From 065fb73f8e4f4f95f511a17612ab9d0646d424f6 Mon Sep 17 00:00:00 2001 From: Jakub Sujak Date: Fri, 29 Nov 2024 09:33:12 +0100 Subject: [PATCH] Contain SME2 assembly inside SMSTART/SMSTOP boundary Regenerate the SME2 GEMV micro-kernel assembly so that it is contained within the SMSTART/SMSTOP boundary, preventing illegal instruction faults when attempting to execute streaming SVE code on a system without SVE support. Signed-off-by: Jakub Sujak --- ...lamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c | 371 +++++++++--------- ...clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c | 310 +++++++-------- 2 files changed, 317 insertions(+), 364 deletions(-) diff --git a/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c b/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c index 563d84a3..47f9f330 100644 --- a/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c +++ b/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla.c @@ -91,45 +91,44 @@ void kai_run_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla( uint64_t flags = 0; __asm__ __volatile__( - "ptrue p8.b\n" ".inst 0xd503477f // SMSTART ZA\n" "mov x9, #0x0\n" - "cntw x27, ALL, MUL #4\n" - "mov x26, %x[B_ptr]\n" - "add x25, %x[N], x27\n" - "mov x24, %x[output_ptr]\n" - "sub x25, x25, #0x1\n" + "mov x27, %x[B_ptr]\n" + "cntw x26, ALL, MUL #4\n" + "mov x25, %x[output_ptr]\n" + "add x24, %x[N], x26\n" "ptrue p1.b\n" - "udiv x25, x25, x27\n" + "sub x24, x24, #0x1\n" ".inst 0x25207811 // ptrue pn9.b\n" - "add x22, x25, #0x3\n" - "mov x21, #0x1\n" - "and x22, x22, #0xfffffffffffffffc\n" - "mul x22, x22, x27\n" - "mul x22, x22, %x[K]\n" - "lsl x22, x22, #0x2\n" + "udiv x24, x24, x26\n" + "mov x22, #0x1\n" + "add x21, x24, #0x3\n" + "and x21, x21, #0xfffffffffffffffc\n" + "mul x21, x21, x26\n" + "mul x21, x21, %x[K]\n" + "lsl x21, x21, #0x2\n" "1:" // RHS size check loop - "cmp x22, #0x200000\n" + "cmp x21, #0x200000\n" "blt 2f\n" - "tbnz x22, #0, 3f\n" - "lsr x22, x22, #0x1\n" - "lsl x21, x21, #0x1\n" + "tbnz x21, #0, 3f\n" + "lsr x21, x21, #0x1\n" + "lsl x22, x22, #0x1\n" "b 1b\n" "2:" // RHS do prefetch - "lsl x20, x22, #0x26\n" - "sub x21, x21, #0x1\n" - "lsl x21, x21, #0x16\n" - "orr x22, x22, x20\n" - "orr x22, x22, x21\n" - ".inst 0xf8b64b5a // rprfm pldonce, x22, [x26]\n" + "lsl x20, x21, #0x26\n" + "sub x22, x22, #0x1\n" + "lsl x22, x22, #0x16\n" + "orr x21, x21, x20\n" + "orr x21, x21, x22\n" + ".inst 0xf8b54b7a // rprfm pldonce, x21, [x27]\n" "3:" // RHS prefetch exit "4:" // Column loop - "cmp x25, #0x4\n" + "cmp x24, #0x4\n" "bge 22f\n" - "cmp x25, #0x2\n" + "cmp x24, #0x2\n" "bgt 16f\n" "beq 10f\n" - ".inst 0xa040c754 // ld1w { z20.s-z23.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c774 // ld1w { z20.s-z23.s }, pn9.b/Z, [x27]\n" "mov x23, %x[K]\n" "mov x21, %x[N]\n" "mov x22, %x[A_ptr]\n" @@ -138,48 +137,46 @@ void kai_run_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla( "cmp x23, #0x4\n" ".inst 0xf8b44ad8 // rprfm pldmany, x20, [x22]\n" ".inst 0xc0042e80 // mova za.d[x9, #0], { z20.d-z23.d }\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" "ble 6f\n" "5:" // Width 1: Multiply loop: Main loop head "whilelt p0.s, XZR, x23\n" - ".inst 0xa040c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n" + "addvl x27, x27, #16\n" "ld1rqw { z2.s }, p0/Z, [x22]\n" "sub x23, x23, #0x4\n" "add x22, x22, #0x10\n" - ".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n" + "addvl x27, x27, #16\n" "cmp x23, #0x4\n" - ".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040c779 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27]\n" + "addvl x27, x27, #16\n" ".inst 0xc152a380 // fmla za.s[x9, 0], { z28.s-z31.s }, z2.s[0]\n" - ".