diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp | 148 |
1 files changed, 74 insertions, 74 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp index 44305f0513..1c97bed317 100644 --- a/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp +++ b/src/core/NEON/kernels/arm_gemm/transforms/sme_transpose_interleave_4VL_2x2_fp32bf16.hpp @@ -45,69 +45,69 @@ void sme_transpose_interleave_4VL_2x2_fp32bf16(bfloat16 *out, const float *in, s "blt 4f\n" "1:" // Main row loop: Head "mov x26, %x[in]\n" - "mov x25, %x[out]\n" - "add x24, x26, %x[in_stride]\n" - "sub %x[height], %x[height], #0x4\n" + "add x25, x26, %x[in_stride]\n" + "add x24, x25, %x[in_stride]\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" + "add %x[in], x23, %x[in_stride]\n" + "mov x22, %x[out]\n" + "sub %x[height], %x[height], #0x4\n" + "mov x21, %x[width]\n" "2:" // Main row loop: Column loop - "mov x20, x22\n" - "decw x22, ALL, MUL #4\n" + "mov x20, x21\n" "whilelt p3.s, XZR, x20\n" + "ld1w { z16.s }, p3/Z, [x26]\n" + ".inst 0x658ab218 // bfcvt z24.h, p4/M, z16.s\n" "decw x20\n" "whilelt p2.s, XZR, x20\n" + "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n" + ".inst 0x658ab217 // bfcvt z23.h, p4/M, z16.s\n" "decw x20\n" - "ld1w { z19.s }, p3/Z, [x26]\n" "whilelt p1.s, XZR, x20\n" + "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n" + ".inst 0x658ab216 // bfcvt z22.h, p4/M, z16.s\n" "decw x20\n" - "ld1w { z18.s }, p2/Z, [x26, #1, MUL VL]\n" "whilelt p0.s, XZR, x20\n" - "ld1w { z17.s }, p1/Z, [x26, #2, MUL VL]\n" - "cmp x22, #0x0\n" "ld1w { z16.s }, p0/Z, [x26, #3, MUL VL]\n" - ".inst 0x658ab27b // bfcvt z27.h, p4/M, z19.s\n" - "addvl x26, x26, #4\n" - "ld1w { z19.s }, p3/Z, [x23]\n" - ".inst 0x658ab25a // bfcvt z26.h, p4/M, z18.s\n" - "ld1w { z18.s }, p2/Z, [x23, #1, MUL VL]\n" - ".inst 0x658ab239 // bfcvt z25.h, p4/M, z17.s\n" - "ld1w { z17.s }, p1/Z, [x23, #2, MUL VL]\n" - ".inst 0x658ab218 // bfcvt z24.h, p4/M, z16.s\n" - "ld1w { z16.s }, p0/Z, [x23, #3, MUL VL]\n" - ".inst 0x658ab277 // bfcvt z23.h, p4/M, z19.s\n" - "addvl x23, x23, #4\n" - ".inst 0x658ab256 // bfcvt z22.h, p4/M, z18.s\n" - "ld1w { z19.s }, p3/Z, [x24]\n" - ".inst 0x658ab235 // bfcvt z21.h, p4/M, z17.s\n" - "ld1w { z18.s }, p2/Z, [x24, #1, MUL VL]\n" + ".inst 0x658ab215 // bfcvt z21.h, p4/M, z16.s\n" + "ld1w { z16.s }, p3/Z, [x24]\n" ".inst 0x658ab214 // bfcvt z20.h, p4/M, z16.s\n" - "ld1w { z17.s }, p1/Z, [x24, #2, MUL VL]\n" + "decw x21, ALL, MUL #4\n" + "cmp x21, #0x0\n" + "ld1w { z16.s }, p2/Z, [x24, #1, MUL