From c71f2f25ebffc11c8c7c18d4bafb1701c49bd22b Mon Sep 17 00:00:00 2001 From: Enrico Degregori Date: Wed, 6 Aug 2025 13:00:33 +0000 Subject: [PATCH] Add multiple Ds instances --- ...bwd_weight_multiple_d_wmma_cshuffle_v3.hpp | 2 +- ..._bwd_weight_two_stage_wmma_cshuffle_v3.hpp | 2 +- ...conv_bwd_weight_wmma_bilinear_instance.hpp | 50 ++++++++++++--- ...ed_conv_bwd_weight_wmma_scale_instance.hpp | 62 ++++++++++++++++--- ...uped_convolution_backward_weight_scale.hpp | 26 ++++++++ .../CMakeLists.txt | 1 + ...ale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp | 49 +++++++++++++++ 7 files changed, 171 insertions(+), 21 deletions(-) create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/wmma/device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_wmma_cshuffle_v3.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_wmma_cshuffle_v3.hpp index e977949d89..3764b07531 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_wmma_cshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_wmma_cshuffle_v3.hpp @@ -273,8 +273,8 @@ struct DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3 using CGridDesc_M_N = remove_cvref_t; using GridwiseGemm = GridwiseGemm_wmma_cshuffle_v3< - tensor_layout::gemm::RowMajor, tensor_layout::gemm::ColumnMajor, + tensor_layout::gemm::RowMajor, Tuple<>, tensor_layout::gemm::RowMajor, Tuple, diff --git a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_wmma_cshuffle_v3.hpp b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_wmma_cshuffle_v3.hpp index d90c58b59f..00d1406a19 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_wmma_cshuffle_v3.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_wmma_cshuffle_v3.hpp @@ -323,8 +323,8 @@ struct DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3 remove_cvref_t())>; using GridwiseGemm = GridwiseGemm_wmma_cshuffle_v3< - tensor_layout::gemm::RowMajor, tensor_layout::gemm::ColumnMajor, + tensor_layout::gemm::RowMajor, Tuple<>, tensor_layout::gemm::RowMajor, Tuple, diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_bilinear_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_bilinear_instance.hpp index 98e8eeab4b..f179f1decf 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_bilinear_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_bilinear_instance.hpp @@ -19,14 +19,6 @@ using BF16 = ck::bhalf_t; using F16 = ck::half_t; using F32 = float; -#ifdef CK_ENABLE_FP8 -using F8 = ck::f8_t; -#endif - -#ifdef CK_ENABLE_BF8 -using BF8 = ck::bf8_t; -#endif - using Empty_Tuple = ck::Tuple<>; template @@ -53,7 +45,27 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_bilinear_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 - DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, true, 1, 1, S<1, 8, 1, 8>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + // 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 + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, 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> // clang-format on >; @@ -68,7 +80,25 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_bilinear_instances = st //#################################################| Dim| | | | | Type| 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| | | | - DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, true, 1, 1, S<1, 8, 1, 8>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + // generic instance + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + // instance for small conv.K + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, 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> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_scale_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_scale_instance.hpp index 0389039e23..92fba58f54 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_scale_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_scale_instance.hpp @@ -19,14 +19,6 @@ using BF16 = ck::bhalf_t; using F16 = ck::half_t; using F32 = float; -#ifdef CK_ENABLE_FP8 -using F8 = ck::f8_t; -#endif - -#ifdef CK_ENABLE_BF8 -using BF8 = ck::bf8_t; -#endif - using Empty_Tuple = ck::Tuple<>; template @@ -53,7 +45,59 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_scale_instances = std::t //#################################################| 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 - DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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, true, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, true, 1, 1, S<1, 8, 1, 8>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // // 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 + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, 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> + // clang-format on + >; + +template +using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_scale_instances = std::tuple< + // clang-format off + //#################################################| Num| InLayout| WeiLayout| OutLayout| DsLayout| InData| WeiData| OutData| AccData| DsData| 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| CBlockTransfer| BlockGemm| BlockGemm| + //#################################################| Dim| | | | | Type| 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 + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // // instance for small conv.K + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_scale.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_scale.hpp index 9762b65b33..ded9cfdddc 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_scale.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_scale.hpp @@ -33,6 +33,22 @@ void add_device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_f16_in Scale, PassThrough>>>& instances); #endif + +#ifdef CK_ENABLE_BF16 +void add_device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( + std::vector, + BF16, + F32, + BF16, + Tuple<>, + PassThrough, + Scale, + PassThrough>>>& instances); +#endif #endif #ifdef CK_USE_XDL @@ -161,6 +177,16 @@ struct DeviceOperationInstanceFactory< add_device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_f16_instances( op_ptrs); } +#endif +#ifdef CK_ENABLE_BF16 + if constexpr(is_same_v && is_same_v && + is_same_v && + is_same_v && + is_same_v) + { + add_device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( + op_ptrs); + } #endif } } diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/CMakeLists.txt index 6327072e83..7543c2096c 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/CMakeLists.txt @@ -11,6 +11,7 @@ endif() list(APPEND GROUPED_CONV3D_BWD_WEIGHT_SCALE wmma/device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp + wmma/device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp ) add_instance_library(device_grouped_conv3d_bwd_weight_scale_instance ${GROUPED_CONV3D_BWD_WEIGHT_SCALE}) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/wmma/device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/wmma/device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp new file mode 100644 index 0000000000..661d3abd93 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight_scale/wmma/device_grouped_conv3d_bwd_weight_wmma_scale_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp @@ -0,0 +1,49 @@ +// 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_wmma_scale_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_scale_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( + std::vector, + BF16, + F32, + BF16, + Tuple<>, + PassThrough, + Scale, + PassThrough>>>& instances) +{ + // 1. Default + add_device_operation_instances( + instances, + device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_scale_instances<3, + NDHWGC, + GKZYXC, + NDHWGK, + ConvBwdWeightDefault>{}); + // 2. Filter1x1Stride1Pad0 + add_device_operation_instances( + instances, + device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_scale_instances< + 3, + NDHWGC, + GKZYXC, + NDHWGK, + ConvBwdWeightFilter1x1Stride1Pad0>{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck