diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp | 101 |
1 files changed, 60 insertions, 41 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp index 0ddca04846..7d746789e4 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_bf16.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020, 2022-2023 Arm Limited. + * Copyright (c) 2017-2020, 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,6 +46,7 @@ #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS #include "kernels/sve_ffhybrid_bf16fp32_mmla_6x4VL.hpp" +#include "kernels/sve_ffinterleaved_bf16fp32_dot_8x3VL.hpp" #include "kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp" #endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS @@ -67,10 +68,10 @@ namespace arm_gemm { -static const GemmImplementation<bfloat16, float> gemm_bf16_methods[] = +static const GemmImplementation<bfloat16, bfloat16, float> gemm_bf16_methods[] = { -#ifdef __aarch64__ #ifdef ARM_COMPUTE_ENABLE_BF16 +#ifdef __aarch64__ #ifdef ARM_COMPUTE_ENABLE_SVE #ifdef ARM_COMPUTE_ENABLE_SME2 // SME kernels @@ -106,36 +107,36 @@ static const GemmImplementation<bfloat16, float> gemm_bf16_methods[] = }, #endif // ARM_COMPUTE_ENABLE_SME2 // gemm_bf16_interleaved -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_interleaved_bf16fp32_mmla_8x3VL", [](const GemmArgs &args) { return args._ci->has_svebf16() && (args._Ksize>4); }, - [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, float>(args); } + [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, bfloat16, bfloat16, float>(args); } ), -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_HYBRID, "sve_hybrid_bf16fp32_mmla_6x4VL", [](const GemmArgs &args) { return args._ci->has_svebf16(); }, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, float>::estimate_cycles<bfloat16>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_mmla_6x4VL, bfloat16, bfloat16, float>(args); } ), -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_HYBRID, "sve_hybrid_bf16fp32_dot_6x4VL", [](const GemmArgs &args) { return args._ci->has_svebf16(); }, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, float>::estimate_cycles<bfloat16>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_bf16fp32_dot_6x4VL, bfloat16, bfloat16, float>(args); } ), -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_interleaved_bf16fp32_dot_8x3VL", [](const GemmArgs &args) { return args._ci->has_svebf16() && (args._Ksize>2); }, - [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, float>(args); } + [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_dot_8x3VL, bfloat16, bfloat16, float>(args); } ), #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_ffinterleaved_bf16fp32_mmla_8x3VL", KernelWeightFormat::VL2VL_BL64, @@ -143,7 +144,15 @@ GemmImplementation<bfloat16, float>::with_estimate( [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); }, [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, bfloat16, float>(args); } ), -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "sve_ffinterleaved_bf16fp32_dot_8x3VL", + KernelWeightFormat::VL1VL_BL32, + [](const GemmArgs &args) { return args._ci->has_svebf16(); }, + [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_dot_8x3VL, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_dot_8x3VL, bfloat16, float>(args); } +), +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_ffhybrid_bf16fp32_mmla_6x4VL", KernelWeightFormat::VL2VL_BL64, @@ -153,36 +162,36 @@ GemmImplementation<bfloat16, float>::with_estimate( ), #endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS #endif // ARM_COMPUTE_ENABLE_SVE -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_hybrid_bf16fp32_mmla_6x16", [](const GemmArgs &args) { return args._ci->has_bf16(); }, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_mmla_6x16, bfloat16, bfloat16, float>(args); } ), -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_interleaved_bf16fp32_mmla_8x12", [](const GemmArgs &args) { return args._ci->has_bf16() && (args._Ksize>4); }, - [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, float>(args); } + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, bfloat16, bfloat16, float>(args); } ), -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_hybrid_bf16fp32_dot_6x16", [](const GemmArgs &args) { return args._ci->has_bf16(); }, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_bf16fp32_dot_6x16, bfloat16, bfloat16, float>(args); } ), -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_interleaved_bf16fp32_dot_8x12", [](const GemmArgs &args) { return args._ci->has_bf16() && (args._Ksize>2); }, - [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, float>(args); } + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_dot_8x12, bfloat16, bfloat16, float>(args); } ), #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_ffinterleaved_bf16fp32_mmla_8x12", KernelWeightFormat::VL256_BL64, @@ -190,7 +199,7 @@ GemmImplementation<bfloat16, float>::with_estimate( [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); }, [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, bfloat16, float>(args); } ), -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_ffhybrid_bf16fp32_mmla_6x16", KernelWeightFormat::VL256_BL64, @@ -198,7 +207,7 @@ GemmImplementation<bfloat16, float>::with_estimate( [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_bf16fp32_mmla_6x16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_bf16fp32_mmla_6x16, bfloat16, float>(args); } ), -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_ffinterleaved_bf16fp32_dot_8x12", KernelWeightFormat::VL128_BL32, @@ -207,15 +216,25 @@ GemmImplementation<bfloat16, float>::with_estimate( [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_dot_8x12, bfloat16, float>(args); } ), #endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS -GemmImplementation<bfloat16, float>::with_estimate( +GemmImplementation<bfloat16, bfloat16, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_sgemm_8x12", nullptr, - [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, float>::estimate_cycles<bfloat16>(args); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, float>(args); } + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, bfloat16, float>::estimate_cycles<bfloat16>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, bfloat16, bfloat16, float>(args); } ), +#elif defined(__arm__) +{ + GemmMethod::GEMM_INTERLEAVED, + "sgemm_8x6", + nullptr, + nullptr, + [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, bfloat16, bfloat16, float>(args); } +}, +#else +# error "Unknown Architecture" +#endif #endif // ARM_COMPUTE_ENABLE_BF16 -#endif // __aarch64__ { GemmMethod::DEFAULT, "", @@ -226,14 +245,14 @@ GemmImplementation<bfloat16, float>::with_estimate( }; template<> -const GemmImplementation<bfloat16, float> *gemm_implementation_list<bfloat16, float>() { +const GemmImplementation<bfloat16, bfloat16, float> *gemm_implementation_list<bfloat16, bfloat16, float>() { return gemm_bf16_methods; } /* Explicitly instantiate the external functions for these types. */ -template UniqueGemmCommon<bfloat16, float> gemm<bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &); -template bool has_opt_gemm<bfloat16, float, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &); -template KernelDescription get_gemm_method<bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &); -template std::vector<KernelDescription> get_compatible_kernels<bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &); +template UniqueGemmCommon<bfloat16, bfloat16, float> gemm<bfloat16, bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &); +template bool has_opt_gemm<bfloat16, bfloat16, float, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &); +template KernelDescription get_gemm_method<bfloat16, bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &); +template std::vector<KernelDescription> get_compatible_kernels<bfloat16, bfloat16, float, Nothing>(const GemmArgs &args, const Nothing &); } // namespace arm_gemm |