diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL')
| -rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp | 560 |
1 files changed, 280 insertions, 280 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp index db29e42ef1..b102e1dea4 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_bf16fp32_dot_16VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -64,21 +64,21 @@ void sme2_gemv_bf16fp32_dot_16VL ( __asm__ __volatile__( "ptrue p8.b\n" ".inst 0xd503477f // SMSTART ZA\n" - "cntw x28, ALL, MUL #4\n" - "add x27, %x[N], x28\n" - "sub x27, x27, #0x1\n" - "udiv x27, x27, x28\n" - "add x22, x27, #0x3\n" - "and x22, x22, #0xfffffffffffffffc\n" - "mul x22, x22, x28\n" - "mul x22, x22, %x[K]\n" "mov x9, #0x0\n" - "mov x26, %x[B_ptr]\n" + "cntw x28, ALL, MUL #4\n" + "mov x27, %x[B_ptr]\n" + "add x26, %x[N], x28\n" "mov x25, %x[output_ptr]\n" + "sub x26, x26, #0x1\n" "ptrue p1.b\n" + "udiv x26, x26, x28\n" ".inst 0x25207811 // ptrue pn9.b\n" - "lsl x22, x22, #0x1\n" + "add x22, x26, #0x3\n" "mov x21, #0x1\n" + "and x22, x22, #0xfffffffffffffffc\n" + "mul x22, x22, x28\n" + "mul x22, x22, %x[K]\n" + "lsl x22, x22, #0x1\n" "1:" // RHS size check loop "cmp x22, #0x200000\n" "blt 2f\n" @@ -92,13 +92,13 @@ void sme2_gemv_bf16fp32_dot_16VL ( "lsl x21, x21, #0x16\n" "orr x22, x22, x20\n" "orr x22, x22, x21\n" - ".inst 0xf8b64b5a // rprfm pldonce, x22, [x26]\n" + ".inst 0xf8b64b7a // rprfm pldonce, x22, [x27]\n" "3:" // RHS prefetch exit "mov x24, %x[bias]\n" "4:" // Column loop - "cmp x27, #0x4\n" + "cmp x26, #0x4\n" "bge 28f\n" - "cmp x27, #0x2\n" + "cmp x26, #0x2\n" "bgt 20f\n" "beq 12f\n" "mov x23, %x[A_ptr]\n" @@ -108,8 +108,8 @@ void sme2_gemv_bf16fp32_dot_16VL ( ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" "cbz x24, 5f\n" - ".inst 0xa040c718 // ld1w { z24.s-z27.s }, pn9.b/Z, [x24]\n" - ".inst 0xc0042f00 // mova za.d[x9, #0], { z24.d-z27.d }\n" + ".inst 0xa040c714 // ld1w { z20.s-z23.s }, pn9.b/Z, [x24]\n" + ".inst 0xc0042e80 // mova za.d[x9, #0], { z20.d-z23.d }\n" "b 6f\n" "5:" // Width 1: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -118,59 +118,59 @@ void sme2_gemv_bf16fp32_dot_16VL ( "ble 8f\n" "7:" // Width 1: Multiply loop: Main loop head "whilelt p0.h, XZR, x22\n" - "ld1rqh { z8.h }, p0/Z, [x23]\n" + ".inst 0xa040a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + "ld1rqh { z2.h }, p0/Z, [x23]\n" "sub x22, x22, #0x8\n" - ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" - ".inst 0xc158b298 // bfdot za.s[x9, 0], { z20.h-z23.h }, z8.h[0]\n" - "addvl x26, x26, #16\n" - "cmp x22, #0x8\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" - ".inst 0xc158b498 // bfdot za.s[x9, 0], { z4.h-z7.h }, z8.h[1]\n" - "addvl x26, x26, #16\n" "add x23, x23, #0x10\n" - ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" - ".inst 0xc158bb98 // bfdot za.s[x9, 0], { z28.h-z31.h }, z8.h[2]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" - ".inst 0xc158bf18 // bfdot za.s[x9, 0], { z24.h-z27.h }, z8.h[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + "cmp x22, #0x8\n" + ".inst 0xa040a779 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + ".inst 0xc152b398 // bfdot za.s[x9, 0], { z28.h-z31.h }, z2.h[0]\n" + ".inst 0xa040a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + ".inst 0xc152b618 // bfdot za.s[x9, 0], { z16.h-z19.h }, z2.h[1]\n" + ".inst 0xc152bb18 // bfdot za.s[x9, 0], { z24.h-z27.h }, z2.h[2]\n" + ".inst 0xc152bd98 // bfdot za.s[x9, 0], { z12.h-z15.h }, z2.h[3]\n" "bgt 7b\n" "8:" // Width 1: Multiply loop: Single iteration only "whilelt p0.h, XZR, x22\n" - "ld1rqh { z11.h }, p0/Z, [x23]\n" + ".inst 0xa040a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" + "ld1rqh { z3.h }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xc15bb398 // bfdot za.s[x9, 0], { z28.h-z31.h }, z11.h[0]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b198 // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[0]\n" "ble 9f\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a765 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xc15bb598 // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[1]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b498 // bfdot za.s[x9, 0], { z4.h-z7.h }, z3.h[1]\n" "ble 9f\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xc15bba18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[2]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[2]\n" "ble 9f\n" - ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" - ".inst 0xc15bbc18 // bfdot za.s[x9, 0], { z0.h-z3.h }, z11.h[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153bd18 // bfdot za.s[x9, 0], { z8.h-z11.h }, z3.h[3]\n" "9:" // Width 1: Multiply loop: multiply skip "tbz %x[flags], #1, 10f\n" "add x21, %x[args_ptr], %[offset_min]\n" "add x20, %x[args_ptr], %[offset_max]\n" - ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - "ld1rw { z3.s }, p1/Z, [x21]\n" - "ld1rw { z29.s }, p1/Z, [x20]\n" - ".inst 0xc1bdc868 // fclamp { z8.s-z11.s }, z3.s, z29.s\n" - ".inst 0xa060c328 // st1w { z8.s-z11.s }, p8, [x25]\n" + ".inst 0xc0062c00 // mova { z0.d-z3.d }, za.d[x9, #0]\n" + "ld1rw { z23.s }, p1/Z, [x21]\n" + "ld1rw { z22.s }, p1/Z, [x20]\n" + ".inst 0xc1b6cae0 // fclamp { z0.s-z3.s }, z23.s, z22.s\n" + ".inst 0xa060c320 // st1w { z0.s-z3.s }, p8, [x25]\n" "addvl x25, x25, #4\n" "b 11f\n" "10:" // Width 1: No activation - ".inst 0xc0062c0c // mova { z12.d-z15.d }, za.d[x9, #0]\n" - ".inst 0xa060c32c // st1w { z12.s-z15.s }, p8, [x25]\n" + ".inst 0xc0062c00 // mova { z0.d-z3.d }, za.d[x9, #0]\n" + ".inst 0xa060c320 // st1w { z0.s-z3.s }, p8, [x25]\n" "addvl x25, x25, #4\n" "11:" // Width 1: Output done "b 36f\n" @@ -182,10 +182,10 @@ void sme2_gemv_bf16fp32_dot_16VL ( ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" "cbz x24, 13f\n" - ".inst 0xa040c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24]\n" - ".inst 0xc0042c80 // mova za.d[x9, #0], { z4.d-z7.d }\n" - ".inst 0xa041c714 // ld1w { z20.s-z23.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0042e81 // mova za.d[x9, #1], { z20.d-z23.d }\n" + ".inst 0xa040c71c // ld1w { z28.s-z31.s }, pn9.b/Z, [x24]\n" + ".inst 0xa041c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xc0042f80 // mova za.d[x9, #0], { z28.d-z31.d }\n" + ".inst 0xc0042c81 // mova za.d[x9, #1], { z4.d-z7.d }\n" "b 14f\n" "13:" // Width 2: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -194,80 +194,80 @@ void sme2_gemv_bf16fp32_dot_16VL ( "ble 16f\n" "15:" // Width 2: Multiply loop: Main loop head "whilelt p0.h, XZR, x22\n" - "ld1rqh { z9.h }, p0/Z, [x23]\n" + ".inst 0xa040a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n" "sub x22, x22, #0x8\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" - ".inst 0xc159b198 // bfdot za.s[x9, 0], { z12.h-z15.h }, z9.h[0]\n" + "ld1rqh { z1.h }, p0/Z, [x23]\n" "cmp x22, #0x8\n" "add x23, x23, #0x10\n" - ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc159b099 // bfdot za.s[x9, 1], { z4.h-z7.h }, z9.h[0]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" - ".inst 0xc159b598 // bfdot za.s[x9, 0], { z12.h-z15.h }, z9.h[1]\n" - ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc159b499 // bfdot za.s[x9, 1], { z4.h-z7.h }, z9.h[1]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" - ".inst 0xc159bb18 // bfdot za.s[x9, 0], { z24.h-z27.h }, z9.h[2]\n" - ".inst 0xa041a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc159b819 // bfdot za.s[x9, 1], { z0.h-z3.h }, z9.h[2]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" - ".inst 0xc159bc18 // bfdot za.s[x9, 0], { z0.h-z3.h }, z9.h[3]\n" - ".inst 0xa041a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc159bf99 // bfdot za.s[x9, 1], { z28.h-z31.h }, z9.h[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xa040a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n" + ".inst 0xc151b398 // bfdot za.s[x9, 0], { z28.h-z31.h }, z1.h[0]\n" + ".inst 0xa041a779 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc151b199 // bfdot za.s[x9, 1], { z12.h-z15.h }, z1.h[0]\n" + ".inst 0xa040a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n" + ".inst 0xa041a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xa040a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n" + ".inst 0xc151b618 // bfdot za.s[x9, 0], { z16.h-z19.h }, z1.h[1]\n" + ".inst 0xa041a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc151b719 // bfdot za.s[x9, 1], { z24.h-z27.h }, z1.h[1]\n" + ".inst 0xc151bb98 // bfdot za.s[x9, 0], { z28.h-z31.h }, z1.h[2]\n" + ".inst 0xc151b999 // bfdot za.s[x9, 1], { z12.h-z15.h }, z1.h[2]\n" + ".inst 0xc151bd18 // bfdot za.s[x9, 0], { z8.h-z11.h }, z1.h[3]\n" + ".inst 0xc151be99 // bfdot za.s[x9, 1], { z20.h-z23.h }, z1.h[3]\n" "bgt 15b\n" "16:" // Width 2: Multiply loop: Single iteration only "whilelt p0.h, XZR, x22\n" - "ld1rqh { z11.h }, p0/Z, [x23]\n" + ".inst 0xa040a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + "ld1rqh { z3.h }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xc15bb198 // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[0]\n" - ".inst 0xa041a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bb019 // bfdot za.s[x9, 1], { z0.h-z3.h }, z11.h[0]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b218 // bfdot za.s[x9, 0], { z16.h-z19.h }, z3.h[0]\n" + ".inst 0xc153b399 // bfdot za.s[x9, 1], { z28.h-z31.h }, z3.h[0]\n" "ble 17f\n" - ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xc15bb718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z11.h[1]\n" - ".inst 0xa041a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bb419 // bfdot za.s[x9, 1], { z0.h-z3.h }, z11.h[1]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b698 // bfdot za.s[x9, 0], { z20.h-z23.h }, z3.h[1]\n" + ".inst 0xc153b619 // bfdot za.s[x9, 1], { z16.h-z19.h }, z3.h[1]\n" "ble 17f\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a765 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xc15bb998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[2]\n" - ".inst 0xa041a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bbb99 // bfdot za.s[x9, 1], { z28.h-z31.h }, z11.h[2]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b898 // bfdot za.s[x9, 0], { z4.h-z7.h }, z3.h[2]\n" + ".inst 0xc153ba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z3.h[2]\n" "ble 17f\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" - ".inst 0xc15bbe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[3]\n" - ".inst 0xa041a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bbe99 // bfdot za.s[x9, 1], { z20.h-z23.h }, z11.h[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n" + ".inst 0xa041a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153bf98 // bfdot za.s[x9, 0], { z28.h-z31.h }, z3.h[3]\n" + ".inst 0xc153bd99 // bfdot za.s[x9, 1], { z12.h-z15.h }, z3.h[3]\n" "17:" // Width 2: Multiply loop: multiply skip "tbz %x[flags], #1, 18f\n" "add x21, %x[args_ptr], %[offset_min]\n" "add x20, %x[args_ptr], %[offset_max]\n" - ".inst 0xc0062c00 // mova { z0.d-z3.d }, za.d[x9, #0]\n" - "ld1rw { z9.s }, p1/Z, [x21]\n" - ".inst 0xc0062c24 // mova { z4.d-z7.d }, za.d[x9, #1]\n" - "ld1rw { z8.s }, p1/Z, [x20]\n" - ".inst 0xc1a8c920 // fclamp { z0.s-z3.s }, z9.s, z8.s\n" - ".inst 0xa060c720 // st1w { z0.s-z3.s }, pn9.b, [x25]\n" - ".inst 0xc1a8c924 // fclamp { z4.s-z7.s }, z9.s, z8.s\n" - ".inst 0xa061c324 // st1w { z4.s-z7.s }, p8, [x25, #0x4, MUL VL]\n" + ".inst 0xc0062c04 // mova { z4.d-z7.d }, za.d[x9, #0]\n" + ".inst 0xc0062c28 // mova { z8.d-z11.d }, za.d[x9, #1]\n" + "ld1rw { z17.s }, p1/Z, [x21]\n" + "ld1rw { z23.s }, p1/Z, [x20]\n" + ".inst 0xc1b7ca24 // fclamp { z4.s-z7.s }, z17.s, z23.s\n" + ".inst 0xc1b7ca28 // fclamp { z8.s-z11.s }, z17.s, z23.s\n" + ".inst 0xa060c724 // st1w { z4.s-z7.s }, pn9.b, [x25]\n" + ".inst 0xa061c328 // st1w { z8.s-z11.s }, p8, [x25, #0x4, MUL VL]\n" "addvl x25, x25, #8\n" "b 19f\n" "18:" // Width 2: No activation - ".inst 0xc0062c10 // mova { z16.d-z19.d }, za.d[x9, #0]\n" - ".inst 0xa060c730 // st1w { z16.s-z19.s }, pn9.b, [x25]\n" - ".inst 0xc0062c2c // mova { z12.d-z15.d }, za.d[x9, #1]\n" - ".inst 0xa061c32c // st1w { z12.s-z15.s }, p8, [x25, #0x4, MUL VL]\n" + ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" + ".inst 0xc0062c30 // mova { z16.d-z19.d }, za.d[x9, #1]\n" + ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" + ".inst 0xa061c330 // st1w { z16.s-z19.s }, p8, [x25, #0x4, MUL VL]\n" "addvl x25, x25, #8\n" "19:" // Width 2: Output done "b 36f\n" @@ -280,12 +280,12 @@ void sme2_gemv_bf16fp32_dot_16VL ( ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" "cbz x24, 21f\n" - ".inst 0xa040c718 // ld1w { z24.s-z27.s }, pn9.b/Z, [x24]\n" - ".inst 0xc0042f00 // mova za.d[x9, #0], { z24.d-z27.d }\n" - ".inst 0xa041c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0042c81 // mova za.d[x9, #1], { z4.d-z7.d }\n" - ".inst 0xa042c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" - ".inst 0xc0042e02 // mova za.d[x9, #2], { z16.d-z19.d }\n" + ".inst 0xa040c708 // ld1w { z8.s-z11.s }, pn9.b/Z, [x24]\n" + ".inst 0xa041c700 // ld1w { z0.s-z3.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xa042c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" + ".inst 0xc0042d00 // mova za.d[x9, #0], { z8.d-z11.d }\n" + ".inst 0xc0042c01 // mova za.d[x9, #1], { z0.d-z3.d }\n" + ".inst 0xc0042c82 // mova za.d[x9, #2], { z4.d-z7.d }\n" "b 22f\n" "21:" // Width 3: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -294,101 +294,101 @@ void sme2_gemv_bf16fp32_dot_16VL ( "ble 24f\n" "23:" // Width 3: Multiply loop: Main loop head "whilelt p0.h, XZR, x22\n" - "ld1rqh { z15.h }, p0/Z, [x23]\n" + ".inst 0xa040a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n" "sub x22, x22, #0x8\n" - ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" - ".inst 0xc15fb018 // bfdot za.s[x9, 0], { z0.h-z3.h }, z15.h[0]\n" + "ld1rqh { z3.h }, p0/Z, [x23]\n" "cmp x22, #0x8\n" "add x23, x23, #0x10\n" - ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15fb099 // bfdot za.s[x9, 1], { z4.h-z7.h }, z15.h[0]\n" - ".inst 0xa042a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15fb01a // bfdot za.s[x9, 2], { z0.h-z3.h }, z15.h[0]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" - ".inst 0xc15fb698 // bfdot za.s[x9, 0], { z20.h-z23.h }, z15.h[1]\n" - ".inst 0xa041a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15fb699 // bfdot za.s[x9, 1], { z20.h-z23.h }, z15.h[1]\n" - ".inst 0xa042a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15fb51a // bfdot za.s[x9, 2], { z8.h-z11.h }, z15.h[1]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26]\n" - ".inst 0xc15fbb18 // bfdot za.s[x9, 0], { z24.h-z27.h }, z15.h[2]\n" - ".inst 0xa041a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15fb919 // bfdot za.s[x9, 1], { z8.h-z11.h }, z15.h[2]\n" - ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15fba1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z15.h[2]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" - ".inst 0xc15fbe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z15.h[3]\n" - ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15fbe19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z15.h[3]\n" - ".inst 0xa042a749 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15fbd1a // bfdot za.s[x9, 2], { z8.h-z11.h }, z15.h[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a765 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b198 // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[0]\n" + ".inst 0xa040a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n" + ".inst 0xc153b119 // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[0]\n" + ".inst 0xa041a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc153b09a // bfdot za.s[x9, 2], { z4.h-z7.h }, z3.h[0]\n" + ".inst 0xa042a779 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xa040a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n" + ".inst 0xc153b618 // bfdot za.s[x9, 0], { z16.h-z19.h }, z3.h[1]\n" + ".inst 0xa041a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc153b699 // bfdot za.s[x9, 1], { z20.h-z23.h }, z3.h[1]\n" + ".inst 0xa042a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b71a // bfdot za.s[x9, 2], { z24.h-z27.h }, z3.h[1]\n" + ".inst 0xa040a765 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27]\n" + ".inst 0xa041a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc153b998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[2]\n" + ".inst 0xa042a779 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153bb99 // bfdot za.s[x9, 1], { z28.h-z31.h }, z3.h[2]\n" + ".inst 0xc153b91a // bfdot za.s[x9, 2], { z8.h-z11.h }, z3.h[2]\n" + ".inst 0xc153bc98 // bfdot za.s[x9, 0], { z4.h-z7.h }, z3.h[3]\n" + ".inst 0xc153be99 // bfdot za.s[x9, 1], { z20.h-z23.h }, z3.h[3]\n" + ".inst 0xc153bf1a // bfdot za.s[x9, 2], { z24.h-z27.h }, z3.h[3]\n" "bgt 23b\n" "24:" // Width 3: Multiply loop: Single iteration only "whilelt p0.h, XZR, x22\n" - "ld1rqh { z11.h }, p0/Z, [x23]\n" + ".inst 0xa040a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xa040a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26]\n" + "ld1rqh { z3.h }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xc15bb398 // bfdot za.s[x9, 0], { z28.h-z31.h }, z11.h[0]\n" - ".inst 0xa041a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bb019 // bfdot za.s[x9, 1], { z0.h-z3.h }, z11.h[0]\n" - ".inst 0xa042a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15bb29a // bfdot za.s[x9, 2], { z20.h-z23.h }, z11.h[0]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a765 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b298 // bfdot za.s[x9, 0], { z20.h-z23.h }, z3.h[0]\n" + ".inst 0xc153b199 // bfdot za.s[x9, 1], { z12.h-z15.h }, z3.h[0]\n" + ".inst 0xc153b09a // bfdot za.s[x9, 2], { z4.h-z7.h }, z3.h[0]\n" "ble 25f\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xc15bb598 // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[1]\n" - ".inst 0xa041a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bb499 // bfdot za.s[x9, 1], { z4.h-z7.h }, z11.h[1]\n" - ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15bb79a // bfdot za.s[x9, 2], { z28.h-z31.h }, z11.h[1]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b698 // bfdot za.s[x9, 0], { z20.h-z23.h }, z3.h[1]\n" + ".inst 0xc153b519 // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[1]\n" + ".inst 0xc153b61a // bfdot za.s[x9, 2], { z16.h-z19.h }, z3.h[1]\n" "ble 25f\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xc15bb898 // bfdot za.s[x9, 0], { z4.h-z7.h }, z11.h[2]\n" - ".inst 0xa041a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bba99 // bfdot za.s[x9, 1], { z20.h-z23.h }, z11.h[2]\n" - ".inst 0xa042a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15bb99a // bfdot za.s[x9, 2], { z12.h-z15.h }, z11.h[2]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a779 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153bb98 // bfdot za.s[x9, 0], { z28.h-z31.h }, z3.h[2]\n" + ".inst 0xc153bb19 // bfdot za.s[x9, 1], { z24.h-z27.h }, z3.h[2]\n" + ".inst 0xc153b99a // bfdot za.s[x9, 2], { z12.h-z15.h }, z3.h[2]\n" "ble 25f\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" - ".inst 0xc15bbd98 // bfdot za.s[x9, 0], { z12.h-z15.h }, z11.h[3]\n" - ".inst 0xa041a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bbe99 // bfdot za.s[x9, 1], { z20.h-z23.h }, z11.h[3]\n" - ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15bbe1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z11.h[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n" + ".inst 0xa041a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153bd18 // bfdot za.s[x9, 0], { z8.h-z11.h }, z3.h[3]\n" + ".inst 0xc153bf99 // bfdot za.s[x9, 1], { z28.h-z31.h }, z3.h[3]\n" + ".inst 0xc153bd9a // bfdot za.s[x9, 2], { z12.h-z15.h }, z3.h[3]\n" "25:" // Width 3: Multiply loop: multiply skip "tbz %x[flags], #1, 26f\n" "add x21, %x[args_ptr], %[offset_min]\n" "add x20, %x[args_ptr], %[offset_max]\n" - ".inst 0xc0062c04 // mova { z4.d-z7.d }, za.d[x9, #0]\n" - "ld1rw { z17.s }, p1/Z, [x21]\n" - ".inst 0xc0062c28 // mova { z8.d-z11.d }, za.d[x9, #1]\n" - "ld1rw { z16.s }, p1/Z, [x20]\n" - ".inst 0xc1b0ca24 // fclamp { z4.s-z7.s }, z17.s, z16.s\n" - ".inst 0xc0062c4c // mova { z12.d-z15.d }, za.d[x9, #2]\n" - ".inst 0xa060c724 // st1w { z4.s-z7.s }, pn9.b, [x25]\n" - ".inst 0xc1b0ca28 // fclamp { z8.s-z11.s }, z17.s, z16.s\n" - ".inst 0xa061c728 // st1w { z8.s-z11.s }, pn9.b, [x25, #0x4, MUL VL]\n" - ".inst 0xc1b0ca2c // fclamp { z12.s-z15.s }, z17.s, z16.s\n" - ".inst 0xa062c32c // st1w { z12.s-z15.s }, p8, [x25, #0x8, MUL VL]\n" + ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" + ".inst 0xc0062c2c // mova { z12.d-z15.d }, za.d[x9, #1]\n" + "ld1rw { z21.s }, p1/Z, [x21]\n" + ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n" + "ld1rw { z20.s }, p1/Z, [x20]\n" + ".inst 0xc1b4caa8 // fclamp { z8.s-z11.s }, z21.s, z20.s\n" + ".inst 0xc1b4caac // fclamp { z12.s-z15.s }, z21.s, z20.s\n" + ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" + ".inst 0xa060c728 // st1w { z8.s-z11.s }, pn9.b, [x25]\n" + ".inst 0xa061c72c // st1w { z12.s-z15.