commit 40943df83026b66356f24e30f31f78c8b9e59c92
author Georgios Pinitas <georgios.pinitas@arm.com> Tue Nov 17 18:46:40 2020 +0000
committer Georgios Pinitas <georgios.pinitas@arm.com> Mon Nov 23 12:43:24 2020 +0000
tree 2aeaa06ab7571a16ba1551bb9110df62b533c0d2
parent 71696baa30fd2cb9ac7493937e65abfb16daa16e

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>

src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp

diff --git a/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp b/src/core/NEON/kernels/arm_gemm/gemm_fp16.cpp
index de2e4f2..75524ff 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 @@
 #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 @@
 #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 @@
 #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__)