ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-16 08:44:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

Re-factor of BF16 instance tuples

Browse Source
This commit is contained in:
Wojciech Laskowski
2025-12-02 14:57:22 +00:00
parent a36fd78205
commit 4fd168ea67
54 changed files with 108 additions and 2626 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_cshufflev3_comp_instance.hpp
View File

@@ -52,43 +52,6 @@ static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_comp_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Compute friendly
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 256, 256, 32, 8, 8, 16, 16, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 256, 256, 32, 8, 8, 16, 16, 8, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 256, 32, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 256, 128, 32, 8, 8, 16, 16, 8, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 64, 64, 8, 8, 16, 16, 4, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 128, 64, 8, 8, 16, 16, 2, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 64, 8, 8, 16, 16, 2, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
// "2x" instances
// DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 16, 16, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> // Assert broken
// "part 2" instances
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 256, 256, 32, 8, 8, 16, 16, 4, 8, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
// AGPR Spill when use permuted lds layout. so, use padding for these two.
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 224, 256, 64, 8, 8, 16, 16, 7, 4, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 256, 224, 64, 8, 8, 16, 16, 4, 7, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 2, 1, S<1, 64, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,

45
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_cshufflev3_instance.hpp
View File

@@ -57,7 +57,26 @@ template <index_t NDimSpatial,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_generic_instances = std::tuple<
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_ngchw_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// generic instance
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 64, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_gnhwc_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
@@ -126,30 +145,6 @@ using device_grouped_conv_fwd_wmma_cshufflev3_bf16_instances = std::tuple<
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_16x16_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 8, 1, 1, 1, S<1, 32, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 256, 64, 64, 32, 8, 8, 16, 16, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,

26
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_wmma_cshufflev3_mem_instance.hpp
View File

@@ -49,32 +49,6 @@ static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding;
static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave;
static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,
typename DsLayout,
typename ELayout,
ConvolutionForwardSpecialization ConvSpec,
BlockGemmPipelineScheduler BlkGemmPipeSched,
typename DsDataTypes = Tuple<>,
typename OutElementOp = PassThrough>
using device_grouped_conv_fwd_wmma_cshufflev3_bf16_mem_instances = std::tuple<
// clang-format off
//########################################| NumDim| A| B| Ds| E| AData| BData| AccData| CShuffle| Ds| EData| A| B| CDE| ConvForward| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MWmma| NWmma| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Pipeline scheduler | Pipeline version |
//########################################| Spatial| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Size| Block| Block| Block| | | WMMA| WMMA| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MWmmaPerWave| NWmmaPerWave| _MBlock_MWaveMPerWmma| ScalarPerVector| | |
//########################################| | | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerWmma| _NWaveNPerWmma| | |
//########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// Latency friendly
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 32, 32, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 2, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>
#ifndef ONE_INSTANCE_PER_LIST
,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 32, 16, 128, 8, 8, 16, 16, 1, 1, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<16, 4, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 64, 32, 16, 64, 8, 8, 16, 16, 1, 1, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 8, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>,
DeviceGroupedConvFwdMultipleABD_Wmma_CShuffle_V3<NDimSpatial, ALayout, BLayout, DsLayout, ELayout, BF16, BF16, F32, BF16, DsDataTypes, BF16, PassThrough, PassThrough, OutElementOp, ConvSpec, GemmMNKPadding, 128, 32, 32, 64, 8, 8, 16, 16, 1, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>
#endif
// clang-format on
>;
template <index_t NDimSpatial,
typename ALayout,
typename BLayout,

32
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View File

@@ -870,18 +870,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_16x16_instances(
op_ptrs);
// add_device_grouped_conv2d_fwd_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_comp_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_INT8
@@ -939,14 +931,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_16x16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_comp_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_mem_intra_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_mem_inter_instances(
op_ptrs);
}
#endif
}
@@ -1094,18 +1078,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instances(
op_ptrs);
// add_device_grouped_conv3d_fwd_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_INT8
@@ -1153,14 +1129,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_16x16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_comp_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_mem_intra_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_mem_inter_instances(
op_ptrs);
}
#endif
}

16
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_clamp.hpp
View File

@@ -285,18 +285,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_16x16_instances(
op_ptrs);
// add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_comp_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
@@ -334,18 +326,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instances(
op_ptrs);
// add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16

113
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_bias_clamp_wmma_cshufflev3.inc
View File

