diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp | 102 |
1 files changed, 51 insertions, 51 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp index 002a12479a..f070d7d322 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2.hpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,49 +45,49 @@ void sme_transpose_interleave_4VL_2x2(uint16_t *out, const uint16_t *in, size_t "blt 4f\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" - "add x23, x24, %x[in_stride]\n" - "add %x[in], x23, %x[in_stride]\n" - "mov x22, %x[out]\n" + "mov x25, %x[out]\n" + "add x24, x26, %x[in_stride]\n" "sub %x[height], %x[height], #0x4\n" - "mov x21, %x[width]\n" + "add x23, x24, %x[in_stride]\n" + "mov x22, %x[width]\n" + "add x21, x23, %x[in_stride]\n" + "add %x[in], x21, %x[in_stride]\n" "2:" // Main row loop: Column loop - "mov x20, x21\n" + "mov x20, x22\n" + "decw x22, ALL, MUL #4\n" "whilelt p1.h, XZR, x20\n" - "ld1h { z19.h }, p1/Z, [x26]\n" "dech x20\n" "whilelt p0.h, XZR, x20\n" - "ld1h { z18.h }, p0/Z, [x26, #1, MUL VL]\n" - "ld1h { z17.h }, p1/Z, [x25]\n" - "decw x21, ALL, MUL #4\n" - "cmp x21, #0x0\n" - "zip1 z24.h, z19.h, z17.h\n" - "ld1h { z16.h }, p0/Z, [x25, #1, MUL VL]\n" + "ld1h { z20.h }, p1/Z, [x26]\n" + "cmp x22, #0x0\n" + "ld1h { z19.h }, p0/Z, [x26, #1, MUL VL]\n" "addvl x26, x26, #2\n" - "addvl x25, x25, #2\n" - "zip2 z23.h, z19.h, z17.h\n" "ld1h { z17.h }, p1/Z, [x24]\n" - "zip1 z22.h, z18.h, z16.h\n" - "zip2 z21.h, z18.h, z16.h\n" - "ld1h { z20.h }, p0/Z, [x24, #1, MUL VL]\n" + "ld1h { z16.h }, p0/Z, [x24, #1, MUL VL]\n" "addvl x24, x24, #2\n" - "ld1h { z16.h }, p1/Z, [x23]\n" - "zip1 z19.h, z17.h, z16.h\n" - "zip2 z18.h, z17.h, z16.h\n" - "ld1h { z16.h }, p0/Z, [x23, #1, MUL VL]\n" + "ld1h { z18.h }, p1/Z, [x23]\n" + "ld1h { z24.h }, p0/Z, [x23, #1, MUL VL]\n" "addvl x23, x23, #2\n" - "zip1 z17.h, z20.h, z16.h\n" - "zip2 z16.h, z20.h, z16.h\n" - "st1h { z24.h }, p2, [x22]\n" - "st1h { z23.h }, p2, [x22, #1, MUL VL]\n" - "st1h { z22.h }, p2, [x22, #2, MUL VL]\n" - "st1h { z21.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" - "st1h { z17.h }, p2, [x22, #6, MUL VL]\n" - "st1h { z16.h }, p2, [x22, #7, MUL VL]\n" - "add x22, x22, %x[out_stride]\n" + "zip1 z23.h, z20.h, z17.h\n" + "zip2 z22.h, z20.h, z17.h\n" + "ld1h { z17.h }, p1/Z, [x21]\n" + "zip1 z21.h, z19.h, z16.h\n" + "zip2 z20.h, z19.h, z16.h\n" + "ld1h { z16.h }, p0/Z, [x21, #1, MUL VL]\n" + "addvl x21, x21, #2\n" + "zip1 z19.h, z18.h, z17.h\n" + "zip2 z18.h, z18.h, z17.h\n" + "st1h { z23.h }, p2, [x25]\n" + "zip1 z17.h, z24.h, z16.h\n" + "zip2 z16.h, z24.h, z16.h\n" + "st1h { z22.h }, p2, [x25, #1, MUL VL]\n" + "st1h { z21.h }, p2, [x25, #2, MUL VL]\n" + "st1h { z20.h }, p2, [x25, #3, MUL VL]\n" + "st1h { z19.h }, p2, [x25, #4, MUL VL]\n" + "st1h { z18.h }, p2, [x25, #5, MUL VL]\n" + "st1h { z17.h }, p2, [x25, #6, MUL VL]\n" + "st1h { z16.h }, p2, [x25, #7, MUL VL]\n" + "add x25, x25, %x[out_stride]\n" "bgt 2b\n" "3:" // Main row loop: Column loop skip "cmp %x[height], #0x4\n" @@ -97,35 +97,35 @@ void sme_transpose_interleave_4VL_2x2(uint16_t *out, const uint16_t *in, size_t "4:" // Main loop skip "5:" // Tail row loop: Head "mov x26, %x[in]\n" - "add x25, x26, %x[in_stride]\n" "cmp %x[height], #0x1\n" - "add %x[in], x25, %x[in_stride]\n" - "mov x22, %x[out]\n" - "csel x25, x25, %x[pad_row], GT\n" + "add x24, x26, %x[in_stride]\n" + "mov x25, %x[out]\n" + "add %x[in], x24, %x[in_stride]\n" + "csel x24, x24, %x[pad_row], GT\n" "sub %x[height], %x[height], #0x2\n" "mov x21, %x[width]\n" "6:" // Tail row loop: Column loop "mov x20, x21\n" + "decw x21, ALL, MUL #4\n" "whilelt p1.h, XZR, x20\n" - "ld1h { z18.h }, p1/Z, [x26]\n" "dech x20\n" "whilelt p0.h, XZR, x20\n" - "ld1h { z20.h }, p0/Z, [x26, #1, MUL VL]\n" - "ld1h { z17.h }, p1/Z, [x25]\n" - "decw x21, ALL, MUL #4\n" + "ld1h { z18.h }, p1/Z, [x26]\n" "cmp x21, #0x0\n" - "zip1 z19.h, z18.h, z17.h\n" - "ld1h { z16.h }, p0/Z, [x25, #1, MUL VL]\n" + "ld1h { z20.h }, p0/Z, [x26, #1, MUL VL]\n" "addvl x26, x26, #2\n" - "addvl x25, x25, #2\n" + "ld1h { z17.h }, p1/Z, [x24]\n" + "ld1h { z16.h }, p0/Z, [x24, #1, MUL VL]\n" + "addvl x24, x24, #2\n" + "zip1 z19.h, z18.h, z17.h\n" "zip2 z18.h, z18.h, z17.h\n" "zip1 z17.h, z20.h, z16.h\n" "zip2 z16.h, z20.h, z16.h\n" - "st1h { z19.h }, p2, [x22]\n" - "st1h { z18.h }, p2, [x22, #1, MUL VL]\n" - "st1h { z17.h }, p2, [x22, #2, MUL VL]\n" - "st1h { z16.h }, p2, [x22, #3, MUL VL]\n" - "add x22, x22, %x[out_stride]\n" + "st1h { z19.h }, p2, [x25]\n" + "st1h { z18.h }, p2, [x25, #1, MUL VL]\n" + "st1h { z17.h }, p2, [x25, #2, MUL VL]\n" + "st1h { z16.h }, p2, [x25, #3, MUL VL]\n" + "add x25, x25, %x[out_stride]\n" "bgt 6b\n" "7:" // Tail row loop: Column loop skip "cmp %x[height], #0x1\n" |