From ea9805b580ac06e9394850f0bf4794754d509dd6 Mon Sep 17 00:00:00 2001 From: Zoltan Lakatos Date: Tue, 10 Jun 2025 17:23:44 +0000 Subject: [PATCH 1/7] added add_fastgelu instances --- .../gpu/gemm_add_fastgelu.hpp | 67 +++++++++++++-- .../gpu/gemm_add_fastgelu/CMakeLists.txt | 5 +- ...e_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp | 76 +++++++++++++++++ ...e_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp | 78 +++++++++++++++++ ...e_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp | 81 ++++++++++++++++++ ...e_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp | 83 +++++++++++++++++++ ...l_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp | 73 ---------------- ...bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp | 2 +- test/gemm_add/CMakeLists.txt | 2 + test/gemm_add/test_gemm_add_fastgelu_wmma.cpp | 5 +- 10 files changed, 390 insertions(+), 82 deletions(-) create mode 100644 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp delete mode 100644 library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp index c93e609b7a..597bb06ef7 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_fastgelu.hpp @@ -97,8 +97,47 @@ void add_device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_in #endif // CK_USE_XDL #if defined(CK_USE_WMMA) -void add_device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances( +void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances( std::vector>>&); + +void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances( + std::vector>>&); + +void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances( + std::vector>>&); + +void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances( + std::vector && is_same_v && is_same_v && is_same_v) { - if constexpr(is_same_v && is_same_v && + if constexpr(is_same_v && is_same_v && is_same_v && is_same_v) { - add_device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances( + add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v && is_same_v) + { + add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v && is_same_v) + { + add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v && is_same_v) + { + add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances( op_ptrs); } } - - // TODO: Add other types and layouts - + #endif // CK_ENABLE_FP16 #endif // CK_USE_WMMA diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt index 13878116c2..46f0c3b9c6 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt @@ -1,6 +1,9 @@ # ONLY XDL_AND_WMMA_KERNELS add_instance_library(device_gemm_add_fastgelu_instance - device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp device_gemm_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp new file mode 100644 index 0000000000..26c2a0e3ef --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp @@ -0,0 +1,76 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise((a * b), d0) +// elementwise(c, d0) = fastgelu(c + d0) +// output: e[m, n] +// input: a[m, k], b[n, k], d0[m, n] + +template +using device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance{}); + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp new file mode 100644 index 0000000000..3b79d73d20 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp @@ -0,0 +1,78 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise((a * b), d0) +// elementwise(c, d0) = fastgelu(c + d0) +// output: e[m, n] +// input: a[m, k], b[n, k], d0[m, n] + +template +using device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances{}); + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp new file mode 100644 index 0000000000..4e3fc98a53 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instance.cpp @@ -0,0 +1,81 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise((a * b), d0) +// elementwise(c, d0) = fastgelu(c + d0) +// output: e[m, n] +// input: a[m, k], b[n, k], d0[m, n] + +template +using device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances{}); + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_kn_mn_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp new file mode 100644 index 0000000000..601df433d0 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp @@ -0,0 +1,83 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise((a * b), d0) +// elementwise(c, d0) = fastgelu(c + d0) +// output: e[m, n] +// input: a[m, k], b[n, k], d0[m, n] + +template +using device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances{}); + add_device_operation_instances( + instances, + device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_mk_nk_mn_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp deleted file mode 100644 index 52edd68752..0000000000 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instance.cpp +++ /dev/null @@ -1,73 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. - -#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" -#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" -#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" -#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" -#include "ck/utility/sequence.hpp" - -namespace ck { -namespace tensor_operation { -namespace device { -namespace instance { - -template -using S = ck::Sequence; - -static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; -static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; - -static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; -static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; - -// e = elementwise((a * b), d0) -// elementwise(c, d0) = fastgelu(c + d0) -// output: e[m, n] -// input: a[m, k], b[n, k], d0[m, n] - -using device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_generic_instances = - std::tuple< - // clang-format off - //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| - //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| - //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | - //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | - DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmMNKPadding, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, BlockGemmPipelineVersion::v1> - // clang-format on - >; - -using device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances = std::tuple< - // clang-format off - //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| - //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| - //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | - //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | - DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Tuple, Row, F16, F16, F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddFastGelu, GemmDefault, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, BlockGemmPipelineVersion::v1> - // clang-format on - >; - -void add_device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances( - std::vector>>& instances) -{ - add_device_operation_instances( - instances, - device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_generic_instances{}); - add_device_operation_instances( - instances, device_gemm_add_fastgelu_wmma_universal_f16_f16_f16_f16_mk_nk_mn_mn_instances{}); -} - -} // namespace instance -} // namespace device -} // namespace tensor_operation -} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp index baaeac618e..0bbbf589a9 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp @@ -47,7 +47,7 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_insta >; void add_device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances( - std::vector decltype(ProfileGemmAddFastgeluImpl) GetImpl() override { return ProfileGemmAddFastgeluImpl; } }; -using KernelTypes = ::testing::Types>; +using KernelTypes = ::testing::Types, + std::tuple, + std::tuple, + std::tuple>; TYPED_TEST_SUITE(TestGemmAddFastgelu, KernelTypes); TYPED_TEST(TestGemmAddFastgelu, Test_BF16FP16) { this->Run(); } From aeca8efdea7003913e3b1606b05b243897d5ae78 Mon Sep 17 00:00:00 2001 From: Zoltan Lakatos Date: Wed, 11 Jun 2025 07:50:53 +0000 Subject: [PATCH 2/7] revert unintended change in xdl add_fastgelu --- ...elu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp index 0bbbf589a9..baaeac618e 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instance.cpp @@ -47,7 +47,7 @@ using device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_insta >; void add_device_gemm_add_fastgelu_xdl_c_shuffle_bf16_i8_bf16_bf16_mk_kn_mn_mn_instances( - std::vector Date: Wed, 11 Jun 2025 12:46:42 +0000 Subject: [PATCH 3/7] created gemm_add_add_fastgelu instances --- .../gpu/gemm_add_add_fastgelu.hpp | 181 +++++++++++++++++- .../gpu/gemm_add_add_fastgelu/CMakeLists.txt | 7 +- ...16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp | 76 ++++++++ ...16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp | 78 ++++++++ ...16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp | 81 ++++++++ ...16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp | 84 ++++++++ .../gpu/gemm_add_fastgelu/CMakeLists.txt | 2 +- profiler/src/CMakeLists.txt | 4 +- 8 files changed, 502 insertions(+), 11 deletions(-) create mode 100644 library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp index 99b2ad1315..5675baaa6b 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2025, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -16,6 +16,7 @@ namespace tensor_operation { namespace device { namespace instance { +#if defined(CK_USE_XDL) void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances( std::vector>>&); +#endif // CK_USE_XDL + +#if defined(CK_USE_WMMA) +void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances( + std::vector>>&); + +void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances( + std::vector>>&); + +void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances( + std::vector>>&); + +void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances( + std::vector>>&); +#endif // CK_USE_WMMA + +// GEMM + Add + FastGelu +// DeviceGemmMultipleDSplitK specialization +template +struct DeviceOperationInstanceFactory< + ck::tensor_operation::device::DeviceGemmMultipleDSplitK, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + PassThrough, + PassThrough, + AddAddFastGelu>> +{ + using DeviceOp = DeviceGemmMultipleDSplitK, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + PassThrough, + PassThrough, + AddAddFastGelu>; + + static auto GetInstances() + { + std::vector> op_ptrs; + +#if defined(CK_USE_XDL) + // No XDL instances for DeviceGemmMultipleDSplitK with AddFastGelu at the moment +#endif // CK_USE_XDL + +#if defined(CK_USE_WMMA) +#if defined(CK_ENABLE_FP16) + constexpr bool IsAllDRowLayout = is_same_v && is_same_v; + constexpr bool IsAllDFloat16 = + is_same_v && is_same_v; + + if constexpr(is_same_v && is_same_v && + is_same_v && IsAllDRowLayout && IsAllDFloat16) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances( + op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances( + op_ptrs); + } + } + +#endif // CK_ENABLE_FP16 +#endif // CK_USE_WMMA + + return op_ptrs; + } +}; // GEMM + Add + Add + FastGelu +// DeviceGemmMultipleD specialization template > op_ptrs; +#if defined(CK_USE_XDL) + constexpr bool IsAllDRowLayout = is_same_v && is_same_v; + constexpr bool IsAllDFloat16 = + is_same_v && is_same_v; + + if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && - is_same_v) + is_same_v && IsAllDRowLayout && IsAllDFloat16) { if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && is_same_v) { add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances( op_ptrs); } else if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && is_same_v) { add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances( op_ptrs); } else if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && is_same_v) { add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances( op_ptrs); } else if constexpr(is_same_v && is_same_v && - is_same_v && is_same_v && is_same_v) { add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances( op_ptrs); } } +#endif // CK_USE_XDL + +#if defined(CK_USE_WMMA) + // Reuse DeviceGemmMultipleDSplitK instances + using Wrapper = DeviceGemmMultipleDSplitKWrapper, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + PassThrough, + PassThrough, + AddAddFastGelu>; + auto new_op_ptrs = + DeviceOperationInstanceFactory::GetInstances(); + for(auto& op_ptr : new_op_ptrs) + { + op_ptrs.emplace_back(std::make_unique(std::move(op_ptr))); + } +#endif // CK_USE_WMMA return op_ptrs; } diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt index 04ae90bc5b..53724bc96d 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt @@ -1,5 +1,10 @@ -# ONLY XDL_KERNELS +# XDL AND WMMA KERNELS add_instance_library(device_gemm_add_add_fastgelu_instance + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp + device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp new file mode 100644 index 0000000000..1695298f99 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp @@ -0,0 +1,76 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise((a * b), d0, d1) +// elementwise(c, d0, d1) = fastgelu(c + d0 + d1) +// output: e[m, n] +// input: a[m, k], b[n, k], d0[m, n], d1[m, n] + +template +using device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance{}); + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp new file mode 100644 index 0000000000..01d2480fd1 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp @@ -0,0 +1,78 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise((a * b), d0) +// elementwise(c, d0) = fastgelu(c + d0) +// output: e[m, n] +// input: a[m, k], b[n, k], d0[m, n] + +template +using device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances{}); + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp new file mode 100644 index 0000000000..daa0e175bd --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instance.cpp @@ -0,0 +1,81 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise((a * b), d0, d1) +// elementwise(c, d0, d1) = fastgelu(c + d0 + d1) +// output: e[m, n] +// input: a[m, k], b[n, k], d0[m, n], d1[m, n] + +template +using device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances{}); + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_kn_mn_mn_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp new file mode 100644 index 0000000000..7b593bd191 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instance.cpp @@ -0,0 +1,84 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise((a * b), d0, d1) +// elementwise(c, d0, d1) = fastgelu(c + d0 + d1) +// output: e[m, n] +// input: a[m, k], b[n, k], d0[m, n], d1[m, n] + +template +using device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Row_Row_Tuple, Row, F16, F16, F16_F16_Tuple, F16, F32, F32, PassThrough, PassThrough, AddAddFastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances{}); + add_device_operation_instances( + instances, + device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_mk_nk_mn_mn_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt index 46f0c3b9c6..e07f83c00d 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt @@ -1,4 +1,4 @@ -# ONLY XDL_AND_WMMA_KERNELS +# XDL AND WMMA KERNELS add_instance_library(device_gemm_add_fastgelu_instance device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index 2929f5a042..baa1c27b73 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -44,7 +44,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") list(APPEND PROFILER_OPS profile_batched_gemm_gemm.cpp) list(APPEND PROFILER_OPS profile_batched_gemm_add_relu_gemm_add.cpp) list(APPEND PROFILER_OPS profile_gemm_add.cpp) - list(APPEND PROFILER_OPS profile_gemm_add_add_fastgelu.cpp) list(APPEND PROFILER_OPS profile_grouped_gemm.cpp) list(APPEND PROFILER_OPS profile_gemm_streamk.cpp) list(APPEND PROFILER_OPS profile_gemm_fastgelu.cpp) @@ -92,6 +91,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1[1 list(APPEND PROFILER_OPS profile_grouped_conv_bwd_weight.cpp) if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) list(APPEND PROFILER_OPS profile_gemm_add_fastgelu.cpp) + list(APPEND PROFILER_OPS profile_gemm_add_add_fastgelu.cpp) endif() endif() @@ -148,7 +148,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") endif() if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) list(APPEND DEVICE_INSTANCES device_gemm_add_instance) - list(APPEND DEVICE_INSTANCES device_gemm_add_add_fastgelu_instance) list(APPEND DEVICE_INSTANCES device_gemm_fastgelu_instance) list(APPEND DEVICE_INSTANCES device_batched_gemm_gemm_instance) list(APPEND DEVICE_INSTANCES device_batched_gemm_add_relu_gemm_add_instance) @@ -205,6 +204,7 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1[1 list(APPEND DEVICE_INSTANCES device_grouped_conv3d_bwd_weight_instance) if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) list(APPEND DEVICE_INSTANCES device_gemm_add_fastgelu_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_add_fastgelu_instance) endif() endif() From 264e1b238a965e402df84dfa13fc9b7951eaeb13 Mon Sep 17 00:00:00 2001 From: Zoltan Lakatos Date: Wed, 11 Jun 2025 14:14:43 +0000 Subject: [PATCH 4/7] created fastegelu instances --- .../gpu/gemm_add_add_fastgelu.hpp | 58 +++--- .../gpu/gemm_add_fastgelu.hpp | 46 +++-- .../gpu/gemm_fastgelu.hpp | 176 ++++++++++++++++-- .../gpu/gemm_fastgelu/CMakeLists.txt | 7 +- ..._shuffle_f16_f16_f16_km_kn_mn_instance.cpp | 76 ++++++++ ..._shuffle_f16_f16_f16_km_nk_mn_instance.cpp | 78 ++++++++ ..._shuffle_f16_f16_f16_mk_kn_mn_instance.cpp | 81 ++++++++ ..._shuffle_f16_f16_f16_mk_nk_mn_instance.cpp | 83 +++++++++ profiler/src/CMakeLists.txt | 8 +- 9 files changed, 541 insertions(+), 72 deletions(-) create mode 100644 library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp diff --git a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp index 5675baaa6b..f2264a491f 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/gemm_add_add_fastgelu.hpp @@ -11,6 +11,7 @@ #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#if defined(CK_ENABLE_FP16) namespace ck { namespace tensor_operation { namespace device { @@ -136,18 +137,17 @@ template -struct