diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp | 206 |
1 files changed, 103 insertions, 103 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp index 382d4af314..284216a337 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_2VL_2x2_fp32bf16.hpp @@ -45,99 +45,99 @@ void sme_transpose_interleave_2VL_2x2_fp32bf16(bfloat16 *out, const float *in, s "blt 6f\n" "1:" // Main row loop: Head "mov x26, %x[in]\n" - "mov x25, %x[width]\n" - "add x24, x26, %x[in_stride]\n" + "add x25, x26, %x[in_stride]\n" + "add x24, x25, %x[in_stride]\n" + "mov x23, %x[width]\n" "cnth x20, ALL, MUL #2\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 x21, x24, %x[in_stride]\n" + "cmp x23, x20\n" "add %x[in], x21, %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 - "ld1w { z19.s }, p2/Z, [x26]\n" - "sub x25, x25, x20\n" - "ld1w { z18.s }, p2/Z, [x26, #1, MUL VL]\n" - "cmp x25, x20\n" - "ld1w { z17.s }, p2/Z, [x23]\n" - "ld1w { z16.s }, p2/Z, [x23, #1, MUL VL]\n" - ".inst 0x658aaa7b // bfcvt z27.h, p2/M, z19.s\n" - "ld1w { z19.s }, p2/Z, [x26, #2, MUL VL]\n" - ".inst 0x658aaa5a // bfcvt z26.h, p2/M, z18.s\n" - "ld1w { z18.s }, p2/Z, [x26, #3, MUL VL]\n" - ".inst 0x658aaa39 // bfcvt z25.h, p2/M, z17.s\n" - "addvl x26, x26, #4\n" - "ld1w { z17.s }, p2/Z, [x23, #2, MUL VL]\n" + "ld1w { z16.s }, p2/Z, [x26]\n" ".inst 0x658aaa18 // bfcvt z24.h, p2/M, z16.s\n" - "ld1w { z16.s }, p2/Z, [x23, #3, MUL VL]\n" - ".inst 0x658aaa77 // bfcvt z23.h, p2/M, z19.s\n" - "addvl x23, x23, #4\n" - "ld1w { z19.s }, p2/Z, [x24]\n" - ".inst 0x658aaa56 // bfcvt z22.h, p2/M, z18.s\n" - "ld1w { z18.s }, p2/Z, [x24, #1, MUL VL]\n" - ".inst 0x658aaa35 // bfcvt z21.h, p2/M, z17.s\n" - "ld1w { z17.s }, p2/Z, [x21]\n" + "sub x23, x23, x20\n" + "cmp x23, x20\n" + "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n" + ".inst 0x658aaa17 // bfcvt z23.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x24]\n" + ".inst 0x658aaa16 // bfcvt z22.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n" + ".inst 0x658aaa15 // bfcvt z21.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x26, #2, MUL VL]\n" ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n" - "ld1w { z16.s }, p2/Z, [x21, #1, MUL VL]\n" - ".inst 0x648aaa7b // bfcvtnt z27.h, p2/M, z19.s\n" - "ld1w { z19.s }, p2/Z, [x24, #2, MUL VL]\n" - ".inst 0x648aaa5a // bfcvtnt z26.h, p2/M, z18.s\n" - "ld1w { z18.s }, p2/Z, [x24, #3, MUL VL]\n" - ".inst 0x648aaa39 // bfcvtnt z25.h, p2/M, z17.s\n" + "ld1w { z16.s }, p2/Z, [x26, #3, MUL VL]\n" + ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n" + "addvl x26, x26, #4\n" + "ld1w { z16.s }, p2/Z, [x24, #2, MUL VL]\n" + ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x24, #3, MUL VL]\n" + ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n" "addvl x24, x24, #4\n" - "ld1w { z17.s }, p2/Z, [x21, #2, MUL VL]\n" + "ld1w { z16.s }, p2/Z, [x25]\n" ".inst 0x648aaa18 // bfcvtnt z24.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n" + ".inst 0x648aaa17 // bfcvtnt z23.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x21]\n" + ".inst 0x648aaa16 // bfcvtnt z22.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x21, #1, MUL VL]\n" + ".inst 0x648aaa15 // bfcvtnt z21.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n" + ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n" + "addvl x25, x25, #4\n" + ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x21, #2, MUL VL]\n" + ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n" "ld1w { z16.s }, p2/Z, [x21, #3, MUL VL]\n" - "st1h { z27.h }, p2, [x22]\n" + "st1h { z24.h }, p2, [x22]\n" "addvl x21, x21, #4\n" - ".inst 0x648aaa77 // bfcvtnt z23.h, p2/M, z19.s\n" - "st1h { z26.h }, p2, [x22, #1, MUL VL]\n" - ".inst 0x648aaa56 // bfcvtnt z22.h, p2/M, z18.s\n" - "st1h { z25.h }, p2, [x22, #2, MUL VL]\n" - ".inst 0x648aaa35 // bfcvtnt z21.h, p2/M, z17.s\n" - "st1h { z24.h }, p2, [x22, #3, MUL VL]\n" + ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\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" "add x22, x22, %x[out_stride]\n" - ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n" - "st1h { z23.h }, p2, [x22]\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 { z20.h }, p2, [x22]\n" + "st1h { z19.h }, p2, [x22, #1, MUL VL]\n" + "st1h { z18.h }, p2, [x22, #2, MUL VL]\n" + "st1h { z17.h }, p2, [x22, #3, MUL VL]\n" "add x22, x22, %x[out_stride]\n" "bge 2b\n" "3:" // Main row loop: Unroll column loop skip - "cbz x25, 5f\n" + "cbz x23, 5f\n" "4:" // Main row loop: Column loop - "mov x20, x25\n" - "decw x25, ALL, MUL #2\n" + "mov x20, x23\n" "whilelt p1.s, XZR, x20\n" + "ld1w { z16.s }, p1/Z, [x26]\n" + ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n" "decw x20\n" "whilelt p0.s, XZR, x20\n" - "ld1w { z19.s }, p1/Z, [x26]\n" - "cmp x25, #0x0\n" - "ld1w { z18.s }, p0/Z, [x26, #1, MUL VL]\n" + "ld1w { z16.s }, p0/Z, [x26, #1, MUL VL]\n" + ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x24]\n" + ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n" + "decw x23, ALL, MUL #2\n" + "cmp x23, #0x0\n" + "ld1w { z16.s }, p0/Z, [x24, #1, MUL VL]\n" + ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n" "addvl x26, x26, #2\n" - "ld1w { z17.s }, p1/Z, [x23]\n" - "ld1w { z16.s }, p0/Z, [x23, #1, MUL VL]\n" - ".inst 0x658aaa77 // bfcvt z23.h, p2/M, z19.s\n" - "addvl x23, x23, #2\n" - ".inst 0x658aaa56 // bfcvt z22.h, p2/M, z18.s\n" - "ld1w { z21.s }, p1/Z, [x24]\n" - ".inst 0x658aaa34 // bfcvt z20.h, p2/M, z17.s\n" - "ld1w { z19.s }, p0/Z, [x24, #1, MUL VL]\n" "addvl x24, x24, #2\n" - ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n" - "ld1w { z17.s }, p1/Z, [x21]\n" + "ld1w { z16.s }, p1/Z, [x25]\n" + ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n" + "addvl x25, x25, #2\n" + ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x21]\n" + ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n" "ld1w { z16.s }, p0/Z, [x21, #1, MUL VL]\n" "addvl x21, x21, #2\n" - ".inst 0x648aaab7 // bfcvtnt z23.h, p2/M, z21.s\n" - ".inst 0x648aaa76 // bfcvtnt z22.h, p2/M, z19.s\n" - ".inst 0x648aaa34 // bfcvtnt z20.h, p2/M, z17.s\n" - ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n" - "st1h { z23.h }, p2, [x22]\n" - "st1h { z22.h }, p2, [x22, #1, MUL VL]\n" - "st1h { z20.h }, p2, [x22, #2, MUL VL]\n" - "st1h { z18.h }, p2, [x22, #3, MUL VL]\n" + ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n" + "st1h { z20.h }, p2, [x22]\n" + "st1h { z19.h }, p2, [x22, #1, MUL VL]\n" + "st1h { z18.h }, p2, [x22, #2, MUL VL]\n" + "st1h { z17.h }, p2, [x22, #3, MUL VL]\n" "add x22, x22, %x[out_stride]\n" "bgt 4b\n" "5:" // Main row loop: Column loop skip @@ -148,65 +148,65 @@ void sme_transpose_interleave_2VL_2x2_fp32bf16(bfloat16 *out, const float *in, s "6:" // Main loop skip "7:" // Tail row loop: Head "mov x26, %x[in]\n" + "add x25, x26, %x[in_stride]\n" "cmp %x[height], #0x1\n" - "add x24, x26, %x[in_stride]\n" "mov x21, %x[width]\n" "cnth x20, ALL, MUL #2\n" - "add %x[in], x24, %x[in_stride]\n" - "csel x24, x24, %x[pad_row], GT\n" + "add %x[in], x25, %x[in_stride]\n" + "csel x25, x25, %x[pad_row], GT\n" "cmp x21, x20\n" "mov x22, %x[out]\n" "sub %x[height], %x[height], #0x2\n" "blt 9f\n" "8:" // Tail row loop: Unroll column loop - "ld1w { z19.s }, p2/Z, [x26]\n" + "ld1w { z16.s }, p2/Z, [x26]\n" + ".inst 0x658aaa14 // bfcvt z20.h, p2/M, z16.s\n" "sub x21, x21, x20\n" - "ld1w { z18.s }, p2/Z, [x26, #1, MUL VL]\n" "cmp x21, x20\n" - "ld1w { z17.s }, p2/Z, [x26, #2, MUL VL]\n" + "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n" + ".inst 0x658aaa13 // bfcvt z19.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x26, #2, MUL VL]\n" + ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n" "ld1w { z16.s }, p2/Z, [x26, #3, MUL VL]\n" - ".inst 0x658aaa77 // bfcvt z23.h, p2/M, z19.s\n" + ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n" "addvl x26, x26, #4\n" - ".inst 0x658aaa56 // bfcvt z22.h, p2/M, z18.s\n" - "ld1w { z21.s }, p2/Z, [x24]\n" - "ld1w { z20.s }, p2/Z, [x24, #1, MUL VL]\n" - ".inst 0x658aaa33 // bfcvt z19.h, p2/M, z17.s\n" - ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n" - "ld1w { z17.s }, p2/Z, [x24, #2, MUL VL]\n" - "ld1w { z16.s }, p2/Z, [x24, #3, MUL VL]\n" - ".inst 0x648aaab7 // bfcvtnt z23.h, p2/M, z21.s\n" - "addvl x24, x24, #4\n" - ".inst 0x648aaa96 // bfcvtnt z22.h, p2/M, z20.s\n" - ".inst 0x648aaa33 // bfcvtnt z19.h, p2/M, z17.s\n" + "ld1w { z16.s }, p2/Z, [x25]\n" + ".inst 0x648aaa14 // bfcvtnt z20.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n" + ".inst 0x648aaa13 // bfcvtnt z19.h, p2/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x25, #2, MUL VL]\n" ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n" - "st1h { z23.h }, p2, [x22]\n" - "st1h { z22.h }, p2, [x22, #1, MUL VL]\n" + "ld1w { z16.s }, p2/Z, [x25, #3, MUL VL]\n" + "st1h { z20.h }, p2, [x22]\n" + "addvl x25, x25, #4\n" + ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n" + "st1h { z19.h }, p2, [x22, #1, MUL VL]\n" "add x22, x22, %x[out_stride]\n" - "st1h { z19.h }, p2, [x22]\n" - "st1h { z18.h }, p2, [x22, #1, MUL VL]\n" + "st1h { z18.h }, p2, [x22]\n" + "st1h { z17.h }, p2, [x22, #1, MUL VL]\n" "add x22, x22, %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 "mov x20, x21\n" - "decw x21, ALL, MUL #2\n" "whilelt p1.s, XZR, x20\n" + "ld1w { z16.s }, p1/Z, [x26]\n" + ".inst 0x658aaa12 // bfcvt z18.h, p2/M, z16.s\n" "decw x20\n" "whilelt p0.s, XZR, x20\n" - "ld1w { z17.s }, p1/Z, [x26]\n" - "cmp x21, #0x0\n" "ld1w { z16.s }, p0/Z, [x26, #1, MUL VL]\n" + ".inst 0x658aaa11 // bfcvt z17.h, p2/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x25]\n" + "decw x21, ALL, MUL #2\n" + "cmp x21, #0x0\n" + ".inst 0x648aaa12 // bfcvtnt z18.h, p2/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x25, #1, MUL VL]\n" "addvl x26, x26, #2\n" - "ld1w { z19.s }, p1/Z, [x24]\n" - ".inst 0x658aaa32 // bfcvt z18.h, p2/M, z17.s\n" - "ld1w { z17.s }, p0/Z, [x24, #1, MUL VL]\n" - "addvl x24, x24, #2\n" - ".inst 0x658aaa10 // bfcvt z16.h, p2/M, z16.s\n" - ".inst 0x648aaa72 // bfcvtnt z18.h, p2/M, z19.s\n" - ".inst 0x648aaa30 // bfcvtnt z16.h, p2/M, z17.s\n" + "addvl x25, x25, #2\n" + ".inst 0x648aaa11 // bfcvtnt z17.h, p2/M, z16.s\n" "st1h { z18.h }, p2, [x22]\n" - "st1h { z16.h }, p2, [x22, #1, MUL VL]\n" + "st1h { z17.h }, p2, [x22, #1, MUL VL]\n" "add x22, x22, %x[out_stride]\n" "bgt 10b\n" "11:" // Tail row loop: Column loop skip |