diff options
| author | Michalis Spyrou <michalis.spyrou@arm.com> | 2019-11-14 14:31:44 +0000 |
|---|---|---|
| committer | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-01-23 14:57:14 +0000 |
| commit | 71ac9037abce1c6c4af42c485d5395dd6fd79a5a (patch) | |
| tree | 7097d94d7760bf8e172fc4c3725a2eff90bea9a1 /src/core/NEON/kernels/arm_gemm/gemm_fp32.cpp | |
| parent | 19bd412fd044197726dbd8c756dbd74a9e33fd2b (diff) | |
| download | ComputeLibrary-71ac9037abce1c6c4af42c485d5395dd6fd79a5a.tar.gz | |
COMPMID-2923 Integrate arm_gemm per channel quantization
Signed-off-by: Michalis Spyrou <michalis.spyrou@arm.com>
Change-Id: I8667e75843fdd6ac75bd8272a86a348b830da28d
Reviewed-on: https://review.mlplatform.org/c/2548
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 cf91ee0652..7f171ec15a 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) 2017-2019 ARM Limited. + * Copyright (c) 2017-2020 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -33,6 +33,7 @@ #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" @@ -106,9 +107,16 @@ static const GemmImplementation<float, float> gemm_fp32_methods[] = }, { 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_HYBRID, "hybrid_fp32_mla_16x4", [](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 ((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); } }, { |