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n" + "addvl x27, x27, #16\n" ".inst 0xc152a600 // fmla za.s[x9, 0], { z16.s-z19.s }, z2.s[1]\n" ".inst 0xc152ab00 // fmla za.s[x9, 0], { z24.s-z27.s }, z2.s[2]\n" ".inst 0xc152ad80 // fmla za.s[x9, 0], { z12.s-z15.s }, z2.s[3]\n" "bgt 5b\n" "6:" // Width 1: Multiply loop: Single iteration only "whilelt p0.s, XZR, x23\n" - ".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" "ld1rqw { z3.s }, p0/Z, [x22]\n" - "add x22, x22, #0x10\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153a180 // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[0]\n" "ble 7f\n" - ".inst 0xa040c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c765 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153a480 // fmla za.s[x9, 0], { z4.s-z7.s }, z3.s[1]\n" "ble 7f\n" - ".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153a980 // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[2]\n" "ble 7f\n" - ".inst 0xa040c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27]\n" ".inst 0xc153ad00 // fmla za.s[x9, 0], { z8.s-z11.s }, z3.s[3]\n" "7:" // Width 1: Multiply loop: multiply skip "tbz %x[flags], #1, 8f\n" @@ -189,50 +186,48 @@ void kai_run_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla( "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 0xa060c300 // st1w { z0.s-z3.s }, p8, [x24]\n" - "addvl x24, x24, #4\n" + ".inst 0xa060c320 // st1w { z0.s-z3.s }, p8, [x25]\n" "b 9f\n" "8:" // Width 1: No activation ".inst 0xc0062c00 // mova { z0.d-z3.d }, za.d[x9, #0]\n" - ".inst 0xa060c300 // st1w { z0.s-z3.s }, p8, [x24]\n" - "addvl x24, x24, #4\n" + ".inst 0xa060c320 // st1w { z0.s-z3.s }, p8, [x25]\n" "9:" // Width 1: Output done "b 28f\n" "10:" // Width 2 - ".inst 0xa040c75c // ld1w { z28.s-z31.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c77c // ld1w { z28.s-z31.s }, pn9.b/Z, [x27]\n" "mov x23, %x[K]\n" - "sub x21, %x[N], x27\n" - ".inst 0xa041c744 // ld1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + "sub x21, %x[N], x26\n" + ".inst 0xa041c764 // ld1w { z4.s-z7.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" "mov x22, %x[A_ptr]\n" "lsl x20, %x[K], #0x2\n" ".inst 0x25b567f0 // whilelt p8.s, XZR, x21, VLx4\n" "cmp x23, #0x4\n" ".inst 0xf8b44ad8 // rprfm pldmany, x20, [x22]\n" ".inst 0xc0042f80 // mova za.d[x9, #0], { z28.d-z31.d }\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #8\n" ".inst 0xc0042c81 // mova za.d[x9, #1], { z4.d-z7.d }\n" "ble 12f\n" "11:" // Width 2: Multiply loop: Main loop head "whilelt p0.s, XZR, x23\n" - ".inst 0xa040c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n" "sub x23, x23, #0x4\n" "ld1rqw { z1.s }, p0/Z, [x22]\n" "cmp x23, #0x4\n" "add x22, x22, #0x10\n" - ".inst 0xa041c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n" + ".inst 0xa041c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xa040c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n" ".inst 0xc151a380 // fmla za.s[x9, 0], { z28.s-z31.s }, z1.s[0]\n" - ".inst 0xa041c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041c779 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc151a181 // fmla za.s[x9, 1], { z12.s-z15.s }, z1.s[0]\n" - ".inst 0xa040c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26]\n" - ".inst 0xa041c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n" + ".inst 0xa041c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xa040c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27]\n" ".inst 0xc151a600 // fmla za.s[x9, 0], { z16.s-z19.s }, z1.s[1]\n" - ".inst 0xa041c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc151a701 // fmla za.s[x9, 1], { z24.s-z27.s }, z1.s[1]\n" ".inst 0xc151ab80 // fmla za.s[x9, 0], { z28.s-z31.s }, z1.s[2]\n" ".inst 0xc151a981 // fmla za.s[x9, 1], { z12.s-z15.s }, z1.s[2]\n" @@ -241,32 +236,30 @@ void kai_run_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla( "bgt 11b\n" "12:" // Width 2: Multiply loop: Single iteration only "whilelt p0.s, XZR, x23\n" - ".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" "ld1rqw { z3.s }, p0/Z, [x22]\n" - "add x22, x22, #0x10\n" - ".inst 0xa041c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc153a200 // fmla za.s[x9, 0], { z16.s-z19.s }, z3.s[0]\n" ".inst 0xc153a381 // fmla za.s[x9, 1], { z28.s-z31.s }, z3.s[0]\n" "ble 13f\n" - ".inst 0xa040c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" - ".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc153a680 // fmla za.s[x9, 0], { z20.s-z23.s }, z3.s[1]\n" ".inst 0xc153a601 // fmla za.s[x9, 1], { z16.s-z19.s }, z3.s[1]\n" "ble 13f\n" - ".inst 0xa040c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c765 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" - ".inst 0xa041c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc153a880 // fmla za.s[x9, 0], { z4.s-z7.s }, z3.s[2]\n" ".inst 0xc153aa01 // fmla za.s[x9, 1], { z16.s-z19.s }, z3.s[2]\n" "ble 13f\n" - ".inst 0xa040c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26]\n" - ".inst 0xa041c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n" + ".inst 0xa041c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" ".inst 0xc153af80 // fmla za.s[x9, 0], { z28.s-z31.s }, z3.s[3]\n" ".inst 0xc153ad81 // fmla za.s[x9, 1], { z12.s-z15.s }, z3.s[3]\n" "13:" // Width 2: Multiply loop: multiply skip @@ -279,64 +272,62 @@ void kai_run_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla( "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 0xa060c704 // st1w { z4.s-z7.s }, pn9.b, [x24]\n" - ".inst 0xa061c308 // st1w { z8.s-z11.s }, p8, [x24, #0x4, MUL VL]\n" - "addvl x24, x24, #8\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" "b 15f\n" "14:" // Width 2: No activation ".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 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n" - ".inst 0xa061c310 // st1w { z16.s-z19.s }, p8, [x24, #0x4, MUL VL]\n" - "addvl x24, x24, #8\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" "15:" // Width 2: Output done "b 28f\n" "16:" // Width 3 "mov x20, #0x2\n" - ".inst 0xa040c748 // ld1w { z8.s-z11.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c768 // ld1w { z8.s-z11.s }, pn9.b/Z, [x27]\n" "mov x23, %x[K]\n" - ".inst 0xa041c740 // ld1w { z0.s-z3.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "msub x21, x27, x20, %x[N]\n" + ".inst 0xa041c760 // ld1w { z0.s-z3.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "msub x21, x26, x20, %x[N]\n" "mov x22, %x[A_ptr]\n" - ".inst 0xa042c744 // ld1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xa042c764 // ld1w { z4.s-z7.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" "lsl x20, %x[K], #0x2\n" ".inst 0x25b567f0 // whilelt p8.s, XZR, x21, VLx4\n" "cmp x23, #0x4\n" ".inst 0xf8b44ad8 // rprfm pldmany, x20, [x22]\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" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc0042c82 // mova za.d[x9, #2], { z4.d-z7.d }\n" "ble 18f\n" "17:" // Width 3: Multiply loop: Main loop head "whilelt p0.s, XZR, x23\n" - ".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n" "sub x23, x23, #0x4\n" "ld1rqw { z3.s }, p0/Z, [x22]\n" "cmp x23, #0x4\n" "add x22, x22, #0x10\n" - ".inst 0xa041c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c765 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc153a180 // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[0]\n" - ".