1 files changed, 115 insertions, 94 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_48.hpp b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_48.hpp
index f35752d5a8..44b6e7c9d4 100644
--- a/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_48.hpp
+++ b/src/core/NEON/kernels/arm_gemm/transforms/a64_transpose_interleave_48.hpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2021, 2023 Arm Limited.
+ * Copyright (c) 2021, 2023-2024 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -34,165 +34,186 @@ void a64_transpose_interleave_48(uint16_t *out, const uint16_t *in, size_t width
 
     __asm__ __volatile__(
       "cmp %x[height], #0x4\n"
-      "blt 10f\n"
+      "blt 11f\n"
       "1:"  // Main row loop: Head
       "mov x25, %x[in]\n"
       "mov x24, %x[width]\n"
-      "add x23, x25, %x[in_stride]\n"
-      "add x22, x23, %x[in_stride]\n"
-      "add x20, x22, %x[in_stride]\n"
+      "mov x23, %x[out]\n"
+      "sub %x[height], %x[height], #0x4\n"
+      "add x22, x25, %x[in_stride]\n"
+      "add x21, x22, %x[in_stride]\n"
+      "add x20, x21, %x[in_stride]\n"
       "cmp x24, #0x18\n"
       "add %x[in], x20, %x[in_stride]\n"
-      "mov x21, %x[out]\n"
-      "sub %x[height], %x[height], #0x4\n"
       "blt 3f\n"
       "2:"  // Main row loop: Column loop
       "ldr q27, [x25], #0x10\n"
-      "ldr q26, [x23], #0x10\n"
+      "ldr q26, [x22], #0x10\n"
       "sub x24, x24, #0x18\n"
-      "cmp x24, #0x18\n"
-      "ldr q25, [x22], #0x10\n"
+      "ldr q25, [x21], #0x10\n"
       "ldr q24, [x20], #0x10\n"
+      "cmp x24, #0x18\n"
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x23], #0x10\n"
-      "ldr q21, [x22], #0x10\n"
+      "ldr q22, [x22], #0x10\n"
+      "ldr q21, [x21], #0x10\n"
       "ldr q20, [x20], #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
+      "ldr q18, [x22], #0x10\n"
+      "ldr q17, [x21], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q27, [x21, #0x0]\n"
-      "str q23, [x21, #0x10]\n"
-      "str q19, [x21, #0x20]\n"
-      "str q26, [x21, #0x30]\n"
-      "str q22, [x21, #0x40]\n"
-      "str q18, [x21, #0x50]\n"
-      "str q25, [x21, #0x60]\n"
-      "str q21, [x21, #0x70]\n"
-      "str q17, [x21, #0x80]\n"
-      "str q24, [x21, #0x90]\n"
-      "str q20, [x21, #0xa0]\n"
-      "str q16, [x21, #0xb0]\n"
-      "add x21, x21, %x[out_stride]\n"
+      "str q27, [x23, #0x0]\n"
+      "str q23, [x23, #0x10]\n"
+      "str q19, [x23, #0x20]\n"
+      "str q26, [x23, #0x30]\n"
+      "str q22, [x23, #0x40]\n"
+      "str q18, [x23, #0x50]\n"
+      "str q25, [x23, #0x60]\n"
+      "str q21, [x23, #0x70]\n"
+      "str q17, [x23, #0x80]\n"
+      "str q24, [x23, #0x90]\n"
+      "str q20, [x23, #0xa0]\n"
+      "str q16, [x23, #0xb0]\n"
+      "add x23, x23, %x[out_stride]\n"
       "bge 2b\n"
       "3:"  // Main row loop: Column loop skip
+      "cbz x24, 10f\n"
       "cmp x24, #0x10\n"
+      "movi v16.8h, #0x0\n"
+      "str q16, [x23, #0x0]\n"
+      "str q16, [x23, #0x10]\n"
+      "str q16, [x23, #0x20]\n"
+      "str q16, [x23, #0x30]\n"
+      "str q16, [x23, #0x40]\n"
+      "str q16, [x23, #0x50]\n"
+      "str q16, [x23, #0x60]\n"
+      "str q16, [x23, #0x70]\n"
+      "str q16, [x23, #0x80]\n"
