1 files changed, 52 insertions, 17 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp
index 2889dd7f0f..c5096ff4ba 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/sve_interleaved_bf16fp32_mmla_8x3VL.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,51 +22,86 @@
  * SOFTWARE.
  */
 #pragma once
+#ifdef ARM_COMPUTE_ENABLE_SVE
 
-#ifdef __ARM_FEATURE_SVE
-
-#include "../bfloat.hpp"
 #include "../std_transforms_sve.hpp"
+#include "../bfloat.hpp"
+#include "../performance_parameters.hpp"
 
-namespace arm_gemm {
+#define ARGLIST  \
+    const bfloat16 *, const bfloat16 *, \
+    float *, int, int, int
 
+namespace arm_gemm
+{
 // Actual kernel implementations
-void sve_interleaved_bf16fp32_mmla_8x3VL(const bfloat16 *, const bfloat16 *, float *, int, int, int);
+void sve_interleaved_bf16fp32_mmla_8x3VL( ARGLIST );
 
-class cls_sve_interleaved_bf16fp32_mmla_8x3VL {
+class cls_sve_interleaved_bf16fp32_mmla_8x3VL
+{
 public:
     typedef bfloat16 operand_type;
     typedef float result_type;
 
-    typedef void (*kern_type)(const bfloat16 *, const bfloat16 *, 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 4;
     }
 
-    // Use the standard fixed size transforms.
+
     StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2> transforms = {};
+    StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2, true> transforms_quantized = {};
+    template<typename T>
+    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
+    {
 
-    kern_type kernel=sve_interleaved_bf16fp32_mmla_8x3VL;
+        if (std::is_same<T, bfloat16>::value) {
+            switch (ci->get_cpu_model()) {
+                default:
+                    return { 31.41, 4.30, 7.14 };
+                case CPUModel::A510:
+                    return { 7.78, 4.01, 2.43 };
+                case CPUModel::V1:
+                    return { 62.50, 5.09, 11.32 };
+            }
+        }
+
+
+        if (std::is_same<T, float>::value) {
+            switch (ci->get_cpu_model()) {
+                default:
+                    return { 30.86, 2.36, 5.28 };
+                case CPUModel::A510:
+                    return { 7.75, 2.47, 2.39 };
+                case CPUModel::V1:
+                    return { 47.63, 5.11, 6.80 };
+            }
+        }
 
+        return { 1.0 };
+    }
+
+    // Default to the generic kernel
+    kern_type kernel=sve_interleaved_bf16fp32_mmla_8x3VL;
     cls_sve_interleaved_bf16fp32_mmla_8x3VL(const CPUInfo *)
     {
-
     }
 };
 
 } // namespace arm_gemm
 
-#endif // __ARM_FEATURE_SVE
+#undef ARGLIST
+#endif // ARM_COMPUTE_ENABLE_SVE