Update GEMM assembly kernels

- Introduce Fp32 kernels with internal calculations in Bfloat16 when
fast_mode is enabled
- Improve kernel selection heuristics

Signed-off-by: Georgios Pinitas <georgios.pinitas@arm.com>
Change-Id: I68a9e7e862b6fd2721b46e0d7cc791091c4ab279
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5965
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>

1 files changed, 24 insertions, 1 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4.hpp
index b68a5f518a..b7af7110ab 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2020 Arm Limited.
+ * Copyright (c) 2017-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -26,6 +26,7 @@
 #ifdef __aarch64__
 
 #include "../std_transforms_fixed.hpp"
+#include "../performance_parameters.hpp"
 
 namespace arm_gemm {
 
@@ -58,6 +59,28 @@ public:
     StdTransformsFixed<operand_type, result_type, 4, 4, 16> transforms = {};
     StdTransformsFixed<operand_type, result_type, 4, 4, 16, true> transforms_quantized = {};
 
+    template<typename T>
+    static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
+        if (std::is_same<T, int32_t>::value) {
+            switch (ci->get_cpu_model()) {
+                case CPUModel::A510:
+                    return { 3.32, 2.56, 2.63 };
+
+                default:
+                    return { 7.97, 3.72, 7.31 };
+            }
+        }
+
+        if (std::is_same<T, int8_t>::value) {
+            switch(ci->get_cpu_model()) {
+                case CPUModel::A510:
+                    return { 3.33, 2.89, 0.09 };
+                default:
+                    return { 7.97, 3.74, 0.34 };
+            }
+        }
+    }
+
     kern_type kernel=a64_gemm_s8_4x4;
 
     cls_a64_gemm_s8_4x4(const CPUInfo *) { }