diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp | 72 |
1 files changed, 36 insertions, 36 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp index f827197ab7..f981624a1d 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -39,34 +39,34 @@ void sme_transpose_interleave_2VL(uint16_t *out, const uint16_t *in, size_t widt "blt 6f\n" "1:" // Main row loop: Head "mov x26, %x[in]\n" - "add x25, x26, %x[in_stride]\n" - "add x24, x25, %x[in_stride]\n" - "mov x23, %x[width]\n" + "mov x25, %x[width]\n" + "add x24, x26, %x[in_stride]\n" "cnth x20, ALL, MUL #4\n" - "add x21, x24, %x[in_stride]\n" - "cmp x23, x20\n" - "add %x[in], x21, %x[in_stride]\n" + "add x23, x24, %x[in_stride]\n" + "cmp x25, x20\n" + "add x21, x23, %x[in_stride]\n" "mov x22, %x[out]\n" + "add %x[in], x21, %x[in_stride]\n" "sub %x[height], %x[height], #0x4\n" "blt 3f\n" "2:" // Main row loop: Unroll column loop - "sub x23, x23, x20\n" + "sub x25, x25, x20\n" "ld1h { z31.h }, p2/Z, [x26]\n" - "cmp x23, x20\n" "ld1h { z30.h }, p2/Z, [x26, #1, MUL VL]\n" + "cmp x25, x20\n" "ld1h { z29.h }, p2/Z, [x26, #2, MUL VL]\n" "ld1h { z28.h }, p2/Z, [x26, #3, MUL VL]\n" "addvl x26, x26, #4\n" - "ld1h { z27.h }, p2/Z, [x25]\n" - "ld1h { z26.h }, p2/Z, [x25, #1, MUL VL]\n" - "ld1h { z25.h }, p2/Z, [x25, #2, MUL VL]\n" - "ld1h { z24.h }, p2/Z, [x25, #3, MUL VL]\n" - "addvl x25, x25, #4\n" - "ld1h { z23.h }, p2/Z, [x24]\n" - "ld1h { z22.h }, p2/Z, [x24, #1, MUL VL]\n" - "ld1h { z21.h }, p2/Z, [x24, #2, MUL VL]\n" - "ld1h { z20.h }, p2/Z, [x24, #3, MUL VL]\n" + "ld1h { z27.h }, p2/Z, [x24]\n" + "ld1h { z26.h }, p2/Z, [x24, #1, MUL VL]\n" + "ld1h { z25.h }, p2/Z, [x24, #2, MUL VL]\n" + "ld1h { z24.h }, p2/Z, [x24, #3, MUL VL]\n" "addvl x24, x24, #4\n" + "ld1h { z23.h }, p2/Z, [x23]\n" + "ld1h { z22.h }, p2/Z, [x23, #1, MUL VL]\n" + "ld1h { z21.h }, p2/Z, [x23, #2, MUL VL]\n" + "ld1h { z20.h }, p2/Z, [x23, #3, MUL VL]\n" + "addvl x23, x23, #4\n" "ld1h { z19.h }, p2/Z, [x21]\n" "ld1h { z18.h }, p2/Z, [x21, #1, MUL VL]\n" "ld1h { z17.h }, p2/Z, [x21, #2, MUL VL]\n" @@ -92,29 +92,29 @@ void sme_transpose_interleave_2VL(uint16_t *out, const uint16_t *in, size_t widt "add x22, x22, %x[out_stride]\n" "bge 2b\n" "3:" // Main row loop: Unroll column loop skip - "cbz x23, 5f\n" + "cbz x25, 5f\n" "4:" // Main row loop: Column loop - "mov x20, x23\n" + "mov x20, x25\n" + "dech x25, ALL, MUL #2\n" "whilelt p1.h, XZR, x20\n" - "ld1h { z23.h }, p1/Z, [x26]\n" "dech x20\n" - "dech x23, ALL, MUL #2\n" - "ld1h { z22.h }, p1/Z, [x25]\n" "whilelt p0.h, XZR, x20\n" - "cmp x23, #0x0\n" - "ld1h { z21.h }, p0/Z, [x26, #1, MUL VL]\n" + "cmp x25, #0x0\n" + "ld1h { z23.h }, p1/Z, [x26]\n" + "ld1h { z22.h }, p0/Z, [x26, #1, MUL VL]\n" "addvl x26, x26, #2\n" - "ld1h { z20.h }, p0/Z, [x25, #1, MUL VL]\n" - "addvl x25, x25, #2\n" - "ld1h { z19.h }, p1/Z, [x24]\n" - "ld1h { z18.h }, p0/Z, [x24, #1, MUL VL]\n" + "ld1h { z21.h }, p1/Z, [x24]\n" + "ld1h { z20.h }, p0/Z, [x24, #1, MUL VL]\n" "addvl x24, x24, #2\n" + "ld1h { z19.h }, p1/Z, [x23]\n" + "ld1h { z18.h }, p0/Z, [x23, #1, MUL VL]\n" + "addvl x23, x23, #2\n" "ld1h { z17.h }, p1/Z, [x21]\n" "ld1h { z16.h }, p0/Z, [x21, #1, MUL VL]\n" "addvl x21, x21, #2\n" "st1h { z23.h }, p2, [x22]\n" - "st1h { z21.h }, p2, [x22, #1, MUL VL]\n" - "st1h { z22.h }, p2, [x22, #2, MUL VL]\n" + "st1h { z22.h }, p2, [x22, #1, MUL VL]\n" + "st1h { z21.h }, p2, [x22, #2, MUL VL]\n" "st1h { z20.h }, p2, [x22, #3, MUL VL]\n" "st1h { z19.h }, p2, [x22, #4, MUL VL]\n" "st1h { z18.h }, p2, [x22, #5, MUL VL]\n" @@ -140,8 +140,8 @@ void sme_transpose_interleave_2VL(uint16_t *out, const uint16_t *in, size_t widt "8:" // Tail row loop: Unroll column loop "sub x21, x21, x20\n" "ld1h { z19.h }, p2/Z, [x26]\n" - "cmp x21, x20\n" "ld1h { z18.h }, p2/Z, [x26, #1, MUL VL]\n" + "cmp x21, x20\n" "ld1h { z17.h }, p2/Z, [x26, #2, MUL VL]\n" "ld1h { z16.h }, p2/Z, [x26, #3, MUL VL]\n" "st1h { z19.h }, p2, [x22]\n" @@ -156,15 +156,15 @@ void sme_transpose_interleave_2VL(uint16_t *out, const uint16_t *in, size_t widt "cbz x21, 11f\n" "10:" // Tail row loop: Column loop "mov x20, x21\n" - "whilelt p0.h, XZR, x20\n" - "ld1h { z17.h }, p0/Z, [x26]\n" - "dech x20\n" "dech x21, ALL, MUL #2\n" + "whilelt p1.h, XZR, x20\n" + "dech x20\n" "whilelt p0.h, XZR, x20\n" "cmp x21, #0x0\n" + "ld1h { z17.h }, p1/Z, [x26]\n" "ld1h { z16.h }, p0/Z, [x26, #1, MUL VL]\n" - "st1h { z17.h }, p2, [x22]\n" "addvl x26, x26, #2\n" + "st1h { z17.h }, p2, [x22]\n" "st1h { z16.h }, p2, [x22, #1, MUL VL]\n" "add x22, x22, %x[out_stride]\n" "bgt 10b\n" |