diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp index e696bfdcdc..038234111a 100644 --- a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include @@ -14,15 +14,12 @@ namespace tensor_operation { namespace device { namespace instance { -using F16 = ck::half_t; using F32 = float; - using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; template -using S = ck::Sequence; - +using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; @@ -34,7 +31,39 @@ using device_gemm_dl_f32_f32_f32_km_kn_mn_instances = std::tuple< // ########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| // ########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | // ########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 1>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<2, 1, 4, 1>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // MPerBlock=128, NPerBlock=128 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=128 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp index d3ad7c60ec..f61ae84bad 100644 --- a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include @@ -14,15 +14,12 @@ namespace tensor_operation { namespace device { namespace instance { -using F16 = ck::half_t; using F32 = float; - using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; template -using S = ck::Sequence; - +using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; @@ -35,7 +32,39 @@ using device_gemm_dl_f32_f32_f32_km_nk_mn_instances = // ########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| // ########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | // ########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 1>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // MPerBlock=128, NPerBlock=128 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=128 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp index a56a36b0ab..2aeaed1fef 100644 --- a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include @@ -14,15 +14,12 @@ namespace tensor_operation { namespace device { namespace instance { -using F16 = ck::half_t; using F32 = float; - using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; template -using S = ck::Sequence; - +using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; @@ -35,7 +32,39 @@ using device_gemm_dl_f32_f32_f32_mk_kn_mn_instances = // ########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| // ########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | // ########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<2, 1, 4, 1>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // MPerBlock=128, NPerBlock=128 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=128, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=128 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 8, 2>, S<8, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<16, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<32, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceGemmDl< F32, F32, F32, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<4, 1, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 1, 2>, S<1, 1, 4, 2>, S<4, 1, 2, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >; diff --git a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp index 63d55e81d6..ff3394d831 100644 --- a/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. #include @@ -14,15 +14,12 @@ namespace tensor_operation { namespace device { namespace instance { -using F16 = ck::half_t; using F32 = float; - using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; template -using S = ck::Sequence; - +using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; @@ -35,7 +32,39 @@ using device_gemm_dl_f32_f32_f32_mk_nk_mn_instances = // ########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Specialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| // ########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | // ########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // MPerBlock=128, NPerBlock=128 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<4, 4>, S<4, 4>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<2, 8>, S<2, 8>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // // MPerBlock=128, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<8, 2>, S<4, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 16, 2, 4, 4, 1, S<2, 8>, S<2, 4>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // // MPerBlock=64, NPerBlock=128 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<4, 2>, S<8, 2>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 16, 2, 4, 4, 1, S<2, 4>, S<2, 8>, S<8, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 2, 2>, S<2, 1, 64, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<2, 4>, S<2, 4>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<8, 1>, S<4, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 8, 2, 4, 4, 1, S<4, 2>, S<8, 1>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 2, 2>, S<2, 1, 32, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=16, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 16, 64, 16, 2, 1, 4, 1, S<4, 2>, S<4, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>, + // MPerBlock=64, NPerBlock=16 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 16, 16, 2, 4, 1, 1, S<4, 2>, S<4, 2>, S<4, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=16, NPerBlock=16 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<2, 2>, S<2, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 16, 16, 16, 16, 2, 2, 2, 1, S<1, 4>, S<1, 4>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 4, 2>, S<4, 1, 4, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=8, NPerBlock=64 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<4, 1>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 8, 64, 32, 2, 1, 2, 1, S<2, 2>, S<8, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + // MPerBlock=64, NPerBlock=8 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<4, 1>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 8, 32, 2, 2, 1, 1, S<8, 2>, S<2, 2>, S<8, 1, 4, 2>, S<4, 1, 16, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<8, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + // MPerBlock=8, NPerBlock=8 + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<4, 1>, S<2, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 1, 2, 1, S<1, 4>, S<1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 2>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<2, 1>, S<4, 1>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1>, + DeviceGemmDl< F32, F32, F32, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 8, 8, 8, 4, 2, 2, 1, 1, S<1, 2>, S<1, 4>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<4, 1, 1, 2>, S<1, 1, 8, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 1> // clang-format on >;