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xa062c330 // st1w { z16.s-z19.s }, p8, [x25, #0x8, MUL VL]\n" "addvl x25, x25, #12\n" "b 27f\n" "26:" // Width 3: No activation - ".inst 0xc0062c14 // mova { z20.d-z23.d }, za.d[x9, #0]\n" - ".inst 0xa060c734 // st1w { z20.s-z23.s }, pn9.b, [x25]\n" + ".inst 0xc0062c04 // mova { z4.d-z7.d }, za.d[x9, #0]\n" ".inst 0xc0062c2c // mova { z12.d-z15.d }, za.d[x9, #1]\n" + ".inst 0xc0062c5c // mova { z28.d-z31.d }, za.d[x9, #2]\n" + ".inst 0xa060c724 // st1w { z4.s-z7.s }, pn9.b, [x25]\n" ".inst 0xa061c72c // st1w { z12.s-z15.s }, pn9.b, [x25, #0x4, MUL VL]\n" - ".inst 0xc0062c4c // mova { z12.d-z15.d }, za.d[x9, #2]\n" - ".inst 0xa062c32c // st1w { z12.s-z15.s }, p8, [x25, #0x8, MUL VL]\n" + ".inst 0xa062c33c // st1w { z28.s-z31.s }, p8, [x25, #0x8, MUL VL]\n" "addvl x25, x25, #12\n" "27:" // Width 3: Output done "b 36f\n" @@ -402,14 +402,14 @@ void sme2_gemv_bf16fp32_dot_16VL ( ".inst 0x25b467f0 // whilelt p8.s, XZR, x20, VLx4\n" "cbz x24, 29f\n" ".inst 0xa040c704 // ld1w { z4.s-z7.s }, pn9.b/Z, [x24]\n" - ".inst 0xc0042c80 // mova za.d[x9, #0], { z4.d-z7.d }\n" - ".inst 0xa041c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0042e01 // mova za.d[x9, #1], { z16.d-z19.d }\n" - ".inst 0xa042c70c // ld1w { z12.s-z15.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" - ".inst 0xc0042d82 // mova za.d[x9, #2], { z12.d-z15.d }\n" + ".inst 0xa041c70c // ld1w { z12.s-z15.s }, pn9.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xa042c71c // ld1w { z28.s-z31.s }, pn9.b/Z, [x24, #0x8, MUL VL]\n" ".inst 0xa043c710 // ld1w { z16.s-z19.s }, pn9.b/Z, [x24, #0xc, MUL VL]\n" - ".inst 0xc0042e03 // mova za.d[x9, #3], { z16.d-z19.d }\n" + ".inst 0xc0042c80 // mova za.d[x9, #0], { z4.d-z7.d }\n" "addvl x24, x24, #16\n" + ".inst 0xc0042d81 // mova za.d[x9, #1], { z12.d-z15.d }\n" + ".inst 0xc0042f82 // mova za.d[x9, #2], { z28.d-z31.d }\n" + ".inst 0xc0042e03 // mova za.d[x9, #3], { z16.d-z19.d }\n" "b 30f\n" "29:" // Width 4: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -418,125 +418,125 @@ void sme2_gemv_bf16fp32_dot_16VL ( "ble 32f\n" "31:" // Width 4: Multiply loop: Main loop head "whilelt p0.h, XZR, x22\n" - "ld1rqh { z8.h }, p0/Z, [x23]\n" + ".inst 0xa040a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n" "sub x22, x22, #0x8\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" - ".inst 0xc158b218 // bfdot za.s[x9, 0], { z16.h-z19.h }, z8.h[0]\n" + "ld1rqh { z3.h }, p0/Z, [x23]\n" "cmp x22, #0x8\n" "add x23, x23, #0x10\n" - ".inst 0xa041a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc158b199 // bfdot za.s[x9, 1], { z12.h-z15.h }, z8.h[0]\n" - ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc158b21a // bfdot za.s[x9, 2], { z16.h-z19.h }, z8.h[0]\n" - ".inst 0xa043a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc158b19b // bfdot za.s[x9, 3], { z12.h-z15.h }, z8.h[0]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a74d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x26]\n" - ".inst 0xc158b598 // bfdot za.s[x9, 0], { z12.h-z15.h }, z8.h[1]\n" - ".inst 0xa041a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc158b699 // bfdot za.s[x9, 1], { z20.h-z23.h }, z8.h[1]\n" - ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc158b61a // bfdot za.s[x9, 2], { z16.h-z19.h }, z8.h[1]\n" - ".inst 0xa043a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc158b69b // bfdot za.s[x9, 3], { z20.h-z23.h }, z8.h[1]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26]\n" - ".inst 0xc158b898 // bfdot za.s[x9, 0], { z4.h-z7.h }, z8.h[2]\n" - ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc158ba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z8.h[2]\n" - ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc158ba1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z8.h[2]\n" - ".inst 0xa043a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc158b81b // bfdot za.s[x9, 3], { z0.h-z3.h }, z8.h[2]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26]\n" - ".inst 0xc158be98 // bfdot za.s[x9, 0], { z20.h-z23.h }, z8.h[3]\n" - ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc158be19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z8.h[3]\n" - ".inst 0xa042a745 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc158bc9a // bfdot za.s[x9, 2], { z4.h-z7.h }, z8.h[3]\n" - ".inst 0xa043a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc158be9b // bfdot za.s[x9, 3], { z20.h-z23.h }, z8.h[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc153b198 // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[0]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b299 // bfdot za.s[x9, 1], { z20.h-z23.h }, z3.h[0]\n" + ".inst 0xa040a779 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n" + ".inst 