+      "str q16, [x23, #0x90]\n"
+      "str q16, [x23, #0xa0]\n"
+      "str q16, [x23, #0xb0]\n"
       "blt 5f\n"
       "4:"  // Main row loop: width 16 loop: loop
       "ldr q23, [x25], #0x10\n"
-      "ldr q22, [x23], #0x10\n"
+      "ldr q22, [x22], #0x10\n"
       "sub x24, x24, #0x10\n"
-      "cmp x24, #0x10\n"
-      "ldr q21, [x22], #0x10\n"
+      "ldr q21, [x21], #0x10\n"
       "ldr q20, [x20], #0x10\n"
+      "cmp x24, #0x10\n"
       "ldr q19, [x25], #0x10\n"
-      "ldr q18, [x23], #0x10\n"
-      "ldr q17, [x22], #0x10\n"
+      "ldr q18, [x22], #0x10\n"
+      "ldr q17, [x21], #0x10\n"
       "ldr q16, [x20], #0x10\n"
-      "str q23, [x21, #0x0]\n"
-      "str q19, [x21, #0x10]\n"
-      "str q22, [x21, #0x30]\n"
-      "str q18, [x21, #0x40]\n"
-      "str q21, [x21, #0x60]\n"
-      "str q17, [x21, #0x70]\n"
-      "str q20, [x21, #0x90]\n"
-      "str q16, [x21, #0xa0]\n"
-      "add x21, x21, #0x20\n"
+      "str q23, [x23, #0x0]\n"
+      "str q19, [x23, #0x10]\n"
+      "str q22, [x23, #0x30]\n"
+      "str q18, [x23, #0x40]\n"
+      "str q21, [x23, #0x60]\n"
+      "str q17, [x23, #0x70]\n"
+      "str q20, [x23, #0x90]\n"
+      "str q16, [x23, #0xa0]\n"
+      "add x23, x23, #0x20\n"
       "bge 4b\n"
       "5:"  // Main row loop: width 16 loop: skip
       "cmp x24, #0x4\n"
       "blt 7f\n"
       "6:"  // Main row loop: width 4 loop: loop
       "ldr d19, [x25], #0x8\n"
-      "ldr d18, [x23], #0x8\n"
+      "ldr d18, [x22], #0x8\n"
       "sub x24, x24, #0x4\n"
-      "cmp x24, #0x4\n"
-      "ldr d17, [x22], #0x8\n"
+      "ldr d17, [x21], #0x8\n"
       "ldr d16, [x20], #0x8\n"
-      "str d19, [x21, #0x0]\n"
-      "str d18, [x21, #0x30]\n"
-      "str d17, [x21, #0x60]\n"
-      "str d16, [x21, #0x90]\n"
-      "add x21, x21, #0x8\n"
+      "cmp x24, #0x4\n"
+      "str d19, [x23, #0x0]\n"
+      "str d18, [x23, #0x30]\n"
+      "str d17, [x23, #0x60]\n"
+      "str d16, [x23, #0x90]\n"
+      "add x23, x23, #0x8\n"
       "bge 6b\n"
       "7:"  // Main row loop: width 4 loop: skip
       "cmp x24, #0x1\n"
       "blt 9f\n"
       "8:"  // Main row loop: width 1 loop: loop
       "ldr h19, [x25], #0x2\n"
-      "ldr h18, [x23], #0x2\n"
+      "ldr h18, [x22], #0x2\n"
       "sub x24, x24, #0x1\n"
-      "cmp x24, #0x1\n"
-      "ldr h17, [x22], #0x2\n"
+      "ldr h17, [x21], #0x2\n"
       "ldr h16, [x20], #0x2\n"
-      "str h19, [x21, #0x0]\n"
-      "str h18, [x21, #0x30]\n"
-      "str h17, [x21, #0x60]\n"
-      "str h16, [x21, #0x90]\n"
-      "add x21, x21, #0x2\n"
+      "cmp x24, #0x1\n"
+      "str h19, [x23, #0x0]\n"
+      "str h18, [x23, #0x30]\n"
+      "str h17, [x23, #0x60]\n"
+      "str h16, [x23, #0x90]\n"
+      "add x23, x23, #0x2\n"
       "bge 8b\n"
       "9:"  // Main row loop: width 1 loop: skip
+      "10:"  // Main row loop: odd col skip
       "cmp %x[height], #0x4\n"
       "add %x[out], %x[out], #0xc0\n"
       "bge 1b\n"
-      "cbz %x[height], 20f\n"
-      "10:"  // Main loop skip
-      "11:"  // Tail row loop: Head
+      "cbz %x[height], 22f\n"
+      "11:"  // Main loop skip
+      "12:"  // Tail row loop: Head
       "mov x20, %x[width]\n"
       "mov x25, %x[in]\n"
+      "mov x23, %x[out]\n"
+      "sub %x[height], %x[height], #0x1\n"
       "cmp x20, #0x18\n"
       "add %x[in], x25, %x[in_stride]\n"
-      "mov x21, %x[out]\n"
-      "sub %x[height], %x[height], #0x1\n"
-      "blt 13f\n"
-      "12:"  // Tail row loop: Column loop
+      "blt 14f\n"
+      "13:"  // Tail row loop: Column loop
       "ldr q18, [x25], #0x10\n"
-      "ldr q17, [x25], #0x10\n"
       "sub x20, x20, #0x18\n"
-      "cmp x20, #0x18\n"
+      "ldr q17, [x25], #0x10\n"
       "ldr q16, [x25], #0x10\n"
-      "str q18, [x21, #0x0]\n"
-      "str q17, [x21, #0x10]\n"
-      "str q16, [x21, #0x20]\n"
-      "add x21, x21, %x[out_stride]\n"
-      "bge 12b\n"
-      "13:"  // Tail row loop: Column loop skip
+      "cmp x20, #0x18\n"
+      "str q18, [x23, #0x0]\n"
+      "str q17, [x23, #0x10]\n"
+      "str q16, [x23, #0x20]\n"
+      "add x23, x23, %x[out_stride]\n"
+      "bge 13b\n"
+      "14:"  // Tail row loop: Column loop skip
+      "cbz x20, 21f\n"
       "cmp x20, #0x10\n"
-      "blt 15f\n"
-      "14:"  // Tail row loop: width 16 loop: loop
+      "movi v16.8h, #0x0\n"
+      "str q16, [x23, #0x0]\n"
+      "str q16, [x23, #0x10]\n"
+      "str q16, [x23, #0x20]\n"
+      "blt 16f\n"
+      "15:"  // Tail row loop: width 16 loop: loop
       "ldr q17, [x25], #0x10\n"
-      "ldr q16, [x25], #0x10\n"
       "sub x20, x20, #0x10\n"
+      "ldr q16, [x25], #0x10\n"
       "cmp x20, #0x10\n"
-      "str q17, [x21, #0x0]\n"
-      "str q16, [x21, #0x10]\n"
-      "add x21, x21, #0x20\n"
-      "bge 14b\n"
-      "15:"  // Tail row loop: width 16 loop: skip
+      "str q17, [x23, #0x0]\n"
+      "str q16, [x23, #0x10]\n"
+      "add x23, x23, #0x20\n"
+      "bge 15b\n"
+      "16:"  // Tail row loop: width 16 loop: skip
       "cmp x20, #0x4\n"
-      "blt 17f\n"
-      "16:"  // Tail row loop: width 4 loop: loop
+      "blt 18f\n"
+      "17:"  // Tail row loop: width 4 loop: loop
       "ldr d16, [x25], #0x8\n"
       "sub x20, x20, #0x4\n"
       "cmp x20, #0x4\n"
-      "str d16, [x21, #0x0]\n"
-      "add x21, x21, #0x8\n"
-      "bge 16b\n"
-      "17:"  // Tail row loop: width 4 loop: skip
+      "str d16, [x23, #0x0]\n"
+      "add x23, x23, #0x8\n"
+      "bge 17b\n"
+      "18:"  // Tail row loop: width 4 loop: skip
       "cmp x20, #0x1\n"
-      "blt 19f\n"
-      "18:"  // Tail row loop: width 1 loop: loop
+      "blt 20f\n"
+      "19:"  // Tail row loop: width 1 loop: loop
       "ldr h16, [x25], #0x2\n"
       "sub x20, x20, #0x1\n"
       "cmp x20, #0x1\n"
-      "str h16, [x21, #0x0]\n"
-      "add x21, x21, #0x2\n"
-      "bge 18b\n"
-      "19:"  // Tail row loop: width 1 loop: skip
+      "str h16, [x23, #0x0]\n"
+      "add x23, x23, #0x2\n"
+      "bge 19b\n"
+      "20:"  // Tail row loop: width 1 loop: skip
+      "21:"  // Tail row loop: odd col skip
       "cmp %x[height], #0x1\n"
       "add %x[out], %x[out], #0x30\n"
-      "bge 11b\n"
-      "20:"  // Done
+      "bge 12b\n"
+      "22:"  // Done
       : [height] "+&r" (height), [in] "+&r" (in), [out] "+&r" (out)
       : [in_stride] "r" (in_stride), [out_stride] "r" (out_stride), [width] "r" (width)
       : "cc", "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "x20", "x21", "x22", "x23", "x24", "x25"