From a61b8b785eae5f4432fbf62e85b3b119cf019ea0 Mon Sep 17 00:00:00 2001 From: zjing14 Date: Wed, 6 Sep 2023 09:59:20 -0500 Subject: [PATCH] Fixed fp8 gemm (#882) * add generic instances; fixed initi with fp8 * fixed comment --------- Co-authored-by: Jing Zhang --- .../library/utility/host_tensor_generator.hpp | 35 ++++++++++++++++ ...f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp | 16 ++++++++ ...f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp | 18 ++++++++- ...dl_splitk_f16_f8_f16_mk_kn_mn_instance.cpp | 12 ++++++ ...dl_splitk_f16_f8_f16_mk_nk_mn_instance.cpp | 40 ++++++++++++------- 5 files changed, 106 insertions(+), 15 deletions(-) diff --git a/library/include/ck/library/utility/host_tensor_generator.hpp b/library/include/ck/library/utility/host_tensor_generator.hpp index 31ff13aec2..2fdb0b141d 100644 --- a/library/include/ck/library/utility/host_tensor_generator.hpp +++ b/library/include/ck/library/utility/host_tensor_generator.hpp @@ -95,6 +95,22 @@ struct GeneratorTensor_2 } }; +#if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8 +template <> +struct GeneratorTensor_2 +{ + int min_value = 0; + int max_value = 1; + + template + ck::f8_t operator()(Is...) + { + float tmp = (std::rand() % (max_value - min_value)) + min_value; + return ck::type_convert(tmp); + } +}; +#endif + template struct GeneratorTensor_3 { @@ -127,6 +143,25 @@ struct GeneratorTensor_3 } }; +#if defined CK_ENABLE_FP8 || defined CK_ENABLE_BF8 +template <> +struct GeneratorTensor_3 +{ + float min_value = 0; + float max_value = 1; + + template + ck::f8_t operator()(Is...) + { + float tmp = float(std::rand()) / float(RAND_MAX); + + float fp32_tmp = min_value + tmp * (max_value - min_value); + + return ck::type_convert(fp32_tmp); + } +}; +#endif + template struct GeneratorTensor_4 { diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp index 7946bafc28..10c4453463 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instance.cpp @@ -33,6 +33,18 @@ using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd; static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; +using device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_generic_instances = + std::tuple< + // clang-format off + // M/N/K padding + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row_Tuple, Row, F16, F8, F32, F32, F32_Tuple, F16, PassThrough, PassThrough, MultiplyAdd, GemmMNKPadding, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 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>, 1> + // clang-format on + >; + using device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances = std::tuple< // clang-format off @@ -73,6 +85,10 @@ void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_m PassThrough, MultiplyAdd>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_generic_instances{}); + add_device_operation_instances( instances, device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_kn_mn_mn_mn_instances{}); diff --git a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp index 99db9f4270..71a3e42496 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_multiply_add/device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instance.cpp @@ -33,6 +33,19 @@ using MultiplyAdd = ck::tensor_operation::element_wise::MultiplyAdd; static constexpr auto GemmMNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; +using device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_generic_instances = + std::tuple< + // clang-format off + // M/N/K padding + // N % 8 == 0 && K % 1 == 0 + //##############################| A| B| Ds| E| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| + //##############################| Layout| Layout| Layout| Layout| Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Specialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| + //##############################| | | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| + //##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F8, F32, F32, F32_Tuple, F16, PassThrough, PassThrough, MultiplyAdd, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, 1, 1, S<1, 2, 1, 32>, 1> + // clang-format on + >; + using device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances = std::tuple< // clang-format off @@ -55,7 +68,6 @@ using device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_i DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F8, F32, F32, F32_Tuple, F16, PassThrough, PassThrough, MultiplyAdd, GemmMNKPadding, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, 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, 2, 1, 64>, 1>, DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F8, F32, F32, F32_Tuple, F16, PassThrough, PassThrough, MultiplyAdd, GemmMNKPadding, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 2, 1, 32>, 1>, DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row_Tuple, Row, F16, F8, F32, F32, F32_Tuple, F16, PassThrough, PassThrough, MultiplyAdd, GemmMNKPadding, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 2, 1, 32>, 1> - // clang-format on >; @@ -72,6 +84,10 @@ void add_device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_m PassThrough, MultiplyAdd>>>& instances) { + add_device_operation_instances( + instances, + device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_generic_instances{}); + add_device_operation_instances( instances, device_gemm_multiply_add_xdl_c_shuffle_f16_f8_f32_f32_f16_mk_nk_mn_mn_mn_instances{}); diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instance.cpp index c0b6ec67bb..6168695b7f 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instance.cpp @@ -30,6 +30,16 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough; // static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr auto GemmMNPadding = ck::tensor_operation::device::GemmSpecialization::MNPadding; +using device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_generic_instances = std::tuple< + // clang-format off + //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| + //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type| + //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmMNPadding, 128, 32, 64, 4, 8, 32, 32, 1, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 1, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 8>, 2, F16> + // clang-format on + >; + // Compilation parameters for a[m, k] * b[k, n] = c[m, n] using device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances = std::tuple< // clang-format off @@ -61,6 +71,8 @@ void add_device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances( DeviceGemmSplitK>>& instances) { + add_device_operation_instances(instances, + device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_generic_instances{}); add_device_operation_instances(instances, device_gemm_xdl_splitk_f16_f8_f16_mk_kn_mn_instances{}); } diff --git a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instance.cpp b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instance.cpp index fa38640ae2..a2dd9532a1 100644 --- a/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instance.cpp +++ b/library/src/tensor_operation_instance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instance.cpp @@ -27,7 +27,17 @@ using S = ck::Sequence; using PassThrough = ck::tensor_operation::element_wise::PassThrough; -static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; +static constexpr auto MNKPadding = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +using device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_generic_instances = std::tuple< + // clang-format off + //#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute| + //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type| + //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | + //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 1, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 2, F16> + // clang-format on + >; // Compilation parameters for a[m, k] * b[k, n] = c[m, n] using device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances = std::tuple< @@ -36,19 +46,19 @@ using device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances = std::tuple< //#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type| //#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| | //#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F16>, - DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F16> + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 32, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F16>, + DeviceGemmXdlSplitKCShuffle< F16, F8, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, MNKPadding, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<1, 4, 16, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 16, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F16> // clang-format on >; @@ -57,6 +67,8 @@ void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances( DeviceGemmSplitK>>& instances) { + add_device_operation_instances(instances, + device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_generic_instances{}); add_device_operation_instances(instances, device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances{}); }