@@ -24,20 +24,6 @@ void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
// void
// add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
@@ -67,48 +53,6 @@ void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_merged_groups_nhwg
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv2d_fwd_bias_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<NHWGK>,
NHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
@@ -123,20 +67,6 @@ void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhw
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
// void
// add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
@@ -165,49 +95,6 @@ void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_merged_groups_ndhw
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
void add_device_grouped_conv3d_fwd_bias_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<NDHWGK>,
NDHWGK,
BF16,
BF16,
Tuple<BF16>,
BF16,
PassThrough,
PassThrough,
AddClamp>>>& instances);
#endif
#ifdef CK_ENABLE_FP16

16
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_clamp.hpp
View File

@@ -281,18 +281,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_16x16_instances(
op_ptrs);
// add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_merged_groups_nhwgc_gkyxc_nhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_comp_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances(
op_ptrs);
add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16
@@ -330,18 +322,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instances(
op_ptrs);
// add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_merged_groups_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
op_ptrs);
add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_FP16

113
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_clamp_wmma_cshufflev3.inc
View File

@@ -24,20 +24,6 @@ void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
// void
// add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_large_tensor_nhwgc_gkyxc_nhwgk_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
@@ -67,48 +53,6 @@ void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_merged_groups_nhwgc_gky
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv2d_fwd_clamp_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Tuple<>,
NHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
@@ -123,20 +67,6 @@ void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
// void
// add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_large_tensor_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
// std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
@@ -165,49 +95,6 @@ void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_merged_groups_ndhwgc_gk
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
void add_device_grouped_conv3d_fwd_clamp_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Tuple<>,
NDHWGK,
BF16,
BF16,
Tuple<>,
BF16,
PassThrough,
PassThrough,
Clamp>>>& instances);
#endif
#ifdef CK_ENABLE_FP16

65
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_comp_wmma_cshufflev3.inc
View File

@@ -9,22 +9,6 @@ namespace device {
namespace instance {
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
@@ -74,39 +58,6 @@ void add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_f16_comp_in
PassThrough>>>& instances);
#endif // CK_ENABLE_FP16
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKCYX,
Empty_Tuple,
NGKHW,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
@@ -140,22 +91,6 @@ void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_f16_comp
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_comp_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NGCDHW,
GKCZYX,
Empty_Tuple,
NGKDHW,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation

65
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_mem_inter_wmma_cshufflev3.inc
View File

@@ -9,22 +9,6 @@ namespace device {
namespace instance {
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
@@ -74,39 +58,6 @@ void add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_f16_mem_int
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKCYX,
Empty_Tuple,
NGKHW,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
@@ -140,22 +91,6 @@ void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_f16_mem_
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_mem_inter_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NGCDHW,
GKCZYX,
Empty_Tuple,
NGKDHW,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation

65
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_mem_intra_wmma_cshufflev3.inc
View File

@@ -9,22 +9,6 @@ namespace device {
namespace instance {
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_f16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
@@ -74,39 +58,6 @@ void add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_f16_mem_int
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKCYX,
Empty_Tuple,
NGKHW,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
// grouped conv3d forward, NDHWGC/GKZYXC/NDHWGK
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_f16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
@@ -140,22 +91,6 @@ void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_f16_mem_
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_mem_intra_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NGCDHW,
GKCZYX,
Empty_Tuple,
NGKDHW,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance
} // namespace device
} // namespace tensor_operation

55
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_wmma_cshufflev3.inc
View File

@@ -105,20 +105,6 @@ void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_instan
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_nhwgc_gkyxc_nhwgk_bf16_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
@@ -261,20 +247,6 @@ void add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_instan
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_wmma_cshufflev3_ngchw_gkcyx_ngkhw_bf16_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<2,
NGCHW,
GKCYX,
Empty_Tuple,
NGKHW,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
@@ -342,20 +314,6 @@ void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_ins
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ndhwgc_gkzyxc_ndhwgk_bf16_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NDHWGC,
GKZYXC,
Empty_Tuple,
NDHWGK,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_FP16
@@ -534,19 +492,6 @@ void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_ins
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_fwd_wmma_cshufflev3_ngcdhw_gkczyx_ngkdhw_bf16_16x16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleABD<3,
NGCDHW,
GKCZYX,
Empty_Tuple,
NGKDHW,
BF16,
BF16,
Empty_Tuple,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
} // namespace instance