diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_int8.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_int8.cpp | 125 |
1 files changed, 83 insertions, 42 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp index 31f225002e..fedda3a47a 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_int8.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -34,101 +34,141 @@ #include "kernels/a64_gemm_s8_8x12.hpp" #include "kernels/a64_gemm_s8_4x4.hpp" #include "kernels/a64_hybrid_s8s32_dot_6x16.hpp" +#include "kernels/a64_hybrid_s8s32_mmla_6x16.hpp" #include "kernels/a64_interleaved_s8s32_mmla_8x12.hpp" #include "kernels/a64_smallK_hybrid_s8s32_dot_6x4.hpp" #include "kernels/a64_smallK_hybrid_s8s32_dot_8x4.hpp" +#ifdef ARM_COMPUTE_ENABLE_SVE +#ifdef ARM_COMPUTE_ENABLE_SME2 +#include "kernels/sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL.hpp" +#include "kernels/sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL.hpp" +#include "kernels/sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL.hpp" +#endif // ARM_COMPUTE_ENABLE_SME2 + #include "kernels/sve_hybrid_s8s32_dot_6x4VL.hpp" +#include "kernels/sve_hybrid_s8s32_mmla_6x4VL.hpp" #include "kernels/sve_interleaved_s8s32_dot_8x3VL.hpp" #include "kernels/sve_interleaved_s8s32_mmla_8x3VL.hpp" -#include "kernels/sve_smallK_hybrid_s8s32_dot_8x1VL.hpp" +#endif // ARM_COMPUTE_ENABLE_SVE namespace arm_gemm { static const GemmImplementation<int8_t, int32_t> gemm_s8_methods[] = { -#ifdef __ARM_FEATURE_SVE -#ifdef MMLA_INT8 +#ifdef ARM_COMPUTE_ENABLE_SVE +#ifdef ARM_COMPUTE_ENABLE_SME2 +// SME kernels { GemmMethod::GEMM_INTERLEAVED, - "sve_interleaved_s8s32_mmla_8x3VL", - [](const GemmArgs &args) { return (args._Ksize>8); }, - nullptr, - [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int32_t>(args); } + "sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL", + [](const GemmArgs &args) { return args._ci->has_sme2(); }, + [](const GemmArgs &args) { const auto VL = sme::get_vector_length<int32_t>(); + return args._Nsize >= 8*VL || args._Msize <= VL || (2*VL < args._Msize && args._Msize <= 3*VL); }, + [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_s8s32_mopa_1VLx4VL, int8_t, int32_t>(args); } }, -#endif { - GemmMethod::GEMM_HYBRID, - "sve_smallK_hybrid_s8s32_dot_8x1VL", - [](const GemmArgs &args) { return args._Ksize<=64 && !args._indirect_input; }, - nullptr, - [](const GemmArgs &args) { return new GemmHybrid<cls_sve_smallK_hybrid_s8s32_dot_8x1VL, int8_t, int32_t>(args); } + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL", + [](const GemmArgs &args) { return args._ci->has_sme2(); }, + [](const GemmArgs &args) { const auto VL = sme::get_vector_length<int32_t>(); + return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); }, + [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_s8s32_mopa_4VLx1VL, int8_t, int32_t>(args); } }, { + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL", + [](const GemmArgs &args) { return args._ci->has_sme2(); }, + nullptr, + [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_s8s32_mopa_2VLx2VL, int8_t, int32_t>(args); } +}, +#endif // ARM_COMPUTE_ENABLE_SME2 +GemmImplementation<int8_t, int32_t>::with_estimate( + GemmMethod::GEMM_HYBRID, + "sve_hybrid_s8s32_mmla_6x4VL", + [](const GemmArgs &args) { return args._ci->has_svei8mm(); }, + [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_mmla_6x4VL, int8_t, int32_t>(args); } +), +GemmImplementation<int8_t, int32_t>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "sve_interleaved_s8s32_mmla_8x3VL", + [](const GemmArgs &args) { return args._ci->has_svei8mm() && (args._Ksize>8); }, + [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_s8s32_mmla_8x3VL, int8_t, int32_t>(args); } +), +GemmImplementation<int8_t, int32_t>::with_estimate( GemmMethod::GEMM_HYBRID, "sve_hybrid_s8s32_dot_6x4VL", - [](const GemmArgs &args) { return args._Ksize>=16; }, - [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, + [](const GemmArgs &args) { return args._ci->has_sve() && args._Ksize>=16; }, + [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); }, [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_s8s32_dot_6x4VL, int8_t, int32_t>(args); } -}, -{ +), +GemmImplementation<int8_t, int32_t>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_interleaved_s8s32_dot_8x3VL", - [](const