DeviceOperationInstanceFactory< - ck::tensor_operation::device::DeviceGemmMultipleDSplitK, - ELayout, - ADataType, - BDataType, - ck::Tuple, - EDataType, - PassThrough, - PassThrough, - AddAddFastGelu>> +struct DeviceOperationInstanceFactory, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + PassThrough, + PassThrough, + AddAddFastGelu>> { using DeviceOp = DeviceGemmMultipleDSplitK && is_same_v; constexpr bool IsAllDFloat16 = is_same_v && is_same_v; @@ -203,8 +202,6 @@ struct DeviceOperationInstanceFactory< op_ptrs); } } - -#endif // CK_ENABLE_FP16 #endif // CK_USE_WMMA return op_ptrs; @@ -223,18 +220,18 @@ template -struct DeviceOperationInstanceFactory, - ELayout, - ADataType, - BDataType, - ck::Tuple, - EDataType, - ck::tensor_operation::element_wise::PassThrough, - ck::tensor_operation::element_wise::PassThrough, - ck::tensor_operation::element_wise::AddAddFastGelu>> +struct DeviceOperationInstanceFactory< + DeviceGemmMultipleD, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + PassThrough, + PassThrough, + AddAddFastGelu>> { using DeviceOp = DeviceGemmMultipleD, EDataType, - ck::tensor_operation::element_wise::PassThrough, - ck::tensor_operation::element_wise::PassThrough, - ck::tensor_operation::element_wise::AddAddFastGelu>; + PassThrough, + PassThrough, + AddAddFastGelu>; static auto GetInstances() { @@ -317,3 +314,4 @@ struct DeviceOperationInstanceFactory -struct DeviceOperationInstanceFactory< - ck::tensor_operation::device::DeviceGemmMultipleDSplitK, - ELayout, - ADataType, - BDataType, - ck::Tuple, - EDataType, - PassThrough, - PassThrough, - AddFastGelu>> +struct DeviceOperationInstanceFactory, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + PassThrough, + PassThrough, + AddFastGelu>> { using DeviceOp = DeviceGemmMultipleDSplitK -struct DeviceOperationInstanceFactory< - ck::tensor_operation::device::DeviceGemmMultipleD, - ELayout, - ADataType, - BDataType, - ck::Tuple, - EDataType, - PassThrough, - PassThrough, - AddFastGelu>> +struct DeviceOperationInstanceFactory, + ELayout, + ADataType, + BDataType, + ck::Tuple, + EDataType, + PassThrough, + PassThrough, + AddFastGelu>> { using DeviceOp = DeviceGemmMultipleD>>&); +#endif // CK_USE_XDL + +#if defined(CK_USE_WMMA) +void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances( + std::vector>>&); + +void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances( + std::vector>>&); + +void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instances( + std::vector>>&); + +void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances( + std::vector>>&); +#endif // CK_USE_WMMA + +// GEMM + Add + FastGelu +// DeviceGemmMultipleDSplitK specialization +template +struct DeviceOperationInstanceFactory> +{ + using DeviceOp = DeviceGemmMultipleDSplitK; + + static auto GetInstances() + { + std::vector> op_ptrs; + +#if defined(CK_USE_XDL) + // No XDL instances for DeviceGemmMultipleDSplitK with AddFastGelu at the moment +#endif // CK_USE_XDL + +#if defined(CK_USE_WMMA) + if constexpr(is_same_v && is_same_v && + is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances(op_ptrs); + } + } +#endif // CK_USE_WMMA + + return op_ptrs; + } +}; // GEMM + FastGelu template -struct DeviceOperationInstanceFactory> +struct DeviceOperationInstanceFactory> { using DeviceOp = DeviceGemmMultipleD> op_ptrs; +#if defined(CK_USE_XDL) if constexpr(is_same_v && is_same_v && is_same_v) { @@ -127,6 +255,28 @@ struct DeviceOperationInstanceFactory; + auto new_op_ptrs = + DeviceOperationInstanceFactory::GetInstances(); + for(auto& op_ptr : new_op_ptrs) + { + op_ptrs.emplace_back(std::make_unique(std::move(op_ptr))); + } +#endif // CK_USE_WMMA return op_ptrs; } @@ -136,4 +286,4 @@ struct DeviceOperationInstanceFactory +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise(a * b) +// elementwise(c) = fastgelu(c) +// output: e[m, n] +// input: a[m, k], b[n, k] + +template +using device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instance = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, 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, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instance{}); + add_device_operation_instances( + instances, + device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instance{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp new file mode 100644 index 0000000000..d1e3267061 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp @@ -0,0 +1,78 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise(a * b) +// elementwise(c) = fastgelu(c) +// output: e[m, n] +// input: a[m, k], b[n, k] + +template +using device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Col, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 64, 64, 32, 64, 8, 8, 16, 16, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances{}); + add_device_operation_instances( + instances, + device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp new file mode 100644 index 0000000000..9ec9f3d120 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp @@ -0,0 +1,81 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise(a * b) +// elementwise(c) = fastgelu(c) +// output: e[m, n] +// input: a[m, k], b[n, k] + +template +using device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances = + std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 0, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Row, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 64, 64, 32, 8, 8, 16, 16, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 4, 1, 1, 1, S<1, 32, 1, 2>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances{}); + add_device_operation_instances( + instances, + device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_kn_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp new file mode 100644 index 0000000000..1fc57e10ee --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp @@ -0,0 +1,83 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_wmma_cshuffle_v3.hpp" +#include "ck/utility/sequence.