diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp | 163 |
1 files changed, 84 insertions, 79 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp index 430d35e06d..dfacb687a8 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_uint8.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 ARM Limited. + * Copyright (c) 2017-2020, 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -28,107 +28,111 @@ #include "gemm_implementation.hpp" #include "gemm_interleaved.hpp" #include "gemm_hybrid.hpp" -#include "gemm_native.hpp" +#include "gemm_hybrid_indirect.hpp" -#include "kernels/a64_gemm_u16_12x8.hpp" -#include "kernels/a64_gemm_u8_12x8.hpp" +#include "kernels/a64_gemm_u16_8x12.hpp" #include "kernels/a64_gemm_u8_4x4.hpp" -#include "kernels/a64_hybrid_u8u32_dot_16x4.hpp" -#include "kernels/a64_interleaved_u8u32_mmla_12x8.hpp" -#include "kernels/a64_smallK_hybrid_u8u32_dot_4x6.hpp" -#include "kernels/a64_smallK_hybrid_u8u32_dot_4x8.hpp" -#include "kernels/sve_hybrid_u8u32_dot_4VLx4.hpp" -#include "kernels/sve_interleaved_u8u32_dot_3VLx8.hpp" -#include "kernels/sve_interleaved_u8u32_mmla_3VLx8.hpp" -#include "kernels/sve_native_u8u32_dot_4VLx4.hpp" -#include "kernels/sve_smallK_hybrid_u8u32_dot_1VLx8.hpp" +#include "kernels/a64_gemm_u8_8x12.hpp" +#include "kernels/a64_hybrid_u8u32_dot_6x16.hpp" +#include "kernels/a64_hybrid_u8u32_mmla_6x16.hpp" +#include "kernels/a64_interleaved_u8u32_mmla_8x12.hpp" +#include "kernels/a64_smallK_hybrid_u8u32_dot_6x4.hpp" +#include "kernels/a64_smallK_hybrid_u8u32_dot_8x4.hpp" + +#include "kernels/sve_hybrid_u8u32_dot_6x4VL.hpp" +#include "kernels/sve_hybrid_u8u32_mmla_6x4VL.hpp" +#include "kernels/sve_interleaved_u8u32_dot_8x3VL.hpp" +#include "kernels/sve_interleaved_u8u32_mmla_8x3VL.hpp" namespace arm_gemm { static const GemmImplementation<uint8_t, uint32_t> gemm_u8_methods[] = { -#ifdef __ARM_FEATURE_SVE -#ifdef V8P6 -{ - GemmMethod::GEMM_INTERLEAVED, - "interleaved_u8u32_mmla_3VLx8", - [](const GemmArgs &args) { return (args._Ksize>8); }, - nullptr, - [](const GemmArgs &args) { return new GemmInterleaved<interleaved_u8u32_mmla_3VLx8, uint8_t, uint32_t>(args); } -}, -#endif -{ +#ifdef ARM_COMPUTE_ENABLE_SVE +GemmImplementation<uint8_t, uint32_t>::with_estimate( GemmMethod::GEMM_HYBRID, - "smallK_hybrid_u8u32_dot_1VLx8", - [](const GemmArgs &args) { return args._Ksize<=64 && !args._trA && args._pretransposed_hint; }, - nullptr, - [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_1VLx8, uint8_t, uint32_t>(args); } -}, -{ + "sve_hybrid_u8u32_mmla_6x4VL", + [](const GemmArgs &args) { return args._ci->has_svei8mm(); }, + [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_mmla_6x4VL, uint8_t, uint32_t>(args); } +), +GemmImplementation<uint8_t, uint32_t>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "sve_interleaved_u8u32_mmla_8x3VL", + [](const GemmArgs &args) { return args._ci->has_svei8mm() && (args._Ksize>8); }, + [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_u8u32_mmla_8x3VL, uint8_t, uint32_t>(args); } +), +GemmImplementation<uint8_t, uint32_t>::with_estimate( GemmMethod::GEMM_HYBRID, - "hybrid_u8u32_dot_4VLx4", - [](const GemmArgs &args) { return args._Ksize>=16 && !args._trA && args._pretransposed_hint; }, - [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, - [](const GemmArgs &args) { return new GemmHybrid<hybrid_u8u32_dot_4VLx4, uint8_t, uint32_t>(args); } -}, -{ - GemmMethod::GEMM_NATIVE, - "native_u8u32_dot_4VLx4", - [](const GemmArgs &args) { return (args._Ksize>=16 && !args._trA && !args._trB); }, - [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)); }, - [](const GemmArgs &args) { return new GemmNative<native_u8u32_dot_4VLx4, uint8_t, uint32_t>(args); } -}, -{ + "sve_hybrid_u8u32_dot_6x4VL", + [](const GemmArgs &args) { return args._ci->has_sve(); }, + [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_u8u32_dot_6x4VL, uint8_t, uint32_t>(args); } +), +GemmImplementation<uint8_t, uint32_t>::with_estimate( GemmMethod::GEMM_INTERLEAVED, - "interleaved_u8u32_dot_3VLx8", - [](const GemmArgs &args) { return (args._Ksize>4); }, - nullptr, - [](const GemmArgs &args) { return new GemmInterleaved<interleaved_u8u32_dot_3VLx8, uint8_t, uint32_t>(args); } -}, -#endif -#ifdef V8P6 -{ + "sve_interleaved_u8u32_dot_8x3VL", + [](const GemmArgs &args) { return args._ci->has_sve() && (args._Ksize>4); }, + [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_u8u32_dot_8x3VL, uint8_t, uint32_t>(args); } +), +#endif // ARM_COMPUTE_ENABLE_SVE +GemmImplementation<uint8_t, uint32_t>::with_estimate( GemmMethod::GEMM_INTERLEAVED, - "interleaved_u8u32_mmla_12x8", - [](const GemmArgs &args) { return (args._Ksize>8); }, - nullptr, - [](const GemmArgs &args) { return new GemmInterleaved<interleaved_u8u32_mmla_12x8, uint8_t, uint32_t>(args); } -}, -#endif -{ + "a64_interleaved_u8u32_mmla_8x12", + [](const GemmArgs &args) { return args._ci->has_i8mm() && (args._Ksize>8); }, + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_u8u32_mmla_8x12, uint8_t, uint32_t>(args); } +), +GemmImplementation<uint8_t, uint32_t>::with_estimate( GemmMethod::GEMM_HYBRID, - "smallK_hybrid_u8u32_dot_4x8", - [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32) && !args._trA && args._pretransposed_hint; }, - nullptr, - [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_4x8, uint8_t, uint32_t>(args); } -}, + "a64_hybrid_u8u32_mmla_6x16", + [](const GemmArgs &args) { return args._ci->has_i8mm(); }, + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_mmla_6x16, uint8_t, uint32_t>(args); } +), { GemmMethod::GEMM_HYBRID, - "smallK_hybrid_u8u32_dot_4x6", - [](const GemmArgs &args) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64) && !args._trA && args._pretransposed_hint; }, - nullptr, - [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_u8u32_dot_4x6, uint8_t, uint32_t>(args); } + "a64_smallK_hybrid_u8u32_dot_8x4", + [](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_u8u32_dot_8x4, uint8_t, uint32_t>(args); } }, { GemmMethod::GEMM_HYBRID, - "hybrid_u8u32_dot_16x4", - [](const GemmArgs &args) { return args._ci->has_dotprod() && args._Ksize>=16 && !args._trA && !args._trB && args._pretransposed_hint; }, - [](const GemmArgs &args) { return args._Nsize<=256 && args._Ksize>128; }, - [](const GemmArgs &args) { return new GemmHybrid<hybrid_u8u32_dot_16x4, uint8_t, uint32_t>(args); } + "a64_smallK_hybrid_u8u32_dot_6x4", + [](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_u8u32_dot_6x4, uint8_t, uint32_t>(args); } }, { GemmMethod::GEMM_INTERLEAVED, - "gemm_u8_12x8", - [](const GemmArgs &args) { return args._ci->has_dotprod(); }, + "a64_gemm_u16_8x12", nullptr, - [](const GemmArgs &args) { return new GemmInterleaved<gemm_u8_12x8, uint8_t, uint32_t>(args); } + [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A53 && args._Msize > 4; }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u16_8x12, uint8_t, uint32_t>(args); }, }, -{ +GemmImplementation<uint8_t, uint32_t>::with_estimate( + GemmMethod::GEMM_HYBRID, + "a64_hybrid_u8u32_dot_6x16", + [](const GemmArgs &args) { return args._ci->has_dotprod(); }, + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_u8u32_dot_6x16, uint8_t, uint32_t>(args); } +), +GemmImplementation<uint8_t, uint32_t>::with_estimate( GemmMethod::GEMM_INTERLEAVED, - "gemm_u8_4x4", - nullptr, + "a64_gemm_u8_8x12", + [](const GemmArgs &args) { return args._ci->has_dotprod(); }, + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_u8_8x12, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u8_8x12, uint8_t, uint32_t>(args); } +), +GemmImplementation<uint8_t, uint32_t>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "a64_gemm_u8_4x4", nullptr, - [](const GemmArgs &args) { return new GemmInterleaved<gemm_u8_4x4, uint8_t, uint32_t>(args); } -}, + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_gemm_u8_4x4, uint8_t, uint32_t>::estimate_cycles<uint32_t>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_gemm_u8_4x4, uint8_t, uint32_t>(args); } +), { GemmMethod::DEFAULT, "", @@ -145,6 +149,7 @@ const GemmImplementation<uint8_t, uint32_t> *gemm_implementation_list<uint8_t, u /* Explicitly instantiate the external functions for these types. */ template UniqueGemmCommon<uint8_t, uint32_t> gemm<uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &); +template bool has_opt_gemm<uint8_t, uint32_t, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &); template KernelDescription get_gemm_method<uint8_t, uint32_t, Nothing>(const GemmArgs &args, const Nothing &); template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint32_t, Nothing> (const GemmArgs &args, const Nothing &); |