GemmArgs &args) { return (args._Ksize>4); }, - nullptr, + [](const GemmArgs &args) { return args._ci->has_sve() && (args._Ksize>4); }, + [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int32_t>::estimate_cycles<int32_t>(args); }, [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_s8s32_dot_8x3VL, int8_t, int32_t>(args); } -}, -#endif // SVE -#ifdef MMLA_INT8 -{ +), +#endif // ARM_COMPUTE_ENABLE_SVE +GemmImplementation<int8_t, int32_t>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_interleaved_s8s32_mmla_8x12", - [](const GemmArgs &args) { return (args._Ksize>8); }, - nullptr, + [](const GemmArgs &args) { return args._ci->has_i8mm() && (args._Ksize>8); }, + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int32_t>::estimate_cycles<int32_t>(args); }, [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_s8s32_mmla_8x12, int8_t, int32_t>(args); } -}, -#endif +), +GemmImplementation<int8_t, int32_t>::with_estimate( + GemmMethod::GEMM_HYBRID, + "a64_hybrid_s8s32_mmla_6x16", + [](const GemmArgs &args) { return args._ci->has_i8mm(); }, + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int32_t>::estimate_cycles<int32_t>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_mmla_6x16, int8_t, int32_t>(args); } +), { GemmMethod::GEMM_HYBRID, "a64_smallK_hybrid_s8s32_dot_8x4", - [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input; }, - nullptr, + [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._indirect_input && !args._accumulate; }, + [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); }, [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_s8s32_dot_8x4, int8_t, int32_t>(args); } }, { GemmMethod::GEMM_HYBRID, "a64_smallK_hybrid_s8s32_dot_6x4", - [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input; }, - nullptr, + [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._indirect_input && !args._accumulate; }, + [](const GemmArgs &args) { return !(args._ci->has_svei8mm() || args._ci->has_i8mm()); }, [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_s8s32_dot_6x4, int8_t, int32_t>(args); } }, { GemmMethod::GEMM_INTERLEAVED, "a64_gemm_s16_8x12", nullptr, - [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Ksize>4; }, + [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && ((args._Msize > 28) || ((args._Msize % 8) > 4)); }, [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s16_8x12, int8_t, int32_t>(args); }, }, -{ +GemmImplementation<int8_t, int32_t>::with_estimate( + GemmMethod::GEMM_HYBRID, "a64_hybrid_s8s32_dot_6x16", [](const GemmArgs &args) { return args._ci->has_dotprod(); }, - [](const GemmArgs &args) { return args._Nsize<=256 && args._Ksize>128; }, + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int32_t>::estimate_cycles<int32_t>(args); }, [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_s8s32_dot_6x16, int8_t, int32_t>(args); } -}, -{ +), +GemmImplementation<int8_t, int32_t>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_gemm_s8_8x12", [](const GemmArgs &args) { return args._ci->has_dotprod(); }, - nullptr, + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_s8_8x12, int8_t, int32_t>::estimate_cycles<int32_t>(args); }, [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s8_8x12, int8_t, int32_t>(args); } -}, -{ +), +GemmImplementation<int8_t, int32_t>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_gemm_s8_4x4", nullptr, - nullptr, + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_s8_4x4, int8_t, int32_t>::estimate_cycles<int32_t>(args); }, [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_s8_4x4, int8_t, int32_t>(args); } -}, +), + { GemmMethod::DEFAULT, "", @@ -145,6 +185,7 @@ const GemmImplementation<int8_t, int32_t> *gemm_implementation_list<int8_t, int3 /* Explicitly instantiate the external functions for these types. */ template UniqueGemmCommon<int8_t, int32_t> gemm<int8_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &); +template bool has_opt_gemm<int8_t, int32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &); template KernelDescription get_gemm_method<int8_t, int32_t, Nothing>(const GemmArgs &args, const Nothing &); template std::vector<KernelDescription> get_compatible_kernels<int8_t, int32_t, Nothing> (const GemmArgs &args, const Nothing &); |