0xc153b21a // bfdot za.s[x9, 2], { z16.h-z19.h }, z3.h[0]\n" + ".inst 0xa041a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc153b11b // bfdot za.s[x9, 3], { z8.h-z11.h }, z3.h[0]\n" + ".inst 0xa042a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043a765 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc153b718 // bfdot za.s[x9, 0], { z24.h-z27.h }, z3.h[1]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b599 // bfdot za.s[x9, 1], { z12.h-z15.h }, z3.h[1]\n" + ".inst 0xa040a779 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27]\n" + ".inst 0xc153b51a // bfdot za.s[x9, 2], { z8.h-z11.h }, z3.h[1]\n" + ".inst 0xa041a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc153b49b // bfdot za.s[x9, 3], { z4.h-z7.h }, z3.h[1]\n" + ".inst 0xa042a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043a765 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc153bb18 // bfdot za.s[x9, 0], { z24.h-z27.h }, z3.h[2]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b919 // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[2]\n" + ".inst 0xa040a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27]\n" + ".inst 0xc153ba1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z3.h[2]\n" + ".inst 0xa041a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc153b89b // bfdot za.s[x9, 3], { z4.h-z7.h }, z3.h[2]\n" + ".inst 0xa042a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc153bd18 // bfdot za.s[x9, 0], { z8.h-z11.h }, z3.h[3]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153bf99 // bfdot za.s[x9, 1], { z28.h-z31.h }, z3.h[3]\n" + ".inst 0xc153bd9a // bfdot za.s[x9, 2], { z12.h-z15.h }, z3.h[3]\n" + ".inst 0xc153be9b // bfdot za.s[x9, 3], { z20.h-z23.h }, z3.h[3]\n" "bgt 31b\n" "32:" // Width 4: Multiply loop: Single iteration only "whilelt p0.h, XZR, x22\n" - "ld1rqh { z11.h }, p0/Z, [x23]\n" + ".inst 0xa040a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" + "ld1rqh { z3.h }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xc15bb218 // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[0]\n" - ".inst 0xa041a755 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bb299 // bfdot za.s[x9, 1], { z20.h-z23.h }, z11.h[0]\n" - ".inst 0xa042a75d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15bb39a // bfdot za.s[x9, 2], { z28.h-z31.h }, z11.h[0]\n" - ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc15bb21b // bfdot za.s[x9, 3], { z16.h-z19.h }, z11.h[0]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a77d // ldnt1h { z28.h-z31.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc153b218 // bfdot za.s[x9, 0], { z16.h-z19.h }, z3.h[0]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b199 // bfdot za.s[x9, 1], { z12.h-z15.h }, z3.h[0]\n" + ".inst 0xc153b39a // bfdot za.s[x9, 2], { z28.h-z31.h }, z3.h[0]\n" + ".inst 0xc153b29b // bfdot za.s[x9, 3], { z20.h-z23.h }, z3.h[0]\n" "ble 33f\n" - ".inst 0xa040a741 // ldnt1h { z0.h-z3.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xc15bb418 // bfdot za.s[x9, 0], { z0.h-z3.h }, z11.h[1]\n" - ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bb619 // bfdot za.s[x9, 1], { z16.h-z19.h }, z11.h[1]\n" - ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15bb61a // bfdot za.s[x9, 2], { z16.h-z19.h }, z11.h[1]\n" - ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc15bb61b // bfdot za.s[x9, 3], { z16.h-z19.h }, z11.h[1]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a765 // ldnt1h { z4.h-z7.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a779 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc153b598 // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[1]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b499 // bfdot za.s[x9, 1], { z4.h-z7.h }, z3.h[1]\n" + ".inst 0xc153b71a // bfdot za.s[x9, 2], { z24.h-z27.h }, z3.h[1]\n" + ".inst 0xc153b69b // bfdot za.s[x9, 3], { z20.h-z23.h }, z3.h[1]\n" "ble 33f\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" + ".inst 0xa040a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n" "subs x22, x22, #0x2\n" - ".inst 0xc15bba18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[2]\n" - ".inst 0xa041a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bba19 // bfdot za.s[x9, 1], { z16.h-z19.h }, z11.h[2]\n" - ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15bba1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z11.h[2]\n" - ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc15bba1b // bfdot za.s[x9, 3], { z16.h-z19.h }, z11.h[2]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc153b998 // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[2]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153b919 // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[2]\n" + ".inst 0xc153ba9a // bfdot za.s[x9, 2], { z20.h-z23.h }, z3.h[2]\n" + ".inst 0xc153ba1b // bfdot za.s[x9, 3], { z16.h-z19.h }, z3.h[2]\n" "ble 33f\n" - ".inst 0xa040a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26]\n" - ".inst 0xc15bbe18 // bfdot za.s[x9, 0], { z16.h-z19.h }, z11.h[3]\n" - ".inst 0xa041a759 // ldnt1h { z24.h-z27.h }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc15bbf19 // bfdot za.s[x9, 1], { z24.h-z27.h }, z11.h[3]\n" - ".inst 0xa042a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc15bbe1a // bfdot za.s[x9, 2], { z16.h-z19.h }, z11.h[3]\n" - ".inst 0xa043a751 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc15bbe1b // bfdot za.s[x9, 3], { z16.h-z19.h }, z11.h[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040a76d // ldnt1h { z12.h-z15.h }, pn9.b/Z, [x27]\n" + ".inst 0xa041a769 // ldnt1h { z8.h-z11.h }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042a775 // ldnt1h { z20.h-z23.h }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043a771 // ldnt1h { z16.h-z19.h }, pn9.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc153bd98 // bfdot za.s[x9, 0], { z12.h-z15.h }, z3.h[3]\n" + "addvl x27, x27, #16\n" + ".inst 0xc153bd19 // bfdot za.s[x9, 1], { z8.h-z11.h }, z3.h[3]\n" + ".inst 0xc153be9a // bfdot za.s[x9, 2], { z20.h-z23.h }, z3.h[3]\n" + ".inst 0xc153be1b // bfdot za.s[x9, 3], { z16.h-z19.h }, z3.h[3]\n" "33:" // Width 4: Multiply loop: multiply skip "tbz %x[flags], #1, 34f\n" "add x21, %x[args_ptr], %[offset_min]\n" "add x20, %x[args_ptr], %[offset_max]\n" - ".inst 0xc0062c0c // mova { z12.d-z15.d }, za.d[x9, #0]\n" + ".inst 0xc0062c04 // mova { z4.d-z7.d }, za.d[x9, #0]\n" + ".inst 0xc0062c20 // mova { z0.d-z3.d }, za.d[x9, #1]\n" "ld1rw { z21.s }, p1/Z, [x21]\n" - ".inst 0xc0062c38 // mova { z24.d-z27.d }, za.d[x9, #1]\n" + ".inst 0xc0062c4c // mova { z12.d-z15.d }, za.d[x9, #2]\n" "ld1rw { z20.s }, p1/Z, [x20]\n" - ".inst 0xc1b4caac // fclamp { z12.s-z15.s }, z21.s, z20.s\n" - ".inst 0xc0062c40 // mova { z0.d-z3.d }, za.d[x9, #2]\n" - ".inst 0xa060c72c // st1w { z12.s-z15.s }, pn9.b, [x25]\n" - ".inst 0xc1b4cab8 // fclamp { z24.s-z27.s }, z21.s, z20.s\n" ".inst 0xc0062c70 // mova { z16.d-z19.d }, za.d[x9, #3]\n" - ".inst 0xa061c738 // st1w { z24.s-z27.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xc1b4caa4 // fclamp { z4.s-z7.s }, z21.s, z20.s\n" ".inst 0xc1b4caa0 // fclamp { z0.s-z3.s }, z21.s, z20.s\n" - ".inst 0xa062c720 // st1w { z0.s-z3.s }, pn9.b, [x25, #0x8, MUL VL]\n" + ".inst 0xc1b4caac // fclamp { z12.s-z15.s }, z21.s, z20.s\n" ".inst 0xc1b4cab0 // fclamp { z16.s-z19.s }, z21.s, z20.s\n" + ".inst 0xa060c724 // st1w { z4.s-z7.s }, pn9.b, [x25]\n" + ".inst 0xa061c720 // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xa062c72c // st1w { z12.s-z15.s }, pn9.b, [x25, #0x8, MUL VL]\n" ".inst 0xa063c330 // st1w { z16.s-z19.s }, p8, [x25, #0xc, MUL VL]\n" "addvl x25, x25, #16\n" "b 35f\n" "34:" // Width 4: No activation ".inst 0xc0062c0c // mova { z12.d-z15.d }, za.d[x9, #0]\n" - ".inst 0xa060c72c // st1w { z12.s-z15.s }, pn9.b, [x25]\n" - ".inst 0xc0062c30 // mova { z16.d-z19.d }, za.d[x9, #1]\n" - ".inst 0xa061c730 // st1w { z16.s-z19.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xc0062c20 // mova { z0.d-z3.d }, za.d[x9, #1]\n" ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n" - ".inst 0xa062c730 // st1w { z16.s-z19.s }, pn9.b, [x25, #0x8, MUL VL]\n" ".inst 0xc0062c64 // mova { z4.d-z7.d }, za.d[x9, #3]\n" + ".inst 0xa060c72c // st1w { z12.s-z15.s }, pn9.b, [x25]\n" + ".inst 0xa061c720 // st1w { z0.s-z3.s }, pn9.b, [x25, #0x4, MUL VL]\n" + ".inst 0xa062c730 // st1w { z16.s-z19.s }, pn9.b, [x25, #0x8, MUL VL]\n" ".inst 0xa063c324 // st1w { z4.s-z7.s }, p8, [x25, #0xc, MUL VL]\n" "addvl x25, x25, #16\n" "35:" // Width 4: Output done - "subs x27, x27, #0x4\n" + "subs x26, x26, #0x4\n" "sub %x[N], %x[N], x28, LSL #2\n" "bgt 4b\n" "36:" // Exit |