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 | 393 |
1 files changed, 300 insertions, 93 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp index 8bef2b7bae..0c1d3a387b 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 ARM Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,153 +24,359 @@ #include "arm_gemm.hpp" #include "gemm_common.hpp" #include "gemm_hybrid.hpp" +#include "gemm_hybrid_indirect.hpp" #include "gemm_implementation.hpp" #include "gemm_interleaved.hpp" -#include "gemm_interleaved_2d.hpp" -#include "gemm_interleaved_pretransposed_2d.hpp" -#include "gemm_native.hpp" #include "gemv_batched.hpp" -#include "gemv_native_transposed.hpp" #include "gemv_pretransposed.hpp" #include "kernels/a32_sgemm_8x6.hpp" -#include "kernels/a64_hybrid_fp32_mla_16x4.hpp" -#include "kernels/a64_hybrid_fp32_mla_4x8.hpp" -#include "kernels/a64_native_fp32_mla_16x4.hpp" -#include "kernels/a64_smallK_hybrid_fp32_mla_4x6.hpp" -#include "kernels/a64_smallK_hybrid_fp32_mla_4x8.hpp" -#include "kernels/a64_sgemm_12x8.hpp" -#include "kernels/a64_sgemv_pretransposed.hpp" -#include "kernels/a64_sgemv_trans.hpp" +#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +#include "kernels/a64_ffhybrid_fp32_mla_6x16.hpp" +#include "kernels/a64_ffhybrid_fp32bf16fp32_mmla_4x24.hpp" +#include "kernels/a64_ffhybrid_fp32bf16fp32_mmla_6x16.hpp" +#include "kernels/a64_ffinterleaved_bf16fp32_mmla_8x12.hpp" +#include "kernels/a64_ffinterleaved_fp32_mla_8x12.hpp" +#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +#include "kernels/a64_hybrid_fp32bf16fp32_mmla_4x24.hpp" +#include "kernels/a64_hybrid_fp32bf16fp32_mmla_6x16.hpp" +#include "kernels/a64_hybrid_fp32_mla_4x24.hpp" +#include "kernels/a64_hybrid_fp32_mla_6x16.hpp" +#include "kernels/a64_hybrid_fp32_mla_8x4.hpp" +#include "kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp" +#include "kernels/a64_sgemm_8x12.hpp" +#include "kernels/a64_sgemm_8x6.hpp" +#include "kernels/a64_smallK_hybrid_fp32_mla_6x4.hpp" +#include "kernels/a64_smallK_hybrid_fp32_mla_8x4.hpp" -#include "kernels/sve_hybrid_fp32_mla_4VLx4.hpp" -#include "kernels/sve_interleaved_fp32_mla_3VLx8.hpp" -#include "kernels/sve_native_fp32_mla_4VLx4.hpp" -#include "kernels/sve_smallK_hybrid_fp32_mla_1VLx8.hpp" +#ifdef ARM_COMPUTE_ENABLE_SVE +#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +#include "kernels/sve_ffhybrid_fp32_mla_6x4VL.hpp" +#include "kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp" +#include "kernels/sve_ffinterleaved_fp32_mla_8x3VL.hpp" +#include "kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp" +#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +#ifdef ARM_COMPUTE_ENABLE_SME2 +#include "kernels/sme2_gemv_fp32_mla_16VL.hpp" +#include "kernels/sme2_gemv_fp32bf16fp32_dot_16VL.hpp" +#include "kernels/sme2_interleaved_nomerge_fp32_mopa_1VLx4VL.hpp" +#include "kernels/sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL.hpp" +#include "kernels/sme2_interleaved_nomerge_fp32_mopa_2VLx2VL.hpp" +#include "kernels/sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL.hpp" +#include "kernels/sme2_interleaved_nomerge_fp32_mopa_4VLx1VL.hpp" +#include "kernels/sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL.hpp" +#endif // ARM_COMPUTE_ENABLE_SME2 + +#include "kernels/sve_ffhybrid_fp32_mla_6x4VL.hpp" +#include "kernels/sve_ffhybrid_fp32bf16fp32_mmla_4x6VL.hpp" +#include "kernels/sve_ffinterleaved_fp32_mla_8x3VL.hpp" +#include "kernels/sve_ffinterleaved_bf16fp32_mmla_8x3VL.hpp" +#include "kernels/sve_hybrid_fp32bf16fp32_mmla_4x6VL.hpp" +#include "kernels/sve_hybrid_fp32bf16fp32_mmla_6x4VL.hpp" +#include "kernels/sve_hybrid_fp32_mla_6x4VL.hpp" +#include "kernels/sve_hybrid_fp32_mla_8x1VL.hpp" +#include "kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp" +#include "kernels/sve_interleaved_fp32_mla_8x3VL.hpp" +#include "kernels/sve_interleaved_fp32_mmla_8x3VL.hpp" +#endif // ARM_COMPUTE_ENABLE_SVE namespace arm_gemm { static const GemmImplementation<float, float> gemm_fp32_methods[] = { +// GEMV cases - starting with 'gemv_batched' wrapper to turn batched GEMV into GEMM. { GemmMethod::GEMV_BATCHED, "gemv_batched", - [](const GemmArgs &args) { return (args._Msize==1) && (args._nbatches>1); }, + [](const GemmArgs &args) { return args._Msize==1 && args._nbatches>1 && !args._indirect_input; }, nullptr, [](const GemmArgs &args) { return new GemvBatched<float, float>(args); } }, #ifdef __aarch64__ +#ifdef ARM_COMPUTE_ENABLE_BF16 +// "fast mode" (BF16) kernels +GemmImplementation<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); } +), + +GemmImplementation<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); } +), +GemmImplementation<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); } +), +#endif // ARM_COMPUTE_ENABLE_BF16 +#ifdef ARM_COMPUTE_ENABLE_SVE +#ifdef ARM_COMPUTE_ENABLE_SME2 +// SME kernels { - GemmMethod::GEMV_PRETRANSPOSED, - "sgemv_pretransposed", - [](const GemmArgs &args) { return (args._Msize==1 && args._pretransposed_hint && args._nbatches==1); }, + GemmMethod::GEMM_HYBRID, + "sme2_gemv_fp32bf16fp32_dot_16VL", + [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2() && args._Msize==1 && args._nbatches==1 && !args._indirect_input && !args._accumulate; }, nullptr, - [](const GemmArgs &args) { return new GemvPretransposed<sgemv_pretransposed, float, float>(args); } + [](const GemmArgs &args) { return new GemvPretransposed<cls_sme2_gemv_fp32bf16fp32_dot_16VL, float, float>(args); } }, { - GemmMethod::GEMV_NATIVE_TRANSPOSED, - "sgemv_trans", - [](const GemmArgs &args) { return (args._Msize==1 && !args._trA && !args._trB && args._nbatches==1); }, + GemmMethod::GEMM_HYBRID, + "sme2_gemv_fp32_mla_16VL", + [](const GemmArgs &args) { return args._ci->has_sme2() && args._Msize==1 && args._nbatches==1 && !args._indirect_input && !args._accumulate; }, nullptr, - [](const GemmArgs &args) { return new GemvNativeTransposed<sgemv_trans, float, float>(args); } + [](const GemmArgs &args) { return new GemvPretransposed<cls_sme2_gemv_fp32_mla_16VL, float, float>(args); } }, - -#ifdef __ARM_FEATURE_SVE -// SVE smallk / native / hybrid methods +#ifdef ARM_COMPUTE_ENABLE_BF16 { - GemmMethod::GEMM_HYBRID, - "smallK_hybrid_fp32_mla_1VLx8", - [](const GemmArgs &args) { return (args._Ksize <= 24) && !args._trA && args._pretransposed_hint; }, - nullptr, - [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_fp32_mla_1VLx8, float, float>(args); } + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_bf16fp32_mopa_1VLx4VL", + [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2() && !args._accumulate; }, + [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>(); + 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_bf16fp32_mopa_1VLx4VL, float, float>(args); } }, +#endif // ARM_COMPUTE_ENABLE_BF16 { - GemmMethod::GEMM_HYBRID, - "hybrid_fp32_mla_4VLx4", - [](const GemmArgs &args) { return (args._Ksize >= 4) && !args._trA && args._pretransposed_hint; }, - [](const GemmArgs &args) { return ((args._Ksize <= 256) && (args._Nsize <= 256)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, - [](const GemmArgs &args) { return new GemmHybrid<hybrid_fp32_mla_4VLx4, float, float>(args); } + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_fp32_mopa_1VLx4VL", + [](const GemmArgs &args) { return args._ci->has_sme2() && !args._accumulate; }, + [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>(); + 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_fp32_mopa_1VLx4VL, float, float>(args); } }, +#ifdef ARM_COMPUTE_ENABLE_BF16 { - GemmMethod::GEMM_NATIVE, - "native_fp32_mla_4VLx4", - [](const GemmArgs &args) { return (args._Ksize>4 && !args._trA && !args._trB); }, - [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, - [](const GemmArgs &args) { return new GemmNative<native_fp32_mla_4VLx4, float, float>(args); } + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL", + [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2() && !args._accumulate; }, + [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>(); + return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); }, + [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_4VLx1VL, float, float>(args); } }, -#endif // __ARM_FEATURE_SVE - -// NEON native / hybrid methods +#endif // ARM_COMPUTE_ENABLE_BF16 { - GemmMethod::GEMM_HYBRID, - "smallK_hybrid_fp32_mla_4x8", - [](const GemmArgs &args) { return (args._Ksize <= 8) && (args._Nsize % 4)==0 && !args._trA && args._pretransposed_hint; }, - nullptr, - [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_fp32_mla_4x8, float, float>(args); } + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_fp32_mopa_4VLx1VL", + [](const GemmArgs &args) { return args._ci->has_sme2() && !args._accumulate; }, + [](const GemmArgs &args) { const auto VL = sme::get_vector_length<float>(); + return args._Nsize <= VL || (2*VL < args._Nsize && args._Nsize <= 3*VL); }, + [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_fp32_mopa_4VLx1VL, float, float>(args); } }, +#ifdef ARM_COMPUTE_ENABLE_BF16 { - GemmMethod::GEMM_HYBRID, - "smallK_hybrid_fp32_mla_4x6", - [](const GemmArgs &args) { return (args._Ksize > 8) && (args._Ksize <= 16) && (args._Nsize % 4)==0 && !args._trA && args._pretransposed_hint; }, + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL", + [](const GemmArgs &args) { return args._fast_mode && args._ci->has_sme2() && !args._accumulate; }, nullptr, - [](const GemmArgs &args) { return new GemmHybrid<smallK_hybrid_fp32_mla_4x6, float, float>(args); } + [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_bf16fp32_mopa_2VLx2VL, float, float>(args); } }, +#endif // ARM_COMPUTE_ENABLE_BF16 { - GemmMethod::GEMM_HYBRID, - "hybrid_fp32_mla_4x8_normal", - [](const GemmArgs &args) { return (args._Ksize >= 4) && !args._trA && args._pretransposed_hint; }, - [](const GemmArgs &args) { return (args._Nsize < 12); }, - [](const GemmArgs &args) { return new GemmHybrid<hybrid_fp32_mla_4x8, float, float>(args); } + GemmMethod::GEMM_INTERLEAVED, + "sme2_interleaved_nomerge_fp32_mopa_2VLx2VL", + [](const GemmArgs &args) { return args._ci->has_sme2() && !args._accumulate; }, + nullptr, + [](const GemmArgs &args) { return new GemmInterleavedNoMerge<cls_sme2_interleaved_nomerge_fp32_mopa_2VLx2VL, float, float>(args); } }, -{ +#endif // ARM_COMPUTE_ENABLE_SME2 +#ifdef ARM_COMPUTE_ENABLE_BF16 +GemmImplementation<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); } +), +GemmImplementation<float, float>::with_estimate( GemmMethod::GEMM_HYBRID, - "hybrid_fp32_mla_16x4_normal", - [](const GemmArgs &args) { return (args._Ksize >= 4) && !args._trA && args._pretransposed_hint; }, - [](const GemmArgs &args) { return ((args._Ksize <= 256) && (args._Nsize <= 256)) || (args._Msize < 16) || (args._nmulti > 1); }, - [](const GemmArgs &args) { return new GemmHybrid<hybrid_fp32_mla_16x4, float, float>(args); } + "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); } +), +GemmImplementation<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); } +), +#endif // ARM_COMPUTE_ENABLE_BF16 +#ifdef ARM_COMPUTE_ENABLE_SVEF32MM +// MMLA next due to higher throughput (which is SVE only) +// Prefer this in all cases, except if fast mode is requested and BF16 is available. +{ + 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); } }, +#endif // ARM_COMPUTE_ENABLE_SVEF32MM +// SVE kernels { - GemmMethod::GEMM_NATIVE, - "native_fp32_mla_16x4", - [](const GemmArgs &args) { return (args._Ksize>4 && (args._Nsize % 16)==0 && !args._trA && !args._trB); }, - [](const GemmArgs &args) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, - [](const GemmArgs &args) { return new GemmNative<native_fp32_mla_16x4, float, float>(args); } + GemmMethod::GEMM_HYBRID, + "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); } }, - -#ifdef __ARM_FEATURE_SVE +GemmImplementation<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); } +), +GemmImplementation<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); } +), + #ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +#ifdef ARM_COMPUTE_ENABLE_BF16 +GemmImplementation<float, float>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "sve_ffinterleaved_bf16fp32_mmla_8x3VL", + KernelWeightFormat::VL2VL_BL64_BF16, + [](const GemmArgs &args) { return args._fast_mode && args._ci->has_svebf16(); }, + [](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( + GemmMethod::GEMM_HYBRID, + "sve_ffhybrid_fp32bf16fp32_mmla_4x6VL", + KernelWeightFormat::VL2VL_BL64_BF16, + [](const GemmArgs &args) { return args._fast_mode && args._ci->has_svebf16(); }, + [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_sve_ffhybrid_fp32bf16fp32_mmla_4x6VL, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_sve_ffhybrid_fp32bf16fp32_mmla_4x6VL, float, float>(args); } +), +#endif +GemmImplementation<float, float>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "sve_ffinterleaved_fp32_mla_8x3VL", + KernelWeightFormat::VL1VL_BL32, + [](const GemmArgs &args) { return args._ci->has_sve(); }, + [](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( + GemmMethod::GEMM_HYBRID, + "sve_ffhybrid_fp32_mla_6x4VL", + KernelWeightFormat::VL1VL_BL32, + [](const GemmArgs &args) { return args._ci->has_sve(); }, + [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_sve_ffhybrid_fp32_mla_6x4VL, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_sve_ffhybrid_fp32_mla_6x4VL, float, float>(args); } +), +#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +#endif // ARM_COMPUTE_ENABLE_SVE +// Cortex-A35 specific kernel - use for any problem on A35, and never in any other cases. { GemmMethod::GEMM_INTERLEAVED, - "interleaved_fp32_mla_3VLx8", - [](const GemmArgs &args) { return (args._Ksize>4); }, + "a64_sgemm_8x6", nullptr, - [](const GemmArgs &args) { return new GemmInterleaved<interleaved_fp32_mla_3VLx8, float, float>(args); } + [](const GemmArgs &args) { return args._ci->get_cpu_model() == CPUModel::A35; }, + [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x6, float, float>(args); } }, -#endif // __ARM_FEATURE_SVE -//Pretranpose, 2D split +// Arm® Neon™ hybrid methods { - GemmMethod::GEMM_INTERLEAVED_2D, - "sgemm_12x8_pretranspose_2d", - [](const GemmArgs &args) { return args._pretransposed_hint; }, - [](const GemmArgs &args) { return args._maxthreads >= 8; }, - [](const GemmArgs &args) { return new GemmInterleavedPretransposed2d<sgemm_12x8, float, float>(args); } + GemmMethod::GEMM_HYBRID, + "a64_smallK_hybrid_fp32_mla_8x4", + [](const GemmArgs &args) { return args._Ksize <= 8 && (args._Nsize % 4)==0 && !args._indirect_input && !args._accumulate; }, + nullptr, + [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_fp32_mla_8x4, float, float>(args); } }, -//Tranpose, 2D split, no blockmanager { - GemmMethod::GEMM_INTERLEAVED_2D, - "sgemm_12x8_2d", + GemmMethod::GEMM_HYBRID, + "a64_smallK_hybrid_fp32_mla_6x4", + [](const GemmArgs &args) { return (args._Ksize > 8 && args._Ksize <= 16) && (args._Nsize % 4)==0 && !args._indirect_input && !args._accumulate; }, nullptr, - [](const GemmArgs &args) { return (!args._pretransposed_hint) && (args._maxthreads >= 8); }, - [](const GemmArgs &args) { return new GemmInterleaved2d<sgemm_12x8, float, float>(args); } + [](const GemmArgs &args) { return new GemmHybrid<cls_a64_smallK_hybrid_fp32_mla_6x4, float, float>(args); } }, -//Tranpose, 1D split, with blockmanager { + GemmMethod::GEMM_HYBRID, + "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); } +}, +GemmImplementation<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); } +), +GemmImplementation<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); } +), +GemmImplementation<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); } +), +#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +#ifdef ARM_COMPUTE_ENABLE_BF16 +// "fast mode" (BF16) kernels +GemmImplementation<float, float>::with_estimate( + GemmMethod::GEMM_INTERLEAVED, + "a64_ffinterleaved_bf16fp32_mmla_8x12", + KernelWeightFormat::VL256_BL64_BF16, + [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); }, + [](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( + GemmMethod::GEMM_HYBRID, + "a64_ffhybrid_fp32bf16fp32_mmla_4x24", + KernelWeightFormat::VL256_BL64_BF16, + [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); }, + [](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( + GemmMethod::GEMM_HYBRID, + "a64_ffhybrid_fp32bf16fp32_mmla_6x16", + KernelWeightFormat::VL256_BL64_BF16, + [](const GemmArgs &args) { return args._fast_mode && args._ci->has_bf16(); }, + [](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( GemmMethod::GEMM_INTERLEAVED, - "sgemm_12x8_1d", + "a64_ffinterleaved_fp32_mla_8x12", + KernelWeightFormat::VL128_BL32, nullptr, + [](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( + GemmMethod::GEMM_HYBRID, + "a64_ffhybrid_fp32_mla_6x16", + KernelWeightFormat::VL128_BL32, nullptr, - [](const GemmArgs &args) { return new GemmInterleaved<sgemm_12x8, float, float>(args); } -}, - + [](const GemmArgs &args) { return GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32_mla_6x16, float, float>::estimate_cycles<float>(args); }, + [](const GemmArgs &args) { return new GemmHybridIndirectFixedFormat<cls_a64_ffhybrid_fp32_mla_6x16, float, float>(args); } +), +#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS #endif // __aarch64__ #ifdef __arm__ @@ -199,6 +405,7 @@ const GemmImplementation<float, float> *gemm_implementation_list<float, float>() /* 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 &); |