diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp | 109 |
1 files changed, 55 insertions, 54 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp index 0c1d3a387b..5da7161671 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp @@ -82,7 +82,7 @@ namespace arm_gemm { -static const GemmImplementation<float, float> gemm_fp32_methods[] = +static const GemmImplementation<float, float, float> gemm_fp32_methods[] = { // GEMV cases - starting with 'gemv_batched' wrapper to turn batched GEMV into GEMM. { @@ -95,27 +95,27 @@ static const GemmImplementation<float, float> gemm_fp32_methods[] = #ifdef __aarch64__ #ifdef ARM_COMPUTE_ENABLE_BF16 // "fast mode" (BF16) kernels -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_interleaved_bf16fp32_mmla_8x12", [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); }, - [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float>(args); } + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_interleaved_bf16fp32_mmla_8x12, float, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_hybrid_fp32bf16fp32_mmla_6x16", [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); }, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_6x16, float, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_hybrid_fp32bf16fp32_mmla_4x24", [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); }, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32bf16fp32_mmla_4x24, float, float, float>(args); } ), #endif // ARM_COMPUTE_ENABLE_BF16 #ifdef ARM_COMPUTE_ENABLE_SVE @@ -189,26 +189,26 @@ GemmImplementation<float, float>::with_estimate( }, #endif // ARM_COMPUTE_ENABLE_SME2 #ifdef ARM_COMPUTE_ENABLE_BF16 -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_interleaved_bf16fp32_mmla_8x3VL", [](const GemmArgs &args) { return args._fast_mode && args._ci->has_svebf16(); }, - [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float>(args); } + [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_bf16fp32_mmla_8x3VL, float, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "sve_hybrid_fp32bf16fp32_mmla_6x4VL", [](const GemmArgs &args) { return args._fast_mode && args._ci->has_svebf16(); }, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_6x4VL, float, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "sve_hybrid_fp32bf16fp32_mmla_4x6VL", [](const GemmArgs &args) { return args._fast_mode && args._ci->has_svebf16(); }, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32bf16fp32_mmla_4x6VL, float, float, float>(args); } ), #endif // ARM_COMPUTE_ENABLE_BF16 #ifdef ARM_COMPUTE_ENABLE_SVEF32MM @@ -218,8 +218,8 @@ GemmImplementation<float, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_interleaved_fp32_mmla_8x3VL", [](const GemmArgs &args) { return args._ci->has_svef32mm() && (args._Ksize>4); }, - [](const GemmArgs &args) { return !(args._fast_mode && args._ci->has_bf16()); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mmla_8x3VL, float, float>(args); } + [](const GemmArgs &args) { return !(args._fast_mode && args._ci->has_svebf16()); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mmla_8x3VL, float, float, float>(args); } }, #endif // ARM_COMPUTE_ENABLE_SVEF32MM // SVE kernels @@ -228,25 +228,25 @@ GemmImplementation<float, float>::with_estimate( "sve_hybrid_fp32_mla_8x1VL", [](const GemmArgs &args) { return args._ci->has_sve(); }, [](const GemmArgs &args) { return (args._Nsize < 12); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_8x1VL, float, float>(args); } + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_8x1VL, float, float, float>(args); } }, -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "sve_hybrid_fp32_mla_6x4VL", [](const GemmArgs &args) { return args._ci->has_sve(); }, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_interleaved_fp32_mla_8x3VL", [](const GemmArgs &args) { return args._ci->has_sve(); }, - [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float>(args); } + [](const GemmArgs &args) { return GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp32_mla_8x3VL, float, float, float>(args); } ), #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS #ifdef ARM_COMPUTE_ENABLE_BF16 -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_ffinterleaved_bf16fp32_mmla_8x3VL", KernelWeightFormat::VL2VL_BL64_BF16, @@ -254,7 +254,7 @@ GemmImplementation<float, float>::with_estimate( [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, float, float>::estimate_cycles<float>(args); }, [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "sve_ffhybrid_fp32bf16fp32_mmla_4x6VL", KernelWeightFormat::VL2VL_BL64_BF16, @@ -263,7 +263,7 @@ GemmImplementation<float, float>::with_estimate( [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_sve_ffhybrid_fp32bf16fp32_mmla_4x6VL, float, float>(args); } ), #endif -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "sve_ffinterleaved_fp32_mla_8x3VL", KernelWeightFormat::VL1VL_BL32, @@ -271,7 +271,7 @@ GemmImplementation<float, float>::with_estimate( [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_sve_ffinterleaved_fp32_mla_8x3VL, float, float>::estimate_cycles<float>(args); }, [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_sve_ffinterleaved_fp32_mla_8x3VL, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "sve_ffhybrid_fp32_mla_6x4VL", KernelWeightFormat::VL1VL_BL32, @@ -287,7 +287,7 @@ GemmImplementation<float, float>::with_estimate( "a64_sgemm_8x6", nullptr, [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A35; }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x6, float, float>(args); } + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x6, float, float, float>(args); } }, // Arm® Neon™ hybrid methods { @@ -309,33 +309,33 @@ GemmImplementation<float, float>::with_estimate( "a64_hybrid_fp32_mla_8x4", nullptr, [](const GemmArgs &args) { return (args._Nsize < 12); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_8x4, float, float>(args); } + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_8x4, float, float, float>(args); } }, -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_hybrid_fp32_mla_4x24", nullptr, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_4x24, float, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_hybrid_fp32_mla_6x16", nullptr, - [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float>(args); } + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp32_mla_6x16, float, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_sgemm_8x12", nullptr, - [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, float, float>::estimate_cycles<float>(args); }, - [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, float, float>(args); } + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_sgemm_8x12, float, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, float, float, float>(args); } ), #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS #ifdef ARM_COMPUTE_ENABLE_BF16 // "fast mode" (BF16) kernels -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_ffinterleaved_bf16fp32_mmla_8x12", KernelWeightFormat::VL256_BL64_BF16, @@ -343,7 +343,7 @@ GemmImplementation<float, float>::with_estimate( [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, float, float>::estimate_cycles<float>(args); }, [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_bf16fp32_mmla_8x12, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_ffhybrid_fp32bf16fp32_mmla_4x24", KernelWeightFormat::VL256_BL64_BF16, @@ -351,7 +351,7 @@ GemmImplementation<float, float>::with_estimate( [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_4x24, float, float>::estimate_cycles<float>(args); }, [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_4x24, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_ffhybrid_fp32bf16fp32_mmla_6x16", KernelWeightFormat::VL256_BL64_BF16, @@ -359,8 +359,9 @@ GemmImplementation<float, float>::with_estimate( [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_6x16, float, float>::estimate_cycles<float>(args); }, [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32bf16fp32_mmla_6x16, float, float>(args); } ), + #endif // BF16 -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_ffinterleaved_fp32_mla_8x12", KernelWeightFormat::VL128_BL32, @@ -368,7 +369,7 @@ GemmImplementation<float, float>::with_estimate( [](const GemmArgs &args) { return GemmInterleavedFixedFormat<cls_a64_ffinterleaved_fp32_mla_8x12, float, float>::estimate_cycles<float>(args); }, [](const GemmArgs &args) { return new GemmInterleavedFixedFormat<cls_a64_ffinterleaved_fp32_mla_8x12, float, float>(args); } ), -GemmImplementation<float, float>::with_estimate( +GemmImplementation<float, float, float>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_ffhybrid_fp32_mla_6x16", KernelWeightFormat::VL128_BL32, @@ -385,7 +386,7 @@ GemmImplementation<float, float>::with_estimate( "sgemm_8x6", nullptr, nullptr, - [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, float, float>(args); } + [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, float, float, float>(args); } }, #endif // __arm__ { @@ -399,14 +400,14 @@ GemmImplementation<float, float>::with_estimate( /* Templated function to return this list. */ template<> -const GemmImplementation<float, float> *gemm_implementation_list<float, float>() { +const GemmImplementation<float, float, float> *gemm_implementation_list<float, float, float>() { return gemm_fp32_methods; } /* Explicitly instantiate the external functions for these types. */ -template UniqueGemmCommon<float, float> gemm<float, float, Nothing>(const GemmArgs &args, const Nothing &); -template bool has_opt_gemm<float, float, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &); -template KernelDescription get_gemm_method<float, float, Nothing>(const GemmArgs &args, const Nothing &); -template std::vector<KernelDescription> get_compatible_kernels<float, float, Nothing> (const GemmArgs &args, const Nothing &); +template UniqueGemmCommon<float, float, float> gemm<float, float, float, Nothing>(const GemmArgs &args, const Nothing &); +template bool has_opt_gemm<float, float, float, Nothing>(WeightFormat &weight_format, const GemmArgs &args, const Nothing &); +template KernelDescription get_gemm_method<float, float, float, Nothing>(const GemmArgs &args, const Nothing &); +template std::vector<KernelDescription> get_compatible_kernels<float, float, float, Nothing> (const GemmArgs &args, const Nothing &); } // namespace arm_gemm |