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +template +using S = ck::Sequence; + +static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +static constexpr auto Interwave = BlockGemmPipelineScheduler::Interwave; +static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; + +static constexpr auto V1 = BlockGemmPipelineVersion::v1; +static constexpr auto V3 = BlockGemmPipelineVersion::v3; + +// e = elementwise(a * b) +// elementwise(c) = fastgelu(c) +// output: e[m, n] +// input: a[m, k], b[n, k] + +template +using device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple< + // clang-format off + //##################################| ALayout| BLayout| DsLayout| ELayout| AData| BData| DsData| EData| AccData| CShuffle| A| B| CDE| GemmSpec| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MRepeat| NRepeat| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CShuffleBlockTransfer| CDEShuffleBlockTransfer| BlkGemm| BlkGemm| + //##################################| | | | | Type| Type| Type| Type| Type| DataType| Elementwise| Elementwise| Elementwise| | Size| Block| Block| Block| | | Wmma| Wmma| | | ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| MRepeat| NRepeat| ClusterLengths| ScalarPerVectors| PipeSched| PipelineVer| + //##################################| | | | | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _MBlock_MPerBlock| | | | + //##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | _NBlock_NPerBlock| | | | + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 256, 64, 8, 8, 16, 16, 4, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 64, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 64, 80, 64, 8, 8, 16, 16, 1, 5, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<8, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 64, 1, 2>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Interwave, V1>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 128, 32, 8, 8, 16, 16, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 64, 64, 8, 8, 16, 16, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 256, 128, 160, 64, 8, 8, 16, 16, 2, 5, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 64, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 128, 128, 128, 32, 8, 8, 16, 16, 4, 4, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, S<8, 8, 8>, Intrawave, V3>, + DeviceGemmMultipleD_Wmma_CShuffleV3< Row, Col, Empty_Tuple, Row, F16, F16, Empty_Tuple, F16, F32, F32, PassThrough, PassThrough, FastGelu, GemmSpec, 64, 32, 64, 64, 8, 8, 16, 16, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 0, 1, 1, S<1, 16, 1, 4>, S<8, 8, 8>, Intrawave, V3> + // clang-format on + >; + +void add_device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances( + std::vector>>& instances) +{ + add_device_operation_instances( + instances, + device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances{}); + add_device_operation_instances( + instances, + device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_mk_nk_mn_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/profiler/src/CMakeLists.txt b/profiler/src/CMakeLists.txt index baa1c27b73..06ce589490 100644 --- a/profiler/src/CMakeLists.txt +++ b/profiler/src/CMakeLists.txt @@ -46,7 +46,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") list(APPEND PROFILER_OPS profile_gemm_add.cpp) list(APPEND PROFILER_OPS profile_grouped_gemm.cpp) list(APPEND PROFILER_OPS profile_gemm_streamk.cpp) - list(APPEND PROFILER_OPS profile_gemm_fastgelu.cpp) list(APPEND PROFILER_OPS profile_gemm_add_relu.cpp) list(APPEND PROFILER_OPS profile_gemm_add_silu.cpp) list(APPEND PROFILER_OPS profile_gemm_add_relu_add_layernorm.cpp) @@ -89,8 +88,9 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1[1 list(APPEND PROFILER_OPS profile_grouped_conv_fwd.cpp) list(APPEND PROFILER_OPS profile_grouped_conv_bwd_data.cpp) list(APPEND PROFILER_OPS profile_grouped_conv_bwd_weight.cpp) + list(APPEND PROFILER_OPS profile_gemm_add_fastgelu.cpp) if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - list(APPEND PROFILER_OPS profile_gemm_add_fastgelu.cpp) + list(APPEND PROFILER_OPS profile_gemm_fastgelu.cpp) list(APPEND PROFILER_OPS profile_gemm_add_add_fastgelu.cpp) endif() endif() @@ -148,7 +148,6 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9") endif() if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) list(APPEND DEVICE_INSTANCES device_gemm_add_instance) - list(APPEND DEVICE_INSTANCES device_gemm_fastgelu_instance) list(APPEND DEVICE_INSTANCES device_batched_gemm_gemm_instance) list(APPEND DEVICE_INSTANCES device_batched_gemm_add_relu_gemm_add_instance) list(APPEND DEVICE_INSTANCES device_grouped_gemm_instance) @@ -202,8 +201,9 @@ if(SUPPORTED_GPU_TARGETS MATCHES "gfx9" OR SUPPORTED_GPU_TARGETS MATCHES "gfx1[1 list(APPEND DEVICE_INSTANCES device_grouped_conv3d_bwd_data_instance) list(APPEND DEVICE_INSTANCES device_grouped_conv2d_fwd_instance) list(APPEND DEVICE_INSTANCES device_grouped_conv3d_bwd_weight_instance) + list(APPEND DEVICE_INSTANCES device_gemm_add_fastgelu_instance) if(DTYPES MATCHES "fp16" OR NOT DEFINED DTYPES) - list(APPEND DEVICE_INSTANCES device_gemm_add_fastgelu_instance) + list(APPEND DEVICE_INSTANCES device_gemm_fastgelu_instance) list(APPEND DEVICE_INSTANCES device_gemm_add_add_fastgelu_instance) endif() endif() From b4d3e4112ce15f71b4c2be27277f89fa069cac1e Mon Sep 17 00:00:00 2001 From: Zoltan Lakatos Date: Thu, 12 Jun 2025 11:43:26 +0000 Subject: [PATCH 5/7] added tests for all splitk fastgelus --- ...16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp | 6 +- test/gemm_add/CMakeLists.txt | 10 ++ .../test_gemm_add_add_fastgelu_wmma.cpp | 38 +++++++ test/gemm_add/test_gemm_add_fastgelu_wmma.cpp | 36 +++--- test/gemm_add/test_gemm_add_fastgelu_xdl.cpp | 16 +++ test/gemm_add/test_gemm_common.hpp | 103 ++++++++++++++++-- test/gemm_add/test_gemm_fastgelu_wmma.cpp | 32 ++++++ 7 files changed, 204 insertions(+), 37 deletions(-) create mode 100644 test/gemm_add/test_gemm_add_add_fastgelu_wmma.cpp create mode 100644 test/gemm_add/test_gemm_fastgelu_wmma.cpp diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp index 01d2480fd1..63d9b73901 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp @@ -24,10 +24,10 @@ static constexpr auto Intrawave = BlockGemmPipelineScheduler::Intrawave; static constexpr auto V1 = BlockGemmPipelineVersion::v1; static constexpr auto V3 = BlockGemmPipelineVersion::v3; -// e = elementwise((a * b), d0) -// elementwise(c, d0) = fastgelu(c + d0) +// e = elementwise((a * b), d0, d1) +// elementwise(c, d0, d1) = fastgelu(c + d0 + d1) // output: e[m, n] -// input: a[m, k], b[n, k], d0[m, n] +// input: a[m, k], b[n, k], d0[m, n], d1[m, n] template using device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instances = diff --git a/test/gemm_add/CMakeLists.txt b/test/gemm_add/CMakeLists.txt index 9e061f7c41..88a0cfd0e2 100644 --- a/test/gemm_add/CMakeLists.txt +++ b/test/gemm_add/CMakeLists.txt @@ -20,7 +20,17 @@ if(result EQUAL 0) target_link_libraries(test_gemm_add_fastgelu_xdl PRIVATE utility device_gemm_add_instance device_gemm_add_fastgelu_instance) endif() +add_gtest_executable(test_gemm_fastgelu_wmma test_gemm_fastgelu_wmma.cpp) +if(result EQUAL 0) + target_link_libraries(test_gemm_fastgelu_wmma PRIVATE utility device_gemm_fastgelu_instance) +endif() + add_gtest_executable(test_gemm_add_fastgelu_wmma test_gemm_add_fastgelu_wmma.cpp) if(result EQUAL 0) target_link_libraries(test_gemm_add_fastgelu_wmma PRIVATE utility device_gemm_add_fastgelu_instance) endif() + +add_gtest_executable(test_gemm_add_add_fastgelu_wmma test_gemm_add_add_fastgelu_wmma.cpp) +if(result EQUAL 0) + target_link_libraries(test_gemm_add_add_fastgelu_wmma PRIVATE utility device_gemm_add_add_fastgelu_instance) +endif() diff --git a/test/gemm_add/test_gemm_add_add_fastgelu_wmma.cpp b/test/gemm_add/test_gemm_add_add_fastgelu_wmma.cpp new file mode 100644 index 0000000000..2cde4c7ea3 --- /dev/null +++ b/test/gemm_add/test_gemm_add_add_fastgelu_wmma.cpp @@ -0,0 +1,38 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "gtest/gtest.h" +#include "ck/ck.hpp" +#include "profiler/profile_gemm_add_add_fastgelu_impl.hpp" +#include "test_gemm_common.hpp" + +template +class TestGemmAddAddFastgelu : public TestGemmD0D1Common +{ + using ProfileCall = typename TestGemmD0D1Common::ProfileCall; + + public: + ProfileCall GetImpl() override + { + return ck::profiler::profile_gemm_add_add_fastgelu_impl< + typename TestGemmD0D1Common::ADataType, + typename TestGemmD0D1Common::BDataType, + typename TestGemmD0D1Common::AccDataType, + typename TestGemmD0D1Common::D0DataType, + typename TestGemmD0D1Common::D1DataType, + typename TestGemmD0D1Common::EDataType, + typename TestGemmD0D1Common::ALayout, + typename TestGemmD0D1Common::BLayout, + typename TestGemmD0D1Common::D0Layout, + typename TestGemmD0D1Common::D1Layout, + typename TestGemmD0D1Common::ELayout>; + } +}; + +using KernelTypes = ::testing::Types, + std::tuple, + std::tuple, + std::tuple>; + +TYPED_TEST_SUITE(TestGemmAddAddFastgelu, KernelTypes); +TYPED_TEST(TestGemmAddAddFastgelu, Test_BF16FP16) { this->Run(); } diff --git a/test/gemm_add/test_gemm_add_fastgelu_wmma.cpp b/test/gemm_add/test_gemm_add_fastgelu_wmma.cpp index dfe9b14969..278922412f 100644 --- a/test/gemm_add/test_gemm_add_fastgelu_wmma.cpp +++ b/test/gemm_add/test_gemm_add_fastgelu_wmma.cpp @@ -9,29 +9,21 @@ template class TestGemmAddFastgelu : public TestGemmD0Common { - private: - using ADataType = std::tuple_element_t<0, Tuple>; - using BDataType = std::tuple_element_t<1, Tuple>; - using AccDataType = std::tuple_element_t<2, Tuple>; - using D0DataType = std::tuple_element_t<3, Tuple>; - using EDataType = std::tuple_element_t<4, Tuple>; - using ALayout = std::tuple_element_t<5, Tuple>; - using BLayout = std::tuple_element_t<6, Tuple>; - using D0Layout = std::tuple_element_t<7, Tuple>; - using ELayout = std::tuple_element_t<8, Tuple>; + using ProfileCall = typename TestGemmD0Common::ProfileCall; - constexpr static auto ProfileGemmAddFastgeluImpl = - ck::profiler::profile_gemm_add_fastgelu_impl; - - decltype(ProfileGemmAddFastgeluImpl) GetImpl() override { return ProfileGemmAddFastgeluImpl; } + ProfileCall GetImpl() override + { + return ck::profiler::profile_gemm_add_fastgelu_impl< + typename TestGemmD0Common::ADataType, + typename TestGemmD0Common::BDataType, + typename TestGemmD0Common::AccDataType, + typename TestGemmD0Common::D0DataType, + typename TestGemmD0Common::EDataType, + typename TestGemmD0Common::ALayout, + typename TestGemmD0Common::BLayout, + typename TestGemmD0Common::D0Layout, + typename TestGemmD0Common::ELayout>; + } }; using KernelTypes = ::testing::Types, diff --git a/test/gemm_add/test_gemm_add_fastgelu_xdl.cpp b/test/gemm_add/test_gemm_add_fastgelu_xdl.cpp index 2c055a8006..79e2349088 100644 --- a/test/gemm_add/test_gemm_add_fastgelu_xdl.cpp +++ b/test/gemm_add/test_gemm_add_fastgelu_xdl.cpp @@ -9,6 +9,22 @@ template class TestGemmAddFastgelu : public TestGemmD0Common { + using ProfileCall = typename TestGemmD0Common::ProfileCall; + + ProfileCall GetImpl() override + { + return ck::profiler::profile_gemm_add_fastgelu_impl< + typename TestGemmD0Common::ADataType, + typename TestGemmD0Common::BDataType, + typename TestGemmD0Common::AccDataType, + typename TestGemmD0Common::D0DataType, + typename TestGemmD0Common::EDataType, + typename TestGemmD0Common::ALayout, + typename TestGemmD0Common::BLayout, + typename TestGemmD0Common::D0Layout, + typename TestGemmD0Common::ELayout>; + } + private: using ADataType = std::tuple_element_t<0, Tuple>; using BDataType = std::tuple_element_t<1, Tuple>; diff --git a/test/gemm_add/test_gemm_common.hpp b/test/gemm_add/test_gemm_common.hpp index 1cf41d7538..ce0f6a66ea 100644 --- a/test/gemm_add/test_gemm_common.hpp +++ b/test/gemm_add/test_gemm_common.hpp @@ -3,7 +3,6 @@ #include "gtest/gtest.h" #include "ck/ck.hpp" -#include "profiler/profile_gemm_add_impl.hpp" using Row = ck::tensor_layout::gemm::RowMajor; using Col = ck::tensor_layout::gemm::ColumnMajor; @@ -13,6 +12,47 @@ using BF16 = ck::bhalf_t; using F16 = ck::half_t; using