diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp | 66 |
1 files changed, 33 insertions, 33 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp index ff171984e7..8ed20a1e48 100644 --- a/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp +++ b/src/core/NEON/kernels/arm_gemm/indirect-interleaves/sme2_interleave1VL_block2_fp32_bf16.hpp @@ -32,103 +32,103 @@ void interleave_block<1, 2, VLType::SME, false>( { __asm__ __volatile__( ".inst 0xd503477f // SMSTART ZA\n" - "sub x28, %x[width], #0x1\n" - "mov x27, #0x0\n" "cntw x22, ALL, MUL #2\n" + "sub x28, %x[width], #0x1\n" "cntw x21, ALL, MUL #2\n" "sub x20, x22, #0x1\n" "whilelt p10.s, XZR, %x[height]\n" "add x28, x28, x21\n" - "ands x26, %x[width], x20\n" + "ands x27, %x[width], x20\n" "udiv x28, x28, x21\n" - "csel x26, x26, x22, NE\n" + "csel x27, x27, x22, NE\n" + "mov x26, #0x0\n" "and x25, x28, #0x1\n" "sub x28, x28, #0x1\n" - "add x26, x26, #0x1\n" + "add x27, x27, #0x1\n" "mov x20, %x[width]\n" "ptrue p0.b\n" "mov x24, %x[outptr_raw]\n" "mov x23, %x[row_offset]\n" "cntw x22\n" "lsr x28, x28, #0x1\n" - "lsr x26, x26, #0x1\n" + "lsr x27, x27, #0x1\n" "mov x12, #0x0\n" - ".inst 0x25b44771 // whilelt pn9.s, x27, x20, VLx2\n" + ".inst 0x25b44751 // whilelt pn9.s, x26, x20, VLx2\n" "mov x21, %x[in]\n" "1:" // Width loop: Preamble: Loop "ldr x20, [x21], #0x8\n" ".inst 0x25306548 // psel p8.s, p9.s/Z, p10.s[w12]\n" - ".inst 0xa0174294 // ld1w { z20.s-z21.s }, pn8.s/Z, [x20, x23, LSL #2]\n" - ".inst 0xc160e294 // bfcvt z20.h, { z20.s-z21.s }\n" - ".inst 0xc0800280 // mova za0h.s[x12], p0/M, z20.s\n" + ".inst 0xa0174286 // ld1w { z6.s-z7.s }, pn8.s/Z, [x20, x23, LSL #2]\n" + ".inst 0xc160e0c6 // bfcvt z6.h, { z6.s-z7.s }\n" + ".inst 0xc08000c0 // mova za0h.s[x12], p0/M, z6.s\n" "add x12, x12, #0x1\n" "cmp x12, x22\n" "blt 1b\n" "incw x23, ALL, MUL #2\n" - "incw x27, ALL, MUL #2\n" + "incw x26, ALL, MUL #2\n" "cbz x28, 5f\n" "2:" // Width loop "mov x20, %x[width]\n" "mov x12, #0x0\n" - ".inst 0x25b44771 // whilelt pn9.s, x27, x20, VLx2\n" + ".inst 0x25b44751 // whilelt pn9.s, x26, x20, VLx2\n" "mov x21, %x[in]\n" "3:" // Width loop: Odd: Loop "ldr x20, [x21], #0x8\n" ".inst 0x25306548 // psel p8.s, p9.s/Z, p10.s[w12]\n" - ".inst 0xc0828018 // mova z24.s, p0/M, za0v.s[x12]\n" - ".inst 0xa0174294 // ld1w { z20.s-z21.s }, pn8.s/Z, [x20, x23, LSL #2]\n" - "st1w { z24.s }, p0, [x24]\n" - "addvl x24, x24, #1\n" - ".inst 0xc160e294 // bfcvt z20.h, { z20.s-z21.s }\n" - ".inst 0xc0800288 // mova za2h.s[x12], p0/M, z20.s\n" + ".inst 0xa017429e // ld1w { z30.s-z31.s }, pn8.s/Z, [x20, x23, LSL #2]\n" + ".inst 0xc160e3de // bfcvt z30.h, { z30.s-z31.s }\n" + ".inst 