ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-06-07 08:15:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add instances for xdl parity (for pipeline v1)

Browse Source
This commit is contained in:
Enrico Degregori
2025-08-01 14:12:22 +00:00
parent b1c6973ad1
commit 0b7f0cbbeb
15 changed files with 489 additions and 32 deletions
Download Patch File Download Diff File Expand all files Collapse all files

36
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp
View File

@@ -49,12 +49,18 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| KPer| ABK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlockGemm| BlockGemm|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| Pipeline| Pipeline|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| MBlock_MPerBlock| _NPerBlock| Scheduler| Version|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
// generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| KPer| ABK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlockGemm| BlockGemm|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| Pipeline| Pipeline|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| MBlock_MPerBlock| _NPerBlock| Scheduler| Version|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
// generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 16, 16, 2, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 16, 16, 2, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 2, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 4, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 2, 5, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 5, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 2, 7, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 7, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>
// clang-format on
>;
@@ -67,12 +73,18 @@ template <ck::index_t NDimSpatial,
BlockGemmPipelineVersion PipelineVersion>
using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances = std::tuple<
// clang-format off
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| KPer| ABK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlockGemm| BlockGemm|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| Pipeline| Pipeline|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| MBlock_MPerBlock| _NPerBlock| Scheduler| Version|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
// generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>
//#########################################| Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData| In| Wei| Out| ConvBackward| Block| MPer| NPer| KPer| ABK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CShuffleBlockTransfer| BlockGemm| BlockGemm|
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| Pipeline| Pipeline|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| MBlock_MPerBlock| _NPerBlock| Scheduler| Version|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
// generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 16, 16, 2, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 16, 16, 2, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 2, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 4, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 2, 5, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 5, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 2, 7, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 7, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion>
//clang-format on
>;

51
library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp
View File

@@ -47,7 +47,7 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_generic_instances = std:
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| MBlock_MPerBlock| _NPerBlock| Scheduler| Version|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
// generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 32, 8, 16, 16, 4, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 32, 8, 16, 16, 4, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 4, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>
// clang-format on
>;
@@ -64,11 +64,11 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances = std::tuple<
//#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVector| Pipeline| Pipeline|
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| MBlock_MPerBlock| _NPerBlock| Scheduler| Version|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
// generic instance
// // generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
// instance for small conv.K
// for fp16 conv.K and conv.C must be divisible by 2
// since half_t atomic_add require scalar_per_x_vector % 2 == 0
// // instance for small conv.K
// // for fp16 conv.K and conv.C must be divisible by 2
// // since half_t atomic_add require scalar_per_x_vector % 2 == 0
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 32, 8, 16, 16, 2, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 16, 16, 2, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
@@ -117,7 +117,24 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances = st
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| MBlock_MPerBlock| _NPerBlock| Scheduler| Version|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
// generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 32, 8, 16, 16, 4, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
// instance for small conv.K
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 32, 8, 16, 16, 2, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 16, 16, 2, 2, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 2, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 32, 8, 16, 16, 4, 4, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 32, 8, 16, 16, 4, 4, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 4, 4, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 32, 8, 16, 16, 4, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 32, 8, 16, 16, 4, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 32, 8, 16, 16, 2, 4, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 32, 8, 16, 16, 4, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 32, 8, 16, 16, 2, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 32, 8, 16, 16, 2, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 1, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 32, 8, 16, 16, 2, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 1, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 16, 16, 2, 2, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>
// clang-format on
>;
@@ -135,7 +152,27 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_instances = std::tuple<
//#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| MBlock_MPerBlock| _NPerBlock| Scheduler| Version|
//#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | |
// generic instance
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 32, 8, 16, 16, 4, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 4, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
// instance for small conv.K
// for bf16 conv.K and conv.C must be divisible by 2
// since half_t atomic_add require scalar_per_x_vector % 2 == 0
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 32, 8, 16, 16, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 16, 16, 2, 2, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 32, 8, 16, 16, 4, 4, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 4, 4, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 32, 8, 16, 16, 4, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 32, 8, 16, 16, 4, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 32, 8, 16, 16, 2, 4, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 32, 8, 16, 16, 4, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 32, 8, 16, 16, 2, 2, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 32, 8, 16, 16, 2, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 1, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 32, 8, 16, 16, 4, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 1, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>,
DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 16, 16, 2, 2, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 2, true, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>
// clang-format on
>;