F32 = float; +template +class TestGemmCommon : public ::testing::Test +{ + protected: + using ADataType = std::tuple_element_t<0, Tuple>; + using BDataType = std::tuple_element_t<1, Tuple>; + using AccDataType = std::tuple_element_t<2, Tuple>; + using EDataType = std::tuple_element_t<3, Tuple>; + using ALayout = std::tuple_element_t<4, Tuple>; + using BLayout = std::tuple_element_t<5, Tuple>; + using ELayout = std::tuple_element_t<6, Tuple>; + + using ProfileCall = bool(*const)(int, int, bool, bool, int, int, int, int, int, int); + + virtual ProfileCall GetImpl() = 0; + + void Run() + { + std::vector> lengths = { + {16, 32, 64}, {2048, 4096, 8192}, {2048, 1024, 16}}; + + bool all_success = true; + + for(auto length : lengths) + { + int M = length[0]; + int N = length[1]; + int K = length[2]; + int StrideA = ck::is_same_v ? K : M; + int StrideB = ck::is_same_v ? N : K; + int StrideE = ck::is_same_v ? N : M; + + all_success = + all_success & + GetImpl()(1, 1, false, false, M, N, K, StrideA, StrideB, StrideE); + } + + EXPECT_TRUE(all_success); + } +}; + template class TestGemmD0Common : public ::testing::Test { @@ -27,17 +67,9 @@ class TestGemmD0Common : public ::testing::Test using D0Layout = std::tuple_element_t<7, Tuple>; using ELayout = std::tuple_element_t<8, Tuple>; - constexpr static auto ProfileGemmAddImpl = ck::profiler::profile_gemm_add_impl; + using ProfileCall = bool(*const)(int, int, bool, bool, int, int, int, int, int, int, int); - virtual decltype(ProfileGemmAddImpl) GetImpl() = 0; + virtual ProfileCall GetImpl() = 0; void Run() { @@ -58,7 +90,54 @@ class TestGemmD0Common : public ::testing::Test all_success = all_success & - GetImpl()(true, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideE); + GetImpl()(1, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideE); + } + + EXPECT_TRUE(all_success); + } +}; + +template +class TestGemmD0D1Common : public ::testing::Test +{ + protected: + using ADataType = std::tuple_element_t<0, Tuple>; + using BDataType = std::tuple_element_t<1, Tuple>; + using AccDataType = std::tuple_element_t<2, Tuple>; + using D0DataType = std::tuple_element_t<3, Tuple>; + using D1DataType = std::tuple_element_t<4, Tuple>; + using EDataType = std::tuple_element_t<5, Tuple>; + using ALayout = std::tuple_element_t<6, Tuple>; + using BLayout = std::tuple_element_t<7, Tuple>; + using D0Layout = std::tuple_element_t<8, Tuple>; + using D1Layout = std::tuple_element_t<9, Tuple>; + using ELayout = std::tuple_element_t<10, Tuple>; + + using ProfileCall = bool(*const)(int, int, bool, bool, int, int, int, int, int, int, int, int); + + virtual ProfileCall GetImpl() = 0; + + void Run() + { + std::vector> lengths = { + {16, 32, 64}, {2048, 4096, 8192}, {2048, 1024, 16}}; + + bool all_success = true; + + for(auto length : lengths) + { + int M = length[0]; + int N = length[1]; + int K = length[2]; + int StrideA = ck::is_same_v ? K : M; + int StrideB = ck::is_same_v ? N : K; + int StrideD0 = ck::is_same_v ? N : M; + int StrideD1 = ck::is_same_v ? N : M; + int StrideE = ck::is_same_v ? N : M; + + all_success = + all_success & + GetImpl()(1, 1, false, false, M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE); } EXPECT_TRUE(all_success); diff --git a/test/gemm_add/test_gemm_fastgelu_wmma.cpp b/test/gemm_add/test_gemm_fastgelu_wmma.cpp new file mode 100644 index 0000000000..d8dd218ec6 --- /dev/null +++ b/test/gemm_add/test_gemm_fastgelu_wmma.cpp @@ -0,0 +1,32 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved. + +#include "gtest/gtest.h" +#include "ck/ck.hpp" +#include "profiler/profile_gemm_fastgelu_impl.hpp" +#include "test_gemm_common.hpp" + +template +class TestGemmFastgelu : public TestGemmCommon +{ + using ProfileCall = typename TestGemmCommon::ProfileCall; + + ProfileCall GetImpl() override + { + return ck::profiler::profile_gemm_fastgelu_impl::ADataType, + typename TestGemmCommon::BDataType, + typename TestGemmCommon::AccDataType, + typename TestGemmCommon::EDataType, + typename TestGemmCommon::ALayout, + typename TestGemmCommon::BLayout, + typename TestGemmCommon::ELayout>; + } +}; + +using KernelTypes = ::testing::Types, + std::tuple, + std::tuple, + std::tuple>; + +TYPED_TEST_SUITE(TestGemmFastgelu, KernelTypes); +TYPED_TEST(TestGemmFastgelu, Test_BF16FP16) { this->Run(); } From 61b6e9a60639acc6e7d3c4c85da4cc62bed00be3 Mon Sep 17 00:00:00 2001 From: Zoltan Lakatos Date: Fri, 13 Jun 2025 09:37:37 +0000 Subject: [PATCH 6/7] splitk xdl test fixes --- test/gemm_add/test_gemm_add_fastgelu_xdl.cpp | 24 ------------- test/gemm_add/test_gemm_add_relu_xdl.cpp | 36 ++++++++------------ test/gemm_add/test_gemm_add_silu_xdl.cpp | 36 ++++++++------------ test/gemm_add/test_gemm_add_xdl.hpp | 35 ++++++++----------- 4 files changed, 42 insertions(+), 89 deletions(-) diff --git a/test/gemm_add/test_gemm_add_fastgelu_xdl.cpp b/test/gemm_add/test_gemm_add_fastgelu_xdl.cpp index 79e2349088..0e034f46b5 100644 --- a/test/gemm_add/test_gemm_add_fastgelu_xdl.cpp +++ b/test/gemm_add/test_gemm_add_fastgelu_xdl.cpp @@ -24,30 +24,6 @@ class TestGemmAddFastgelu : public TestGemmD0Common typename TestGemmD0Common::D0Layout, typename TestGemmD0Common::ELayout>; } - - private: - using ADataType = std::tuple_element_t<0, Tuple>; - using BDataType = std::tuple_element_t<1, Tuple>; - using AccDataType = std::tuple_element_t<2, Tuple>; - using D0DataType = std::tuple_element_t<3, Tuple>; - using EDataType = std::tuple_element_t<4, Tuple>; - using ALayout = std::tuple_element_t<5, Tuple>; - using BLayout = std::tuple_element_t<6, Tuple>; - using D0Layout = std::tuple_element_t<7, Tuple>; - using ELayout = std::tuple_element_t<8, Tuple>; - - constexpr static auto ProfileGemmAddFastgeluImpl = - ck::profiler::profile_gemm_add_fastgelu_impl; - - decltype(ProfileGemmAddFastgeluImpl) GetImpl() override { return ProfileGemmAddFastgeluImpl; } }; using KernelTypes = ::testing::Types, diff --git a/test/gemm_add/test_gemm_add_relu_xdl.cpp b/test/gemm_add/test_gemm_add_relu_xdl.cpp index 35aaba96b1..4b445e8e41 100644 --- a/test/gemm_add/test_gemm_add_relu_xdl.cpp +++ b/test/gemm_add/test_gemm_add_relu_xdl.cpp @@ -9,29 +9,21 @@ template class TestGemmAddRelu : public TestGemmD0Common { - private: - using ADataType = std::tuple_element_t<0, Tuple>; - using BDataType = std::tuple_element_t<1, Tuple>; - using AccDataType = std::tuple_element_t<2, Tuple>; - using D0DataType = std::tuple_element_t<3, Tuple>; - using EDataType = std::tuple_element_t<4, Tuple>; - using ALayout = std::tuple_element_t<5, Tuple>; - using BLayout = std::tuple_element_t<6, Tuple>; - using D0Layout = std::tuple_element_t<7, Tuple>; - using ELayout = std::tuple_element_t<8, Tuple>; + using ProfileCall = typename TestGemmD0Common::ProfileCall; - constexpr static auto ProfileGemmAddReluImpl = - ck::profiler::profile_gemm_add_relu_impl; - - decltype(ProfileGemmAddReluImpl) GetImpl() override { return ProfileGemmAddReluImpl; } + ProfileCall GetImpl() override + { + return ck::profiler::profile_gemm_add_relu_impl< + typename TestGemmD0Common::ADataType, + typename TestGemmD0Common::BDataType, + typename TestGemmD0Common::AccDataType, + typename TestGemmD0Common::D0DataType, + typename TestGemmD0Common::EDataType, + typename TestGemmD0Common::ALayout, + typename TestGemmD0Common::BLayout, + typename TestGemmD0Common::D0Layout, + typename TestGemmD0Common::ELayout>; + } }; using KernelTypes = ::testing::Types, diff --git a/test/gemm_add/test_gemm_add_silu_xdl.cpp b/test/gemm_add/test_gemm_add_silu_xdl.cpp index 8d242869c6..6bd0ee422d 100644 --- a/test/gemm_add/test_gemm_add_silu_xdl.cpp +++ b/test/gemm_add/test_gemm_add_silu_xdl.cpp @@ -9,29 +9,21 @@ template class TestGemmAddSilu : public TestGemmD0Common { - private: - using ADataType = std::tuple_element_t<0, Tuple>; - using BDataType = std::tuple_element_t<1, Tuple>; - using AccDataType = std::tuple_element_t<2, Tuple>; - using D0DataType = std::tuple_element_t<3, Tuple>; - using EDataType = std::tuple_element_t<4, Tuple>; - using ALayout = std::tuple_element_t<5, Tuple>; - using BLayout = std::tuple_element_t<6, Tuple>; - using D0Layout = std::tuple_element_t<7, Tuple>; - using ELayout = std::tuple_element_t<8, Tuple>; + using ProfileCall = typename TestGemmD0Common::ProfileCall; - constexpr static auto ProfileGemmAddSiluImpl = - ck::profiler::profile_gemm_add_silu_impl; - - decltype(ProfileGemmAddSiluImpl) GetImpl() override { return ProfileGemmAddSiluImpl; } + ProfileCall GetImpl() override + { + return ck::profiler::profile_gemm_add_silu_impl< + typename TestGemmD0Common::ADataType, + typename TestGemmD0Common::BDataType, + typename TestGemmD0Common::AccDataType, + typename TestGemmD0Common::D0DataType, + typename TestGemmD0Common::EDataType, + typename TestGemmD0Common::ALayout, + typename TestGemmD0Common::BLayout, + typename TestGemmD0Common::D0Layout, + typename TestGemmD0Common::ELayout>; + } }; using KernelTypes = ::testing::Types, diff --git a/test/gemm_add/test_gemm_add_xdl.hpp b/test/gemm_add/test_gemm_add_xdl.hpp index 3cc5405b5f..6df3892883 100644 --- a/test/gemm_add/test_gemm_add_xdl.hpp +++ b/test/gemm_add/test_gemm_add_xdl.hpp @@ -9,28 +9,21 @@ template class TestGemmAdd : public TestGemmD0Common { - private: - using ADataType = std::tuple_element_t<0, Tuple>; - using BDataType = std::tuple_element_t<1, Tuple>; - using AccDataType = std::tuple_element_t<2, Tuple>; - using D0DataType = std::tuple_element_t<3, Tuple>; - using EDataType = std::tuple_element_t<4, Tuple>; - using ALayout = std::tuple_element_t<5, Tuple>; - using BLayout = std::tuple_element_t<6, Tuple>; - using D0Layout = std::tuple_element_t<7, Tuple>; - using ELayout = std::tuple_element_t<8, Tuple>; + using ProfileCall = typename TestGemmD0Common::ProfileCall; - constexpr static auto ProfileGemmAddImpl = ck::profiler::profile_gemm_add_impl; - - decltype(ProfileGemmAddImpl) GetImpl() override { return ProfileGemmAddImpl; } + ProfileCall GetImpl() override + { + return ck::profiler::profile_gemm_add_impl< + typename TestGemmD0Common::ADataType, + typename TestGemmD0Common::BDataType, + typename TestGemmD0Common::AccDataType, + typename TestGemmD0Common::D0DataType, + typename TestGemmD0Common::EDataType, + typename TestGemmD0Common::ALayout, + typename TestGemmD0Common::BLayout, + typename TestGemmD0Common::D0Layout, + typename TestGemmD0Common::ELayout>; + } }; using KernelTypes = ::testing::Types, From 7424b4a0f8ced6ccfbb5f061e762d74cb087ca01 Mon Sep 17 00:00:00 2001 From: Zoltan Lakatos Date: Tue, 17 Jun 2025 19:50:07 +0000 Subject: [PATCH 7/7] fixed ONLY_XDL_AND_WMMA_KERNELS tag --- .../gpu/gemm_add_add_fastgelu/CMakeLists.txt | 2 +- .../gpu/gemm_add_fastgelu/CMakeLists.txt | 2 +- .../tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt index 53724bc96d..ab8023d1ba 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt @@ -1,4 +1,4 @@ -# XDL AND WMMA KERNELS +# ONLY XDL_AND_WMMA_KERNELS add_instance_library(device_gemm_add_add_fastgelu_instance device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_kn_mn_mn_mn_instance.cpp device_gemm_add_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_f16_km_nk_mn_mn_mn_instance.cpp diff --git a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt index e07f83c00d..46f0c3b9c6 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_add_fastgelu/CMakeLists.txt @@ -1,4 +1,4 @@ -# XDL AND WMMA KERNELS +# ONLY XDL_AND_WMMA_KERNELS add_instance_library(device_gemm_add_fastgelu_instance device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_kn_mn_mn_instance.cpp device_gemm_add_fastgelu_wmma_c_shuffle_f16_f16_f16_f16_km_nk_mn_mn_instance.cpp diff --git a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt index 5e28a705f8..f3273fb8ed 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/gemm_fastgelu/CMakeLists.txt @@ -1,4 +1,4 @@ -# XDL AND WMMA KERNELS +# ONLY XDL_AND_WMMA_KERNELS add_instance_library(device_gemm_fastgelu_instance device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp device_gemm_fastgelu_wmma_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp