Member List

• ck
• GridwiseBatchedGemmSoftmaxGemm_Wmma
• SharedMemTrait

ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait Member List

This is the complete list of members for ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait, including all inherited members.

a_block_space_offset	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static
a_block_space_size_aligned	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static
b0_block_space_offset	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static
b0_block_space_size_aligned	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static
b1_block_space_offset	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static
b1_block_space_size_aligned	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static
c_block_space_size	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static
max_lds_align	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static
reduction_space_offset	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static
reduction_space_size_aligned	ck::GridwiseBatchedGemmSoftmaxGemm_Wmma< ADataType, B0DataType, Acc0DataType, B1DataType, Acc1DataType, CShuffleDataType, CDataType, AElementwiseOperation, B0ElementwiseOperation, AccElementwiseOperation, B1ElementwiseOperation, CElementwiseOperation, CGlobalMemoryDataOperation, AGridDesc, B0GridDesc, B1GridDesc, CGridDesc_M_N, MPerBlock, LPerBlock, KPerBlock, AK1Value, BK1Value, NPerBlock, LTilePerBlock, L1Value, MPerWmma, LPerWmma, NPerWmma, MRepeat, LRepeat, NRepeat, BlockSize, ABlockTransferThreadClusterLengths_K0_M_K1, ABlockTransferThreadClusterArrangeOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1, AThreadTransferSrcResetCoordinateAfterRun, AEnableLds, ABlockLdsExtraM, B0BlockTransferThreadClusterLengths_K0_L_K1, B0BlockTransferThreadClusterArrangeOrder, B0BlockTransferSrcAccessOrder, B0BlockTransferSrcVectorDim, B0BlockTransferSrcScalarPerVector, B0BlockTransferDstScalarPerVector_K1, B0ThreadTransferSrcResetCoordinateAfterRun, B0EnableLds, B0BlockLdsExtraL, B1BlockTransferThreadClusterLengths_L0_N_L1, B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferSrcAccessOrder, B1BlockTransferSrcVectorDim, B1BlockTransferSrcScalarPerVector, B1BlockTransferDstScalarPerVector_L1, B1ThreadTransferSrcResetCoordinateAfterRun, B1EnableLds, B1BlockLdsExtraN, CShuffleMRepeatPerShuffle, CShuffleNRepeatPerShuffle, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock, PadN, MaskOutUpperTriangle, NumGemmKPrefetchStage, LoopSched, PipelineVer >::SharedMemTrait	static