COMPMID-1380: Pre-work for SVE support.

This patch makes the needed infrastructure changes to allow SVE
kernels to be added later on.

Change-Id: Ide5bccac2f47278e93fff3d648231aee2d5f8c2e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/139070
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>

9 files changed, 138 insertions, 89 deletions

diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6.hpp
index 01bf1f9297..06e62456dc 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6.hpp
@@ -25,6 +25,8 @@
 
 #ifdef __arm__
 
+#include "../std_transforms_fixed.hpp"
+
 namespace arm_gemm {
 
 // Actual kernel implementations
@@ -47,20 +49,21 @@ public:
 
     typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
 
-    /* Describes the data layout for A input */
-    static const int A_interleave = 6;
-    static const int A_block = 1;
-    static const int A_transpose = 0;
+    /* Kernel blocking parameters */
+    static int out_width() {
+        return 8;
+    }
 
-    /* Same for B input */
-    static const int B_interleave = 8;
-    static const int B_block = 1;
-    static const int B_transpose = 1;
+    static int out_height() {
+        return 6;
+    }
 
-    /* Kernel blocking parameters */
-    static const int out_width = 8;
-    static const int out_height = 6;
-    static const int k_unroll = 1;
+    static int k_unroll() {
+        return 1;
+    }
+
+    // Use the standard fixed size transforms.
+    StdTransformsFixed<operand_type, result_type, 6, 8> transforms = {};
 
     kern_type kernel = a32_sgemm_8x6;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_12x8.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_12x8.hpp
index 27700b47d1..95a2bc2fbc 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_12x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s16_12x8.hpp
@@ -25,6 +25,8 @@
 
 #ifdef __aarch64__
 
+#include "../std_transforms_fixed.hpp"
+
 namespace arm_gemm {
 
 // Actual kernel implementations
@@ -45,20 +47,21 @@ public:
 
     typedef void (*kern_type)(const int16_t *, const int16_t *, int32_t *, int, int, int);
 
-    /* Describes the data layout for A input */
-    static const int A_interleave = 8;
-    static const int A_block = 1;
-    static const int A_transpose = 0;
+    /* Kernel blocking parameters */
+    static int out_width() {
+        return 12;
+    }
+
+    static int out_height() {
+        return 8;
+    }
 
-    /* Same for B input */
-    static const int B_interleave = 12;
-    static const int B_block = 1;
-    static const int B_transpose = 1;
+    static int k_unroll() {
+        return 1;
+    }
 
-    /* Kernel blocking parameters */
-    static const int out_width = 12;
-    static const int out_height = 8;
-    static const int k_unroll = 1;
+    // Use the standard fixed size transforms.
+    StdTransformsFixed<operand_type, result_type, 8, 12> transforms = {};
 
     kern_type kernel = a64_gemm_s16_asimd_12x8;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8.hpp
index cb97270c24..fdc0200435 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_12x8.hpp
@@ -27,6 +27,8 @@
 
 #include "arm_gemm.hpp"
 
+#include "../std_transforms_fixed.hpp"
+
 namespace arm_gemm {
 
 // Load the actual kernel
@@ -40,20 +42,21 @@ public:
 
     typedef void (*kern_type)(const int8_t *, const int8_t *, int32_t *, int, int, int);
 
-    /* Describes the data layout for A input */
-    static const int A_interleave = 8;
-    static const int A_block = 4;
-    static const bool A_transpose = false;
+    /* Kernel blocking parameters */
+    static int out_width() {
+        return 12;
+    }
 
-    /* Same for B input */
-    static const int B_interleave = 12;
-    static const int B_block = 4;
-    static const bool B_transpose = true;
+    static int out_height() {
+        return 8;
+    }
 
-    /* Kernel blocking parameters */
-    static const int out_width = 12;
-    static const int out_height = 8;
-    static const int k_unroll = 4;
+    static int k_unroll() {
+        return 4;
+    }
+
+    // Use the standard fixed size transforms.
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4> transforms = {};
 
     kern_type kernel = a64_gemm_s8_12x8;
 
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 b5b07b2c56..be7ead9f48 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
@@ -25,6 +25,8 @@
 
 #ifdef __aarch64__
 
+#include "../std_transforms_fixed.hpp"
+
 namespace arm_gemm {
 
 // Load the actual kernel
@@ -39,20 +41,21 @@ public:
 
     typedef void (*kern_type)(const int8_t *, const int8_t *, int32_t *, int, int, int);
 
-    /* Describes the data layout for A input */
-    static const int A_interleave = 4;
-    static const int A_block = 16;
-    static const bool A_transpose = false;
+    /* Kernel blocking parameters */
+    static int out_width() {
+        return 4;
+    }
+
+    static int out_height() {
+        return 4;
+    }
 
-    /* Same for B input */
-    static const int B_interleave = 4;
-    static const int B_block = 16;
-    static const bool B_transpose = true;
+    static int k_unroll() {
+        return 16;
+    }
 
-    /* Kernel blocking parameters */
-    static const int out_width = 4;
-    static const int out_height = 4;
-    static const int k_unroll = 16;
+    // Use the standard fixed size transforms.
+    StdTransformsFixed<operand_type, result_type, 4, 4, 16> transforms = {};
 
     kern_type kernel=a64_gemm_s8_4x4;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_12x8.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_12x8.hpp
index 13dd570677..d2692ba77f 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_12x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u16_12x8.hpp
@@ -25,6 +25,8 @@
 
 #ifdef __aarch64__
 
+#include "../std_transforms_fixed.hpp"
+
 namespace arm_gemm {
 
 // Actual kernel implementations
@@ -45,20 +47,21 @@ public:
 
     typedef void (*kern_type)(const uint16_t *, const uint16_t *, uint32_t *, int, int, int);
 
-    /* Describes the data layout for A input */
-    static const int A_interleave = 8;
-    static const int A_block = 1;
-    static const int A_transpose = 0;
+    /* Kernel blocking parameters */
+    static int out_width() {
+        return 12;
+    }
+
+    static int out_height() {
+        return 8;
+    }
 
-    /* Same for B input */
-    static const int B_interleave = 12;
-    static const int B_block = 1;
-    static const int B_transpose = 1;
+    static int k_unroll() {
+        return 1;
+    }
 
-    /* Kernel blocking parameters */
-    static const int out_width = 12;
-    static const int out_height = 8;
-    static const int k_unroll = 1;
+    // Use the standard fixed size transforms.
+    StdTransformsFixed<operand_type, result_type, 8, 12> transforms = {};
 
     kern_type kernel = a64_gemm_u16_asimd_12x8;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8.hpp
index c67aed7275..a252abfd3e 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_12x8.hpp
@@ -27,6 +27,8 @@
 
 #include "arm_gemm.hpp"
 
+#include "../std_transforms_fixed.hpp"
+
 namespace arm_gemm {
 
 // Load the actual kernel
@@ -51,9 +53,20 @@ public:
     static const bool B_transpose = true;
 
     /* Kernel blocking parameters */
-    static const int out_width = 12;
-    static const int out_height = 8;
-    static const int k_unroll = 4;
+    static int out_width() {
+        return 12;
+    }
+
+    static int out_height() {
+        return 8;
+    }
+
+    static int k_unroll() {
+        return 4;
+    }
+
+    // Use the standard fixed sized transforms.
+    StdTransformsFixed<operand_type, result_type, 8, 12, 4> transforms = {};
 
     kern_type kernel = a64_gemm_u8_12x8;
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp
index 23f4c1d84f..2da3ecd4f8 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_4x4.hpp
@@ -25,6 +25,8 @@
 
 #ifdef __aarch64__
 
+#include "../std_transforms_fixed.hpp"
+
 namespace arm_gemm {
 
 // Kernel definition
@@ -48,14 +50,24 @@ public:
     static const bool B_transpose = true;
 
     /* Kernel blocking parameters */
-    static const int out_width = 4;
-    static const int out_height = 4;
-    static const int k_unroll = 16;
+    static int out_width() {
+        return 4;
+    }
+
+    static int out_height() {
+        return 4;
+    }
+
+    static int k_unroll() {
+        return 16;
+    }
+
+    // Use the standard fixed size transforms.
+    StdTransformsFixed<operand_type, result_type, 4, 4, 16> transforms = {};
 
-    kern_type kernel = nullptr;
+    kern_type kernel = a64_gemm_u8_4x4;
 
     gemm_u8_4x4(const CPUInfo *ci) {
-        kernel = a64_gemm_u8_4x4;
     }
 };
 
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_24x8.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_24x8.hpp
index fe74b994f5..911a4ebb01 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_24x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_24x8.hpp
@@ -27,6 +27,8 @@
 
 #include "arm_gemm.hpp"
 
+#include "../std_transforms_fixed.hpp"
+
 namespace arm_gemm {
 
 // Actual kernel implementations
@@ -44,17 +46,21 @@ public:
 
     typedef void (*kern_type)(const __fp16 *, const __fp16 *, __fp16 *, int, int, int);
 
-    static const int A_block = 1;
-    static const int A_interleave = 8;
-    static const bool A_transpose = false;
+    /* Kernel blocking parameters */
+    static int out_width() {
+        return 24;
+    }
 
-    static const int B_block = 1;
-    static const int B_interleave = 24;
-    static const bool B_transpose = true;
+    static int out_height() {
+        return 8;
+    }
+
+    static int k_unroll() {
+        return 1;
+    }
 
-    static const int out_width = 24;
-    static const int out_height = 8;
-    static const int k_unroll = 1;
+    // Use the standard fixed size transforms.
+    StdTransformsFixed<operand_type, result_type, 8, 24> transforms = {};
 
     // Default to the generic kernel
     kern_type kernel = a64_hgemm_asimd_24x8;
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8.hpp b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8.hpp
index c91d50469f..10d1069417 100644
--- a/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8.hpp
+++ b/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_12x8.hpp
@@ -25,6 +25,8 @@
 
 #ifdef __aarch64__
 
+#include "../std_transforms_fixed.hpp"
+
 namespace arm_gemm {
 
 // Actual kernel implementations
@@ -48,20 +50,21 @@ public:
 
     typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
 
-    /* Describes the data layout for A input */
-    static const int A_interleave = 8;
-    static const int A_block = 1;
-    static const int A_transpose = 0;
+    /* Kernel blocking parameters */
+    static int out_width() {
+        return 12;
+    }
 
-    /* Same for B input */
-    static const int B_interleave = 12;
-    static const int B_block = 1;
-    static const int B_transpose = 1;
+    static int out_height() {
+        return 8;
+    }
 
-    /* Kernel blocking parameters */
-    static const int out_width = 12;
-    static const int out_height = 8;
-    static const int k_unroll = 1;
+    static int k_unroll() {
+        return 1;
+    }
+
+    // Use the standard fixed size transforms.
+    StdTransformsFixed<operand_type, result_type, 8, 12> transforms = {};
 
     kern_type kernel=a64_sgemm_asimd_12x8;