VL]\n" + ".inst 0x658ab213 // bfcvt z19.h, p4/M, z16.s\n" + "addvl x26, x26, #4\n" + "ld1w { z16.s }, p1/Z, [x24, #2, MUL VL]\n" + ".inst 0x658ab212 // bfcvt z18.h, p4/M, z16.s\n" "ld1w { z16.s }, p0/Z, [x24, #3, MUL VL]\n" + ".inst 0x658ab211 // bfcvt z17.h, p4/M, z16.s\n" "addvl x24, x24, #4\n" - ".inst 0x648ab27b // bfcvtnt z27.h, p4/M, z19.s\n" - "ld1w { z19.s }, p3/Z, [x21]\n" - ".inst 0x648ab25a // bfcvtnt z26.h, p4/M, z18.s\n" - "ld1w { z18.s }, p2/Z, [x21, #1, MUL VL]\n" - ".inst 0x648ab239 // bfcvtnt z25.h, p4/M, z17.s\n" - "ld1w { z17.s }, p1/Z, [x21, #2, MUL VL]\n" + "ld1w { z16.s }, p3/Z, [x25]\n" ".inst 0x648ab218 // bfcvtnt z24.h, p4/M, z16.s\n" - "ld1w { z16.s }, p0/Z, [x21, #3, MUL VL]\n" - "addvl x21, x21, #4\n" - ".inst 0x648ab277 // bfcvtnt z23.h, p4/M, z19.s\n" - "st1h { z27.h }, p4, [x25]\n" - ".inst 0x648ab256 // bfcvtnt z22.h, p4/M, z18.s\n" - "st1h { z26.h }, p4, [x25, #1, MUL VL]\n" - ".inst 0x648ab235 // bfcvtnt z21.h, p4/M, z17.s\n" - "st1h { z25.h }, p4, [x25, #2, MUL VL]\n" + "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n" + ".inst 0x648ab217 // bfcvtnt z23.h, p4/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n" + ".inst 0x648ab216 // bfcvtnt z22.h, p4/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x25, #3, MUL VL]\n" + "addvl x25, x25, #4\n" + ".inst 0x648ab215 // bfcvtnt z21.h, p4/M, z16.s\n" + "ld1w { z16.s }, p3/Z, [x23]\n" ".inst 0x648ab214 // bfcvtnt z20.h, p4/M, z16.s\n" - "st1h { z24.h }, p4, [x25, #3, MUL VL]\n" - "st1h { z23.h }, p4, [x25, #4, MUL VL]\n" - "st1h { z22.h }, p4, [x25, #5, MUL VL]\n" - "st1h { z21.h }, p4, [x25, #6, MUL VL]\n" - "st1h { z20.h }, p4, [x25, #7, MUL VL]\n" - "add x25, x25, %x[out_stride]\n" + "ld1w { z16.s }, p2/Z, [x23, #1, MUL VL]\n" + ".inst 0x648ab213 // bfcvtnt z19.h, p4/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x23, #2, MUL VL]\n" + ".inst 0x648ab212 // bfcvtnt z18.h, p4/M, z16.s\n" + "ld1w { z16.s }, p0/Z, [x23, #3, MUL VL]\n" + "addvl x23, x23, #4\n" + ".inst 0x648ab211 // bfcvtnt z17.h, p4/M, z16.s\n" + "st1h { z24.h }, p4, [x22]\n" + "st1h { z23.h }, p4, [x22, #1, MUL VL]\n" + "st1h { z22.h }, p4, [x22, #2, MUL VL]\n" + "st1h { z21.h }, p4, [x22, #3, MUL VL]\n" + "st1h { z20.h }, p4, [x22, #4, MUL VL]\n" + "st1h { z19.h }, p4, [x22, #5, MUL VL]\n" + "st1h { z18.h }, p4, [x22, #6, MUL VL]\n" + "st1h { z17.h }, p4, [x22, #7, MUL VL]\n" + "add x22, x22, %x[out_stride]\n" "bgt 2b\n" "3:" // Main row loop: Column loop skip "cmp %x[height], #0x4\n" @@ -117,47 +117,47 @@ void sme_transpose_interleave_4VL_2x2_fp32bf16(bfloat16 *out, const float *in, s "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 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" + "add %x[in], x25, %x[in_stride]\n" + "mov x22, %x[out]\n" + "csel x25, x25, %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 p3.s, XZR, x20\n" + "ld1w { z16.s }, p3/Z, [x26]\n" + ".inst 0x658ab214 // bfcvt z20.h, p4/M, z16.s\n" "decw x20\n" "whilelt p2.s, XZR, x20\n" + "ld1w { z16.s }, p2/Z, [x26, #1, MUL VL]\n" + ".inst 0x658ab213 // bfcvt z19.h, p4/M, z16.s\n" "decw x20\n" - "ld1w { z19.s }, p3/Z, [x26]\n" "whilelt p1.s, XZR, x20\n" + "ld1w { z16.s }, p1/Z, [x26, #2, MUL VL]\n" + ".inst 0x658ab212 // bfcvt z18.h, p4/M, z16.s\n" "decw x20\n" - "ld1w { z18.s }, p2/Z, [x26, #1, MUL VL]\n" "whilelt p0.s, XZR, x20\n" - "ld1w { z17.s }, p1/Z, [x26, #2, MUL VL]\n" - "cmp x21, #0x0\n" "ld1w { z16.s }, p0/Z, [x26, #3, MUL VL]\n" - ".inst 0x658ab277 // bfcvt z23.h, p4/M, z19.s\n" + ".inst 0x658ab211 // bfcvt z17.h, p4/M, z16.s\n" + "ld1w { z16.s }, p3/Z, [x25]\n" + "decw x21, ALL, MUL #4\n" + "cmp x21, #0x0\n" + ".inst 0x648ab214 // bfcvtnt z20.h, p4/M, z16.s\n" + "ld1w { z16.s }, p2/Z, [x25, #1, MUL VL]\n" "addvl x26, x26, #4\n" - ".inst 0x658ab256 // bfcvt z22.h, p4/M, z18.s\n" - "ld1w { z21.s }, p3/Z, [x24]\n" - ".inst 0x658ab234 // bfcvt z20.h, p4/M, z17.s\n" - "ld1w { z19.s }, p2/Z, [x24, #1, MUL VL]\n" - ".inst 0x658ab212 // bfcvt z18.h, p4/M, z16.s\n" - "ld1w { z17.s }, p1/Z, [x24, #2, MUL VL]\n" - "ld1w { z16.s }, p0/Z, [x24, #3, MUL VL]\n" - "addvl x24, x24, #4\n" - ".inst 0x648ab2b7 // bfcvtnt z23.h, p4/M, z21.s\n" - ".inst 0x648ab276 // bfcvtnt z22.h, p4/M, z19.s\n" - ".inst 0x648ab234 // bfcvtnt z20.h, p4/M, z17.s\n" + ".inst 0x648ab213 // bfcvtnt z19.h, p4/M, z16.s\n" + "ld1w { z16.s }, p1/Z, [x25, #2, MUL VL]\n" ".inst 0x648ab212 // bfcvtnt z18.h, p4/M, z16.s\n" - "st1h { z23.h }, p4, [x25]\n" - "st1h { z22.h }, p4, [x25, #1, MUL VL]\n" - "st1h { z20.h }, p4, [x25, #2, MUL VL]\n" - "st1h { z18.h }, p4, [x25, #3, MUL VL]\n" - "add x25, x25, %x[out_stride]\n" + "ld1w { z16.s }, p0/Z, [x25, #3, MUL VL]\n" + "addvl x25, x25, #4\n" + ".inst 0x648ab211 // bfcvtnt z17.h, p4/M, z16.s\n" + "st1h { z20.h }, p4, [x22]\n" + "st1h { z19.h }, p4, [x22, #1, MUL VL]\n" + "st1h { z18.h }, p4, [x22, #2, MUL VL]\n" + "st1h { z17.h }, p4, [x22, #3, MUL VL]\n" + "add x22, x22, %x[out_stride]\n" "bgt 6b\n" "7:" // Tail row loop: Column loop skip "cmp %x[height], #0x1\n" |