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, 32 insertions, 4 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL.hpp
index f98ccdc7d3..6db9c0cdf3 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL.hpp
@@ -25,6 +25,7 @@
 #ifdef ARM_COMPUTE_ENABLE_SVE
 
 #include "../std_transforms_sve.hpp"
+#include "../performance_parameters.hpp"
 
 #define ARGLIST  \
     unsigned int, const unsigned int *, \
@@ -38,11 +39,13 @@ namespace arm_gemm
 {
 // Actual kernel implementations
 void sve_hybrid_fp16_mla_6x4VL( ARGLIST );
+void sve_hybrid_fp16_mla_6x4VL_a64fx( ARGLIST );
 
 class cls_sve_hybrid_fp16_mla_6x4VL
 {
 public:
-    typedef __fp16 operand_type;
+    typedef __fp16 lhs_operand_type;
+    typedef __fp16 rhs_operand_type;
     typedef __fp16 result_type;
 
     typedef void (*kern_type)( ARGLIST );
@@ -68,16 +71,41 @@ public:
         return true;
     }
 
-    StdTransformsSVE<operand_type, result_type, 6, 4, 1> transforms = {};
+    StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 1> transforms = {};
+    template<typename T>
+    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
+    {
+
+        if (std::is_same<T, __fp16>::value) {
+            switch (ci->get_cpu_model()) {
+                default:
+                    return { 15.90 };
+                case CPUModel::A510:
+                    return { 12.44 };
+                case CPUModel::V1:
+                    return { 31.51 };
+            }
+        }
+
+        return { 1.0 };
+    }
 
     // Default to the generic kernel
     kern_type kernel=sve_hybrid_fp16_mla_6x4VL;
-    cls_sve_hybrid_fp16_mla_6x4VL(const CPUInfo *)
+    cls_sve_hybrid_fp16_mla_6x4VL(const CPUInfo *ci)
     {
+        switch(ci->get_cpu_model()) {
+            default:
+                break;
+            case CPUModel::A64FX:
+                kernel=sve_hybrid_fp16_mla_6x4VL_a64fx;
+                break;
+        }
     }
 };
 
 } // namespace arm_gemm
 
 #undef ARGLIST
-#endif // ARM_COMPUTE_ENABLE_SVE
+
+#endif // __aarch64__