1 files changed, 50 insertions, 18 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL.hpp
index b84ba83b6a..23ab7ce10a 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_fp32_mla_8x3VL.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,51 +22,83 @@
  * SOFTWARE.
  */
 #pragma once
-
-#ifdef __ARM_FEATURE_SVE
-
+#ifdef ARM_COMPUTE_ENABLE_SVE
 
 #include "../std_transforms_sve.hpp"
+#include "../performance_parameters.hpp"
 
-namespace arm_gemm {
+#define ARGLIST  \
+    const float *, const float *, \
+    float *, int, int, int
 
+namespace arm_gemm
+{
 // Actual kernel implementations
-void sve_interleaved_fp32_mla_8x3VL(const float *, const float *, float *, int, int, int);
+void sve_interleaved_fp32_mla_8x3VL( ARGLIST );
+void sve_interleaved_fp32_mla_8x3VL_a64fx( ARGLIST );
 
-class cls_sve_interleaved_fp32_mla_8x3VL {
+class cls_sve_interleaved_fp32_mla_8x3VL
+{
 public:
     typedef float operand_type;
     typedef float result_type;
 
-    typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
+    typedef void (*kern_type)( ARGLIST );
 
     /* Kernel blocking parameters */
-    static unsigned int out_width()
+    static constexpr unsigned int out_height()
     {
-        return get_vector_length<float>() * 3;
+        return 8;
     }
 
-    static unsigned int out_height()
+    static unsigned int out_width()
     {
-        return 8;
+        return get_vector_length<float>() * 3;
     }
 
-    static unsigned int k_unroll()
+    static constexpr unsigned int k_unroll()
     {
         return 1;
     }
 
-    // Use the standard fixed size transforms.
+
     StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 3, 1, 1, true> transforms_quantized = {};
+    template<typename T>
+    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
+    {
 
-    kern_type kernel=sve_interleaved_fp32_mla_8x3VL;
+        if (std::is_same<T, float>::value) {
+            switch (ci->get_cpu_model()) {
+                default:
+                    return { 7.2307, 3.876, 2.932 };
+                case CPUModel::A64FX:
+                    return { 26.52, 3.42, 4.59 };
+                case CPUModel::A510:
+                    return { 6.25, 3.84, 2.47 };
+                case CPUModel::V1:
+                    return { 15.15, 9.24, 6.42 };
+            }
+        }
+
+        return { 1.0 };
+    }
 
-    cls_sve_interleaved_fp32_mla_8x3VL(const CPUInfo *)
+    // Default to the generic kernel
+    kern_type kernel=sve_interleaved_fp32_mla_8x3VL;
+    cls_sve_interleaved_fp32_mla_8x3VL(const CPUInfo *ci)
     {
-
+        switch(ci->get_cpu_model()) {
+            default:
+                break;
+            case CPUModel::A64FX:
+                kernel=sve_interleaved_fp32_mla_8x3VL_a64fx;
+                break;
+        }
     }
 };
 
 } // namespace arm_gemm
 
-#endif // __ARM_FEATURE_SVE
+#undef ARGLIST
+#endif // ARM_COMPUTE_ENABLE_SVE