1 files changed, 59 insertions, 16 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp
index b2c2407b28..66c2b92a34 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_interleaved_bf16fp32_mmla_8x12.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -22,51 +22,94 @@
  * SOFTWARE.
  */
 #pragma once
-
 #ifdef __aarch64__
 
-#include "../bfloat.hpp"
 #include "../std_transforms_fixed.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 a64_interleaved_bf16fp32_mmla_8x12(const bfloat16 *, const bfloat16 *, float *, int, int, int);
+void a64_interleaved_bf16fp32_mmla_8x12( ARGLIST );
+void a64_interleaved_bf16fp32_mmla_8x12_a510( ARGLIST );
 
-class cls_a64_interleaved_bf16fp32_mmla_8x12 {
+class cls_a64_interleaved_bf16fp32_mmla_8x12
+{
 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 12;
+        return 8;
     }
 
-    static unsigned int out_height()
+    static unsigned int out_width()
     {
-        return 8;
+        return 12;
     }
 
-    static unsigned int k_unroll()
+    static constexpr unsigned int k_unroll()
     {
         return 4;
     }
 
-    // Use the standard fixed size transforms.
+
     StdTransformsFixed<operand_type, result_type, 8, 12, 4> transforms = {};
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
+    template<typename T>
+    static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
+    {
 
-    kern_type kernel=a64_interleaved_bf16fp32_mmla_8x12;
+        if (std::is_same<T, bfloat16>::value) {
+            switch (ci->get_cpu_model()) {
+                default:
+                    return { 31.54, 4.30, 7.33 };
+                case CPUModel::V1:
+                    return { 59.94, 5.08, 9.83 };
+                case CPUModel::A510:
+                    return { 7.82, 4.05, 3.07 };
+            }
+        }
 
-    cls_a64_interleaved_bf16fp32_mmla_8x12(const CPUInfo *)
-    {
 
+        if (std::is_same<T, float>::value) {
+            switch (ci->get_cpu_model()) {
+                default:
+                    return { 31.15, 2.51, 5.25 };
+                case CPUModel::V1:
+                    return { 41.44, 5.01, 5.64 };
+                case CPUModel::A510:
+                    return { 7.83, 2.53, 2.71 };
+            }
+        }
+
+        return { 1.0 };
+    }
+
+    // Default to the generic kernel
+    kern_type kernel=a64_interleaved_bf16fp32_mmla_8x12;
+    cls_a64_interleaved_bf16fp32_mmla_8x12(const CPUInfo *ci)
+    {
+        switch(ci->get_cpu_model()) {
+            default:
+                break;
+            case CPUModel::A510:
+                kernel=a64_interleaved_bf16fp32_mmla_8x12_a510;
+                break;
+        }
     }
 };
 
 } // namespace arm_gemm
 
+#undef ARGLIST
 #endif // __aarch64__