14
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp
View File

@@ -848,6 +848,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<OutDataType, half_t> && is_same_v<ComputeTypeA, half_t> &&
is_same_v<ComputeTypeB, half_t>)
{
add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instances(
@@ -877,6 +879,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
op_ptrs);
add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_pad0_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_instances(
op_ptrs);
@@ -905,6 +909,8 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
op_ptrs);
add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_pad0_pipev1_instances(
op_ptrs);
add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_instances(
op_ptrs);
@@ -1007,6 +1013,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_default_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev1_instances(
op_ptrs);
}
#endif
#ifdef CK_ENABLE_BF16
@@ -1026,6 +1036,10 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_default_pipev1_instances(
op_ptrs);
add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pad0_pipev1_instances(
op_ptrs);
}
#endif
}

108
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_wmma.inc
View File

@@ -39,6 +39,18 @@ void add_device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instances(
// conv2d backward weight
#ifdef CK_ENABLE_FP16
void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
GNHWC,
GKYXC,
GNHWK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
GNHWC,
@@ -63,6 +75,18 @@ void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
@@ -75,6 +99,18 @@ void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pip
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_pad0_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
@@ -158,18 +194,6 @@ void add_device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instances(
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
#endif
#ifdef CK_ENABLE_BF16
@@ -209,6 +233,18 @@ void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pi
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_pad0_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
@@ -332,6 +368,30 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_default_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NGCDHW,
@@ -369,6 +429,30 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instanc
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_default_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pad0_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,

3
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt
View File

@@ -68,15 +68,18 @@ list(APPEND GROUPED_CONV2D_BWD_WEIGHT
wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instance.cpp
wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instance.cpp
wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp
wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_pad0_pipev1_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_pad0_pipev1_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instance.cpp
wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instance.cpp

10
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instance.cpp
View File

@@ -31,6 +31,16 @@ void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_i
GKYXC,
GNHWK,
ConvBwdWeightDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_instances<
2,
GNHWC,
GKYXC,
GNHWK,
ConvBwdWeightFilter1x1Stride1Pad0>{});
}
} // namespace instance

48
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp Normal file
View File

@@ -0,0 +1,48 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[g, n, hi, wi, c] * wei[g, k, y, x, c] = out[g, n, ho, wo, k]
void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
GNHWC,
GKYXC,
GNHWK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
// 1. Default
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_generic_instances<
2,
GNHWC,
GKYXC,
GNHWK,
ConvBwdWeightDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_generic_instances<
2,
GNHWC,
GKYXC,
GNHWK,
ConvBwdWeightFilter1x1Stride1Pad0>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

40
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_pad0_pipev1_instance.cpp Normal file
View File

@@ -0,0 +1,40 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_pad0_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<
2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightFilter1x1Stride1Pad0,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

39
library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_pad0_pipev1_instance.cpp Normal file
View File

@@ -0,0 +1,39 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_pad0_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<2,
NHWGC,
GKYXC,
NHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<
2,
NHWGC,
GKYXC,
NHWGK,
ConvBwdWeightFilter1x1Stride1Pad0,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

6
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt
View File

@@ -64,9 +64,13 @@ list(APPEND GROUPED_CONV3D_BWD_WEIGHT
wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instance.cpp
wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_default_pipev1_instance.cpp
wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev1_instance
wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp
wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_default_pipev1_instance.cpp
wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pad0_pipev1_instance
wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp
wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp

10
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instance.cpp
View File

@@ -30,6 +30,16 @@ void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_bf16_f32_bf1
GKZYXC,
GNDHWK,
ConvBwdWeightDefault>{});
// 2. Filter1x1Stride1Pad0
add_device_operation_instances(
instances,
device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_instances<
3,
GNDHWC,
GKZYXC,
GNDHWK,
ConvBwdWeightFilter1x1Stride1Pad0>{});
}
} // namespace instance

39
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_default_pipev1_instance.cpp Normal file
View File

@@ -0,0 +1,39 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_default_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

39
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pad0_pipev1_instance.cpp Normal file
View File

@@ -0,0 +1,39 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pad0_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
BF16,
BF16,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightFilter1x1Stride1Pad0,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

39
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_default_pipev1_instance.cpp Normal file
View File

@@ -0,0 +1,39 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_default_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightDefault,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck

39
library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev1_instance.cpp Normal file
View File

@@ -0,0 +1,39 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// Compilation parameters for in[n, hi, wi, g, c] * wei[g, k, y, x, c] = out[n, ho, wo, g, k]
void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev1_instances(
std::vector<std::unique_ptr<DeviceGroupedConvBwdWeight<3,
NDHWGC,
GKZYXC,
NDHWGK,
F16,
F16,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances)
{
add_device_operation_instances(instances,
device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<
3,
NDHWGC,
GKZYXC,
NDHWGK,
ConvBwdWeightFilter1x1Stride1Pad0,
BlockGemmPipelineScheduler::Intrawave,
BlockGemmPipelineVersion::v1>{});
}
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck