From 7a516b8e9908e4f68d22dbdb7bba724cfbccf259 Mon Sep 17 00:00:00 2001 From: kiefer Date: Mon, 20 Oct 2025 13:34:56 +0000 Subject: [PATCH 01/18] Adapt all grouped conv bwd weight vanilla Xdl instances to 16x16. MRepeat doubled for all but 12 of them (some static assert failure). Also added custom reduced profiler target for building grouped conv bwd weight vanilla only profiler. Verified with gtest test. --- ...p_gemm_xdl_universal_km_kn_mn_instance.hpp | 20 +- ...conv_bwd_weight_two_stage_xdl_instance.hpp | 124 ++++++------ ...rouped_conv_bwd_weight_v3_xdl_instance.hpp | 22 +-- ...e_grouped_conv_bwd_weight_xdl_instance.hpp | 176 +++++++++--------- profiler/src/CMakeLists.txt | 19 ++ 5 files changed, 190 insertions(+), 171 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp index f9d77ca451..80691b91c6 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp @@ -45,18 +45,18 @@ using device_gemm_xdl_universal_km_kn_mn_comp_instances = //#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 32, 32, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 32, 32, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 32, 32, 4, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 32, 32, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 32, 32, 4, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 2, 2, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, // Can we support this kind of odd case? 224(256) = 28*8 + (4*8) //DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 8, 16, 16, 7, 8, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Interwave, 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_two_stage_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp index 3633015a78..9e3bb0f961 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp @@ -64,15 +64,15 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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>, 1, Scheduler, PipelineVersion, 8> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -89,11 +89,11 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_part2_ins //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> @@ -154,15 +154,15 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instance //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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>, 1, Scheduler, PipelineVersion, 8> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -180,11 +180,11 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_part2_in //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> @@ -247,27 +247,27 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8 ,1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8 ,1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> // clang-format on >; @@ -284,9 +284,9 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_part2_ins //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, F16, F16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, F16, F16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 4, 4>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 2, 4>, @@ -331,27 +331,27 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instance //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8 ,1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8 ,1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> // clang-format on >; @@ -370,9 +370,9 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_part2_in //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, BF16, BF16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, BF16, BF16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 4, 4>, DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 2, 4>, diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp index 143d857333..990bf5423e 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp @@ -55,7 +55,7 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_instances = std::tuple //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion> + // DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion> // clang-format on >; @@ -91,11 +91,11 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f16_instances = std::tuple //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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_Xdl_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_Xdl_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_Xdl_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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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 @@ -115,11 +115,11 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_bf16_instances = std::tupl //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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_Xdl_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_Xdl_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_Xdl_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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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 diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp index 22f1912198..0c9154360e 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp @@ -53,7 +53,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_generic_instances = std:: //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> +// DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 16, 16, 4, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> // clang-format on >; @@ -71,29 +71,29 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances = std::tuple< //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // instances for small conv.K and conv.C - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 16, 16, 4, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // // instances for small conv.K and conv.C + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 16, 16, 4, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 16, 16, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 16, 16, 8, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 16, 16, 4, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 16, 16, 8, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 16, 16, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 16, 16, 4, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 16, 16, 4, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 16, 16, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 16, 16, 4, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 16, 16, 4, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 16, 16, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 16, 16, 2, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 16, 16, 2, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 16, 16, 2, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 16, 16, 2, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -165,7 +165,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_generic_instances = std:: //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2> // clang-format on >; @@ -183,26 +183,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances = std::tuple< //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -218,7 +218,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_generic_instanc //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> // clang-format on >; @@ -236,24 +236,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_instances = std //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // instance for small conv.K - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -271,26 +271,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances = std::tuple< //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -309,26 +309,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_comp_bf8_f8_instances = s //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | #if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8 // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> #endif // clang-format on >; diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 5833fc3626..8b1090d0ff 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -283,3 +283,22 @@ message(VERBOSE "ckProfiler libs: ${PROFILER_LIBS}") target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE ${PROFILER_LIBS}) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) + +## Defining specific operation targets + +macro(define_profiler_target NAME SOURCES LIBS) + add_executable(${NAME} profiler.cpp ${SOURCES}) + target_compile_options(${NAME} PRIVATE -Wno-global-constructors) + + if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) + target_compile_options(${NAME} PRIVATE --offload-compress) + endif() + + target_link_libraries(${NAME} PRIVATE utility getopt::getopt ${LIBS}) + + rocm_install(TARGETS ${NAME} COMPONENT profiler) +endmacro() + +define_profiler_target(ckProfiler_grouped_conv_bwd_weight_vanilla + "profile_grouped_conv_bwd_weight.cpp" + "device_grouped_convnd_bwd_weight_instance;device_grouped_conv1d_bwd_weight_instance;device_grouped_conv2d_bwd_weight_instance;device_grouped_conv3d_bwd_weight_instance") \ No newline at end of file From 133f5538e34e66e122a50a0a553a3c6f84178fd8 Mon Sep 17 00:00:00 2001 From: kiefer Date: Mon, 1 Dec 2025 11:25:03 +0000 Subject: [PATCH 02/18] Reduce instances to only the tuned wmma V3 ones for implicit v1 intra and explicit v1 intra pad/nopad. --- ..._gemm_wmma_universal_km_kn_mn_instance.hpp | 98 +++++---- ...p_gemm_xdl_universal_km_kn_mn_instance.hpp | 160 +++++++------- ...ce_grouped_conv_bwd_weight_dl_instance.hpp | 6 +- ...onv_bwd_weight_two_stage_wmma_instance.hpp | 188 ++++++++--------- ...conv_bwd_weight_two_stage_xdl_instance.hpp | 192 ++++++++--------- ...ouped_conv_bwd_weight_v3_wmma_instance.hpp | 36 ++-- ...rouped_conv_bwd_weight_v3_xdl_instance.hpp | 30 +-- ...conv_bwd_weight_wmma_bilinear_instance.hpp | 74 +++---- ..._grouped_conv_bwd_weight_wmma_instance.hpp | 112 +++++----- ...ed_conv_bwd_weight_wmma_scale_instance.hpp | 72 +++---- ..._conv_bwd_weight_xdl_bilinear_instance.hpp | 180 ++++++++-------- ...e_grouped_conv_bwd_weight_xdl_instance.hpp | 196 +++++++++--------- ...ped_conv_bwd_weight_xdl_scale_instance.hpp | 178 ++++++++-------- ...c_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp | 20 +- ..._gkyxc_nhwgk_bf16_pad0_pipev1_instance.cpp | 21 +- ...c_gkyxc_nhwgk_f16_pad0_pipev1_instance.cpp | 20 +- ...kzyxc_ndhwgk_bf16_pad0_pipev1_instance.cpp | 20 +- ...gkzyxc_ndhwgk_f16_pad0_pipev1_instance.cpp | 20 +- ...16_bf16_bf16_exp_comp_default_instance.cpp | 4 +- ...bf16_bf16_exp_comp_mnkpadding_instance.cpp | 4 +- ..._bf16_bf16_exp_mem_v1_default_instance.cpp | 52 ++--- ...16_bf16_exp_mem_v1_mnkpadding_instance.cpp | 56 ++--- ..._bf16_bf16_exp_mem_v2_default_instance.cpp | 52 ++--- ...16_bf16_exp_mem_v2_mnkpadding_instance.cpp | 56 ++--- ..._f16_f16_f16_exp_comp_default_instance.cpp | 4 +- ...6_f16_f16_exp_comp_mnkpadding_instance.cpp | 4 +- ...16_f16_f16_exp_mem_v1_default_instance.cpp | 52 ++--- ...f16_f16_exp_mem_v1_mnkpadding_instance.cpp | 56 ++--- ...16_f16_f16_exp_mem_v2_default_instance.cpp | 52 ++--- ...f16_f16_exp_mem_v2_mnkpadding_instance.cpp | 56 ++--- 30 files changed, 1045 insertions(+), 1026 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp index daf221bb5a..da7208bce0 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp @@ -37,46 +37,58 @@ static constexpr auto GemmMNKPadding = GemmSpecialization::MNKPadding; static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; -template -using device_gemm_wmma_universal_km_kn_mn_comp_instances = std::tuple< +template +using device_gemm_wmma_universal_km_kn_mn_GemmDefault_instances = std::tuple< // clang-format off //#####################################| ALayout| BLayout| DsLayout |ELayout| ADataType| BDataType| DsDataType| CDataType| AccDataType| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| BlockwiseGemm| BlockwiseGemm| //#####################################| | | | | | | | | | DataType| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock_NBlock_NPerBlock| ScalarPerVectors| Pipeline| Pipeline| //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4,4,4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4,4,4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 2, 2, 16, 16, 4, 2, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4,4,4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4,4,4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4,4,4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S< 8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4,4,4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4,4,4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4,4,4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4,4,4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> - + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 48, 96, 64, 8, 8, 16, 16, 3, 3, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 64, 128, 8, 8, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 192, 48, 96, 192, 8, 8, 16, 16, 3, 1, S<24, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<24, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 12>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 48, 64, 64, 8, 8, 16, 16, 3, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 96, 128, 64, 8, 8, 16, 16, 6, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 192, 32, 96, 192, 8, 8, 16, 16, 2, 1, S<24, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<24, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, 1, 1, S<1, 16, 1, 12>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 192, 32, 96, 192, 8, 8, 16, 16, 2, 1, S<24, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<24, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 12>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 32, 8, 8, 16, 16, 4, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 32, 64, 8, 8, 16, 16, 2, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // clang-format on >; -template -using device_gemm_wmma_universal_km_kn_mn_mem_instances = std::tuple< +template +using device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances = std::tuple< // clang-format off - //#####################################| ALayout| BLayout| DsLayout |ELayout| ADataType| BDataType| DsDataType| CDataType| AccDataType| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| BlockwiseGemm| BlockwiseGemm| - //#####################################| | | | | | | | | | DataType| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock_NBlock_NPerBlock| ScalarPerVectors| Pipeline| Pipeline| - //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| - //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - // Latency friendly - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 8, 1, 8>, S<2,2,2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, S<2,2,2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 8, 1, 4>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 8, 1, 8>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> - // Memory friendly - // TODO: add once v2 is implemented + //#####################################| ALayout| BLayout| DsLayout |ELayout| ADataType| BDataType| DsDataType| CDataType| AccDataType| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| BlockwiseGemm| BlockwiseGemm| + //#####################################| | | | | | | | | | DataType| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock_NBlock_NPerBlock| ScalarPerVectors| Pipeline| Pipeline| + //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| + //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 128, 48, 64, 128, 8, 8, 16, 16, 3, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 48, 64, 64, 8, 8, 16, 16, 3, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 96, 64, 32, 8, 8, 16, 16, 6, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 48, 32, 128, 8, 8, 16, 16, 3, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 96, 64, 96, 48, 8, 8, 16, 16, 4, 2, S<6, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<6, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // clang-format on >; +// device_gemm_wmma_universal_km_kn_mn_comp_instances +// device_gemm_wmma_universal_km_kn_mn_mem_instances + template @@ -87,11 +99,11 @@ using device_gemm_wmma_universal_km_kn_mn_irregular_odd_m_instances = std::tuple //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 8, 1, 8>, S<2,2,2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, S<2,2,2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 8, 1, 4>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 8, 1, 8>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 8, 1, 8>, S<2,2,2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, S<2,2,2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 8, 1, 4>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 8, 1, 8>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> // Memory friendly // TODO: add once v2 is implemented // clang-format on @@ -107,11 +119,11 @@ using device_gemm_wmma_universal_km_kn_mn_odd_n_instances = std::tuple< //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 4>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 4>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> // Memory friendly // TODO: add once v2 is implemented // clang-format on @@ -127,11 +139,11 @@ using device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances = std::tupl //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 4>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 4>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> // Memory friendly // TODO: add once v2 is implemented // clang-format on diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp index 80691b91c6..04bd56d98e 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp @@ -45,18 +45,18 @@ using device_gemm_xdl_universal_km_kn_mn_comp_instances = //#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 2, 2, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - // Can we support this kind of odd case? 224(256) = 28*8 + (4*8) - //DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 8, 16, 16, 7, 8, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 2, 2, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + // // Can we support this kind of odd case? 224(256) = 28*8 + (4*8) + // //DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 8, 16, 16, 7, 8, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> // clang-format on >; @@ -70,23 +70,23 @@ using device_gemm_xdl_universal_km_kn_mn_mem_instances = std::tuple< //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // Memory friendly - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // // Memory friendly + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; @@ -100,23 +100,23 @@ using device_gemm_xdl_universal_km_kn_mn_irregular_odd_m_instances = std::tuple< //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // Memory friendly - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // // Memory friendly + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; @@ -130,23 +130,23 @@ using device_gemm_xdl_universal_km_kn_mn_odd_n_instances = std::tuple< //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // Memory friendly - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // // Memory friendly + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; @@ -160,23 +160,23 @@ using device_gemm_xdl_universal_km_kn_mn_irregular_odd_mn_instances = std::tuple //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // Memory friendly - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // // Memory friendly + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp index 0493a55e2f..b1c28a18a0 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp @@ -43,7 +43,7 @@ using device_grouped_conv_bwd_weight_dl_f32_instances = //############################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | _K0_M0_M1_K1| _K0_M0_M1_K1| ArrangeOrder| | _K0_M0_M1_K1| ContiguousDimOrder| _K0_M0_M1_K1| _K0_N0_N1_K1| _K0_N0_N1_K1| ArrangeOrder| | _K0_N0_N1_K1| ContiguousDimOrder| _K0_N0_N1_K1| | | | //############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >; @@ -60,7 +60,7 @@ using device_grouped_conv_bwd_weight_dl_f16_instances = //############################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | _K0_M0_M1_K1| _K0_M0_M1_K1| ArrangeOrder| | _K0_M0_M1_K1| ContiguousDimOrder| _K0_M0_M1_K1| _K0_N0_N1_K1| _K0_N0_N1_K1| ArrangeOrder| | _K0_N0_N1_K1| ContiguousDimOrder| _K0_N0_N1_K1| | | | //############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >; @@ -77,7 +77,7 @@ using device_grouped_conv_bwd_weight_dl_bf16_instances = //############################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | _K0_M0_M1_K1| _K0_M0_M1_K1| ArrangeOrder| | _K0_M0_M1_K1| ContiguousDimOrder| _K0_M0_M1_K1| _K0_N0_N1_K1| _K0_N0_N1_K1| ArrangeOrder| | _K0_N0_N1_K1| ContiguousDimOrder| _K0_N0_N1_K1| | | | //############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> + // DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp index 0a1167b1bd..ecae19070e 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp @@ -45,16 +45,16 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_instance //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_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, 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>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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, 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8>, - DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8> + // DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -73,14 +73,14 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_in //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4> // clang-format on >; @@ -99,11 +99,11 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregula //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 2, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 2, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 2, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 2, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 2, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 2, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -120,16 +120,16 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_instanc //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_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, 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>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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, 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8>, - DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8> + // DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -147,14 +147,14 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_i //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4> // clang-format on >; @@ -172,11 +172,11 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregul //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 2, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 2, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_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<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>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_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<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>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 2, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 2, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 2, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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<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>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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<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>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 2, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -194,7 +194,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_ //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1> // clang-format on >; @@ -212,29 +212,29 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_instance //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> // clang-format on >; @@ -252,17 +252,17 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_in //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, F16, F16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 2, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 2, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 2, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 2, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4> // clang-format on >; @@ -280,7 +280,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1> // clang-format on >; @@ -297,29 +297,29 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_instanc //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1>, - DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> + // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> // clang-format on >; @@ -337,17 +337,17 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_i //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, BF16, BF16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 2, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 2, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 2, 4>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 2, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp index 9e3bb0f961..570ffaae50 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp @@ -46,7 +46,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_generic_i //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -63,16 +63,16 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 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>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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>, 1, Scheduler, PipelineVersion, 8> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -89,14 +89,14 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_part2_ins //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> // clang-format on >; @@ -114,11 +114,11 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_irregular //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 4, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 4, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 4, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 4, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 4, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 4, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -136,7 +136,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_generic_ //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -153,16 +153,16 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instance //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 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>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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>, 1, Scheduler, PipelineVersion, 8> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -180,14 +180,14 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_part2_in //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> // clang-format on >; @@ -205,11 +205,11 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_irregula //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 4, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 4, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 4, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 4, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 4, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 4, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -227,7 +227,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_i //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1> // clang-format on >; @@ -245,29 +245,29 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8 ,1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8 ,1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> // clang-format on >; @@ -284,17 +284,17 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_part2_ins //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, F16, F16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, F16, F16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 2, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 2, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 2, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 2, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4> // clang-format on >; @@ -312,7 +312,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_ //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1> // clang-format on >; @@ -329,29 +329,29 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instance //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8 ,1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8 ,1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> // clang-format on >; @@ -370,17 +370,17 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_part2_in //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, BF16, BF16, 2, 2>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, BF16, BF16, 2, 2>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 2, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 2, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 2, 4>, - DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4> + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 2, 4>, + // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp index 531a0408b0..ca94bf625c 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp @@ -54,14 +54,17 @@ using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances = std::tupl //#########################################| 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 + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, Scheduler, PipelineVersion> + // clang-format on >; template , 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> + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, Scheduler, PipelineVersion>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, Scheduler, PipelineVersion> //clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp index 990bf5423e..4bdf6e7dc0 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp @@ -73,7 +73,7 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances = std:: //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion, TF32, TF32> + // DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion, TF32, TF32> // clang-format on >; @@ -91,13 +91,13 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f16_instances = std::tuple //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_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_Xdl_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_Xdl_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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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> + // DeviceGroupedConvBwdWeight_Xdl_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_Xdl_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_Xdl_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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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 >; @@ -115,13 +115,13 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_bf16_instances = std::tupl //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_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_Xdl_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_Xdl_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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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> + // DeviceGroupedConvBwdWeight_Xdl_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_Xdl_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_Xdl_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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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 >; 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 9f8a315c59..f430170da6 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 @@ -45,27 +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, 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, 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> + // 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 >; @@ -81,24 +81,24 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_bilinear_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 - 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, 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> + // 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_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp index 46e9f34988..f870b7dc9b 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp @@ -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, 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> + // 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 >; @@ -65,25 +65,25 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_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, 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 - 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>, + // 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 + // 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>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, 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, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, 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, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, 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, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, 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, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> // clang-format on >; @@ -99,7 +99,7 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_instan //#########################################| 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, 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> + // 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> // clang-format on >; @@ -117,24 +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, 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, 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> + // 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 >; @@ -152,26 +152,26 @@ 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, 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, 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> + // 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 >; 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 c2b752188f..a41b468189 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 @@ -45,26 +45,26 @@ 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, 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, 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> + // 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 >; @@ -80,24 +80,24 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_scale_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, 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, 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> + // 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_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp index 75655dda8d..063234d426 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp @@ -54,24 +54,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_bilinear_instances = std: //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - // instances for small conv.K and conv.C - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + // // instances for small conv.K and conv.C + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4> // clang-format on >; @@ -87,24 +87,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_bilinear_instances = //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>, - // instances for small conv.K and conv.C - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>, + // // instances for small conv.K and conv.C + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32> // clang-format on >; @@ -120,27 +120,27 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_bilinear_instances = std: //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2>, - // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2>, + // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8> // clang-format on >; @@ -156,24 +156,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_bilinear_instances = std //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - // instance for small conv.K - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + // // instance for small conv.K + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4> // clang-format on >; @@ -190,26 +190,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_comp_bf8_f8_bilinear_inst //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | #if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8 // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8>, - // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8>, + // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> #endif // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp index 0c9154360e..24847b9d69 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp @@ -109,7 +109,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_generic_instances = //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32> +// DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32> // clang-format on >; @@ -127,29 +127,29 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_instances = std::tup //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // instances for small conv.K and conv.C - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // // instances for small conv.K and conv.C + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -165,7 +165,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_generic_instances = std:: //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2> +// DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2> // clang-format on >; @@ -183,26 +183,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances = std::tuple< //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -218,7 +218,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_generic_instanc //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> +// DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> // clang-format on >; @@ -236,24 +236,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_instances = std //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // instance for small conv.K - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // // instance for small conv.K + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -271,26 +271,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances = std::tuple< //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -309,26 +309,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_comp_bf8_f8_instances = s //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | #if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8 // generic instance - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> #endif // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp index ed1cb3997f..ab50f44f33 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp @@ -54,24 +54,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_scale_instances = std::tu //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - // instances for small conv.K and conv.C - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + // // instances for small conv.K and conv.C + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4> // clang-format on >; @@ -87,24 +87,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_scale_instances = st //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>, - // instances for small conv.K and conv.C - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>, + // // instances for small conv.K and conv.C + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32> // clang-format on >; @@ -120,26 +120,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_scale_instances = std::tu //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2>, - // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2>, + // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8> // clang-format on >; @@ -155,24 +155,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_scale_instances = std::t //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - // instance for small conv.K - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + // // instance for small conv.K + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4> // clang-format on >; @@ -189,26 +189,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_comp_bf8_f8_scale_instanc //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | #if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8 // generic instance - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8>, - // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8>, + // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> #endif // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp index 365feeafd4..0136625318 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp @@ -11,7 +11,7 @@ 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_pad0_pipev1_instances( - std::vector>>& instances) { - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - 2, - GNHWC, - GKYXC, - GNHWK, - ConvBwdWeightFilter1x1Stride1Pad0, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances(instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< + // 2, + // GNHWC, + // GKYXC, + // GNHWK, + // ConvBwdWeightFilter1x1Stride1Pad0, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/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 b/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 index 85ce996a3b..7d04517257 100644 --- a/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 +++ b/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 @@ -11,7 +11,7 @@ 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>>& 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>{}); + + // 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 diff --git a/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 b/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 index 1fee43c50d..724fa6ef2a 100644 --- a/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 +++ b/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 @@ -11,7 +11,7 @@ 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>>& 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>{}); + // 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 diff --git a/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 b/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 index e5784c3885..202849ee37 100644 --- a/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 +++ b/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 @@ -11,7 +11,7 @@ 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>>& 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>{}); + // 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 diff --git a/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 b/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 index 09f6e42afb..96671737ef 100644 --- a/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 +++ b/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 @@ -11,7 +11,7 @@ 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>>& 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>{}); + // 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 diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_default_instance.cpp index 4db9c93b32..894063e081 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_default_instance.cpp @@ -32,7 +32,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_default_i PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_comp_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmDefault_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_default_instances( @@ -58,7 +58,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_default_i PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_comp_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmDefault_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instance.cpp index a4b726e6bf..3a60d2f4c9 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instance.cpp @@ -32,7 +32,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpaddin PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_comp_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instances( @@ -58,7 +58,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpaddin PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_comp_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instance.cpp index c98858ad1e..f238ce5b38 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instance.cpp @@ -10,7 +10,7 @@ namespace device { namespace instance { void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); + // add_explicit_gemm_device_operation_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // BF16, + // BF16, + // BF16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); + // add_explicit_gemm_device_operation_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // BF16, + // BF16, + // BF16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instance.cpp index d03a824c92..1a55e2658f 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instance.cpp @@ -10,7 +10,7 @@ namespace device { namespace instance { void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>( - instances); + // add_explicit_gemm_device_operation_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // BF16, + // BF16, + // BF16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>( + // instances); } void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>( - instances); + // add_explicit_gemm_device_operation_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // BF16, + // BF16, + // BF16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>( + // instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instance.cpp index f8121658a3..4de541abff 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instance.cpp @@ -10,7 +10,7 @@ namespace device { namespace instance { void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); + // add_explicit_gemm_device_operation_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // BF16, + // BF16, + // BF16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); + // add_explicit_gemm_device_operation_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // BF16, + // BF16, + // BF16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instance.cpp index bad4c278a1..f9470a632a 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instance.cpp @@ -10,7 +10,7 @@ namespace device { namespace instance { void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>( - instances); + // add_explicit_gemm_device_operation_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // BF16, + // BF16, + // BF16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>( + // instances); } void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>( - instances); + // add_explicit_gemm_device_operation_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // BF16, + // BF16, + // BF16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>( + // instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_default_instance.cpp index 4bbccaed72..38e98e719e 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_default_instance.cpp @@ -32,7 +32,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_default_inst PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_comp_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmDefault_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_default_instances( @@ -58,7 +58,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_default_inst PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_comp_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmDefault_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instance.cpp index 0f4d35f4c6..7bf57a9505 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instance.cpp @@ -32,7 +32,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_i PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_comp_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instances( @@ -58,7 +58,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_i PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_comp_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instance.cpp index 0e819b87e6..21e27e24ff 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instance.cpp @@ -10,7 +10,7 @@ namespace device { namespace instance { void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); + // add_explicit_gemm_device_operation_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // F16, + // F16, + // F16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); + // add_explicit_gemm_device_operation_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // F16, + // F16, + // F16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instance.cpp index 9ec81e7d7b..27e5ed1d36 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instance.cpp @@ -10,7 +10,7 @@ namespace device { namespace instance { void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>( - instances); + // add_explicit_gemm_device_operation_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // F16, + // F16, + // F16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>( + // instances); } void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>( - instances); + // add_explicit_gemm_device_operation_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // F16, + // F16, + // F16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>( + // instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instance.cpp index c8a344d343..d4f9148c24 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instance.cpp @@ -10,7 +10,7 @@ namespace device { namespace instance { void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); + // add_explicit_gemm_device_operation_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // F16, + // F16, + // F16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); + // add_explicit_gemm_device_operation_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // F16, + // F16, + // F16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instance.cpp index 2f31e05e87..361aa80b0f 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instance.cpp @@ -10,7 +10,7 @@ namespace device { namespace instance { void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>( - instances); + // add_explicit_gemm_device_operation_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // F16, + // F16, + // F16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>( + // instances); } void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instances( - std::vector>>& instances) { - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_mem_instances>( - instances); + // add_explicit_gemm_device_operation_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // F16, + // F16, + // F16, + // PassThrough, + // PassThrough, + // PassThrough, + // device_gemm_wmma_universal_km_kn_mn_mem_instances>( + // instances); } } // namespace instance From 8e8866083471fd9db03fe8877dd18b7d49fc5986 Mon Sep 17 00:00:00 2001 From: kiefer Date: Tue, 2 Dec 2025 14:14:04 +0000 Subject: [PATCH 03/18] Add explicit oddMN support with custom tuned instances --- ..._gemm_wmma_universal_km_kn_mn_instance.hpp | 20 ++++++++++++------- ...mma_bf16_bf16_bf16_exp_odd_mn_instance.cpp | 6 ++---- ...t_wmma_f16_f16_f16_exp_odd_mn_instance.cpp | 6 ++---- 3 files changed, 17 insertions(+), 15 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp index da7208bce0..7d5486e708 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp @@ -130,8 +130,7 @@ using device_gemm_wmma_universal_km_kn_mn_odd_n_instances = std::tuple< >; template + BlockGemmPipelineScheduler BlkGemmPipeSched> using device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances = std::tuple< // clang-format off //#####################################| ALayout| BLayout| DsLayout |ELayout| ADataType| BDataType| DsDataType| CDataType| AccDataType| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| BlockwiseGemm| BlockwiseGemm| @@ -139,11 +138,18 @@ using device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances = std::tupl //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 4>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 32, 64, 8, 8, 16, 16, 1, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 32, 64, 8, 8, 16, 16, 1, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 32, 128, 8, 8, 16, 16, 1, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 48, 128, 8, 8, 16, 16, 1, 3, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 64, 32, 8, 8, 16, 16, 1, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 64, 32, 8, 8, 16, 16, 1, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 64, 64, 8, 8, 16, 16, 1, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 64, 128, 8, 8, 16, 16, 1, 4, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 96, 64, 8, 8, 16, 16, 1, 6, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 96, 128, 8, 8, 16, 16, 1, 6, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 128, 192, 32, 8, 8, 16, 16, 1, 12, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 256, 256, 96, 64, 8, 8, 16, 16, 2, 6, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 16>, S<1, 1, 1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> // Memory friendly // TODO: add once v2 is implemented // clang-format on diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instance.cpp index 98eb2851c1..8db0818f67 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instance.cpp @@ -33,8 +33,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instanc PassThrough, PassThrough, device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); + Intrawave>>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instances( @@ -61,8 +60,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instanc PassThrough, PassThrough, device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); + Intrawave>>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp index 54485d7298..1129d1c9fe 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp @@ -33,8 +33,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instances( PassThrough, PassThrough, device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); + Intrawave>>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instances( @@ -61,8 +60,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instances( PassThrough, PassThrough, device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); + Intrawave>>(instances); } } // namespace instance From 4fe6c7ddcb023e5f5bf1fdce4b656f80909e7a76 Mon Sep 17 00:00:00 2001 From: kiefer Date: Tue, 2 Dec 2025 15:50:59 +0000 Subject: [PATCH 04/18] Add two stage instances based on the parameters from the tuned cshuffle V3 instances. CShuffleBlockTranserScalarPerVector adapted to 4, and mergegroups fixed to 1 for now. No more special instance lists. --- ...onv_bwd_weight_two_stage_wmma_instance.hpp | 297 +++--------------- ...gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp | 22 +- ..._gkcyx_ngkhw_f16_pipev1_part2_instance.cpp | 22 +- ...ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp | 22 +- ..._ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp | 22 +- ...c_nhwgk_bf16_pipev1_irregular_instance.cpp | 22 +- ...gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp | 22 +- ...xc_nhwgk_f16_pipev1_irregular_instance.cpp | 22 +- ..._gkyxc_nhwgk_f16_pipev1_part2_instance.cpp | 22 +- ..._ndhwgk_bf16_pipev1_irregular_instance.cpp | 22 +- ...zyxc_ndhwgk_bf16_pipev1_part2_instance.cpp | 22 +- ...c_ndhwgk_f16_pipev1_irregular_instance.cpp | 22 +- ...kzyxc_ndhwgk_f16_pipev1_part2_instance.cpp | 22 +- ...czyx_ngkdhw_bf16_pipev1_part2_instance.cpp | 22 +- ...kczyx_ngkdhw_f16_pipev1_part2_instance.cpp | 22 +- ...dhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp | 22 +- ...cdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp | 22 +- 17 files changed, 225 insertions(+), 424 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp index ecae19070e..c3a7859b06 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp @@ -45,67 +45,22 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_instance //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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, 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>, 1, Scheduler, PipelineVersion, 2>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8> - + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // clang-format on >; -template -using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, +// device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_instances - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4> - - // clang-format on - >; - -template -using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregular_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 2, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 2, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 2, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> - // clang-format on - >; +// device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregular_instances template , S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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, 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>, 1, Scheduler, PipelineVersion, 2>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 8> + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // clang-format on >; -template -using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, +// device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_instances - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 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, 2, 1, 16>, 1, Scheduler, PipelineVersion, 4> - // clang-format on - >; +// device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_instances -template -using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 2, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 2, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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<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>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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<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>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 2, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> - // clang-format on - >; - -template -using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1> - // clang-format on - >; +// device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances // NGCHW requires transpose, we use vector loads and stores params for them template , S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> - // clang-format on + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> + // clang-format on >; -template -using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, F16, F16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, +// device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 2, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 4>, - - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 2, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4> - // clang-format on - >; - -template -using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1> - // clang-format on - >; +// device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances template , S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1>, - - // DeviceGroupedConvBwdWeightTwoStage_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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - // DeviceGroupedConvBwdWeightTwoStage_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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // clang-format on >; -template -using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, BF16, BF16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, - - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 2, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 8, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 4>, - - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 2, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 4, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4> - // clang-format on - >; - +// device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_instances } // namespace instance } // namespace device } // namespace tensor_operation diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp index 677e03b196..c5ecdc087e 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_instances< - 2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_instances< + // 2, + // NGCHW, + // GKCYX, + // NGKHW, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instance.cpp index a1e8ab5a1e..b55f3c7c74 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances< - 2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances< + // 2, + // NGCHW, + // GKCYX, + // NGKHW, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp index e222e8c663..6168df220d 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< - 2, - NGCHW, - GKYXC, - NGKHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< + // 2, + // NGCHW, + // GKYXC, + // NGKHW, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp index 85d395c7c2..e9b2752f85 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< - 2, - NGCHW, - GKYXC, - NGKHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< + // 2, + // NGCHW, + // GKYXC, + // NGKHW, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instance.cpp index ef8ba7109d..2578205be1 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp index c6787c8200..9102cd49bf 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instance.cpp index c71f60416e..f919b9ebd3 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregular_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregular_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instance.cpp index 35e7dccce4..416f157391 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instance.cpp index ed27879639..1803e85006 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instance.cpp index 32e3988eb1..f4975aa1ba 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instance.cpp index 5e044213ce..97cdcfc196 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregular_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregular_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instance.cpp index d872fa1fc8..0ddcc2911d 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instance.cpp index 49a7ce52b8..d7a22c93ae 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_instances< - 3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_instances< + // 3, + // NGCDHW, + // GKCZYX, + // NGKDHW, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instance.cpp index cb446b5c3a..d996737e9a 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances< - 3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances< + // 3, + // NGCDHW, + // GKCZYX, + // NGKDHW, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp index 39d8ef3091..0ae3433c4e 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< - 3, - NGCDHW, - GKZYXC, - NGKDHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< + // 3, + // NGCDHW, + // GKZYXC, + // NGKDHW, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp index 76b6fc9b0d..bc4f5c8cec 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp @@ -11,7 +11,7 @@ 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_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances( - std::vector>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< - 3, - NGCDHW, - GKZYXC, - NGKDHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< + // 3, + // NGCDHW, + // GKZYXC, + // NGKDHW, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance From 9d5810942a69be52ec5ac07a37fbfa2b4388ba5f Mon Sep 17 00:00:00 2001 From: kiefer Date: Tue, 2 Dec 2025 16:02:40 +0000 Subject: [PATCH 05/18] Replace cshuffle non-v3 lists with v3 lists, making sure to not have duplications. Also removing stride1pad0 support for NHWGC since we can use explicit for those cases. --- ...ouped_conv_bwd_weight_v3_wmma_instance.hpp | 10 +++--- ...eight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp | 14 ++++---- ...kyxc_gnhwk_f16_default_pipev1_instance.cpp | 20 +++++------ ...ht_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp | 28 +++++++-------- ...t_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp | 32 ++++++++--------- ...ht_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp | 32 ++++++++--------- ...yxc_nhwgk_bf16_default_pipev1_instance.cpp | 20 +++++------ ...t_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp | 26 +++++++------- ...kyxc_nhwgk_f16_default_pipev1_instance.cpp | 20 +++++------ ...ht_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp | 26 +++++++------- ...wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp | 28 +++++++-------- ...xc_ndhwgk_bf16_default_pipev1_instance.cpp | 20 +++++------ ...mma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp | 26 +++++++------- ...yxc_ndhwgk_f16_default_pipev1_instance.cpp | 20 +++++------ ...wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp | 26 +++++++------- ...mma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp | 34 ++++++++----------- ...wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp | 32 ++++++++--------- 17 files changed, 199 insertions(+), 215 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp index ca94bf625c..1263b13e5f 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp @@ -45,15 +45,14 @@ template + BlockGemmPipelineScheduler Scheduler = BlockGemmPipelineScheduler::Intrawave, + BlockGemmPipelineVersion PipelineVersion = BlockGemmPipelineVersion::v1> 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 8, Scheduler, PipelineVersion>, @@ -72,15 +71,14 @@ template + BlockGemmPipelineScheduler Scheduler = BlockGemmPipelineScheduler::Intrawave, + BlockGemmPipelineVersion PipelineVersion = BlockGemmPipelineVersion::v1> 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 8, Scheduler, PipelineVersion>, diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp index 724a5f409f..b19f31da22 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -24,14 +24,14 @@ void add_device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instances( // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances<1, - GNWC, - GKXC, - GNWK, - ConvBwdWeightDefault>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<1, + GNWC, + GKXC, + GNWK, + ConvBwdWeightDefault>{}); // 2. Filter1x1Stride1Pad0 add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances< + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< 1, GNWC, GKXC, diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instance.cpp index 78b2d7b93d..c43e4d4fb5 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instance.cpp @@ -11,7 +11,7 @@ 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_default_pipev1_instances( - std::vector>>& instances) { - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - 2, - GNHWC, - GKYXC, - GNHWK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances(instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< + // 2, + // GNHWC, + // GKYXC, + // GNHWK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance } // namespace device diff --git a/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 b/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 index 131c505468..445fc8dd94 100644 --- a/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 +++ b/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 @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,21 +25,19 @@ void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_instances( // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_generic_instances< - 2, - GNHWC, - GKYXC, - GNHWK, - ConvBwdWeightDefault>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_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>{}); + add_device_operation_instances(instances, + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< + 2, + GNHWC, + GKYXC, + GNHWK, + ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp index 67db6f5f1f..f852a707b7 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,22 +25,20 @@ void add_device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instances( // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_instances<2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault, - 1, - 1>{}); - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_instances<2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault, - 4, - 4>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<2, + NGCHW, + GKCYX, + NGKHW, + ConvBwdWeightDefault>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<2, + // NGCHW, + // GKCYX, + // NGKHW, + // ConvBwdWeightDefault, + // 4, + // 4>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp index e4d306c41a..dc80707fbc 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,22 +25,20 @@ void add_device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instances( // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances<2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault, - 1, - 1>{}); - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances<2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault, - 4, - 4>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<2, + NGCHW, + GKCYX, + NGKHW, + ConvBwdWeightDefault>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<2, + // NGCHW, + // GKCYX, + // NGKHW, + // ConvBwdWeightDefault, + // 4, + // 4>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instance.cpp index 6768244bb3..ad0de5e9f4 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instance.cpp @@ -11,7 +11,7 @@ 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_default_pipev1_instances( - std::vector>>& instances) { - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances(instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp index 78b29101ed..93c1aef2ca 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,19 +25,19 @@ void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instances( // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_instances<2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightDefault>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<2, + NHWGC, + GKYXC, + NHWGK, + ConvBwdWeightDefault>{}); // 2. Filter1x1Stride1Pad0 - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); + // add_device_operation_instances(instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instance.cpp index e323cd3d72..e2fe192584 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instance.cpp @@ -11,7 +11,7 @@ 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_default_pipev1_instances( - std::vector>>& instances) { - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); + // add_device_operation_instances(instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // ConvBwdWeightDefault, + // BlockGemmPipelineScheduler::Intrawave, + // BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp index c4c40f49be..fc9a710374 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,19 +25,19 @@ void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instances( // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances<2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightDefault>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<2, + NHWGC, + GKYXC, + NHWGK, + ConvBwdWeightDefault>{}); // 2. Filter1x1Stride1Pad0 - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); + // add_device_operation_instances(instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< + // 2, + // NHWGC, + // GKYXC, + // NHWGK, + // ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp index aa82de070f..7cd9d5655f 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -24,21 +24,19 @@ void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_generic_instances< - 3, - GNDHWC, - GKZYXC, - GNDHWK, - ConvBwdWeightDefault>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<3, + GNDHWC, + GKZYXC, + GNDHWK, + ConvBwdWeightDefault>{}); // 2. Filter1x1Stride1Pad0 - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_generic_instances< - 3, - GNDHWC, - GKZYXC, - GNDHWK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); + add_device_operation_instances(instances, + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< + 3, + GNDHWC, + GKZYXC, + GNDHWK, + ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/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 b/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 index 828936212c..e2ec8928f4 100644 --- a/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 +++ b/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 @@ -11,7 +11,7 @@ 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>>& 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>{}); + // 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 diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp index 3b08d9f1ad..d5fa41ae4a 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,19 +25,19 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instanc // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_instances<3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightDefault>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<3, + NDHWGC, + GKZYXC, + NDHWGK, + ConvBwdWeightDefault>{}); // 2. Filter1x1Stride1Pad0 - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); + // add_device_operation_instances(instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/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 b/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 index a9126540ff..413a254644 100644 --- a/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 +++ b/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 @@ -11,7 +11,7 @@ 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>>& 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>{}); + // 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 diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp index 5374dc1be9..1f0286e5ff 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,19 +25,19 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances<3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightDefault>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<3, + NDHWGC, + GKZYXC, + NDHWGK, + ConvBwdWeightDefault>{}); // 2. Filter1x1Stride1Pad0 - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); + // add_device_operation_instances(instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< + // 3, + // NDHWGC, + // GKZYXC, + // NDHWGK, + // ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp index 4aff904534..4b58393704 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -23,24 +23,20 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instanc PassThrough>>>& instances) { // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_instances<3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault, - 1, - 1>{}); - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_instances<3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault, - 4, - 4>{}); + add_device_operation_instances(instances, + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances < + 3, + NGCDHW, + GKCZYX, + NGKDHW, + ConvBwdWeightDefault>{}); + // add_device_operation_instances(instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances < + // 3, + // NGCDHW, + // GKCZYX, + // NGKDHW, + // ConvBwdWeightDefault); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp index 2cfbeb7a96..ee5ceba586 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,22 +25,20 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances<3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault, - 1, - 1>{}); - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances<3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault, - 4, - 4>{}); + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<3, + NGCDHW, + GKCZYX, + NGKDHW, + ConvBwdWeightDefault>{}); + // add_device_operation_instances( + // instances, + // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<3, + // NGCDHW, + // GKCZYX, + // NGKDHW, + // ConvBwdWeightDefault, + // 4, + // 4>{}); } } // namespace instance From c9c05bfc7deafe0239ea3e6ebed2bbb361d4717d Mon Sep 17 00:00:00 2001 From: kiefer Date: Thu, 4 Dec 2025 08:27:08 +0000 Subject: [PATCH 06/18] Remove some instances that give incorrect results (f16 NHWGC) --- ..._gemm_wmma_universal_km_kn_mn_instance.hpp | 54 +++++++++++++------ ...onv_bwd_weight_two_stage_wmma_instance.hpp | 8 +-- ...ouped_conv_bwd_weight_v3_wmma_instance.hpp | 12 ++--- ...bf16_bf16_exp_comp_mnkpadding_instance.cpp | 4 +- ...6_f16_f16_exp_comp_mnkpadding_instance.cpp | 4 +- 5 files changed, 51 insertions(+), 31 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp index 7d5486e708..6c88a9c970 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp @@ -63,26 +63,47 @@ using device_gemm_wmma_universal_km_kn_mn_GemmDefault_instances = std::tuple< // clang-format on >; -template -using device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances = std::tuple< +using device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_f16_instances = std::tuple< // clang-format off //#####################################| ALayout| BLayout| DsLayout |ELayout| ADataType| BDataType| DsDataType| CDataType| AccDataType| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| BlockwiseGemm| BlockwiseGemm| //#####################################| | | | | | | | | | DataType| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock_NBlock_NPerBlock| ScalarPerVectors| Pipeline| Pipeline| //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 128, 48, 64, 128, 8, 8, 16, 16, 3, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 48, 64, 64, 8, 8, 16, 16, 3, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 96, 64, 32, 8, 8, 16, 16, 6, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 48, 32, 128, 8, 8, 16, 16, 3, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 96, 64, 96, 48, 8, 8, 16, 16, 4, 2, S<6, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<6, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 128, 48, 64, 128, 8, 8, 16, 16, 3, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 48, 64, 64, 8, 8, 16, 16, 3, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 96, 64, 32, 8, 8, 16, 16, 6, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 48, 32, 128, 8, 8, 16, 16, 3, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 96, 64, 96, 48, 8, 8, 16, 16, 4, 2, S<6, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<6, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, // Incorrect results for f16 + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, F16, F16, Tuple<>, F16, F32, F16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + // clang-format on + >; + +using device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_bf16_instances = std::tuple< + // clang-format off + //#####################################| ALayout| BLayout| DsLayout |ELayout| ADataType| BDataType| DsDataType| CDataType| AccDataType| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| BlockwiseGemm| BlockwiseGemm| + //#####################################| | | | | | | | | | DataType| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock_NBlock_NPerBlock| ScalarPerVectors| Pipeline| Pipeline| + //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| + //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 128, 48, 64, 128, 8, 8, 16, 16, 3, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 64, 64, 8, 8, 16, 16, 4, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 64, 96, 32, 8, 8, 16, 16, 4, 3, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 48, 64, 64, 8, 8, 16, 16, 3, 2, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 96, 64, 32, 8, 8, 16, 16, 6, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 48, 32, 128, 8, 8, 16, 16, 3, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 96, 64, 8, 8, 16, 16, 2, 3, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 96, 64, 96, 48, 8, 8, 16, 16, 4, 2, S<6, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<6, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, // Incorrect results for f16 + DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, BF16, BF16, Tuple<>, BF16, F32, BF16, PassThrough, PassThrough, PassThrough, GemmMNKPadding, 64, 32, 32, 128, 8, 8, 16, 16, 2, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // clang-format on >; @@ -129,8 +150,7 @@ using device_gemm_wmma_universal_km_kn_mn_odd_n_instances = std::tuple< // clang-format on >; -template +template using device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances = std::tuple< // clang-format off //#####################################| ALayout| BLayout| DsLayout |ELayout| ADataType| BDataType| DsDataType| CDataType| AccDataType| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| BlockwiseGemm| BlockwiseGemm| diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp index c3a7859b06..1a6266a840 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp @@ -52,9 +52,9 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_instance DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, // Incorrect results for at least GemmDefault + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // Incorrect results for at least GemmDefault // clang-format on >; @@ -118,7 +118,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_instance DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> - // clang-format on + // clang-format on >; // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp index 1263b13e5f..df666f413d 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp @@ -45,7 +45,7 @@ template using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances = std::tuple< // clang-format off @@ -60,10 +60,10 @@ using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances = std::tupl DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, Scheduler, PipelineVersion>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, Scheduler, PipelineVersion>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, Scheduler, PipelineVersion> - // clang-format on + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, Scheduler, PipelineVersion> + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, Scheduler, PipelineVersion>, // Incorrect results for at least GemmDefault + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, Scheduler, PipelineVersion> // Incorrect results for at least GemmDefault + // clang-format on >; template using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances = std::tuple< // clang-format off diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instance.cpp index 3a60d2f4c9..a3b16e4216 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instance.cpp @@ -32,7 +32,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpaddin PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_bf16_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instances( @@ -58,7 +58,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpaddin PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_bf16_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instance.cpp index 7bf57a9505..b0a8998562 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instance.cpp @@ -32,7 +32,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_i PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_f16_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instances( @@ -58,7 +58,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_i PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_f16_instances>(instances); } } // namespace instance From 93ab2c4a8cc334b550f10e8197a6428214b3eb6b Mon Sep 17 00:00:00 2001 From: kiefer Date: Thu, 4 Dec 2025 09:55:23 +0000 Subject: [PATCH 07/18] Add bf16 f32 bf16 instances based on tuned b16 NHWGC GKYXC instances. --- ..._grouped_conv_bwd_weight_wmma_instance.hpp | 44 +++++-------------- ...hwc_gkyxc_gnhwk_bf16_f32_bf16_instance.cpp | 4 +- ...c_gkzyxc_gndhwk_bf16_f32_bf16_instance.cpp | 4 +- 3 files changed, 15 insertions(+), 37 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp index f870b7dc9b..d39cdf4628 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp @@ -87,21 +87,7 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_instances = std::tuple< // clang-format on >; -template -using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_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, 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> - // clang-format on - >; +// device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_instances template , 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> + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 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, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 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, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> // clang-format on >; diff --git a/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 b/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 index b2a78136f7..68e76ac150 100644 --- a/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 +++ b/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 @@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_i // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_instances< + device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< 2, GNHWC, GKYXC, @@ -35,7 +35,7 @@ void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_i // 2. Filter1x1Stride1Pad0 add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_instances< + device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< 2, GNHWC, GKYXC, diff --git a/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 b/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 index b05ca0788f..673d2b253d 100644 --- a/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 +++ b/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 @@ -24,7 +24,7 @@ void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_bf16_f32_bf1 // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_instances< + device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< 3, GNDHWC, GKZYXC, @@ -34,7 +34,7 @@ void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_bf16_f32_bf1 // 2. Filter1x1Stride1Pad0 add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_instances< + device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< 3, GNDHWC, GKZYXC, From c2fab9d6777e7ed47cacd21bbdd8c80e5554869c Mon Sep 17 00:00:00 2001 From: kiefer Date: Fri, 5 Dec 2025 13:35:22 +0000 Subject: [PATCH 08/18] Add back some generic instances to make sure we have the same shape / layout / datatype support as before the instance selection process. --- ...onv_bwd_weight_two_stage_wmma_instance.hpp | 40 +++++++++++++++++- ...ouped_conv_bwd_weight_v3_wmma_instance.hpp | 2 + ..._grouped_conv_bwd_weight_wmma_instance.hpp | 1 + ...ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp | 41 ++++++++++--------- ..._ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp | 41 ++++++++++--------- ...dhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp | 41 ++++++++++--------- ...cdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp | 41 ++++++++++--------- 7 files changed, 125 insertions(+), 82 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp index 1a6266a840..f41f4191f7 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp @@ -45,6 +45,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_instance //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, @@ -75,6 +76,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_instanc //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, @@ -92,7 +94,23 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_instanc // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_instances -// device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances +template +using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_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| NumGroups| + //################################################| 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| ToMerge| + //################################################| 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| Sched| Ver| | + //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1> + // clang-format on + >; // NGCHW requires transpose, we use vector loads and stores params for them template , S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, @@ -123,7 +142,23 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_instance // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances -// device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances +template +using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_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| NumGroups| + //################################################| 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| ToMerge| + //################################################| 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| Sched| Ver| | + //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1> + // clang-format on + >; template , S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp index df666f413d..761b07ea60 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp @@ -53,6 +53,7 @@ using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances = std::tupl //#########################################| 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| | | | + 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, 0, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, 0, 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 8, Scheduler, PipelineVersion>, @@ -79,6 +80,7 @@ using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances = std::tup //#########################################| 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| | | | + 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, 0, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, 0, 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, Scheduler, PipelineVersion>, DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 8, Scheduler, PipelineVersion>, diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp index d39cdf4628..60746e37ce 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp @@ -103,6 +103,7 @@ 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, 64, 64, 32, 8, 16, 16, 4, 2, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 16, 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp index 6168df220d..bdafbeba8f 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp @@ -11,28 +11,29 @@ 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_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instances( - [[maybe_unused]]std::vector>>& instances) + [[maybe_unused]] std::vector>>& + instances) { // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< - // 2, - // NGCHW, - // GKYXC, - // NGKHW, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); + add_device_operation_instances( + instances, + device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< + 2, + NGCHW, + GKYXC, + NGKHW, + ConvBwdWeightDefault, + BlockGemmPipelineScheduler::Intrawave, + BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp index e9b2752f85..a663d608d0 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp @@ -11,28 +11,29 @@ 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_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instances( - [[maybe_unused]]std::vector>>& instances) + [[maybe_unused]] std::vector>>& + instances) { // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< - // 2, - // NGCHW, - // GKYXC, - // NGKHW, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); + add_device_operation_instances( + instances, + device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< + 2, + NGCHW, + GKYXC, + NGKHW, + ConvBwdWeightDefault, + BlockGemmPipelineScheduler::Intrawave, + BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp index 0ae3433c4e..234970d8ac 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp @@ -11,28 +11,29 @@ 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_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances( - [[maybe_unused]]std::vector>>& instances) + [[maybe_unused]] std::vector>>& + instances) { // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< - // 3, - // NGCDHW, - // GKZYXC, - // NGKDHW, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); + add_device_operation_instances( + instances, + device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< + 3, + NGCDHW, + GKZYXC, + NGKDHW, + ConvBwdWeightDefault, + BlockGemmPipelineScheduler::Intrawave, + BlockGemmPipelineVersion::v1>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp index bc4f5c8cec..a9b148b688 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp @@ -11,28 +11,29 @@ 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_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances( - [[maybe_unused]]std::vector>>& instances) + [[maybe_unused]] std::vector>>& + instances) { // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< - // 3, - // NGCDHW, - // GKZYXC, - // NGKDHW, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); + add_device_operation_instances( + instances, + device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< + 3, + NGCDHW, + GKZYXC, + NGKDHW, + ConvBwdWeightDefault, + BlockGemmPipelineScheduler::Intrawave, + BlockGemmPipelineVersion::v1>{}); } } // namespace instance From 6380f9ab812e9b54a952e953fe5ff5dfa7dbcef8 Mon Sep 17 00:00:00 2001 From: kiefer Date: Fri, 5 Dec 2025 13:37:51 +0000 Subject: [PATCH 09/18] Add instances for scale and bilinear based on the bf16 NHWGC GKYXC tuning. Keep generic instances for support. --- ...conv_bwd_weight_wmma_bilinear_instance.hpp | 67 ++++++++----------- ...ed_conv_bwd_weight_wmma_scale_instance.hpp | 67 ++++++++----------- 2 files changed, 56 insertions(+), 78 deletions(-) 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 f430170da6..8045bba070 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 @@ -45,27 +45,21 @@ 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, 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> + 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, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 1, 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, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 4>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + // 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, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, 0, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, 0, 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 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, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 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, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, // Presumably doesn't produce correct results for f16 + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // Presumably doesn't produce correct results for f16 // clang-format on >; @@ -81,24 +75,19 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_bilinear_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 - // 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> + 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, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + // other instances + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, 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, 0, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, 2, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, 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 a41b468189..c1e0c1046b 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 @@ -45,27 +45,21 @@ 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, 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 + 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, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 4, 1, 1, 1, S<1, 16, 1, 4>, 2, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + // 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, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, 0, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, 0, 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 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, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 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, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, // Presumably doesn't produce correct results for fp16 + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // Presumably doesn't produce correct results for fp16 + // clang-format on >; template , 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> + 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, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + // other instances + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, 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, 0, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, 2, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, 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, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 8, 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, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, 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, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // clang-format on >; From 6a5ea18a65e4f40e7038714088ef646fb68eed78 Mon Sep 17 00:00:00 2001 From: kiefer Date: Fri, 5 Dec 2025 15:34:14 +0000 Subject: [PATCH 10/18] Disable two stage f16 instances which produce incorrect results. --- ...vice_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp index f41f4191f7..09b828c342 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp @@ -134,9 +134,9 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_instance DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, // Incorrect results for at least GemmDefault + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // Incorrect results for at least GemmDefault // clang-format on >; From 5470b80774daf308472ac29e60633b7fb9931988 Mon Sep 17 00:00:00 2001 From: kiefer Date: Mon, 8 Dec 2025 14:38:49 +0000 Subject: [PATCH 11/18] Remove more instances which fail verification, for bf16_f32_bf16 and for f16 scale / bilinear. --- ...evice_grouped_conv_bwd_weight_wmma_bilinear_instance.hpp | 6 +++--- .../device_grouped_conv_bwd_weight_wmma_instance.hpp | 6 +++--- .../device_grouped_conv_bwd_weight_wmma_scale_instance.hpp | 6 +++--- 3 files changed, 9 insertions(+), 9 deletions(-) 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 8045bba070..f254628f73 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 @@ -85,9 +85,9 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_bilinear_instances = st DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, // Verification failure + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // Verification failure // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp index 60746e37ce..8c66baf475 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp @@ -111,9 +111,9 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances = st DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 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, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, // Verification failure + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> // Verification failture // 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 c1e0c1046b..218995a984 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 @@ -84,9 +84,9 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_scale_instances = std:: DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, 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, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, - DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, // Verification failure + // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // Verification failure // clang-format on >; From 31f08a016979fa36c53d0d699eed556c2a754e6d Mon Sep 17 00:00:00 2001 From: kiefer Date: Mon, 8 Dec 2025 15:07:15 +0000 Subject: [PATCH 12/18] Disable all non-generic two-stage instances in the instance lists for NHWGC. They are never faster and support is already carried by CShuffleV3 and Explicit. --- ...onv_bwd_weight_two_stage_wmma_instance.hpp | 40 +++++++++---------- 1 file changed, 20 insertions(+), 20 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp index 09b828c342..395cba8133 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp @@ -45,15 +45,15 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_instance //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1> + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1> // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, // Incorrect results for at least GemmDefault // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // Incorrect results for at least GemmDefault // clang-format on @@ -76,17 +76,17 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_instanc //################################################| 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| ToMerge| //################################################| 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| Sched| Ver| | //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> + DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1> + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, + // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // clang-format on >; From c2faf36db277c1727ad4b8f0d2b43cce89e9dc6a Mon Sep 17 00:00:00 2001 From: kiefer Date: Tue, 9 Dec 2025 13:02:48 +0000 Subject: [PATCH 13/18] Remove unused instance lists and related add_x_instance() functions, fwd declarations, cmakelists entries. Also merge the "wmma" and "wmma v3" instance list files, which are both v3. --- ..._gemm_wmma_universal_km_kn_mn_instance.hpp | 43 --- ...onv_bwd_weight_two_stage_wmma_instance.hpp | 8 - ...ouped_conv_bwd_weight_v3_wmma_instance.hpp | 28 ++ ..._grouped_conv_bwd_weight_wmma_instance.hpp | 161 ---------- .../grouped_convolution_backward_weight.hpp | 95 ------ ...volution_backward_weight_explicit_wmma.inc | 288 ------------------ ...ouped_convolution_backward_weight_wmma.inc | 264 ---------------- ..._gnwc_gkxc_gnwk_bf16_f32_bf16_instance.cpp | 6 +- .../grouped_conv2d_bwd_weight/CMakeLists.txt | 12 - ...hwc_gkyxc_gnhwk_bf16_f32_bf16_instance.cpp | 6 +- ...kyxc_gnhwk_f16_default_pipev1_instance.cpp | 38 --- ...c_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp | 39 --- ...gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp | 41 --- ..._gkcyx_ngkhw_f16_pipev1_part2_instance.cpp | 41 --- ...t_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp | 9 - ...ht_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp | 9 - ...c_nhwgk_bf16_pipev1_irregular_instance.cpp | 41 --- ...gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp | 41 --- ...xc_nhwgk_f16_pipev1_irregular_instance.cpp | 41 --- ..._gkyxc_nhwgk_f16_pipev1_part2_instance.cpp | 41 --- ...yxc_nhwgk_bf16_default_pipev1_instance.cpp | 39 --- ...wgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp | 6 +- ...t_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp | 8 - ..._gkyxc_nhwgk_bf16_pad0_pipev1_instance.cpp | 40 --- ...kyxc_nhwgk_f16_default_pipev1_instance.cpp | 39 --- ...ht_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp | 8 - ...c_gkyxc_nhwgk_f16_pad0_pipev1_instance.cpp | 39 --- .../grouped_conv3d_bwd_weight/CMakeLists.txt | 10 - ...c_gkzyxc_gndhwk_bf16_f32_bf16_instance.cpp | 6 +- ..._ndhwgk_bf16_pipev1_irregular_instance.cpp | 41 --- ...zyxc_ndhwgk_bf16_pipev1_part2_instance.cpp | 41 --- ...c_ndhwgk_f16_pipev1_irregular_instance.cpp | 41 --- ...kzyxc_ndhwgk_f16_pipev1_part2_instance.cpp | 41 --- ...xc_ndhwgk_bf16_default_pipev1_instance.cpp | 39 --- ...c_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp | 6 +- ...mma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp | 8 - ...kzyxc_ndhwgk_bf16_pad0_pipev1_instance.cpp | 39 --- ...yxc_ndhwgk_f16_default_pipev1_instance.cpp | 39 --- ...wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp | 8 - ...gkzyxc_ndhwgk_f16_pad0_pipev1_instance.cpp | 39 --- ...czyx_ngkdhw_bf16_pipev1_part2_instance.cpp | 41 --- ...kczyx_ngkdhw_f16_pipev1_part2_instance.cpp | 41 --- ...mma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp | 21 +- ...wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp | 9 - .../grouped_convnd_bwd_weight/CMakeLists.txt | 13 - ..._bf16_bf16_exp_mem_v1_default_instance.cpp | 67 ---- ...16_bf16_exp_mem_v1_mnkpadding_instance.cpp | 69 ----- ..._bf16_bf16_exp_mem_v2_default_instance.cpp | 67 ---- ...16_bf16_exp_mem_v2_mnkpadding_instance.cpp | 69 ----- ...wmma_bf16_bf16_bf16_exp_odd_m_instance.cpp | 71 ----- ...wmma_bf16_bf16_bf16_exp_odd_n_instance.cpp | 69 ----- ...16_f16_f16_exp_mem_v1_default_instance.cpp | 67 ---- ...f16_f16_exp_mem_v1_mnkpadding_instance.cpp | 69 ----- ...16_f16_f16_exp_mem_v2_default_instance.cpp | 67 ---- ...f16_f16_exp_mem_v2_mnkpadding_instance.cpp | 69 ----- ...ht_wmma_f16_f16_f16_exp_odd_m_instance.cpp | 71 ----- ...ht_wmma_f16_f16_f16_exp_odd_n_instance.cpp | 69 ----- 57 files changed, 50 insertions(+), 2688 deletions(-) delete mode 100644 library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instance.cpp delete mode 100644 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 delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instance.cpp delete mode 100644 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 delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instance.cpp delete mode 100644 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 delete mode 100644 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 delete mode 100644 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 delete mode 100644 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 delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_m_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_n_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_m_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_n_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp index 6c88a9c970..48c9f10312 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp @@ -107,49 +107,6 @@ using device_gemm_wmma_universal_km_kn_mn_GemmMNKPadding_bf16_instances = std::t // clang-format on >; -// device_gemm_wmma_universal_km_kn_mn_comp_instances -// device_gemm_wmma_universal_km_kn_mn_mem_instances - -template -using device_gemm_wmma_universal_km_kn_mn_irregular_odd_m_instances = std::tuple< - // clang-format off - //#####################################| ALayout| BLayout| DsLayout |ELayout| ADataType| BDataType| DsDataType| CDataType| AccDataType| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| BlockwiseGemm| BlockwiseGemm| - //#####################################| | | | | | | | | | DataType| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock_NBlock_NPerBlock| ScalarPerVectors| Pipeline| Pipeline| - //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| - //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - // Latency friendly - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 8, 1, 8>, S<2,2,2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, S<2,2,2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 8, 1, 4>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 8, 1, 8>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 8, 1, 8>, S<4,4,4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> - // Memory friendly - // TODO: add once v2 is implemented - // clang-format on - >; - -template -using device_gemm_wmma_universal_km_kn_mn_odd_n_instances = std::tuple< - // clang-format off - //#####################################| ALayout| BLayout| DsLayout |ELayout| ADataType| BDataType| DsDataType| CDataType| AccDataType| CShuffle| A| B| CDE| GEMM| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransferClusterLengths| CShuffleBlockTransfer| BlockwiseGemm| BlockwiseGemm| - //#####################################| | | | | | | | | | DataType| Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeat| NRepeat| _MBlock_MPerBlock_NBlock_NPerBlock| ScalarPerVectors| Pipeline| Pipeline| - //#####################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| | | Scheduler| Verision| - //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - // Latency friendly - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 32, 16, 16, 64, 4, 4, 16, 16, 1, 1, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S< 8, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 4>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 4, 4, 16, 16, 1, 1, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S< 8, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Wmma_CShuffleV3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 8, 1, 8>, S<1,1,1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1> - // Memory friendly - // TODO: add once v2 is implemented - // clang-format on - >; - template using device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances = std::tuple< // clang-format off diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp index 395cba8133..afd1b3650b 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp @@ -59,10 +59,6 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_instance // clang-format on >; -// device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_instances - -// device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregular_instances - template ; -// device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_instances - -// device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_instances - template ; +template +using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_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, 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, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 16, 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 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, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 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, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, // Verification failure + // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> // Verification failture + // clang-format on + >; + } // namespace instance } // namespace device } // namespace tensor_operation diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp deleted file mode 100644 index 8c66baf475..0000000000 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_wmma_instance.hpp +++ /dev/null @@ -1,161 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#pragma once - -#include "ck/ck.hpp" -#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" -#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" -#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_wmma_cshuffle_v3.hpp" -#include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -using BF16 = ck::bhalf_t; -using F16 = ck::half_t; -using F32 = float; -using I8 = int8_t; -using I32 = int32_t; - -template -using S = ck::Sequence; - -using namespace ck::tensor_layout::convolution; - -using PassThrough = ck::tensor_operation::element_wise::PassThrough; - -static constexpr auto ConvBwdWeightDefault = - ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default; - -static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 = - ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0; - -template -using device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_generic_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, 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 - >; - -template -using device_grouped_conv_bwd_weight_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, 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 - // 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>, - - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, 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, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, 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, 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, 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, F16, F16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> - // clang-format on - >; - -// device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_generic_instances - -template -using device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_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, 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, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 16, 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 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, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 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, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, // Verification failure - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> // Verification failture - // clang-format on - >; - -template -using device_grouped_conv_bwd_weight_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, 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 - >; - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp index 573a42e743..5e88bc5c4c 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp @@ -894,10 +894,6 @@ struct DeviceOperationInstanceFactory>>& instances); -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_m_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_n_instances( - std::vector>>& instances); - void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instances( std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instances( - std::vector>>& instances); #endif #ifdef CK_ENABLE_FP16 @@ -145,54 +73,6 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_i PassThrough, PassThrough>>>& instances); -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instances( - std::vector>>& instances); - void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instances( std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_m_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_n_instances( - std::vector>>& instances); #endif // 3D @@ -245,42 +101,6 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_default_i PassThrough, PassThrough>>>& instances); -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_m_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_n_instances( - std::vector>>& instances); - void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instances( std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instances( - std::vector>>& instances); #endif #ifdef CK_ENABLE_FP16 @@ -368,54 +152,6 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_i PassThrough, PassThrough>>>& instances); -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instances( - std::vector>>& instances); - void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instances( std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_m_instances( - std::vector>>& instances); - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_n_instances( - std::vector>>& instances); #endif } // namespace instance diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_wmma.inc b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_wmma.inc index 822a46b16f..d7f70e14c5 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_wmma.inc +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_wmma.inc @@ -51,30 +51,6 @@ void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_instances( PassThrough, PassThrough>>>& instances); -void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instances( std::vector>>& instances); -void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_pad0_pipev1_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_instances( std::vector>>& instances); -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instances( std::vector>>& instances); -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instances( std::vector>>& instances); -void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_pad0_pipev1_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_instances( std::vector>>& instances); -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances( std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instances( - std::vector>>& instances); #endif // conv3d backward weight @@ -368,30 +224,6 @@ 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>>& instances); - -void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev1_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instances( std::vector>>& instances); -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instances( - std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instances( std::vector>>& instances); -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances( std::vector>>& instances); -void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_default_pipev1_instances( - std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pad0_pipev1_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instances( std::vector>>& instances); -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instances( - std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( std::vector>>& instances); -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances( std::vector>>& instances) -{ - // add_device_operation_instances(instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - // 2, - // GNHWC, - // GKYXC, - // GNHWK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp deleted file mode 100644 index 0136625318..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/gnhwc_gkyxc_gnhwk/device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instance.cpp +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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[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_pad0_pipev1_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_device_operation_instances(instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - // 2, - // GNHWC, - // GKYXC, - // GNHWK, - // ConvBwdWeightFilter1x1Stride1Pad0, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp deleted file mode 100644 index c5ecdc087e..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_part2_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_instances< - // 2, - // NGCHW, - // GKCYX, - // NGKHW, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instance.cpp deleted file mode 100644 index b55f3c7c74..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_part2_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances< - // 2, - // NGCHW, - // GKCYX, - // NGKHW, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp index f852a707b7..81f71d459d 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp @@ -30,15 +30,6 @@ void add_device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instances( GKCYX, NGKHW, ConvBwdWeightDefault>{}); - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<2, - // NGCHW, - // GKCYX, - // NGKHW, - // ConvBwdWeightDefault, - // 4, - // 4>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp index dc80707fbc..d173b0afac 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp @@ -30,15 +30,6 @@ void add_device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instances( GKCYX, NGKHW, ConvBwdWeightDefault>{}); - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<2, - // NGCHW, - // GKCYX, - // NGKHW, - // ConvBwdWeightDefault, - // 4, - // 4>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instance.cpp deleted file mode 100644 index 2578205be1..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_irregular_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp deleted file mode 100644 index 9102cd49bf..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_part2_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instance.cpp deleted file mode 100644 index f919b9ebd3..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_irregular_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregular_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instance.cpp deleted file mode 100644 index 416f157391..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_part2_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instance.cpp deleted file mode 100644 index ad0de5e9f4..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_default_pipev1_instance.cpp +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_default_pipev1_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_device_operation_instances(instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp index db422cd571..2cc474005e 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,7 +25,7 @@ void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_i // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< 2, NHWGC, GKYXC, @@ -34,7 +34,7 @@ void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_i // 2. Filter1x1Stride1Pad0 add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< 2, NHWGC, GKYXC, diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp index 93c1aef2ca..adc9de3a3d 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instance.cpp @@ -30,14 +30,6 @@ void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instances( GKYXC, NHWGK, ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - // add_device_operation_instances(instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/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 b/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 deleted file mode 100644 index 7d04517257..0000000000 --- a/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 +++ /dev/null @@ -1,40 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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( - [[maybe_unused]]std::vector>>& 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 diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instance.cpp deleted file mode 100644 index e2fe192584..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_default_pipev1_instance.cpp +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_default_pipev1_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_device_operation_instances(instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp index fc9a710374..f304d1bba4 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instance.cpp @@ -30,14 +30,6 @@ void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instances( GKYXC, NHWGK, ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - // add_device_operation_instances(instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/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 b/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 deleted file mode 100644 index 724fa6ef2a..0000000000 --- a/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 +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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( - [[maybe_unused]]std::vector>>& 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 diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt index 2f5b86c9ee..6d313ab2f5 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt @@ -73,25 +73,15 @@ list(APPEND GROUPED_CONV3D_BWD_WEIGHT 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_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.cpp 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.cpp wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instance.cpp - wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instance.cpp - wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instance.cpp wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instance.cpp - wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instance.cpp - wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instance.cpp 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 wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_instance.cpp - wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instance.cpp wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_instance.cpp - wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instance.cpp wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp diff --git a/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 b/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 index 673d2b253d..98e8547fb4 100644 --- a/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 +++ b/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 @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -24,7 +24,7 @@ void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_bf16_f32_bf1 // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< 3, GNDHWC, GKZYXC, @@ -34,7 +34,7 @@ void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_bf16_f32_bf1 // 2. Filter1x1Stride1Pad0 add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< 3, GNDHWC, GKZYXC, diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instance.cpp deleted file mode 100644 index 1803e85006..