlockTransferSrcAccessOrder, ck::index_t B0BlockTransferSrcVectorDim, ck::index_t B0BlockTransferSrcScalarPerVector, ck::index_t B0BlockTransferDstScalarPerVector_K1, bool B0BlockLdsAddExtraL, typename B1BlockTransferThreadClusterLengths_L0_N_L1, typename B1BlockTransferThreadClusterArrangeOrder, typename B1BlockTransferSrcAccessOrder, ck::index_t B1BlockTransferSrcVectorDim, ck::index_t B1BlockTransferSrcScalarPerVector, ck::index_t B1BlockTransferDstScalarPerVector_L1, bool B1BlockLdsAddExtraN, index_t CShuffleMRepeatPerShuffle, index_t CShuffleNRepeatPerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpecialization MaskingSpec, ck::LoopScheduler LoopSched = make_default_loop_scheduler(), ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>

ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg::SelfAttnArg	(	const ADataType *	p_qkv_grid,
		CDataType *	p_out_grid,
		index_t	batch_size,
		index_t	sequence_length,
		index_t	head_count,
		index_t	head_size,
		float	alpha )

inline

Member Data Documentation

◆ alpha_

template<index_t NumDimG, index_t NumDimM, index_t NumDimL, index_t NumDimK, index_t NumDimN, typename ADataType, typename B0DataType, typename B1DataType, typename CDataType, typename Acc0BiasDataType, typename Acc0DataType, typename Acc1BiasDataType, typename Acc1DataType, typename CShuffleDataType, typename AElementwiseOperation, typename B0ElementwiseOperation, typename AccElementwiseOperation, typename B1ElementwiseOperation, typename CElementwiseOperation, GemmSpecialization GemmSpec, TensorSpecialization ASpec, TensorSpecialization B0Spec, TensorSpecialization B1Spec, TensorSpecialization CSpec, ck::index_t NumPrefetch, ck::index_t BlockSize, ck::index_t MPerBlock, ck::index_t LPerBlock, ck::index_t KPerBlock, ck::index_t AK1, ck::index_t BK1, ck::index_t NPerBlock, ck::index_t LTilePerBlock, ck::index_t L1, ck::index_t MPerWmma, ck::index_t LPerWmma, ck::index_t NPerWmma, ck::index_t MRepeat, ck::index_t LRepeat, ck::index_t NRepeat, typename ABlockTransferThreadClusterLengths_K0_M_K1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, ck::index_t ABlockTransferSrcVectorDim, ck::index_t ABlockTransferSrcScalarPerVector, ck::index_t ABlockTransferDstScalarPerVector_K1, bool ABlockLdsAddExtraM, typename B0BlockTransferThreadClusterLengths_K0_L_K1, typename B0BlockTransferThreadClusterArrangeOrder, typename B0BlockTransferSrcAccessOrder, ck::index_t B0BlockTransferSrcVectorDim, ck::index_t B0BlockTransferSrcScalarPerVector, ck::index_t B0BlockTransferDstScalarPerVector_K1, bool B0BlockLdsAddExtraL, typename B1BlockTransferThreadClusterLengths_L0_N_L1, typename B1BlockTransferThreadClusterArrangeOrder, typename B1BlockTransferSrcAccessOrder, ck::index_t B1BlockTransferSrcVectorDim, ck::index_t B1BlockTransferSrcScalarPerVector, ck::index_t B1BlockTransferDstScalarPerVector_L1, bool B1BlockLdsAddExtraN, index_t CShuffleMRepeatPerShuffle, index_t CShuffleNRepeatPerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpecialization MaskingSpec, ck::LoopScheduler LoopSched = make_default_loop_scheduler(), ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>

float ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg::alpha_

◆ batch_size_

template<index_t NumDimG, index_t NumDimM, index_t NumDimL, index_t NumDimK, index_t NumDimN, typename ADataType, typename B0DataType, typename B1DataType, typename CDataType, typename Acc0BiasDataType, typename Acc0DataType, typename Acc1BiasDataType, typename Acc1DataType, typename CShuffleDataType, typename AElementwiseOperation, typename B0ElementwiseOperation, typename AccElementwiseOperation, typename B1ElementwiseOperation, typename CElementwiseOperation, GemmSpecialization GemmSpec, TensorSpecialization ASpec, TensorSpecialization B0Spec, TensorSpecialization B1Spec, TensorSpecialization CSpec, ck::index_t NumPrefetch, ck::index_t BlockSize, ck::index_t MPerBlock, ck::index_t LPerBlock, ck::index_t KPerBlock, ck::index_t AK1, ck::index_t BK1, ck::index_t NPerBlock, ck::index_t LTilePerBlock, ck::index_t L1, ck::index_t MPerWmma, ck::index_t LPerWmma, ck::index_t NPerWmma, ck::index_t MRepeat, ck::index_t LRepeat, ck::index_t NRepeat, typename ABlockTransferThreadClusterLengths_K0_M_K1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, ck::index_t ABlockTransferSrcVectorDim, ck::index_t ABlockTransferSrcScalarPerVector, ck::index_t ABlockTransferDstScalarPerVector_K1, bool ABlockLdsAddExtraM, typename B0BlockTransferThreadClusterLengths_K0_L_K1, typename B0BlockTransferThreadClusterArrangeOrder, typename B0BlockTransferSrcAccessOrder, ck::index_t B0BlockTransferSrcVectorDim, ck::index_t B0BlockTransferSrcScalarPerVector, ck::index_t B0BlockTransferDstScalarPerVector_K1, bool B0BlockLdsAddExtraL, typename B1BlockTransferThreadClusterLengths_L0_N_L1, typename B1BlockTransferThreadClusterArrangeOrder, typename B1BlockTransferSrcAccessOrder, ck::index_t B1BlockTransferSrcVectorDim, ck::index_t B1BlockTransferSrcScalarPerVector, ck::index_t B1BlockTransferDstScalarPerVector_L1, bool B1BlockLdsAddExtraN, index_t CShuffleMRepeatPerShuffle, index_t CShuffleNRepeatPerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpecialization MaskingSpec, ck::LoopScheduler LoopSched = make_default_loop_scheduler(), ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>

index_t ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg::batch_size_

◆ head_count_

template<index_t NumDimG, index_t NumDimM, index_t NumDimL, index_t NumDimK, index_t NumDimN, typename ADataType, typename B0DataType, typename B1DataType, typename CDataType, typename Acc0BiasDataType, typename Acc0DataType, typename Acc1BiasDataType, typename Acc1DataType, typename CShuffleDataType, typename AElementwiseOperation, typename B0ElementwiseOperation, typename AccElementwiseOperation, typename B1ElementwiseOperation, typename CElementwiseOperation, GemmSpecialization GemmSpec, TensorSpecialization ASpec, TensorSpecialization B0Spec, TensorSpecialization B1Spec, TensorSpecialization CSpec, ck::index_t NumPrefetch, ck::index_t BlockSize, ck::index_t MPerBlock, ck::index_t LPerBlock, ck::index_t KPerBlock, ck::index_t AK1, ck::index_t BK1, ck::index_t NPerBlock, ck::index_t LTilePerBlock, ck::index_t L1, ck::index_t MPerWmma, ck::index_t LPerWmma, ck::index_t NPerWmma, ck::index_t MRepeat, ck::index_t LRepeat, ck::index_t NRepeat, typename ABlockTransferThreadClusterLengths_K0_M_K1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, ck::index_t ABlockTransferSrcVectorDim, ck::index_t ABlockTransferSrcScalarPerVector, ck::index_t ABlockTransferDstScalarPerVector_K1, bool ABlockLdsAddExtraM, typename B0BlockTransferThreadClusterLengths_K0_L_K1, typename B0BlockTransferThreadClusterArrangeOrder, typename B0BlockTransferSrcAccessOrder, ck::index_t B0BlockTransferSrcVectorDim, ck::index_t B0BlockTransferSrcScalarPerVector, ck::index_t B0BlockTransferDstScalarPerVector_K1, bool B0BlockLdsAddExtraL, typename B1BlockTransferThreadClusterLengths_L0_N_L1, typename B1BlockTransferThreadClusterArrangeOrder, typename B1BlockTransferSrcAccessOrder, ck::index_t B1BlockTransferSrcVectorDim, ck::index_t B1BlockTransferSrcScalarPerVector, ck::index_t B1BlockTransferDstScalarPerVector_L1, bool B1BlockLdsAddExtraN, index_t CShuffleMRepeatPerShuffle, index_t CShuffleNRepeatPerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpecialization MaskingSpec, ck::LoopScheduler LoopSched = make_default_loop_scheduler(), ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>

index_t ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg::head_count_

◆ head_size_

template<index_t NumDimG, index_t NumDimM, index_t NumDimL, index_t NumDimK, index_t NumDimN, typename ADataType, typename B0DataType, typename B1DataType, typename CDataType, typename Acc0BiasDataType, typename Acc0DataType, typename Acc1BiasDataType, typename Acc1DataType, typename CShuffleDataType, typename AElementwiseOperation, typename B0ElementwiseOperation, typename AccElementwiseOperation, typename B1ElementwiseOperation, typename CElementwiseOperation, GemmSpecialization GemmSpec, TensorSpecialization ASpec, TensorSpecialization B0Spec, TensorSpecialization B1Spec, TensorSpecialization CSpec, ck::index_t NumPrefetch, ck::index_t BlockSize, ck::index_t MPerBlock, ck::index_t LPerBlock, ck::index_t KPerBlock, ck::index_t AK1, ck::index_t BK1, ck::index_t NPerBlock, ck::index_t LTilePerBlock, ck::index_t L1, ck::index_t MPerWmma, ck::index_t LPerWmma, ck::index_t NPerWmma, ck::index_t MRepeat, ck::index_t LRepeat, ck::index_t NRepeat, typename ABlockTransferThreadClusterLengths_K0_M_K1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, ck::index_t ABlockTransferSrcVectorDim, ck::index_t ABlockTransferSrcScalarPerVector, ck::index_t ABlockTransferDstScalarPerVector_K1, bool ABlockLdsAddExtraM, typename B0BlockTransferThreadClusterLengths_K0_L_K1, typename B0BlockTransferThreadClusterArrangeOrder, typename B0BlockTransferSrcAccessOrder, ck::index_t B0BlockTransferSrcVectorDim, ck::index_t B0BlockTransferSrcScalarPerVector, ck::index_t B0BlockTransferDstScalarPerVector_K1, bool B0BlockLdsAddExtraL, typename B1BlockTransferThreadClusterLengths_L0_N_L1, typename B1BlockTransferThreadClusterArrangeOrder, typename B1BlockTransferSrcAccessOrder, ck::index_t B1BlockTransferSrcVectorDim, ck::index_t B1BlockTransferSrcScalarPerVector, ck::index_t B1BlockTransferDstScalarPerVector_L1, bool B1BlockLdsAddExtraN, index_t CShuffleMRepeatPerShuffle, index_t CShuffleNRepeatPerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpecialization MaskingSpec, ck::LoopScheduler LoopSched = make_default_loop_scheduler(), ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>

index_t ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg::head_size_

◆ p_out_grid_

template<index_t NumDimG, index_t NumDimM, index_t NumDimL, index_t NumDimK, index_t NumDimN, typename ADataType, typename B0DataType, typename B1DataType, typename CDataType, typename Acc0BiasDataType, typename Acc0DataType, typename Acc1BiasDataType, typename Acc1DataType, typename CShuffleDataType, typename AElementwiseOperation, typename B0ElementwiseOperation, typename AccElementwiseOperation, typename B1ElementwiseOperation, typename CElementwiseOperation, GemmSpecialization GemmSpec, TensorSpecialization ASpec, TensorSpecialization B0Spec, TensorSpecialization B1Spec, TensorSpecialization CSpec, ck::index_t NumPrefetch, ck::index_t BlockSize, ck::index_t MPerBlock, ck::index_t LPerBlock, ck::index_t KPerBlock, ck::index_t AK1, ck::index_t BK1, ck::index_t NPerBlock, ck::index_t LTilePerBlock, ck::index_t L1, ck::index_t MPerWmma, ck::index_t LPerWmma, ck::index_t NPerWmma, ck::index_t MRepeat, ck::index_t LRepeat, ck::index_t NRepeat, typename ABlockTransferThreadClusterLengths_K0_M_K1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, ck::index_t ABlockTransferSrcVectorDim, ck::index_t ABlockTransferSrcScalarPerVector, ck::index_t ABlockTransferDstScalarPerVector_K1, bool ABlockLdsAddExtraM, typename B0BlockTransferThreadClusterLengths_K0_L_K1, typename B0BlockTransferThreadClusterArrangeOrder, typename B0BlockTransferSrcAccessOrder, ck::index_t B0BlockTransferSrcVectorDim, ck::index_t B0BlockTransferSrcScalarPerVector, ck::index_t B0BlockTransferDstScalarPerVector_K1, bool B0BlockLdsAddExtraL, typename B1BlockTransferThreadClusterLengths_L0_N_L1, typename B1BlockTransferThreadClusterArrangeOrder, typename B1BlockTransferSrcAccessOrder, ck::index_t B1BlockTransferSrcVectorDim, ck::index_t B1BlockTransferSrcScalarPerVector, ck::index_t B1BlockTransferDstScalarPerVector_L1, bool B1BlockLdsAddExtraN, index_t CShuffleMRepeatPerShuffle, index_t CShuffleNRepeatPerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpecialization MaskingSpec, ck::LoopScheduler LoopSched = make_default_loop_scheduler(), ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>

CDataType* ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg::p_out_grid_

◆ p_qkv_grid_

template<index_t NumDimG, index_t NumDimM, index_t NumDimL, index_t NumDimK, index_t NumDimN, typename ADataType, typename B0DataType, typename B1DataType, typename CDataType, typename Acc0BiasDataType, typename Acc0DataType, typename Acc1BiasDataType, typename Acc1DataType, typename CShuffleDataType, typename AElementwiseOperation, typename B0ElementwiseOperation, typename AccElementwiseOperation, typename B1ElementwiseOperation, typename CElementwiseOperation, GemmSpecialization GemmSpec, TensorSpecialization ASpec, TensorSpecialization B0Spec, TensorSpecialization B1Spec, TensorSpecialization CSpec, ck::index_t NumPrefetch, ck::index_t BlockSize, ck::index_t MPerBlock, ck::index_t LPerBlock, ck::index_t KPerBlock, ck::index_t AK1, ck::index_t BK1, ck::index_t NPerBlock, ck::index_t LTilePerBlock, ck::index_t L1, ck::index_t MPerWmma, ck::index_t LPerWmma, ck::index_t NPerWmma, ck::index_t MRepeat, ck::index_t LRepeat, ck::index_t NRepeat, typename ABlockTransferThreadClusterLengths_K0_M_K1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, ck::index_t ABlockTransferSrcVectorDim, ck::index_t ABlockTransferSrcScalarPerVector, ck::index_t ABlockTransferDstScalarPerVector_K1, bool ABlockLdsAddExtraM, typename B0BlockTransferThreadClusterLengths_K0_L_K1, typename B0BlockTransferThreadClusterArrangeOrder, typename B0BlockTransferSrcAccessOrder, ck::index_t B0BlockTransferSrcVectorDim, ck::index_t B0BlockTransferSrcScalarPerVector, ck::index_t B0BlockTransferDstScalarPerVector_K1, bool B0BlockLdsAddExtraL, typename B1BlockTransferThreadClusterLengths_L0_N_L1, typename B1BlockTransferThreadClusterArrangeOrder, typename B1BlockTransferSrcAccessOrder, ck::index_t B1BlockTransferSrcVectorDim, ck::index_t B1BlockTransferSrcScalarPerVector, ck::index_t B1BlockTransferDstScalarPerVector_L1, bool B1BlockLdsAddExtraN, index_t CShuffleMRepeatPerShuffle, index_t CShuffleNRepeatPerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpecialization MaskingSpec, ck::LoopScheduler LoopSched = make_default_loop_scheduler(), ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>

const ADataType* ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg::p_qkv_grid_

◆ sequence_length_

template<index_t NumDimG, index_t NumDimM, index_t NumDimL, index_t NumDimK, index_t NumDimN, typename ADataType, typename B0DataType, typename B1DataType, typename CDataType, typename Acc0BiasDataType, typename Acc0DataType, typename Acc1BiasDataType, typename Acc1DataType, typename CShuffleDataType, typename AElementwiseOperation, typename B0ElementwiseOperation, typename AccElementwiseOperation, typename B1ElementwiseOperation, typename CElementwiseOperation, GemmSpecialization GemmSpec, TensorSpecialization ASpec, TensorSpecialization B0Spec, TensorSpecialization B1Spec, TensorSpecialization CSpec, ck::index_t NumPrefetch, ck::index_t BlockSize, ck::index_t MPerBlock, ck::index_t LPerBlock, ck::index_t KPerBlock, ck::index_t AK1, ck::index_t BK1, ck::index_t NPerBlock, ck::index_t LTilePerBlock, ck::index_t L1, ck::index_t MPerWmma, ck::index_t LPerWmma, ck::index_t NPerWmma, ck::index_t MRepeat, ck::index_t LRepeat, ck::index_t NRepeat, typename ABlockTransferThreadClusterLengths_K0_M_K1, typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferSrcAccessOrder, ck::index_t ABlockTransferSrcVectorDim, ck::index_t ABlockTransferSrcScalarPerVector, ck::index_t ABlockTransferDstScalarPerVector_K1, bool ABlockLdsAddExtraM, typename B0BlockTransferThreadClusterLengths_K0_L_K1, typename B0BlockTransferThreadClusterArrangeOrder, typename B0BlockTransferSrcAccessOrder, ck::index_t B0BlockTransferSrcVectorDim, ck::index_t B0BlockTransferSrcScalarPerVector, ck::index_t B0BlockTransferDstScalarPerVector_K1, bool B0BlockLdsAddExtraL, typename B1BlockTransferThreadClusterLengths_L0_N_L1, typename B1BlockTransferThreadClusterArrangeOrder, typename B1BlockTransferSrcAccessOrder, ck::index_t B1BlockTransferSrcVectorDim, ck::index_t B1BlockTransferSrcScalarPerVector, ck::index_t B1BlockTransferDstScalarPerVector_L1, bool B1BlockLdsAddExtraN, index_t CShuffleMRepeatPerShuffle, index_t CShuffleNRepeatPerShuffle, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpecialization MaskingSpec, ck::LoopScheduler LoopSched = make_default_loop_scheduler(), ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>

index_t ck::tensor_operation::device::DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle< NumDimG, NumDimM, NumDimL, NumDimK, NumDimN, ADataType, B0DataType, B1DataType, CDataType, Acc0BiasDataType, Acc0DataType, Acc1BiasDataType, Acc1DataType, CShuffleDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, GemmSpec, ASpec, B0Spec, B1Spec, CSpec, NumPrefetch, BlockSize, MPerBlock, LPerBlock, KPerBlock, AK1, BK1, NPerBlock, LTilePerBlock, L1, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, ABlockLdsAddExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0BlockLdsAddExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1BlockLdsAddExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, LoopSched, PipelineVer >::SelfAttnArg::sequence_length_

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

device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp