diff options
| author | Georgios Pinitas <georgios.pinitas@arm.com> | 2020-11-17 18:46:40 +0000 |
|---|---|---|
| committer | Michele Di Giorgio <michele.digiorgio@arm.com> | 2020-11-23 13:50:40 +0000 |
| commit | 573c7bbb15c8f205e93397105e66fbc5102882e4 (patch) | |
| tree | 00f34dfe387f3a264e6eb74d0665b2190ab4d3bb /src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp | |
| parent | 34c0c8a7d841111a0147ef84e29627f3e610d273 (diff) | |
| download | ComputeLibrary-573c7bbb15c8f205e93397105e66fbc5102882e4.tar.gz | |
Update tuning numbers for A55 for both fp16 and fp32
Resolves: COMPMID-3974
Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I6d5189e44ebeda1575a80dd14ec3a09c75f68e03
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4521
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
(cherry picked from commit 40943df83026b66356f24e30f31f78c8b9e59c92)
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/4169
Tested-by: Michele Di Giorgio <michele.digiorgio@arm.com>
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp | 14 |
1 files changed, 7 insertions, 7 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp index de2e4f2c2b..75524fff97 100644 --- a/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp +++ b/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp @@ -62,7 +62,7 @@ static const GemmImplementation<__fp16, __fp16> gemm_fp16_methods[] = { #endif #if defined(__aarch64__) && (defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) || defined(FP16_KERNELS)) -{ +GemmImplementation<__fp16, __fp16>::with_estimate( GemmMethod::GEMM_HYBRID, "a64_hybrid_fp16_mla_6x32", #ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC @@ -70,10 +70,10 @@ static const GemmImplementation<__fp16, __fp16> gemm_fp16_methods[] = { #else nullptr, #endif - [](const GemmArgs &args) { return ((args._Ksize <= 256) && (args._Nsize <= 256)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); }, + [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp16_mla_6x32, __fp16, __fp16>::estimate_cycles(args, cls_a64_hybrid_fp16_mla_6x32::get_performance_parameters(args._ci)); }, [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp16_mla_6x32, __fp16, __fp16>(args); } -}, -{ +), +GemmImplementation<__fp16, __fp16>::with_estimate( GemmMethod::GEMM_INTERLEAVED, "a64_hgemm_8x24", #ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC @@ -81,16 +81,16 @@ static const GemmImplementation<__fp16, __fp16> gemm_fp16_methods[] = { #else nullptr, #endif - nullptr, + [](const GemmArgs &args) { return GemmInterleaved<cls_a64_hgemm_8x24, __fp16, __fp16>::estimate_cycles(args, cls_a64_hgemm_8x24::get_performance_parameters(args._ci)); }, [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_hgemm_8x24, __fp16, __fp16>(args); } -}, +), #endif // aarch64 && FP16 #ifdef __aarch64__ { GemmMethod::GEMM_INTERLEAVED, "a64_sgemm_8x12", nullptr, - nullptr, + [](const GemmArgs &args) { return !args._ci->has_fp16(); }, [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, __fp16, __fp16>(args); } }, #elif defined(__arm__) |