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_irregular_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_irregular_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instance.cpp deleted file mode 100644 index f4975aa1ba..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_part2_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_part2_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instance.cpp deleted file mode 100644 index 97cdcfc196..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_irregular_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_irregular_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instance.cpp deleted file mode 100644 index 0ddcc2911d..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_part2_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_f16_part2_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/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 b/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 deleted file mode 100644 index e2ec8928f4..0000000000 --- a/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 +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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( - [[maybe_unused]]std::vector>>& 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 diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp index 98c0b71bef..5bf7fa3a9a 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp @@ -2,7 +2,7 @@ // SPDX-License-Identifier: MIT #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" +#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 { @@ -25,7 +25,7 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf1 // 1. Default add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< 3, NDHWGC, GKZYXC, @@ -34,7 +34,7 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf1 // 2. Filter1x1Stride1Pad0 add_device_operation_instances( instances, - device_grouped_conv_bwd_weight_wmma_c_shuffle_bf16_f32_bf16_instances< + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< 3, NDHWGC, GKZYXC, diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp index d5fa41ae4a..728f514f9a 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp @@ -30,14 +30,6 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instanc GKZYXC, NDHWGK, ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - // add_device_operation_instances(instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/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 b/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 deleted file mode 100644 index 202849ee37..0000000000 --- a/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 +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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( - [[maybe_unused]]std::vector>>& 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 diff --git a/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 b/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 deleted file mode 100644 index 413a254644..0000000000 --- a/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 +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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( - [[maybe_unused]]std::vector>>& 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 diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp index 1f0286e5ff..f929196ddb 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp @@ -30,14 +30,6 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instance GKZYXC, NDHWGK, ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - // add_device_operation_instances(instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // ConvBwdWeightFilter1x1Stride1Pad0>{}); } } // namespace instance diff --git a/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 b/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 deleted file mode 100644 index 96671737ef..0000000000 --- a/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 +++ /dev/null @@ -1,39 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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( - [[maybe_unused]]std::vector>>& 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 diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instance.cpp deleted file mode 100644 index d7a22c93ae..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_part2_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_instances< - // 3, - // NGCDHW, - // GKCZYX, - // NGKDHW, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instance.cpp deleted file mode 100644 index d996737e9a..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_part2_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // 1. Default - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances< - // 3, - // NGCDHW, - // GKCZYX, - // NGKDHW, - // ConvBwdWeightDefault, - // BlockGemmPipelineScheduler::Intrawave, - // BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp index 4b58393704..060578c09a 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp @@ -23,20 +23,13 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instanc PassThrough>>>& instances) { // 1. Default - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances < - 3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault>{}); - // add_device_operation_instances(instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances < - // 3, - // NGCDHW, - // GKCZYX, - // NGKDHW, - // ConvBwdWeightDefault); + add_device_operation_instances( + instances, + device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<3, + NGCDHW, + GKCZYX, + NGKDHW, + ConvBwdWeightDefault>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp index ee5ceba586..a993497fab 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp @@ -30,15 +30,6 @@ void add_device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance GKCZYX, NGKDHW, ConvBwdWeightDefault>{}); - // add_device_operation_instances( - // instances, - // device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<3, - // NGCDHW, - // GKCZYX, - // NGKDHW, - // ConvBwdWeightDefault, - // 4, - // 4>{}); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/CMakeLists.txt index 7ac458e93c..08f95601f7 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/CMakeLists.txt @@ -28,23 +28,10 @@ set(GROUPED_CONVND_EXP_BWD_WEIGHT explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_default_instance.cpp explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_comp_mnkpadding_instance.cpp - explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instance.cpp - explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_m_instance.cpp - explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_n_instance.cpp explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instance.cpp - explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instance.cpp - explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instance.cpp - explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instance.cpp explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_default_instance.cpp explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_comp_mnkpadding_instance.cpp - explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instance.cpp - explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instance.cpp - explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instance.cpp - explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instance.cpp - explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp - explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_n_instance.cpp - explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_m_instance.cpp ) add_instance_library(device_grouped_convnd_bwd_weight_instance ${GROUPED_CONVND_EXP_BWD_WEIGHT}) diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instance.cpp deleted file mode 100644 index f238ce5b38..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instance.cpp +++ /dev/null @@ -1,67 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // BF16, - // BF16, - // BF16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_default_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // BF16, - // BF16, - // BF16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instance.cpp deleted file mode 100644 index 1a55e2658f..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instance.cpp +++ /dev/null @@ -1,69 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // BF16, - // BF16, - // BF16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>( - // instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v1_mnkpadding_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // BF16, - // BF16, - // BF16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>( - // instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instance.cpp deleted file mode 100644 index 4de541abff..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instance.cpp +++ /dev/null @@ -1,67 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // BF16, - // BF16, - // BF16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_default_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // BF16, - // BF16, - // BF16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instance.cpp deleted file mode 100644 index f9470a632a..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instance.cpp +++ /dev/null @@ -1,69 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // BF16, - // BF16, - // BF16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>( - // instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_mem_v2_mnkpadding_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // BF16, - // BF16, - // BF16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>( - // instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_m_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_m_instance.cpp deleted file mode 100644 index 7bfaff0a51..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_m_instance.cpp +++ /dev/null @@ -1,71 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_m_instances( - std::vector>>& instances) -{ - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_irregular_odd_m_instances>(instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_m_instances( - std::vector>>& instances) -{ - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_irregular_odd_m_instances>(instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_n_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_n_instance.cpp deleted file mode 100644 index f34d511866..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/bf16_bf16_bf16/device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_n_instance.cpp +++ /dev/null @@ -1,69 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_n_instances( - std::vector>>& instances) -{ - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_odd_n_instances>( - instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_n_instances( - std::vector>>& instances) -{ - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - BF16, - BF16, - BF16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_odd_n_instances>( - instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instance.cpp deleted file mode 100644 index 21e27e24ff..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instance.cpp +++ /dev/null @@ -1,67 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // F16, - // F16, - // F16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_default_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // F16, - // F16, - // F16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instance.cpp deleted file mode 100644 index 27e5ed1d36..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instance.cpp +++ /dev/null @@ -1,69 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // F16, - // F16, - // F16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>( - // instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v1_mnkpadding_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // F16, - // F16, - // F16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>( - // instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instance.cpp deleted file mode 100644 index d4f9148c24..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instance.cpp +++ /dev/null @@ -1,67 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // F16, - // F16, - // F16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_default_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // F16, - // F16, - // F16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>(instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instance.cpp deleted file mode 100644 index 361aa80b0f..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instance.cpp +++ /dev/null @@ -1,69 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 2, - // NHWGC, - // GKYXC, - // NHWGK, - // F16, - // F16, - // F16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>( - // instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_mem_v2_mnkpadding_instances( - [[maybe_unused]]std::vector>>& instances) -{ - // add_explicit_gemm_device_operation_instances< - // 3, - // NDHWGC, - // GKZYXC, - // NDHWGK, - // F16, - // F16, - // F16, - // PassThrough, - // PassThrough, - // PassThrough, - // device_gemm_wmma_universal_km_kn_mn_mem_instances>( - // instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_m_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_m_instance.cpp deleted file mode 100644 index b042058853..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_m_instance.cpp +++ /dev/null @@ -1,71 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_m_instances( - std::vector>>& instances) -{ - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_irregular_odd_m_instances>(instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_m_instances( - std::vector>>& instances) -{ - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_irregular_odd_m_instances>(instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_n_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_n_instance.cpp deleted file mode 100644 index b2f8d4c3e3..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_n_instance.cpp +++ /dev/null @@ -1,69 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_wmma_universal_km_kn_mn_instance.hpp" -#include "ck/host_utility/device_prop.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_n_instances( - std::vector>>& instances) -{ - add_explicit_gemm_device_operation_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_odd_n_instances>( - instances); -} - -void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_n_instances( - std::vector>>& instances) -{ - add_explicit_gemm_device_operation_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - F16, - F16, - F16, - PassThrough, - PassThrough, - PassThrough, - device_gemm_wmma_universal_km_kn_mn_odd_n_instances>( - instances); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck From 9021fd68b67251d7188787edeae09f0e17fb6e56 Mon Sep 17 00:00:00 2001 From: kiefer Date: Tue, 9 Dec 2025 13:13:49 +0000 Subject: [PATCH 14/18] Re-enable all xdl instances (un-16x16-adapted) and dl instances. Remove custom ckProfiler target. --- ...p_gemm_xdl_universal_km_kn_mn_instance.hpp | 160 ++++++------ ...ce_grouped_conv_bwd_weight_dl_instance.hpp | 6 +- ...conv_bwd_weight_two_stage_xdl_instance.hpp | 192 +++++++------- ...rouped_conv_bwd_weight_v3_xdl_instance.hpp | 32 +-- ..._conv_bwd_weight_xdl_bilinear_instance.hpp | 180 ++++++------- ...e_grouped_conv_bwd_weight_xdl_instance.hpp | 240 +++++++++--------- ...ped_conv_bwd_weight_xdl_scale_instance.hpp | 178 ++++++------- profiler/src/CMakeLists.txt | 19 -- 8 files changed, 494 insertions(+), 513 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp index 04bd56d98e..f9d77ca451 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_exp_gemm_xdl_universal_km_kn_mn_instance.hpp @@ -45,18 +45,18 @@ using device_gemm_xdl_universal_km_kn_mn_comp_instances = //#########################| | | | Type| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Pipeline| Pipeline| //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 2, 2, 16, 16, 4, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 16, 16, 8, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 16, 16, 8, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - // // Can we support this kind of odd case? 224(256) = 28*8 + (4*8) - // //DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 8, 16, 16, 7, 8, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 32, 32, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 32, 32, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 32, 32, 4, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 4, 4, 32, 32, 4, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 256, 32, 2, 2, 32, 32, 4, 4, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + // Can we support this kind of odd case? 224(256) = 28*8 + (4*8) + //DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 224, 256, 64, 8, 8, 16, 16, 7, 8, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, 1, 2, S<1, 32, 1, 8>, S<8>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1> // clang-format on >; @@ -70,23 +70,23 @@ using device_gemm_xdl_universal_km_kn_mn_mem_instances = std::tuple< //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // // Memory friendly - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // Memory friendly + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; @@ -100,23 +100,23 @@ using device_gemm_xdl_universal_km_kn_mn_irregular_odd_m_instances = std::tuple< //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // // Memory friendly - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // Memory friendly + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 32, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 16, 1, 8>, S<2>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 4>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 8>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, InOutDataType, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, 1, 1, S<1, 16, 1, 16>, S<4>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; @@ -130,23 +130,23 @@ using device_gemm_xdl_universal_km_kn_mn_odd_n_instances = std::tuple< //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // // Memory friendly - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // Memory friendly + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; @@ -160,23 +160,23 @@ using device_gemm_xdl_universal_km_kn_mn_irregular_odd_mn_instances = std::tuple //#########################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| Scheduler| Verision| //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // Latency friendly - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, - // // Memory friendly - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, - // DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 2, 2, 16, 16, 1, 1, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v1>, + // Memory friendly + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 8, 2, 16, 16, 4, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 256, 16, 64, 2, 2, 16, 16, 4, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 32, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 128, 16, 64, 8, 4, 16, 16, 4, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 64, 16, 64, 4, 4, 16, 16, 2, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 32, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 64, 16, 16, 64, 4, 4, 16, 16, 1, 1, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 4>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 32, 64, 4, 4, 16, 16, 1, 1, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 64, 64, 4, 4, 16, 16, 1, 2, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 128, 16, 128, 64, 4, 4, 16, 16, 1, 4, S<16, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 8>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 4, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2>, + DeviceBatchedGemmMultiD_Xdl_CShuffle_V3< Col, Row, Tuple<>, Row, InOutDataType, InOutDataType, Tuple<>, F32, F32, InOutDataType, PassThrough, PassThrough, PassThrough, GemmSpec, 256, 16, 256, 64, 2, 2, 16, 16, 1, 4, S<32, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, 0, 1, 1, S<1, 16, 1, 16>, S<1>, BlkGemmPipeSched, BlockGemmPipelineVersion::v2> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp index b1c28a18a0..0493a55e2f 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_dl_instance.hpp @@ -43,7 +43,7 @@ using device_grouped_conv_bwd_weight_dl_f32_instances = //############################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | _K0_M0_M1_K1| _K0_M0_M1_K1| ArrangeOrder| | _K0_M0_M1_K1| ContiguousDimOrder| _K0_M0_M1_K1| _K0_N0_N1_K1| _K0_N0_N1_K1| ArrangeOrder| | _K0_N0_N1_K1| ContiguousDimOrder| _K0_N0_N1_K1| | | | //############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - // DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> + DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >; @@ -60,7 +60,7 @@ using device_grouped_conv_bwd_weight_dl_f16_instances = //############################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | _K0_M0_M1_K1| _K0_M0_M1_K1| ArrangeOrder| | _K0_M0_M1_K1| ContiguousDimOrder| _K0_M0_M1_K1| _K0_N0_N1_K1| _K0_N0_N1_K1| ArrangeOrder| | _K0_N0_N1_K1| ContiguousDimOrder| _K0_N0_N1_K1| | | | //############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - // DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> + DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >; @@ -77,7 +77,7 @@ using device_grouped_conv_bwd_weight_dl_bf16_instances = //############################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | _K0_M0_M1_K1| _K0_M0_M1_K1| ArrangeOrder| | _K0_M0_M1_K1| ContiguousDimOrder| _K0_M0_M1_K1| _K0_N0_N1_K1| _K0_N0_N1_K1| ArrangeOrder| | _K0_N0_N1_K1| ContiguousDimOrder| _K0_N0_N1_K1| | | | //############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | // generic instance - // DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> + DeviceGroupedConvBwdWeight_Dl< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<1, 8, 1, 1, 1>, S<1, 2, 1, 128, 1>, S<0, 2, 3, 1, 4>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<0, 2, 3, 1, 4>, S<1, 1, 1, 1, 1>, S<1, 1, 1, 8, 1>, S<1, 16, 1, 16, 1>, S<0, 1, 4, 2, 3>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 4, 2, 3>, S<1, 1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp index 570ffaae50..3633015a78 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_xdl_instance.hpp @@ -46,7 +46,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_generic_i //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -63,16 +63,16 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_instances //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 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>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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>, 1, Scheduler, PipelineVersion, 8> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -89,14 +89,14 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_part2_ins //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> // clang-format on >; @@ -114,11 +114,11 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_f16_irregular //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 4, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 4, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 4, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 4, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 4, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 4, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -136,7 +136,7 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_generic_ //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -153,16 +153,16 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_instance //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 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>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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>, 1, Scheduler, PipelineVersion, 8> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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>, 1, Scheduler, PipelineVersion, 8> // clang-format on >; @@ -180,14 +180,14 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_part2_in //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 4, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4> // clang-format on >; @@ -205,11 +205,11 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_xdl_c_shuffle_bf16_irregula //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 4, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 4, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 4, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 48, 64, 32, 8, 16, 16, 3, 4, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 3, 4, false, S<4, 16, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 4, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 48, 32, 8, 16, 16, 4, 3, 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, 3, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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>, 1, Scheduler, PipelineVersion, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 208, 32, 8, 16, 16, 4, 13, 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, 13, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1> // clang-format on >; @@ -227,7 +227,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_generic_i //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1> // clang-format on >; @@ -245,29 +245,29 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_instances //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 1, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8 ,1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, F16, F16, 2, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8 ,1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, F16, F16, 8, 1> // clang-format on >; @@ -284,17 +284,17 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_f16_part2_ins //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, F16, F16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, F16, F16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, F16, F16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 2, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 2, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, F16, F16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 2, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 2, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, F16, F16, 1, 4> // clang-format on >; @@ -312,7 +312,7 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_generic_ //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1> // clang-format on >; @@ -329,29 +329,29 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_instance //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 16, 16, 2, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 1, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 2, 2, false, 1, 1, S<1, 8, 1, 8>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 1, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 8>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8 ,1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 2, BF16, BF16, 2, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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, 4, 1, 8>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8 ,1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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, 4>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 8, 1> // clang-format on >; @@ -370,17 +370,17 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_xdl_c_shuffle_bf16_part2_in //#########################################| Dim| | | | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Weight| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| ClusterLengths| ScalarPerVector| Pipeline| Pipeline| ToMerge| //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| | //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, BF16, BF16, 2, 2>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 16, 16, 2, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, S<8, 8, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 8, 8, false, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1, BF16, BF16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 2, 8, false, 1, 1, S<1, 8, 1, 32>, 2, Scheduler, PipelineVersion, 1, BF16, BF16, 2, 2>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 64, 8, 32, 32, 1, 1, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, S<8, 32, 1>, S<2, 0, 1>, S<2, 0, 1>, 1, 1, 8, false, 1, 1, S<1, 4, 1, 64>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 2, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 2, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 256, 32, 8, 16, 16, 1, 16, S<4, 2, 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, 4, 1, 16>, 1, Scheduler, PipelineVersion, 8, BF16, BF16, 1, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 2, 4>, - // DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4> + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 4, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 2, 4>, + DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 128, 32, 8, 16, 16, 1, 8, S<4, 4, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, S<4, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, false, 1, 1, S<1, 4, 1, 16>, 1, Scheduler, PipelineVersion, 4, BF16, BF16, 1, 4> // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp index 4bdf6e7dc0..143d857333 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp @@ -55,7 +55,7 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_instances = std::tuple //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - // DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion> + DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion> // clang-format on >; @@ -73,7 +73,7 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_tf32_instances = std:: //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - // DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion, TF32, TF32> + DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 8, 1, 8>, 2, Scheduler, PipelineVersion, TF32, TF32> // clang-format on >; @@ -91,13 +91,13 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f16_instances = std::tuple //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - // DeviceGroupedConvBwdWeight_Xdl_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_Xdl_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_Xdl_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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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> + DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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 >; @@ -115,13 +115,13 @@ using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_bf16_instances = std::tupl //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| Scheduler| Version| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | // generic instance - // DeviceGroupedConvBwdWeight_Xdl_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_Xdl_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_Xdl_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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 16, 16, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 16, 16, 8, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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> + DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 32, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 128, 32, 8, 32, 32, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 32, 8, 32, 32, 2, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 128, 32, 32, 8, 32, 32, 4, 1, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 80, 32, 8, 16, 16, 4, 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_Xdl_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 112, 32, 8, 16, 16, 4, 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 >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp index 063234d426..75655dda8d 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_bilinear_instance.hpp @@ -54,24 +54,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_bilinear_instances = std: //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - // // instances for small conv.K and conv.C - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + // instances for small conv.K and conv.C + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4> // clang-format on >; @@ -87,24 +87,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_bilinear_instances = //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>, - // // instances for small conv.K and conv.C - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>, + // instances for small conv.K and conv.C + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F32, F32, F32, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32> // clang-format on >; @@ -120,27 +120,27 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_bilinear_instances = std: //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2>, - // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2>, + // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8> // clang-format on >; @@ -156,24 +156,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_bilinear_instances = std //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - // // instance for small conv.K - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + // instance for small conv.K + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, BF16, F32, BF16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4> // clang-format on >; @@ -190,26 +190,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_comp_bf8_f8_bilinear_inst //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | #if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8 // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8>, - // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8>, + // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Tuple, F16, F16, F16, F32, Tuple, PassThrough, Bilinear, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> #endif // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp index 24847b9d69..22f1912198 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_instance.hpp @@ -53,7 +53,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_generic_instances = std:: //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance -// DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 16, 16, 4, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> // clang-format on >; @@ -71,29 +71,29 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_instances = std::tuple< //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 16, 16, 4, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // // instances for small conv.K and conv.C - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 16, 16, 4, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 16, 16, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // instances for small conv.K and conv.C + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 16, 16, 8, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 16, 16, 4, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 16, 16, 8, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 16, 16, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 16, 16, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 16, 16, 4, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 16, 16, 4, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 16, 16, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 16, 16, 4, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 16, 16, 4, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 16, 16, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 16, 16, 2, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 16, 16, 2, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 16, 16, 2, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 16, 16, 2, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, F32, F32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -109,7 +109,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_generic_instances = //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance -// DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32> // clang-format on >; @@ -127,29 +127,29 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_instances = std::tup //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // // instances for small conv.K and conv.C - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // instances for small conv.K and conv.C + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 16>, 4, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 64, 8, 8, 32, 32, 1, 1, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, S<1, 8, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, 2, 1, 4, true, 1, 1, S<1, 4, 1, 64>, 1, TF32, TF32, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -165,7 +165,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_generic_instances = std:: //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance -// DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2> // clang-format on >; @@ -183,26 +183,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_instances = std::tuple< //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, F16, F16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -218,7 +218,7 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_generic_instanc //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance -// DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1> // clang-format on >; @@ -236,24 +236,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_f32_bf16_instances = std //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // // instance for small conv.K - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // instance for small conv.K + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -271,26 +271,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_instances = std::tuple< //#########################################| Spatial| | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| | | ScalarPerVector| ScalarPerVector| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | // generic instance - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF16, BF16, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> // clang-format on >; @@ -309,26 +309,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_comp_bf8_f8_instances = s //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | | #if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8 // generic instance - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 16, 16, 8, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 16, 16, 4, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 16, 16, 4, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 16, 16, 2, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 16, 16, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, // Didn't double MRepeat - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 16, 16, 4, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, - // DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 16, 16, 2, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector>, + DeviceGroupedConvBwdWeight_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8, MaxTransposeTransferSrcScalarPerVector, MaxTransposeTransferDstScalarPerVector> #endif // clang-format on >; diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp index ab50f44f33..ed1cb3997f 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_xdl_scale_instance.hpp @@ -54,24 +54,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_scale_instances = std::tu //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - // // instances for small conv.K and conv.C - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + // instances for small conv.K and conv.C + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4> // clang-format on >; @@ -87,24 +87,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f32_tf32_scale_instances = st //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>, - // // instances for small conv.K and conv.C - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1, TF32, TF32>, + // instances for small conv.K and conv.C + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 4, 32, 32, 2, 4, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 64, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 8>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 4, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 4, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 1, true, S<1, 4, 32, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 32, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 4, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F32, F32, F32, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 4, 32, 32, 1, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 2, true, S<1, 4, 16, 1>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 4, 4, true, 1, 1, S<1, 16, 1, 4>, 4, TF32, TF32> // clang-format on >; @@ -120,26 +120,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_scale_instances = std::tu //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2>, - // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2>, + // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8> // clang-format on >; @@ -155,24 +155,24 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_bf16_scale_instances = std::t //#########################################| Spatial| | | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| MBlock_MPerBlock| NWaveNPerXdl| //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, - // // instance for small conv.K - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 4, true, 1, 1, S<1, 16, 1, 4>, 1>, + // instance for small conv.K + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 1, true, 1, 1, S<1, 32, 1, 4>, 1>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 1, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, BF16, F32, BF16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 4> // clang-format on >; @@ -189,26 +189,26 @@ using device_grouped_conv_bwd_weight_xdl_c_shuffle_f16_comp_bf8_f8_scale_instanc //#########################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | #if defined CK_ENABLE_FP8 && defined CK_ENABLE_BF8 // generic instance - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8>, - // // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 4, true, 1, 1, S<1, 16, 1, 4>, 2, BF8, F8>, + // 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_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 1, true, 1, 1, S<1, 32, 1, 4>, 2, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 2, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, - // DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 32, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 8>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 8, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 8>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 1, true, S<1, 4, 16, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 32, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8>, + DeviceGroupedConvBwdWeightMultipleD_Xdl_CShuffle< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 4, 4>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 2, true, S<1, 4, 8, 2>, S<0, 3, 1, 2>, S<0, 2, 1, 3>, 2, 8, 4, true, 1, 1, S<1, 16, 1, 4>, 8, BF8, F8> #endif // clang-format on >; diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 8b1090d0ff..5833fc3626 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -283,22 +283,3 @@ message(VERBOSE "ckProfiler libs: ${PROFILER_LIBS}") target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE ${PROFILER_LIBS}) rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler) - -## Defining specific operation targets - -macro(define_profiler_target NAME SOURCES LIBS) - add_executable(${NAME} profiler.cpp ${SOURCES}) - target_compile_options(${NAME} PRIVATE -Wno-global-constructors) - - if(NOT WIN32 AND ${hip_VERSION_FLAT} GREATER 600241132) - target_compile_options(${NAME} PRIVATE --offload-compress) - endif() - - target_link_libraries(${NAME} PRIVATE utility getopt::getopt ${LIBS}) - - rocm_install(TARGETS ${NAME} COMPONENT profiler) -endmacro() - -define_profiler_target(ckProfiler_grouped_conv_bwd_weight_vanilla - "profile_grouped_conv_bwd_weight.cpp" - "device_grouped_convnd_bwd_weight_instance;device_grouped_conv1d_bwd_weight_instance;device_grouped_conv2d_bwd_weight_instance;device_grouped_conv3d_bwd_weight_instance") \ No newline at end of file From aba5959e11e4934d1cb9748ce963736d6decaca1 Mon Sep 17 00:00:00 2001 From: kiefer Date: Tue, 9 Dec 2025 13:24:17 +0000 Subject: [PATCH 15/18] Remove straggler comments --- ...device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp | 4 ---- 1 file changed, 4 deletions(-) diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp index afd1b3650b..ba0685015f 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp @@ -132,8 +132,6 @@ using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_instance // clang-format on >; -// device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_part2_instances - template , S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // clang-format on >; - -// device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_part2_instances } // namespace instance } // namespace device } // namespace tensor_operation From 2154ee899f18b6a9641e15d2680f7dc3ff2b4f52 Mon Sep 17 00:00:00 2001 From: kiefer Date: Tue, 9 Dec 2025 13:38:55 +0000 Subject: [PATCH 16/18] Remove [[maybe_unused]] --- ...ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp | 21 +++++++++---------- ..._ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp | 21 +++++++++---------- ...dhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp | 21 +++++++++---------- ...cdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp | 21 +++++++++---------- 4 files changed, 40 insertions(+), 44 deletions(-) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp index bdafbeba8f..e222e8c663 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp @@ -11,17 +11,16 @@ 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_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instances( - [[maybe_unused]] std::vector>>& - instances) + std::vector>>& instances) { // 1. Default add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp index a663d608d0..85d395c7c2 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp @@ -11,17 +11,16 @@ 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_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instances( - [[maybe_unused]] std::vector>>& - instances) + std::vector>>& instances) { // 1. Default add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp index 234970d8ac..39d8ef3091 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp @@ -11,17 +11,16 @@ 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_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances( - [[maybe_unused]] std::vector>>& - instances) + std::vector>>& instances) { // 1. Default add_device_operation_instances( diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp index a9b148b688..76b6fc9b0d 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp @@ -11,17 +11,16 @@ 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_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances( - [[maybe_unused]] std::vector>>& - instances) + std::vector>>& instances) { // 1. Default add_device_operation_instances( From 5d9c3705d507909705f2ae6d9462561d3cc67c95 Mon Sep 17 00:00:00 2001 From: kiefer Date: Wed, 10 Dec 2025 11:08:14 +0000 Subject: [PATCH 17/18] Fix clang format --- .../device_grouped_conv_bwd_weight_wmma_scale_instance.hpp | 2 +- ...d_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instance.cpp | 6 ++---- ...nvnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp | 6 ++---- 3 files changed, 5 insertions(+), 9 deletions(-) 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 218995a984..e893c92d1d 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 @@ -59,7 +59,7 @@ using device_grouped_conv_bwd_weight_wmma_c_shuffle_f16_scale_instances = std::t DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1>, // Presumably doesn't produce correct results for fp16 // DeviceGroupedConvBwdWeightMultipleD_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, Empty_Tuple, F16, F16, F16, F32, Empty_Tuple, PassThrough, Scale, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 8, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1> // Presumably doesn't produce correct results for fp16 - // clang-format on + // clang-format on >; template >(instances); + device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instances( @@ -59,8 +58,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_bf16_bf16_bf16_exp_odd_mn_instanc PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); } } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp index 1129d1c9fe..ace411ea68 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/grouped_convnd_bwd_weight/explicit_wmma/fp16_fp16_fp16/device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instance.cpp @@ -32,8 +32,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instances( PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); } void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instances( @@ -59,8 +58,7 @@ void add_device_grouped_convnd_bwd_weight_wmma_f16_f16_f16_exp_odd_mn_instances( PassThrough, PassThrough, PassThrough, - device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); + device_gemm_wmma_universal_km_kn_mn_irregular_odd_mn_instances>(instances); } } // namespace instance From 94f6c66f8c23af696f949b4fe79279070f6f0a61 Mon Sep 17 00:00:00 2001 From: kiefer Date: Wed, 10 Dec 2025 16:22:42 +0000 Subject: [PATCH 18/18] Remove unwanted instances. This includes all instances which are not NHWGCxGKYXC and F16 or BF16 (no mixed in-out types). --- ...onv_bwd_weight_two_stage_wmma_instance.hpp | 91 ------- ...ouped_conv_bwd_weight_v3_wmma_instance.hpp | 28 -- .../grouped_convolution_backward_weight.hpp | 192 -------------- ...ouped_convolution_backward_weight_wmma.inc | 246 +----------------- .../grouped_conv1d_bwd_weight/CMakeLists.txt | 4 - ..._gnwc_gkxc_gnwk_bf16_f32_bf16_instance.cpp | 47 ---- ...eight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp | 45 ---- .../grouped_conv2d_bwd_weight/CMakeLists.txt | 12 - ...hwc_gkyxc_gnhwk_bf16_f32_bf16_instance.cpp | 49 ---- ...ht_wmma_gnhwc_gkyxc_gnhwk_f16_instance.cpp | 46 ---- ...ngchw_gkcyx_ngkhw_bf16_pipev1_instance.cpp | 41 --- ..._ngchw_gkcyx_ngkhw_f16_pipev1_instance.cpp | 41 --- ...t_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp | 38 --- ...ht_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp | 38 --- ...ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp | 41 --- ..._ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp | 41 --- ...wgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp | 48 ---- .../grouped_conv3d_bwd_weight/CMakeLists.txt | 12 - ...c_gkzyxc_gndhwk_bf16_f32_bf16_instance.cpp | 48 ---- ...wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp | 45 ---- ...c_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp | 48 ---- ...dhw_gkczyx_ngkdhw_bf16_pipev1_instance.cpp | 41 --- ...cdhw_gkczyx_ngkdhw_f16_pipev1_instance.cpp | 41 --- ...mma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp | 38 --- ...wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp | 38 --- ...dhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp | 41 --- ...cdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp | 41 --- 27 files changed, 1 insertion(+), 1440 deletions(-) delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_bf16_f32_bf16_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp delete mode 100644 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 delete mode 100644 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 delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp delete mode 100644 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 delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp index ba0685015f..b77c7348db 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_two_stage_wmma_instance.hpp @@ -85,97 +85,6 @@ using device_grouped_conv_bwd_weight_two_stage_nhwgc_wmma_c_shuffle_bf16_instanc // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // clang-format on >; - -template -using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, F16, F16, 1, 1> - // clang-format on - >; - -// NGCHW requires transpose, we use vector loads and stores params for them -template -using device_grouped_conv_bwd_weight_two_stage_ngchw_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1> - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, // Incorrect results for at least GemmDefault - // DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> // Incorrect results for at least GemmDefault - // clang-format on - >; - -template -using device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, false, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1, BF16, BF16, 1, 1> - // clang-format on - >; - -template -using device_grouped_conv_bwd_weight_two_stage_ngchw_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| NumGroups| - //################################################| 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| ToMerge| - //################################################| 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| Sched| Ver| | - //################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | NBlock_NPerBlock| | | | | - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 32, 16, 16, 32, 8, 16, 16, 1, 1, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 4, 1, 8>, 1, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 128, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 128, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 8>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 256, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1>, - DeviceGroupedConvBwdWeightTwoStage_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, BF16, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, Scheduler, PipelineVersion, 1> - // clang-format on - >; } // namespace instance } // namespace device } // namespace tensor_operation diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp index d942ce82b2..761b07ea60 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_wmma_instance.hpp @@ -94,34 +94,6 @@ using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances = std::tup //clang-format on >; -template -using device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_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, 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, 0, S<4, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 1, 4, 0, 1, 1, S<1, 16, 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, 128, 128, 128, 32, 8, 16, 16, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 8, 2, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<4, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 64, 64, 8, 16, 16, 4, 2, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<8, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 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, 256, 64, 8, 16, 16, 8, 2, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 4, 8, 1, S<8, 32, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 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, 48, 64, 128, 8, 16, 16, 3, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 64, 8, 16, 16, 6, 2, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<8, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 64, 64, 128, 8, 16, 16, 4, 1, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, S<16, 8, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 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, 96, 128, 128, 8, 16, 16, 6, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<16, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 8, 8, 0, 1, 1, S<1, 16, 1, 16>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector>, // Verification failure - // DeviceGroupedConvBwdWeight_Wmma_CShuffleV3< NDimSpatial, ALayout, BLayout, ELayout, BF16, F32, BF16, F32, PassThrough, PassThrough, PassThrough, ConvSpec, 96, 96, 96, 48, 8, 16, 16, 6, 2, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 1, S<6, 16, 1>, S<2, 0, 1>, S<1, 0, 2>, 1, 6, 8, 0, 1, 1, S<1, 16, 1, 6>, 4, BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v1, BF16, BF16, TransposeTransferSrcScalarPerVector, TransposeTransferDstScalarPerVector> // Verification failture - // clang-format on - >; - } // namespace instance } // namespace device } // namespace tensor_operation diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp index 5e88bc5c4c..6dd8758eb7 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp @@ -857,56 +857,8 @@ struct DeviceOperationInstanceFactory && is_same_v && - is_same_v) - { -#ifdef CK_ENABLE_FP16 - if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && - is_same_v) - { - add_device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_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_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_bf16_f32_bf16_instances( - op_ptrs); - } -#endif - } - } if constexpr(NumDimSpatial == 2) { - if constexpr(is_same_v && is_same_v && - is_same_v) - { -#ifdef CK_ENABLE_FP16 - if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && - is_same_v) - { - add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_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_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances( - op_ptrs); - } -#endif - } if constexpr(is_same_v && is_same_v && is_same_v) { @@ -929,14 +881,6 @@ struct DeviceOperationInstanceFactory && is_same_v && - is_same_v && - is_same_v && - is_same_v) - { - add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances( - op_ptrs); - } if constexpr(is_same_v && is_same_v && is_same_v && @@ -945,7 +889,6 @@ struct DeviceOperationInstanceFactory && is_same_v && - is_same_v) - { -#ifdef CK_ENABLE_FP16 - if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && - is_same_v) - { - add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_instances( - op_ptrs); - add_device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_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_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instances( - op_ptrs); - add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_instances( - op_ptrs); - } -#endif - } - if constexpr(is_same_v && is_same_v && - is_same_v) - { -#ifdef CK_ENABLE_FP16 - if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && - is_same_v) - { - add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_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_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instances( - op_ptrs); - } #endif } } if constexpr(NumDimSpatial == 3) { - if constexpr(is_same_v && is_same_v && - is_same_v) - { -#ifdef CK_ENABLE_FP16 - if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && - is_same_v) - { - add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_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_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( - op_ptrs); - } -#endif - } if constexpr(is_same_v && is_same_v && is_same_v) { @@ -1058,14 +926,6 @@ struct DeviceOperationInstanceFactory && is_same_v && - is_same_v && - is_same_v && - is_same_v) - { - add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( - op_ptrs); - } if constexpr(is_same_v && is_same_v && is_same_v && @@ -1084,58 +944,6 @@ struct DeviceOperationInstanceFactory && is_same_v && - is_same_v) - { -#ifdef CK_ENABLE_FP16 - if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && - is_same_v) - { - add_device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instances( - op_ptrs); - add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_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_ngcdhw_gkczyx_ngkdhw_bf16_instances( - op_ptrs); - add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_instances( - op_ptrs); - } -#endif - } - if constexpr(is_same_v && is_same_v && - is_same_v) - { -#ifdef CK_ENABLE_FP16 - if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && - is_same_v) - { - add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_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_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances( - op_ptrs); - } #endif } } diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_wmma.inc b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_wmma.inc index d7f70e14c5..06247019f1 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_wmma.inc +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight_wmma.inc @@ -8,49 +8,8 @@ namespace tensor_operation { namespace device { namespace instance { -// conv1d backward weight -#ifdef CK_ENABLE_BF16 -void add_device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_bf16_f32_bf16_instances( - std::vector>>& instances); -#endif - -#ifdef CK_ENABLE_FP16 -void add_device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instances( - std::vector>>& instances); -#endif - // conv2d backward weight #ifdef CK_ENABLE_FP16 -void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_f16_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_f16_instances( std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instances( - std::vector>>& instances); #endif #ifdef CK_ENABLE_BF16 -void add_device_grouped_conv2d_bwd_weight_wmma_gnhwc_gkyxc_gnhwk_bf16_f32_bf16_instances( - std::vector>>& instances); - void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_instances( std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instances( - std::vector>>& instances); - -void add_device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_instances( - std::vector>>& instances); #endif // conv3d backward weight #ifdef CK_ENABLE_FP16 -void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instances( - std::vector>>& instances); - void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_f16_instances( std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instances( - std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_instances( - std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances( - std::vector>>& instances); #endif -#ifdef CK_ENABLE_BF16 -void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( - std::vector>>& instances); +#ifdef CK_ENABLE_BF16 void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_instances( std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instances( - std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instances( - std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_instances( - std::vector>>& instances); - -void add_device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances( - std::vector>>& instances); #endif } // namespace instance diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt index eaf1aa6b30..56a9d16623 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/CMakeLists.txt @@ -17,8 +17,4 @@ if(DL_KERNELS) dl/device_grouped_conv1d_bwd_weight_dl_nwgc_gkxc_nwgk_bf16_f32_bf16_instance.cpp) endif() -list(APPEND GROUPED_CONV1D_BWD_WEIGHT - wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp - wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_bf16_f32_bf16_instance.cpp) - add_instance_library(device_grouped_conv1d_bwd_weight_instance ${GROUPED_CONV1D_BWD_WEIGHT}) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_bf16_f32_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_bf16_f32_bf16_instance.cpp deleted file mode 100644 index d855f574f3..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_bf16_f32_bf16_instance.cpp +++ /dev/null @@ -1,47 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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 { - -void add_device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_bf16_f32_bf16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 1, - GNWC, - GKXC, - GNWK, - ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 1, - GNWC, - GKXC, - GNWK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp deleted file mode 100644 index b19f31da22..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv1d_bwd_weight/wmma/device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instance.cpp +++ /dev/null @@ -1,45 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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 { - -void add_device_grouped_conv1d_bwd_weight_wmma_gnwc_gkxc_gnwk_f16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<1, - GNWC, - GKXC, - GNWK, - ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - 1, - GNWC, - GKXC, - GNWK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt index c8fcf642eb..ec9e7da391 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/CMakeLists.txt @@ -73,20 +73,8 @@ if(DL_KERNELS) endif() 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_instance.cpp - - wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp - wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp - wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_instance.cpp - wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_instance.cpp - - wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp - wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_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_two_stage_wmma_nhwgc_gkyxc_nhwgk_bf16_pipev1_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_two_stage_wmma_nhwgc_gkyxc_nhwgk_f16_pipev1_instance.cpp ) diff --git a/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 b/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 deleted file mode 100644 index c99c44c0a6..0000000000 --- a/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 +++ /dev/null @@ -1,49 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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[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_bf16_f32_bf16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 2, - GNHWC, - GKYXC, - GNHWK, - ConvBwdWeightDefault>{}); - - // 2. Filter1x1Stride1Pad0 - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 2, - GNHWC, - GKYXC, - GNHWK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/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 b/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 deleted file mode 100644 index 445fc8dd94..0000000000 --- a/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 +++ /dev/null @@ -1,46 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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[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>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<2, - GNHWC, - GKYXC, - GNHWK, - ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - 2, - GNHWC, - GKYXC, - GNHWK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_instance.cpp deleted file mode 100644 index 1edaa86214..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngchw_gkcyx_ngkhw_bf16_pipev1_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_instances< - 2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_instance.cpp deleted file mode 100644 index 43039fe11c..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngchw_gkcyx_ngkhw_f16_pipev1_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_instances< - 2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp deleted file mode 100644 index 81f71d459d..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_bf16_instance.cpp +++ /dev/null @@ -1,38 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_ngchw_gkcyx_ngkhw_bf16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp deleted file mode 100644 index d173b0afac..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkcyx_ngkhw/device_grouped_conv2d_bwd_weight_wmma_ngchw_gkcyx_ngkhw_f16_instance.cpp +++ /dev/null @@ -1,38 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_ngchw_gkcyx_ngkhw_f16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<2, - NGCHW, - GKCYX, - NGKHW, - ConvBwdWeightDefault>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp deleted file mode 100644 index e222e8c663..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngchw_gkyxc_ngkhw_bf16_pipev1_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< - 2, - NGCHW, - GKYXC, - NGKHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp deleted file mode 100644 index 85d395c7c2..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/ngchw_gkyxc_ngkhw/device_grouped_conv2d_bwd_weight_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngchw_gkyxc_ngkhw_f16_pipev1_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< - 2, - NGCHW, - GKYXC, - NGKHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp deleted file mode 100644 index 2cc474005e..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_weight/wmma/nhwgc_gkyxc_nhwgk/device_grouped_conv2d_bwd_weight_wmma_nhwgc_gkyxc_nhwgk_bf16_f32_bf16_instance.cpp +++ /dev/null @@ -1,48 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_f32_bf16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 2, - NHWGC, - GKYXC, - NHWGK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt index 6d313ab2f5..b246b87178 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/CMakeLists.txt @@ -69,22 +69,10 @@ if(DL_KERNELS) endif() 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_f16_instance.cpp - 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_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instance.cpp wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_two_stage_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instance.cpp - - 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 - wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_instance.cpp - wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_instance.cpp - - wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp - wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp ) if((DTYPES MATCHES "fp8" AND DTYPES MATCHES "bf8" AND DTYPES MATCHES "fp16") OR NOT DEFINED DTYPES) diff --git a/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 b/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 deleted file mode 100644 index 98e8547fb4..0000000000 --- a/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 +++ /dev/null @@ -1,48 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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 { - -void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_bf16_f32_bf16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 3, - GNDHWC, - GKZYXC, - GNDHWK, - ConvBwdWeightDefault>{}); - - // 2. Filter1x1Stride1Pad0 - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 3, - GNDHWC, - GKZYXC, - GNDHWK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp deleted file mode 100644 index 7cd9d5655f..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/gndhwc_gkzyxc_gndhwk/device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instance.cpp +++ /dev/null @@ -1,45 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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 { - -void add_device_grouped_conv3d_bwd_weight_wmma_gndhwc_gkzyxc_gndhwk_f16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<3, - GNDHWC, - GKZYXC, - GNDHWK, - ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - add_device_operation_instances(instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances< - 3, - GNDHWC, - GKZYXC, - GNDHWK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp deleted file mode 100644 index 5bf7fa3a9a..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ndhwgc_gkzyxc_ndhwgk/device_grouped_conv3d_bwd_weight_wmma_ndhwgc_gkzyxc_ndhwgk_bf16_f32_bf16_instance.cpp +++ /dev/null @@ -1,48 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_f32_bf16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightDefault>{}); - // 2. Filter1x1Stride1Pad0 - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_f32_bf16_instances< - 3, - NDHWGC, - GKZYXC, - NDHWGK, - ConvBwdWeightFilter1x1Stride1Pad0>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_instance.cpp deleted file mode 100644 index 1ddeacbc79..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_bf16_pipev1_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_instances< - 3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_instance.cpp deleted file mode 100644 index eaffb670cc..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngcdhw_gkczyx_ngkdhw_f16_pipev1_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_instances< - 3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp deleted file mode 100644 index 060578c09a..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_bf16_instance.cpp +++ /dev/null @@ -1,38 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_ngcdhw_gkczyx_ngkdhw_bf16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_bf16_instances<3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp deleted file mode 100644 index a993497fab..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkczyx_ngkdhw/device_grouped_conv3d_bwd_weight_wmma_ngcdhw_gkczyx_ngkdhw_f16_instance.cpp +++ /dev/null @@ -1,38 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_ngcdhw_gkczyx_ngkdhw_f16_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_v3_wmma_c_shuffle_f16_instances<3, - NGCDHW, - GKCZYX, - NGKDHW, - ConvBwdWeightDefault>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp deleted file mode 100644 index 39d8ef3091..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_bf16_pipev1_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_bf16_generic_instances< - 3, - NGCDHW, - GKZYXC, - NGKDHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp deleted file mode 100644 index 76b6fc9b0d..0000000000 --- a/library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_weight/wmma/ngcdhw_gkzyxc_ngkdhw/device_grouped_conv3d_bwd_weight_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instance.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// Copyright (c) Advanced Micro Devices, Inc., or its affiliates. -// SPDX-License-Identifier: MIT - -#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_two_stage_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_two_stage_wmma_ngcdhw_gkzyxc_ngkdhw_f16_pipev1_instances( - std::vector>>& instances) -{ - // 1. Default - add_device_operation_instances( - instances, - device_grouped_conv_bwd_weight_two_stage_ngchw_wmma_c_shuffle_f16_generic_instances< - 3, - NGCDHW, - GKZYXC, - NGKDHW, - ConvBwdWeightDefault, - BlockGemmPipelineScheduler::Intrawave, - BlockGemmPipelineVersion::v1>{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck