diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp | 76 |
1 files changed, 38 insertions, 38 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp index b33c4f6c2d..e37879e19b 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/sve_transpose_interleave_1VL.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021, 2023 Arm Limited. + * Copyright (c) 2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -40,59 +40,59 @@ void sve_transpose_interleave_1VL(uint32_t *out, const uint32_t *in, size_t widt "mov x26, %x[in]\n" "mov x25, %x[width]\n" "cntw x24, ALL, MUL #2\n" - "add x23, x26, %x[in_stride]\n" - "add x21, x23, %x[in_stride]\n" + "mov x23, %x[out]\n" + "sub %x[height], %x[height], #0x4\n" + "add x22, x26, %x[in_stride]\n" + "add x21, x22, %x[in_stride]\n" "add x20, x21, %x[in_stride]\n" "cmp x25, x24\n" "add %x[in], x20, %x[in_stride]\n" - "mov x22, %x[out]\n" - "sub %x[height], %x[height], #0x4\n" "blt 3f\n" "2:" // Main row loop: Unroll column loop "sub x25, x25, x24\n" "ld1w { z23.s }, p1/Z, [x26]\n" "ld1w { z22.s }, p1/Z, [x26, #1, MUL VL]\n" - "cmp x25, x24\n" - "ld1w { z21.s }, p1/Z, [x23]\n" - "ld1w { z20.s }, p1/Z, [x23, #1, MUL VL]\n" "addvl x26, x26, #2\n" - "addvl x23, x23, #2\n" + "ld1w { z21.s }, p1/Z, [x22]\n" + "ld1w { z20.s }, p1/Z, [x22, #1, MUL VL]\n" + "cmp x25, x24\n" + "addvl x22, x22, #2\n" "ld1w { z19.s }, p1/Z, [x21]\n" "ld1w { z18.s }, p1/Z, [x21, #1, MUL VL]\n" "addvl x21, x21, #2\n" "ld1w { z17.s }, p1/Z, [x20]\n" "ld1w { z16.s }, p1/Z, [x20, #1, MUL VL]\n" - "st1w { z23.s }, p1, [x22]\n" + "st1w { z23.s }, p1, [x23]\n" "addvl x20, x20, #2\n" - "st1w { z21.s }, p1, [x22, #1, MUL VL]\n" - "st1w { z19.s }, p1, [x22, #2, MUL VL]\n" - "st1w { z17.s }, p1, [x22, #3, MUL VL]\n" - "add x22, x22, %x[out_stride]\n" - "st1w { z22.s }, p1, [x22]\n" - "st1w { z20.s }, p1, [x22, #1, MUL VL]\n" - "st1w { z18.s }, p1, [x22, #2, MUL VL]\n" - "st1w { z16.s }, p1, [x22, #3, MUL VL]\n" - "add x22, x22, %x[out_stride]\n" + "st1w { z21.s }, p1, [x23, #1, MUL VL]\n" + "st1w { z19.s }, p1, [x23, #2, MUL VL]\n" + "st1w { z17.s }, p1, [x23, #3, MUL VL]\n" + "add x23, x23, %x[out_stride]\n" + "st1w { z22.s }, p1, [x23]\n" + "st1w { z20.s }, p1, [x23, #1, MUL VL]\n" + "st1w { z18.s }, p1, [x23, #2, MUL VL]\n" + "st1w { z16.s }, p1, [x23, #3, MUL VL]\n" + "add x23, x23, %x[out_stride]\n" "bge 2b\n" "3:" // Main row loop: Unroll column loop skip "cbz x25, 5f\n" "4:" // Main row loop: Column loop "whilelt p0.s, XZR, x25\n" "decw x25\n" - "ld1w { z19.s }, p0/Z, [x26]\n" - "ld1w { z18.s }, p0/Z, [x23]\n" "cmp x25, #0x0\n" + "ld1w { z19.s }, p0/Z, [x26]\n" "addvl x26, x26, #1\n" + "ld1w { z18.s }, p0/Z, [x22]\n" + "addvl x22, x22, #1\n" "ld1w { z17.s }, p0/Z, [x21]\n" - "ld1w { z16.s }, p0/Z, [x20]\n" - "addvl x23, x23, #1\n" "addvl x21, x21, #1\n" - "st1w { z19.s }, p1, [x22]\n" + "ld1w { z16.s }, p0/Z, [x20]\n" "addvl x20, x20, #1\n" - "st1w { z18.s }, p1, [x22, #1, MUL VL]\n" - "st1w { z17.s }, p1, [x22, #2, MUL VL]\n" - "st1w { z16.s }, p1, [x22, #3, MUL VL]\n" - "add x22, x22, %x[out_stride]\n" + "st1w { z19.s }, p1, [x23]\n" + "st1w { z18.s }, p1, [x23, #1, MUL VL]\n" + "st1w { z17.s }, p1, [x23, #2, MUL VL]\n" + "st1w { z16.s }, p1, [x23, #3, MUL VL]\n" + "add x23, x23, %x[out_stride]\n" "bgt 4b\n" "5:" // Main row loop: Column loop skip "cmp %x[height], #0x4\n" @@ -104,32 +104,32 @@ void sve_transpose_interleave_1VL(uint32_t *out, const uint32_t *in, size_t widt "mov x21, %x[width]\n" "cntw x20, ALL, MUL #2\n" "mov x26, %x[in]\n" + "mov x23, %x[out]\n" + "sub %x[height], %x[height], #0x1\n" "cmp x21, x20\n" "add %x[in], x26, %x[in_stride]\n" - "mov x22, %x[out]\n" - "sub %x[height], %x[height], #0x1\n" "blt 9f\n" "8:" // Tail row loop: Unroll column loop "sub x21, x21, x20\n" "ld1w { z17.s }, p1/Z, [x26]\n" "ld1w { z16.s }, p1/Z, [x26, #1, MUL VL]\n" - "st1w { z17.s }, p1, [x22]\n" - "add x22, x22, %x[out_stride]\n" - "cmp x21, x20\n" - "st1w { z16.s }, p1, [x22]\n" "addvl x26, x26, #2\n" - "add x22, x22, %x[out_stride]\n" + "cmp x21, x20\n" + "st1w { z17.s }, p1, [x23]\n" + "add x23, x23, %x[out_stride]\n" + "st1w { z16.s }, p1, [x23]\n" + "add x23, x23, %x[out_stride]\n" "bge 8b\n" "9:" // Tail row loop: Unroll column loop skip "cbz x21, 11f\n" "10:" // Tail row loop: Column loop "whilelt p0.s, XZR, x21\n" "decw x21\n" - "ld1w { z16.s }, p0/Z, [x26]\n" - "st1w { z16.s }, p1, [x22]\n" "cmp x21, #0x0\n" + "ld1w { z16.s }, p0/Z, [x26]\n" "addvl x26, x26, #1\n" - "add x22, x22, %x[out_stride]\n" + "st1w { z16.s }, p1, [x23]\n" + "add x23, x23, %x[out_stride]\n" "bgt 10b\n" "11:" // Tail row loop: Column loop skip "cmp %x[height], #0x1\n" |