cpu: aarch64: Expand ARM SVE support for 1x1 convolution

include/oneapi/dnnl/dnnl.hpp

@@ -1,5 +1,6 @@
 /*******************************************************************************
 * Copyright 2016-2024 Intel Corporation
+* Copyright 2024 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -1412,6 +1413,7 @@ struct memory : public handle<dnnl_memory_t> {
         aBdeC8b4c = dnnl_aBdeC8b4c,
         aBdefc16b = dnnl_aBdefc16b,
         aCBdef16c16b = dnnl_aCBdef16c16b,
+        aCBdef8b8c = dnnl_aCBdef8b8c,
         aCBdef16b16c = dnnl_aCBdef16b16c,
         aBdefc4b = dnnl_aBdefc4b,
         aBdefc8b = dnnl_aBdefc8b,
@@ -1422,8 +1424,10 @@ struct memory : public handle<dnnl_memory_t> {
         Acb8a = dnnl_Acb8a,
         AcB8a2b = dnnl_AcB8a2b,
         AcB8a4b = dnnl_AcB8a4b,
+        aCBd8b8c = dnnl_aCBd8b8c,
         aCBd16b16c = dnnl_aCBd16b16c,
         aCBd16c16b = dnnl_aCBd16c16b,
+        aCBde8b8c = dnnl_aCBde8b8c,
         aCBde16b16c = dnnl_aCBde16b16c,
         aCBde16c16b = dnnl_aCBde16c16b,
         Acdb16a = dnnl_Acdb16a,
@@ -1436,12 +1440,15 @@ struct memory : public handle<dnnl_memory_t> {
         Acdeb8a = dnnl_Acdeb8a,
         AcdeB8a2b = dnnl_AcdeB8a2b,
         AcdeB8a4b = dnnl_AcdeB8a4b,
+        BAc8a8b = dnnl_BAc8a8b,
         BAc16a16b = dnnl_BAc16a16b,
         BAc16b16a = dnnl_BAc16b16a,
+        BAcd8a8b = dnnl_BAcd8a8b,
         BAcd16a16b = dnnl_BAcd16a16b,
         BAcd16b16a = dnnl_BAcd16b16a,
         ABcd32a32b = dnnl_ABcd32a32b,
         BAcde16b16a = dnnl_BAcde16b16a,
+        BAcde8a8b = dnnl_BAcde8a8b,
         BAcde16a16b = dnnl_BAcde16a16b,
         aBdec32b = dnnl_aBdec32b,
         Abcdef16a = dnnl_Abcdef16a,
@@ -1692,6 +1699,7 @@ struct memory : public handle<dnnl_memory_t> {
         gIOhw16i16o = dnnl_gIOhw16i16o,
         gOhwi32o = dnnl_gOhwi32o,
         Goidhw16g = dnnl_Goidhw16g,
+        IOw8o8i = dnnl_IOw8o8i,
         IOw16o16i = dnnl_IOw16o16i,
         OIw16i16o = dnnl_OIw16i16o,
         OwI16i16o = dnnl_OwI16i16o,
@@ -1748,6 +1756,7 @@ struct memory : public handle<dnnl_memory_t> {
         Owi8o = dnnl_Owi8o,
         OwI8o2i = dnnl_OwI8o2i,
         OwI8o4i = dnnl_OwI8o4i,
+        IOhw8o8i = dnnl_IOhw8o8i,
         IOhw16o16i = dnnl_IOhw16o16i,
         Ohwi16o = dnnl_Ohwi16o,
         OhwI16o2i = dnnl_OhwI16o2i,
@@ -1793,6 +1802,7 @@ struct memory : public handle<dnnl_memory_t> {
         OIhw8o8i = dnnl_OIhw8o8i,
         OIhw8o4i = dnnl_OIhw8o4i,
         OIhw2i8o4i = dnnl_OIhw2i8o4i,
+        IOdhw8o8i = dnnl_IOdhw8o8i,
         IOdhw16o16i = dnnl_IOdhw16o16i,
         Odhwi16o = dnnl_Odhwi16o,
         OdhwI16o2i = dnnl_OdhwI16o2i,
@@ -1846,6 +1856,7 @@ struct memory : public handle<dnnl_memory_t> {
         OdhwI8i8o = dnnl_OdhwI8i8o,
         OIdhw8o8i = dnnl_OIdhw8o8i,
         OIdhw8o4i = dnnl_OIdhw8o4i,
+        gIOw8o8i = dnnl_gIOw8o8i,
         gIOw16o16i = dnnl_gIOw16o16i,
         gOIw16i16o = dnnl_gOIw16i16o,
         gOIw16o16i = dnnl_gOIw16o16i,
@@ -1874,6 +1885,7 @@ struct memory : public handle<dnnl_memory_t> {
         gOwI8o4i = dnnl_gOwI8o4i,
         Goiw8g = dnnl_Goiw8g,
         Goiw16g = dnnl_Goiw16g,
+        gIOhw8o8i = dnnl_gIOhw8o8i,
         gIOhw16o16i = dnnl_gIOhw16o16i,
         gOhwi16o = dnnl_gOhwi16o,
         gOhwI16o2i = dnnl_gOhwI16o2i,
@@ -1920,6 +1932,7 @@ struct memory : public handle<dnnl_memory_t> {
         gOIhw8o8i = dnnl_gOIhw8o8i,
         gOIhw8o4i = dnnl_gOIhw8o4i,
         gIOdhw16i16o = dnnl_gIOdhw16i16o,
+        gIOdhw8o8i = dnnl_gIOdhw8o8i,
         gIOdhw16o16i = dnnl_gIOdhw16o16i,
         gOdhwi16o = dnnl_gOdhwi16o,
         gOdhwI16o2i = dnnl_gOdhwI16o2i,

include/oneapi/dnnl/dnnl_types.h

@@ -1,5 +1,6 @@
 /*******************************************************************************
 * Copyright 2016-2024 Intel Corporation
+* Copyright 2024 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -1038,6 +1039,12 @@ typedef enum {
     dnnl_cabd,
     dnnl_dabc,
     dnnl_Ab32a,
+    dnnl_aCBd8b8c,
+    dnnl_aCBde8b8c,
+    dnnl_BAc8a8b,
+    dnnl_BAcd8a8b,
+    dnnl_BAcde8a8b,
+    dnnl_aCBdef8b8c,
 
     /// Just a sentinel, not real memory format tag. Must be changed after new
     /// format tag is added.
@@ -1258,6 +1265,7 @@ typedef enum {
     dnnl_OI8i8o = dnnl_AB8b8a,
 
     // weights, 3D
+    dnnl_IOw8o8i = dnnl_BAc8a8b,
     dnnl_IOw16o16i = dnnl_BAc16a16b,
     dnnl_IOw16i16o = dnnl_BAc16b16a,
     dnnl_OIw16i16o = dnnl_ABc16b16a,
@@ -1328,6 +1336,7 @@ typedef enum {
 
     // weights, 4D
     dnnl_IOhw16i16o = dnnl_BAcd16b16a,
+    dnnl_IOhw8o8i = dnnl_BAcd8a8b,
     dnnl_IOhw16o16i = dnnl_BAcd16a16b,
     dnnl_Ohwi16o = dnnl_Acdb16a,
     dnnl_OhwI16o2i = dnnl_AcdB16a2b,
@@ -1460,6 +1469,7 @@ typedef enum {
     dnnl_OIdhw8o4i = dnnl_ABcde8a4b,
     dnnl_IOdhw16i16o = dnnl_BAcde16b16a,
     dnnl_OIdhw4o8i8o4i = dnnl_ABcde4a8b8a4b,
+    dnnl_IOdhw8o8i = dnnl_BAcde8a8b,
     dnnl_IOdhw16o16i = dnnl_BAcde16a16b,
     dnnl_OIdhw16o16i2o = dnnl_ABcde16a16b2a,
     dnnl_OIdhw8i32o = dnnl_ABcde8b32a,
@@ -1473,6 +1483,7 @@ typedef enum {
     dnnl_Goiw16g = dnnl_Abcd16a,
     dnnl_Goiw8g = dnnl_Abcd8a,
     dnnl_Goiw4g = dnnl_Abcd4a,
+    dnnl_gIOw8o8i = dnnl_aCBd8b8c,
     dnnl_gIOw16o16i = dnnl_aCBd16b16c,
     dnnl_gIOw16i16o = dnnl_aCBd16c16b,
     dnnl_gOIw16i16o = dnnl_aBCd16c16b,
@@ -1518,6 +1529,7 @@ typedef enum {
 
     // weights w/ groups, 4D
     dnnl_gIOhw16i16o = dnnl_aCBde16c16b,
+    dnnl_gIOhw8o8i = dnnl_aCBde8b8c,
     dnnl_gIOhw16o16i = dnnl_aCBde16b16c,
     dnnl_gOhwi16o = dnnl_aBdec16b,
     dnnl_gOhwI16o2i = dnnl_aBdeC16b2c,
@@ -1585,6 +1597,7 @@ typedef enum {
 
     // weights w/ groups, 6D
     dnnl_gIOdhw16i16o = dnnl_aCBdef16c16b,
+    dnnl_gIOdhw8o8i = dnnl_aCBdef8b8c,
     dnnl_gIOdhw16o16i = dnnl_aCBdef16b16c,
     dnnl_gOdhwi16o = dnnl_aBdefc16b,
     dnnl_gOdhwI16o2i = dnnl_aBdefC16b2c,

src/common/c_types_map.hpp

@@ -1,5 +1,6 @@
 /*******************************************************************************
 * Copyright 2016-2024 Intel Corporation
+* Copyright 2024 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -620,6 +621,7 @@ const format_tag_t aBdefc16b = dnnl_aBdefc16b;
 const format_tag_t aBdefC16b2c = dnnl_aBdefC16b2c;
 const format_tag_t aBdefC16b4c = dnnl_aBdefC16b4c;
 const format_tag_t aCBdef16c16b = dnnl_aCBdef16c16b;
+const format_tag_t aCBdef8b8c = dnnl_aCBdef8b8c;
 const format_tag_t aCBdef16b16c = dnnl_aCBdef16b16c;
 const format_tag_t aBdefc4b = dnnl_aBdefc4b;
 const format_tag_t aBdefc8b = dnnl_aBdefc8b;
@@ -634,8 +636,10 @@ const format_tag_t Acb4a = dnnl_Acb4a;
 const format_tag_t Acb8a = dnnl_Acb8a;
 const format_tag_t AcB8a2b = dnnl_AcB8a2b;
 const format_tag_t AcB8a4b = dnnl_AcB8a4b;
+const format_tag_t aCBd8b8c = dnnl_aCBd8b8c;
 const format_tag_t aCBd16b16c = dnnl_aCBd16b16c;
 const format_tag_t aCBd16c16b = dnnl_aCBd16c16b;
+const format_tag_t aCBde8b8c = dnnl_aCBde8b8c;
 const format_tag_t aCBde16b16c = dnnl_aCBde16b16c;
 const format_tag_t aCBde16c16b = dnnl_aCBde16c16b;
 const format_tag_t Acdb16a = dnnl_Acdb16a;
@@ -654,7 +658,9 @@ const format_tag_t AcdeB8a2b = dnnl_AcdeB8a2b;
 const format_tag_t AcdeB8a4b = dnnl_AcdeB8a4b;
 const format_tag_t Acedb16a = dnnl_Acedb16a;
 const format_tag_t Adcb16a = dnnl_Adcb16a;
+const format_tag_t BAc8a8b = dnnl_BAc8a8b;
 const format_tag_t BAc16a16b = dnnl_BAc16a16b;
+const format_tag_t BAcd8a8b = dnnl_BAcd8a8b;
 const format_tag_t BAcd16a16b = dnnl_BAcd16a16b;
 const format_tag_t ABc32a16b = dnnl_ABc32a16b;
 const format_tag_t ABcd32a16b = dnnl_ABcd32a16b;
@@ -663,6 +669,7 @@ const format_tag_t ABc40a16b = dnnl_ABc40a16b;
 const format_tag_t ABcd40a16b = dnnl_ABcd40a16b;
 const format_tag_t ABcde40a16b = dnnl_ABcde40a16b;
 const format_tag_t ABc32a32b = dnnl_ABc32a32b;
+const format_tag_t BAcde8a8b = dnnl_BAcde8a8b;
 const format_tag_t BAcde16a16b = dnnl_BAcde16a16b;
 const format_tag_t ABcd32a32b = dnnl_ABcd32a32b;
 const format_tag_t ABcde32a32b = dnnl_ABcde32a32b;
@@ -1169,6 +1176,7 @@ const format_tag_t Ohwi32o = dnnl_Ohwi32o;
 const format_tag_t gIOhw16i16o = dnnl_gIOhw16i16o;
 const format_tag_t gOhwi32o = dnnl_gOhwi32o;
 const format_tag_t Goidhw16g = dnnl_Goidhw16g;
+const format_tag_t IOw8o8i = dnnl_IOw8o8i;
 const format_tag_t IOw16o16i = dnnl_IOw16o16i;
 const format_tag_t IOw16i16o = dnnl_IOw16i16o;
 const format_tag_t gIOw16i16o = dnnl_gIOw16i16o;
@@ -1224,7 +1232,9 @@ const format_tag_t Owi4o = dnnl_Owi4o;
 const format_tag_t Owi8o = dnnl_Owi8o;
 const format_tag_t OwI8o2i = dnnl_OwI8o2i;
 const format_tag_t OwI8o4i = dnnl_OwI8o4i;
+const format_tag_t IOdhw8o8i = dnnl_IOdhw8o8i;
 const format_tag_t IOdhw16o16i = dnnl_IOdhw16o16i;
+const format_tag_t IOhw8o8i = dnnl_IOhw8o8i;
 const format_tag_t IOhw16o16i = dnnl_IOhw16o16i;
 const format_tag_t Ohwi16o = dnnl_Ohwi16o;
 const format_tag_t OhwI16o2i = dnnl_OhwI16o2i;
@@ -1332,6 +1342,7 @@ const format_tag_t OIdhw8i8o = dnnl_OIdhw8i8o;
 const format_tag_t OdhwI8i8o = dnnl_OdhwI8i8o;
 const format_tag_t OIdhw8o8i = dnnl_OIdhw8o8i;
 const format_tag_t OIdhw8o4i = dnnl_OIdhw8o4i;
+const format_tag_t gIOw8o8i = dnnl_gIOw8o8i;
 const format_tag_t gIOw16o16i = dnnl_gIOw16o16i;
 const format_tag_t Goiw16g = dnnl_Goiw16g;
 const format_tag_t Goiw8g = dnnl_Goiw8g;
@@ -1360,7 +1371,9 @@ const format_tag_t gOwi4o = dnnl_gOwi4o;
 const format_tag_t gOwi8o = dnnl_gOwi8o;
 const format_tag_t gOwI8o2i = dnnl_gOwI8o2i;
 const format_tag_t gOwI8o4i = dnnl_gOwI8o4i;
+const format_tag_t gIOdhw8o8i = dnnl_gIOdhw8o8i;
 const format_tag_t gIOdhw16o16i = dnnl_gIOdhw16o16i;
+const format_tag_t gIOhw8o8i = dnnl_gIOhw8o8i;
 const format_tag_t gIOhw16o16i = dnnl_gIOhw16o16i;
 const format_tag_t gOhwi16o = dnnl_gOhwi16o;
 const format_tag_t gOhwI16o2i = dnnl_gOhwI16o2i;

src/common/dnnl_debug_autogenerated.cpp

@@ -1,5 +1,6 @@
 /*******************************************************************************
 * Copyright 2018-2024 Intel Corporation
+* Copyright 2024 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -96,8 +97,8 @@ const char *dnnl_engine_kind2str(dnnl_engine_kind_t v) {
 const char *dnnl_sparse_encoding2str(dnnl_sparse_encoding_t v) {
     if (v == dnnl_sparse_encoding_undef) return "undef";
     if (v == dnnl_csr) return "csr";
-    if (v == dnnl_coo) return "coo";
     if (v == dnnl_packed) return "packed";
+    if (v == dnnl_coo) return "coo";
     assert(!"unknown sparse_encoding");
     return "unknown sparse_encoding";
 }
@@ -927,7 +928,6 @@ const char *dnnl_fmt_tag2str(dnnl_format_tag_t v) {
     if (v == dnnl_AcdeB4b8a4b) return "AcdeB4b8a4b";
     if (v == dnnl_Ab4a) return "Ab4a";
     if (v == dnnl_Ab8a) return "Ab8a";
-    if (v == dnnl_Ab32a) return "Ab32a";
     if (v == dnnl_BA4b4a) return "BA4b4a";
     if (v == dnnl_BA8b4a) return "BA8b4a";
     if (v == dnnl_BA2a24b) return "BA2a24b";
@@ -943,6 +943,13 @@ const char *dnnl_fmt_tag2str(dnnl_format_tag_t v) {
     if (v == dnnl_bcad) return "bcad";
     if (v == dnnl_cabd) return "cabd";
     if (v == dnnl_dabc) return "dabc";
+    if (v == dnnl_Ab32a) return "Ab32a";
+    if (v == dnnl_aCBd8b8c) return "aCBd8b8c";
+    if (v == dnnl_aCBde8b8c) return "aCBde8b8c";
+    if (v == dnnl_BAc8a8b) return "BAc8a8b";
+    if (v == dnnl_BAcd8a8b) return "BAcd8a8b";
+    if (v == dnnl_BAcde8a8b) return "BAcde8a8b";
+    if (v == dnnl_aCBdef8b8c) return "aCBdef8b8c";
     if (v == dnnl_format_tag_last) return "format_tag_last";
     if (v == dnnl_x) return "x";
     if (v == dnnl_nc) return "nc";
@@ -1048,6 +1055,7 @@ const char *dnnl_fmt_tag2str(dnnl_format_tag_t v) {
     if (v == dnnl_OI8i24o) return "OI8i24o";
     if (v == dnnl_OI8i16o) return "OI8i16o";
     if (v == dnnl_OI8i8o) return "OI8i8o";
+    if (v == dnnl_IOw8o8i) return "IOw8o8i";
     if (v == dnnl_IOw16o16i) return "IOw16o16i";
     if (v == dnnl_IOw16i16o) return "IOw16i16o";
     if (v == dnnl_OIw16i16o) return "OIw16i16o";
@@ -1116,6 +1124,7 @@ const char *dnnl_fmt_tag2str(dnnl_format_tag_t v) {
     if (v == dnnl_OwI8i16o) return "OwI8i16o";
     if (v == dnnl_OwI8o4i) return "OwI8o4i";
     if (v == dnnl_IOhw16i16o) return "IOhw16i16o";
+    if (v == dnnl_IOhw8o8i) return "IOhw8o8i";
     if (v == dnnl_IOhw16o16i) return "IOhw16o16i";
     if (v == dnnl_Ohwi16o) return "Ohwi16o";
     if (v == dnnl_OhwI16o2i) return "OhwI16o2i";
@@ -1246,6 +1255,7 @@ const char *dnnl_fmt_tag2str(dnnl_format_tag_t v) {
     if (v == dnnl_OIdhw8o4i) return "OIdhw8o4i";
     if (v == dnnl_IOdhw16i16o) return "IOdhw16i16o";
     if (v == dnnl_OIdhw4o8i8o4i) return "OIdhw4o8i8o4i";
+    if (v == dnnl_IOdhw8o8i) return "IOdhw8o8i";
     if (v == dnnl_IOdhw16o16i) return "IOdhw16o16i";
     if (v == dnnl_OIdhw16o16i2o) return "OIdhw16o16i2o";
     if (v == dnnl_OIdhw8i32o) return "OIdhw8i32o";
@@ -1257,6 +1267,7 @@ const char *dnnl_fmt_tag2str(dnnl_format_tag_t v) {
     if (v == dnnl_Goiw16g) return "Goiw16g";
     if (v == dnnl_Goiw8g) return "Goiw8g";
     if (v == dnnl_Goiw4g) return "Goiw4g";
+    if (v == dnnl_gIOw8o8i) return "gIOw8o8i";
     if (v == dnnl_gIOw16o16i) return "gIOw16o16i";
     if (v == dnnl_gIOw16i16o) return "gIOw16i16o";
     if (v == dnnl_gOIw16i16o) return "gOIw16i16o";
@@ -1300,6 +1311,7 @@ const char *dnnl_fmt_tag2str(dnnl_format_tag_t v) {
     if (v == dnnl_goIw4i) return "goIw4i";
     if (v == dnnl_goIw32i) return "goIw32i";
     if (v == dnnl_gIOhw16i16o) return "gIOhw16i16o";
+    if (v == dnnl_gIOhw8o8i) return "gIOhw8o8i";
     if (v == dnnl_gIOhw16o16i) return "gIOhw16o16i";
     if (v == dnnl_gOhwi16o) return "gOhwi16o";
     if (v == dnnl_gOhwI16o2i) return "gOhwI16o2i";
@@ -1363,6 +1375,7 @@ const char *dnnl_fmt_tag2str(dnnl_format_tag_t v) {
     if (v == dnnl_gOIhw4i8o2i) return "gOIhw4i8o2i";
     if (v == dnnl_gOIhw4o8i2o) return "gOIhw4o8i2o";
     if (v == dnnl_gIOdhw16i16o) return "gIOdhw16i16o";
+    if (v == dnnl_gIOdhw8o8i) return "gIOdhw8o8i";
     if (v == dnnl_gIOdhw16o16i) return "gIOdhw16o16i";
     if (v == dnnl_gOdhwi16o) return "gOdhwi16o";
     if (v == dnnl_gOdhwI16o2i) return "gOdhwI16o2i";

src/common/memory_desc_wrapper.cpp

@@ -1,5 +1,6 @@
 /*******************************************************************************
 * Copyright 2016-2024 Intel Corporation
+* Copyright 2024 FUJITSU LIMITED
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
@@ -516,7 +517,9 @@ status_t memory_desc_wrapper::compute_blocking(
         C(Acb8a, {0, 2, 1}, {8}, {0});
         C(AcB8a2b, {0, 2, 1}, {8, 2}, {0, 1});
         C(AcB8a4b, {0, 2, 1}, {8, 4}, {0, 1});
+        C(aCBd8b8c, {0, 2, 1, 3}, {8, 8}, {1, 2});
         C(aCBd16b16c, {0, 2, 1, 3}, {16, 16}, {1, 2});
+        C(aCBde8b8c, {0, 2, 1, 3, 4}, {8, 8}, {1, 2});
         C(aCBde16b16c, {0, 2, 1, 3, 4}, {16, 16}, {1, 2});
         C(Acdb16a, {0, 2, 3, 1}, {16}, {0});
         C(AcdB16a2b, {0, 2, 3, 1}, {16, 2}, {0, 1});
@@ -532,7 +535,9 @@ status_t memory_desc_wrapper::compute_blocking(
         C(AcdeB8a4b, {0, 2, 3, 4, 1}, {8, 4}, {0, 1});
         C(Acedb16a, {0, 2, 4, 3, 1}, {16}, {0});
         C(Adcb16a, {0, 3, 2, 1}, {16}, {0});
+        C(BAc8a8b, {1, 0, 2}, {8, 8}, {0, 1});
         C(BAc16a16b, {1, 0, 2}, {16, 16}, {0, 1});
+        C(BAcd8a8b, {1, 0, 2, 3}, {8, 8}, {0, 1});
         C(BAcd16a16b, {1, 0, 2, 3}, {16, 16}, {0, 1});
         C(ABc32a16b, {0, 1, 2}, {32, 16}, {0, 1});
         C(ABcd32a16b, {0, 1, 2, 3}, {32, 16}, {0, 1});
@@ -585,13 +590,15 @@ status_t memory_desc_wrapper::compute_blocking(
         C(aBCde2b8c8b2c, {0, 1, 2, 3, 4}, {2, 8, 8, 2}, {1, 2, 1, 2});
         C(aBdec32b, {0, 1, 3, 4, 2}, {32}, {1});
         C(aCBdef16c16b, {0, 2, 1, 3, 4, 5}, {16, 16}, {2, 1});
+        C(aCBdef8b8c, {0, 2, 1, 3, 4, 5}, {8, 8}, {1, 2});
         C(aCBdef16b16c, {0, 2, 1, 3, 4, 5}, {16, 16}, {1, 2});
         C(Abcdef16a, {0, 1, 2, 3, 4, 5}, {16}, {0});
         C(Abcdef32a, {0, 1, 2, 3, 4, 5}, {32}, {0});
         C(aCBd16c16b, {0, 2, 1, 3}, {16, 16}, {2, 1});
         C(aCBde16c16b, {0, 2, 1, 3, 4}, {16, 16}, {2, 1});
         C(Acdb32a, {0, 2, 3, 1}, {32}, {0});
         C(BAcd16b16a, {1, 0, 2, 3}, {16, 16}, {1, 0});
+        C(BAcde8a8b, {1, 0, 2, 3, 4}, {8, 8}, {0, 1});
         C(BAcde16a16b, {1, 0, 2, 3, 4}, {16, 16}, {0, 1});
         C(BAc16b16a, {1, 0, 2}, {16, 16}, {1, 0});
         C(aBCd2b4c2b, {0, 1, 2, 3}, {2, 4, 2}, {1, 2, 1});