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n" ".inst 0xc153a101 // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[0]\n" - ".inst 0xa041c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xa041c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" ".inst 0xc153a082 // fmla za.s[x9, 2], { z4.s-z7.s }, z3.s[0]\n" - ".inst 0xa042c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n" + ".inst 0xa042c779 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xa040c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n" ".inst 0xc153a600 // fmla za.s[x9, 0], { z16.s-z19.s }, z3.s[1]\n" - ".inst 0xa041c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xa041c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" ".inst 0xc153a681 // fmla za.s[x9, 1], { z20.s-z23.s }, z3.s[1]\n" - ".inst 0xa042c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa042c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc153a702 // fmla za.s[x9, 2], { z24.s-z27.s }, z3.s[1]\n" - ".inst 0xa040c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26]\n" - ".inst 0xa041c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xa040c765 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27]\n" + ".inst 0xa041c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" ".inst 0xc153a980 // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[2]\n" - ".inst 0xa042c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa042c779 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc153ab81 // fmla za.s[x9, 1], { z28.s-z31.s }, z3.s[2]\n" ".inst 0xc153a902 // fmla za.s[x9, 2], { z8.s-z11.s }, z3.s[2]\n" ".inst 0xc153ac80 // fmla za.s[x9, 0], { z4.s-z7.s }, z3.s[3]\n" @@ -345,39 +336,37 @@ void kai_run_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla( "bgt 17b\n" "18:" // Width 3: Multiply loop: Single iteration only "whilelt p0.s, XZR, x23\n" - ".inst 0xa040c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" "ld1rqw { z3.s }, p0/Z, [x22]\n" - "add x22, x22, #0x10\n" - ".inst 0xa041c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c765 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc153a280 // fmla za.s[x9, 0], { z20.s-z23.s }, z3.s[0]\n" ".inst 0xc153a181 // fmla za.s[x9, 1], { z12.s-z15.s }, z3.s[0]\n" ".inst 0xc153a082 // fmla za.s[x9, 2], { z4.s-z7.s }, z3.s[0]\n" "ble 19f\n" - ".inst 0xa040c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" - ".inst 0xa041c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc153a680 // fmla za.s[x9, 0], { z20.s-z23.s }, z3.s[1]\n" ".inst 0xc153a501 // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[1]\n" ".inst 0xc153a602 // fmla za.s[x9, 2], { z16.s-z19.s }, z3.s[1]\n" "ble 19f\n" - ".inst 0xa040c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" - ".inst 0xa041c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa041c779 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" ".inst 0xc153ab80 // fmla za.s[x9, 0], { z28.s-z31.s }, z3.s[2]\n" ".inst 0xc153ab01 // fmla za.s[x9, 1], { z24.s-z27.s }, z3.s[2]\n" ".inst 0xc153a982 // fmla za.s[x9, 2], { z12.s-z15.s }, z3.s[2]\n" "ble 19f\n" - ".inst 0xa040c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26]\n" - ".inst 0xa041c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - "addvl x26, x26, #16\n" + ".inst 0xa040c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27]\n" + ".inst 0xa041c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" ".inst 0xc153ad00 // fmla za.s[x9, 0], { z8.s-z11.s }, z3.s[3]\n" ".inst 0xc153af81 // fmla za.s[x9, 1], { z28.s-z31.s }, z3.s[3]\n" ".inst 0xc153ad82 // fmla za.s[x9, 2], { z12.s-z15.s }, z3.s[3]\n" @@ -393,126 +382,123 @@ void kai_run_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla( ".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 0xa060c708 // st1w { z8.s-z11.s }, pn9.b, [x24]\n" - ".inst 0xa061c70c // st1w { z12.s-z15.s }, pn9.b, [x24, #0x4, MUL VL]\n" - ".inst 0xa062c310 // st1w { z16.s-z19.s }, p8, [x24, #0x8, MUL VL]\n" - "addvl x24, x24, #12\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" "b 21f\n" "20:" // Width 3: No activation ".