0xc08003c8 // mova za2h.s[x12], p0/M, z30.s\n" + ".inst 0xc082800f // mova z15.s, p0/M, za0v.s[x12]\n" "add x12, x12, #0x1\n" "cmp x12, x22\n" + "st1w { z15.s }, p0, [x24]\n" + "addvl x24, x24, #1\n" "blt 3b\n" - "incw x27, ALL, MUL #2\n" + "incw x26, ALL, MUL #2\n" "mov x20, %x[width]\n" "incw x23, ALL, MUL #2\n" "mov x12, #0x0\n" - ".inst 0x25b44771 // whilelt pn9.s, x27, x20, VLx2\n" + ".inst 0x25b44751 // whilelt pn9.s, x26, x20, VLx2\n" "mov x21, %x[in]\n" "4:" // Width loop: Even: Loop "ldr x20, [x21], #0x8\n" ".inst 0x25306548 // psel p8.s, p9.s/Z, p10.s[w12]\n" - ".inst 0xc082810c // mova z12.s, p0/M, za2v.s[x12]\n" - ".inst 0xa0174284 // ld1w { z4.s-z5.s }, pn8.s/Z, [x20, x23, LSL #2]\n" - "st1w { z12.s }, p0, [x24]\n" - "addvl x24, x24, #1\n" - ".inst 0xc160e084 // bfcvt z4.h, { z4.s-z5.s }\n" - ".inst 0xc0800080 // mova za0h.s[x12], p0/M, z4.s\n" + ".inst 0xa0174298 // ld1w { z24.s-z25.s }, pn8.s/Z, [x20, x23, LSL #2]\n" + ".inst 0xc160e318 // bfcvt z24.h, { z24.s-z25.s }\n" + ".inst 0xc0800300 // mova za0h.s[x12], p0/M, z24.s\n" + ".inst 0xc0828110 // mova z16.s, p0/M, za2v.s[x12]\n" "add x12, x12, #0x1\n" "cmp x12, x22\n" + "st1w { z16.s }, p0, [x24]\n" + "addvl x24, x24, #1\n" "blt 4b\n" "subs x28, x28, #0x1\n" "incw x23, ALL, MUL #2\n" - "incw x27, ALL, MUL #2\n" + "incw x26, ALL, MUL #2\n" "bgt 2b\n" "5:" // Width loop: Tails "cbnz x25, 8f\n" "mov x20, %x[width]\n" "mov x12, #0x0\n" - ".inst 0x25b44771 // whilelt pn9.s, x27, x20, VLx2\n" + ".inst 0x25b44751 // whilelt pn9.s, x26, x20, VLx2\n" "mov x21, %x[in]\n" "6:" // Width loop: Tails: Even: Odd: Loop "ldr x20, [x21], #0x8\n" ".inst 0x25306548 // psel p8.s, p9.s/Z, p10.s[w12]\n" - ".inst 0xc0828010 // mova z16.s, p0/M, za0v.s[x12]\n" ".inst 0xa017428e // ld1w { z14.s-z15.s }, pn8.s/Z, [x20, x23, LSL #2]\n" - "st1w { z16.s }, p0, [x24]\n" - "addvl x24, x24, #1\n" ".inst 0xc160e1ce // bfcvt z14.h, { z14.s-z15.s }\n" ".inst 0xc08001c8 // mova za2h.s[x12], p0/M, z14.s\n" + ".inst 0xc0828010 // mova z16.s, p0/M, za0v.s[x12]\n" "add x12, x12, #0x1\n" "cmp x12, x22\n" + "st1w { z16.s }, p0, [x24]\n" + "addvl x24, x24, #1\n" "blt 6b\n" "mov x12, #0x0\n" "7:" // Width loop: Tails: Even: Even: Loop ".inst 0xc0828110 // mova z16.s, p0/M, za2v.s[x12]\n" "add x12, x12, #0x1\n" - "cmp x12, x26\n" + "cmp x12, x27\n" "st1w { z16.s }, p0, [x24]\n" "addvl x24, x24, #1\n" "blt 7b\n" @@ -138,7 +138,7 @@ void interleave_block<1, 2, VLType::SME, false>( "9:" // Width loop: Tails: Odd: Loop ".inst 0xc0828010 // mova z16.s, p0/M, za0v.s[x12]\n" "add x12, x12, #0x1\n" - "cmp x12, x26\n" + "cmp x12, x27\n" "st1w { z16.s }, p0, [x24]\n" "addvl x24, x24, #1\n" "blt 9b\n" |