diff options
| author | Michalis Spyrou <michalis.spyrou@arm.com> | 2021-04-20 12:15:52 +0100 |
|---|---|---|
| committer | Michalis Spyrou <michalis.spyrou@arm.com> | 2021-04-20 17:10:48 +0000 |
| commit | 778b95cb755880ab4adc972fbf3b7022a99b63f9 (patch) | |
| tree | 13a4f6060e20a0989c558d7750386b8f32827e4b /src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp | |
| parent | dcf4c87cf78a5f1667699c1a3511d09356938660 (diff) | |
| download | ComputeLibrary-778b95cb755880ab4adc972fbf3b7022a99b63f9.tar.gz | |
Update assembly code
This patch brings performance uplift on Cortex-A35.
Resolves: COMPMID-4316
Change-Id: I2b9c02a599373f780dd1b981b821e33bd59a3422
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5461
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp | 12 |
1 files changed, 10 insertions, 2 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp index 6c3743dce7..d94814fb4c 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) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -35,6 +35,7 @@ #include "kernels/a64_hybrid_fp32_mla_6x16.hpp" #include "kernels/a64_hybrid_fp32_mla_8x4.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" @@ -110,7 +111,14 @@ static const GemmImplementation<float, float> gemm_fp32_methods[] = [](const GemmArgs &args) { return new GemmHybridIndirect<cls_sve_hybrid_fp32_mla_6x4VL, float, float>(args); } }, #endif // __ARM_FEATURE_SVE - +// Cortex-A35 specific kernel - use for any problem on A35, and never in any other cases. +{ + GemmMethod::GEMM_INTERLEAVED, + "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); } +}, // Arm® Neon™ hybrid methods { GemmMethod::GEMM_HYBRID, |