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 0xa060c704 // st1w { z4.s-z7.s }, pn9.b, [x24]\n" - ".inst 0xa061c70c // st1w { z12.s-z15.s }, pn9.b, [x24, #0x4, MUL VL]\n" - ".inst 0xa062c31c // st1w { z28.s-z31.s }, p8, [x24, #0x8, MUL VL]\n" - "addvl x24, x24, #12\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 0xa062c33c // st1w { z28.s-z31.s }, p8, [x25, #0x8, MUL VL]\n" "21:" // Width 3: Output done "b 28f\n" "22:" // Width 4 "mov x20, #0x3\n" - ".inst 0xa040c744 // ld1w { z4.s-z7.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c764 // ld1w { z4.s-z7.s }, pn9.b/Z, [x27]\n" "mov x23, %x[K]\n" - ".inst 0xa041c74c // ld1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - "msub x21, x27, x20, %x[N]\n" + ".inst 0xa041c76c // ld1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + "msub x21, x26, x20, %x[N]\n" "mov x22, %x[A_ptr]\n" - ".inst 0xa042c75c // ld1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" + ".inst 0xa042c77c // ld1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" "lsl x20, %x[K], #0x2\n" ".inst 0x25b567f0 // whilelt p8.s, XZR, x21, VLx4\n" - ".inst 0xa043c750 // ld1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xa043c770 // ld1w { z16.s-z19.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n" "cmp x23, #0x4\n" ".inst 0xf8b44ad8 // rprfm pldmany, x20, [x22]\n" ".inst 0xc0042c80 // mova za.d[x9, #0], { z4.d-z7.d }\n" ".inst 0xc0042d81 // mova za.d[x9, #1], { z12.d-z15.d }\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\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" "ble 24f\n" "23:" // Width 4: Multiply loop: Main loop head "whilelt p0.s, XZR, x23\n" - ".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n" "sub x23, x23, #0x4\n" "ld1rqw { z3.s }, p0/Z, [x22]\n" "cmp x23, #0x4\n" "add x22, x22, #0x10\n" - ".inst 0xa041c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xa041c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n" ".inst 0xc153a180 // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[0]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153a281 // fmla za.s[x9, 1], { z20.s-z23.s }, z3.s[0]\n" - ".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c779 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27]\n" ".inst 0xc153a202 // fmla za.s[x9, 2], { z16.s-z19.s }, z3.s[0]\n" - ".inst 0xa041c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xa041c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" ".inst 0xc153a103 // fmla za.s[x9, 3], { z8.s-z11.s }, z3.s[0]\n" - ".inst 0xa042c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xa042c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043c765 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n" ".inst 0xc153a700 // fmla za.s[x9, 0], { z24.s-z27.s }, z3.s[1]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153a581 // fmla za.s[x9, 1], { z12.s-z15.s }, z3.s[1]\n" - ".inst 0xa040c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c779 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27]\n" ".inst 0xc153a502 // fmla za.s[x9, 2], { z8.s-z11.s }, z3.s[1]\n" - ".inst 0xa041c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xa041c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" ".inst 0xc153a483 // fmla za.s[x9, 3], { z4.s-z7.s }, z3.s[1]\n" - ".inst 0xa042c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xa042c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043c765 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n" ".inst 0xc153ab00 // fmla za.s[x9, 0], { z24.s-z27.s }, z3.s[2]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153a901 // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[2]\n" - ".inst 0xa040c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27]\n" ".inst 0xc153aa02 // fmla za.s[x9, 2], { z16.s-z19.s }, z3.s[2]\n" - ".inst 0xa041c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" + ".inst 0xa041c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" ".inst 0xc153a883 // fmla za.s[x9, 3], { z4.s-z7.s }, z3.s[2]\n" - ".inst 0xa042c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xa042c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n" ".inst 0xc153ad00 // fmla za.s[x9, 0], { z8.s-z11.s }, z3.s[3]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153af81 // fmla za.s[x9, 1], { z28.s-z31.s }, z3.s[3]\n" ".inst 0xc153ad82 // fmla za.s[x9, 2], { z12.s-z15.s }, z3.s[3]\n" ".inst 0xc153ae83 // fmla za.s[x9, 3], { z20.s-z23.s }, z3.s[3]\n" "bgt 23b\n" "24:" // Width 4: Multiply loop: Single iteration only "whilelt p0.s, XZR, x23\n" - ".inst 0xa040c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" "ld1rqw { z3.s }, p0/Z, [x22]\n" - "add x22, x22, #0x10\n" - ".inst 0xa041c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c75d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xa041c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c77d // ldnt1w { z28.s-z31.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n" ".inst 0xc153a200 // fmla za.s[x9, 0], { z16.s-z19.s }, z3.s[0]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153a181 // fmla za.s[x9, 1], { z12.s-z15.s }, z3.s[0]\n" ".inst 0xc153a382 // fmla za.s[x9, 2], { z28.s-z31.s }, z3.s[0]\n" ".inst 0xc153a283 // fmla za.s[x9, 3], { z20.s-z23.s }, z3.s[0]\n" "ble 25f\n" - ".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" - ".inst 0xa041c745 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c759 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xa041c765 // ldnt1w { z4.s-z7.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c779 // ldnt1w { z24.s-z27.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n" ".inst 0xc153a580 // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[1]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153a481 // fmla za.s[x9, 1], { z4.s-z7.s }, z3.s[1]\n" ".inst 0xc153a702 // fmla za.s[x9, 2], { z24.s-z27.s }, z3.s[1]\n" ".inst 0xc153a683 // fmla za.s[x9, 3], { z20.s-z23.s }, z3.s[1]\n" "ble 25f\n" - ".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n" + ".inst 0xa040c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n" "subs x23, x23, #0x1\n" - ".inst 0xa041c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xa041c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n" ".inst 0xc153a980 // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[2]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153a901 // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[2]\n" ".inst 0xc153aa82 // fmla za.s[x9, 2], { z20.s-z23.s }, z3.s[2]\n" ".inst 0xc153aa03 // fmla za.s[x9, 3], { z16.s-z19.s }, z3.s[2]\n" "ble 25f\n" - ".inst 0xa040c74d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x26]\n" - ".inst 0xa041c749 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xa042c755 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xa043c751 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x26, #0xc, MUL VL]\n" + ".inst 0xa040c76d // ldnt1w { z12.s-z15.s }, pn9.b/Z, [x27]\n" + ".inst 0xa041c769 // ldnt1w { z8.s-z11.s }, pn9.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042c775 // ldnt1w { z20.s-z23.s }, pn9.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043c771 // ldnt1w { z16.s-z19.s }, pn9.b/Z, [x27, #0xc, MUL VL]\n" ".inst 0xc153ad80 // fmla za.s[x9, 0], { z12.s-z15.s }, z3.s[3]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" ".inst 0xc153ad01 // fmla za.s[x9, 1], { z8.s-z11.s }, z3.s[3]\n" ".inst 0xc153ae82 // fmla za.s[x9, 2], { z20.s-z23.s }, z3.s[3]\n" ".inst 0xc153ae03 // fmla za.s[x9, 3], { z16.s-z19.s }, z3.s[3]\n" @@ -530,37 +516,36 @@ void kai_run_matmul_clamp_f32_f32_f32p16vlx1b_1x16vl_sme2_mla( ".inst 0xc1b4caa0 // fclamp { z0.s-z3.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 0xa060c704 // st1w { z4.s-z7.s }, pn9.b, [x24]\n" - ".inst 0xa061c700 // st1w { z0.s-z3.s }, pn9.b, [x24, #0x4, MUL VL]\n" - ".inst 0xa062c70c // st1w { z12.s-z15.s }, pn9.b, [x24, #0x8, MUL VL]\n" - ".inst 0xa063c310 // st1w { z16.s-z19.s }, p8, [x24, #0xc, MUL VL]\n" - "addvl x24, x24, #16\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 27f\n" "26:" // Width 4: No activation ".inst 0xc0062c0c // mova { z12.d-z15.d }, za.d[x9, #0]\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 0xc0062c64 // mova { z4.d-z7.d }, za.d[x9, #3]\n" - ".inst 0xa060c70c // st1w { z12.s-z15.s }, pn9.b, [x24]\n" - ".inst 0xa061c700 // st1w { z0.s-z3.s }, pn9.b, [x24, #0x4, MUL VL]\n" - ".inst 0xa062c710 // st1w { z16.s-z19.s }, pn9.b, [x24, #0x8, MUL VL]\n" - ".inst 0xa063c304 // st1w { z4.s-z7.s }, p8, [x24, #0xc, MUL VL]\n" - "addvl x24, x24, #16\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" "27:" // Width 4: Output done - "subs x25, x25, #0x4\n" - "sub %x[N], %x[N], x27, LSL #2\n" + "subs x24, x24, #0x4\n" + "sub %x[N], %x[N], x26, LSL #2\n" "bgt 4b\n" "28:" // Exit ".inst 0xd503467f // SMSTOP\n" - "ptrue p8.b\n" : [N] "+&r"(N) : [A_ptr] "r"(A_ptr), [B_ptr] "r"(B_ptr), [K] "r"(K), [args_ptr] "r"(&ka), [flags] "r"(flags), [offset_max] "I"(offsetof(KernelArgs, maxval)), [offset_min] "I"(offsetof(KernelArgs, minval)), [output_ptr] "r"(output_ptr) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", - "p15", "x9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "z0", "z1", "z2", "z3", "z4", "z5", "z6", - "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", - "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"); + : "cc", "memory", "p0", "p1", "p10", "p11", "p12", "p13", "p14", "p15", "p2", "p3", "p4", "p5", "p6", "p7", + "p8", "p9", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x9", "z0", "z1", "z10", "z11", "z12", + "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z2", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", + "z28", "z29", "z3", "z30", "z31", "z4", "z5", "z6", "z7", "z8", "z9"); } #endif // Architectural features check. diff --git a/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c b/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c index 8809e5a6..bf9d9bb4 100644 --- a/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c +++ b/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla.c @@ -92,37 +92,36 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( uint64_t flags = 0; __asm__ __volatile__( - "ptrue p8.b\n" ".inst 0xd503477f // SMSTART ZA\n" "mov x8, #0x0\n" - "cntw x16, ALL, MUL #4\n" - "mov x15, %x[B_ptr]\n" - "add x14, %x[N], x16\n" - "mov x13, %x[output_ptr]\n" - "sub x14, x14, #0x1\n" + "mov x16, %x[B_ptr]\n" + "cntw x15, ALL, MUL #4\n" + "mov x14, %x[output_ptr]\n" + "add x13, %x[N], x15\n" "ptrue p1.b\n" - "udiv x14, x14, x16\n" + "sub x13, x13, #0x1\n" ".inst 0x25207811 // ptrue pn9.b\n" - "add x22, x14, #0x3\n" - "mov x21, #0x1\n" - "and x22, x22, #0xfffffffffffffffc\n" - "mul x22, x22, x16\n" - "mul x22, x22, %x[K]\n" - "lsl x22, x22, #0x2\n" + "udiv x13, x13, x15\n" + "mov x22, #0x1\n" + "add x21, x13, #0x3\n" + "and x21, x21, #0xfffffffffffffffc\n" + "mul x21, x21, x15\n" + "mul x21, x21, %x[K]\n" + "lsl x21, x21, #0x2\n" "1:" // RHS size check loop - "cmp x22, #0x200000\n" + "cmp x21, #0x200000\n" "blt 2f\n" - "tbnz x22, #0, 3f\n" - "lsr x22, x22, #0x1\n" - "lsl x21, x21, #0x1\n" + "tbnz x21, #0, 3f\n" + "lsr x21, x21, #0x1\n" + "lsl x22, x22, #0x1\n" "b 