aarch64: matmul: addition of JIT int8 kernel

Author: Shreyas-fuj

include/oneapi/dnnl/dnnl.hpp

@@ -1,6 +1,6 @@
 /*******************************************************************************
 * Copyright 2016-2025 Intel Corporation
-* Copyright 2024 FUJITSU LIMITED
+* Copyright 2024-2025 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -1615,6 +1615,9 @@ struct memory : public handle<dnnl_memory_t> {
         BA16a32b4a = dnnl_BA16a32b4a,
         BA16a48b4a = dnnl_BA16a48b4a,
         BA16a64b4a = dnnl_BA16a64b4a,
+        BA24b8a = dnnl_BA24b8a,
+        aCB24c8b = dnnl_aCB24c8b,
+        abDC24d8c = dnnl_abDC24d8c,
         decbA16a = dnnl_decbA16a,
         decbA8a = dnnl_decbA8a,
         defcbA16a = dnnl_defcbA16a,

include/oneapi/dnnl/dnnl_types.h

@@ -1,6 +1,6 @@
 /*******************************************************************************
 * Copyright 2016-2025 Intel Corporation
-* Copyright 2024 FUJITSU LIMITED
+* Copyright 2024-2025 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -1047,6 +1047,9 @@ typedef enum {
     dnnl_aCBdef8b8c,
     dnnl_abdEC16e4c,
     dnnl_abDC16d4c,
+    dnnl_BA24b8a,
+    dnnl_aCB24c8b,
+    dnnl_abDC24d8c,
 
     /// Just a sentinel, not real memory format tag. Must be changed after new
     /// format tag is added.

src/common/c_types_map.hpp

@@ -1,6 +1,6 @@
 /*******************************************************************************
 * Copyright 2016-2025 Intel Corporation
-* Copyright 2024 FUJITSU LIMITED
+* Copyright 2024-2025 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -372,6 +372,9 @@ const format_tag_t aCB16b16c = dnnl_aCB16b16c;
 const format_tag_t aCB16b32c = dnnl_aCB16b32c;
 const format_tag_t aCB16b48c = dnnl_aCB16b48c;
 const format_tag_t aCB16b64c = dnnl_aCB16b64c;
+const format_tag_t BA24b8a = dnnl_BA24b8a;
+const format_tag_t aCB24c8b = dnnl_aCB24c8b;
+const format_tag_t abDC24d8c = dnnl_abDC24d8c;
 const format_tag_t aCB16b16c2b = dnnl_aCB16b16c2b;
 const format_tag_t aCB16b32c2b = dnnl_aCB16b32c2b;
 const format_tag_t aCB16b48c2b = dnnl_aCB16b48c2b;

src/common/dnnl_debug_autogenerated.cpp

@@ -1,6 +1,6 @@
 /*******************************************************************************
 * Copyright 2018-2025 Intel Corporation
-* Copyright 2024 FUJITSU LIMITED
+* Copyright 2024-2025 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -621,6 +621,9 @@ const char *dnnl_fmt_tag2str(dnnl_format_tag_t v) {
     if (v == dnnl_aBC16b32c) return "aBC16b32c";
     if (v == dnnl_AB16a16b) return "AB16a16b";
     if (v == dnnl_AB16a32b) return "AB16a32b";
+    if (v == dnnl_BA24b8a) return "BA24b8a";
+    if (v == dnnl_aCB24c8b) return "aCB24c8b";
+    if (v == dnnl_abDC24d8c) return "abDC24d8c";
     if (v == dnnl_ABcde16a16b2a) return "ABcde16a16b2a";
     if (v == dnnl_aBCdef16b16c2b) return "aBCdef16b16c2b";
     if (v == dnnl_Acedb16a) return "Acedb16a";

src/common/memory_desc_wrapper.cpp

@@ -1,6 +1,6 @@
 /*******************************************************************************
 * Copyright 2016-2025 Intel Corporation
-* Copyright 2024 FUJITSU LIMITED
+* Copyright 2024-2025 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -202,6 +202,9 @@ status_t memory_desc_wrapper::compute_blocking(
         C(BA16a32b, {1, 0}, {16, 32}, {0, 1});
         C(BA16a48b, {1, 0}, {16, 48}, {0, 1});
         C(BA16a64b, {1, 0}, {16, 64}, {0, 1});
+        C(BA24b8a, {1, 0}, {24, 8}, {1, 0});
+        C(aCB24c8b, {0, 2, 1}, {24, 8}, {2, 1});
+        C(abDC24d8c, {0, 1, 3, 2}, {24, 8}, {3, 2});
         C(BA16a16b2a, {1, 0}, {16, 16, 2}, {0, 1, 0});
         C(BA16a32b2a, {1, 0}, {16, 32, 2}, {0, 1, 0});
         C(BA16a48b2a, {1, 0}, {16, 48, 2}, {0, 1, 0});