1b\n" "2:" // RHS do prefetch - "lsl x20, x22, #0x26\n" - "sub x21, x21, #0x1\n" - "lsl x21, x21, #0x16\n" - "orr x22, x22, x20\n" - "orr x22, x22, x21\n" - ".inst 0xf8b649fa // rprfm pldonce, x22, [x15]\n" + "lsl x20, x21, #0x26\n" + "sub x22, x22, #0x1\n" + "lsl x22, x22, #0x16\n" + "orr x21, x21, x20\n" + "orr x21, x21, x22\n" + ".inst 0xf8b54a1a // rprfm pldonce, x21, [x16]\n" "3:" // RHS prefetch exit "mov x12, %x[K]\n" "cntw x20, ALL, MUL #2\n" @@ -130,17 +129,17 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "add x12, x12, #0x4\n" "mul x12, x12, x20\n" "4:" // Column loop - "cmp x14, #0x4\n" + "cmp x13, #0x4\n" "bge 22f\n" - "cmp x14, #0x2\n" + "cmp x13, #0x2\n" "bgt 16f\n" "beq 10f\n" "cntw x20, ALL, MUL #2\n" - "add x22, x15, x12\n" - ".inst 0xa04045f4 // ld1w { z20.s-z21.s }, pn9.b/Z, [x15]\n" + "add x22, x16, x12\n" + ".inst 0xa0404614 // ld1w { z20.s-z21.s }, pn9.b/Z, [x16]\n" "cmp %x[N], x20\n" "mov x11, %x[K]\n" - "csel x22, x22, x15, GT\n" + "csel x22, x22, x16, GT\n" "mov x21, %x[N]\n" ".inst 0xa04046d6 // ld1w { z22.s-z23.s }, pn9.b/Z, [x22]\n" "mov x10, %x[A_ptr]\n" @@ -148,31 +147,31 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0x25b567f0 // whilelt p8.s, XZR, x21, VLx4\n" "cmp x11, #0x4\n" ".inst 0xf8b44958 // rprfm pldmany, x20, [x10]\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" "addvl x22, x22, #2\n" ".inst 0xc0040e80 // mova za.d[x8, #0], { z20.d-z23.d }\n" "ble 6f\n" "5:" // Width 1: Multiply loop: Main loop head "whilelt p0.s, XZR, x11\n" - ".inst 0xa04045e5 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404605 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" "ld1rqw { z15.s }, p0/Z, [x10]\n" "sub x11, x11, #0x4\n" "add x10, x10, #0x10\n" ".inst 0xa04046c7 // ldnt1w { z6.s-z7.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" "cmp x11, #0x4\n" - ".inst 0xa04045fd // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa040461d // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa04046df // ldnt1w { z30.s-z31.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xc15f8080 // fmla za.s[x8, 0], { z4.s-z7.s }, z15.s[0]\n" - ".inst 0xa04045e1 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404601 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa04046c3 // ldnt1w { z2.s-z3.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" - ".inst 0xa04045f5 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404615 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa04046d7 // ldnt1w { z22.s-z23.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xc15f8780 // fmla za.s[x8, 0], { z28.s-z31.s }, z15.s[1]\n" @@ -181,33 +180,30 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "bgt 5b\n" "6:" // Width 1: Multiply loop: Single iteration only "whilelt p0.s, XZR, x11\n" - ".inst 0xa04045e1 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x15]\n" + ".inst 0xa0404601 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" "ld1rqw { z8.s }, p0/Z, [x10]\n" - "add x10, x10, #0x10\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa04046c3 // ldnt1w { z2.s-z3.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xc1588000 // fmla za.s[x8, 0], { z0.s-z3.s }, z8.s[0]\n" "ble 7f\n" - ".inst 0xa04045f1 // ldnt1w { z16.s-z17.s }, pn9.b/Z, [x15]\n" + ".inst 0xa0404611 // ldnt1w { z16.s-z17.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa04046d3 // ldnt1w { z18.s-z19.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xc1588600 // fmla za.s[x8, 0], { z16.s-z19.s }, z8.s[1]\n" "ble 7f\n" - ".inst 0xa04045f5 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x15]\n" + ".inst 0xa0404615 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa04046d7 // ldnt1w { z22.s-z23.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xc1588a80 // fmla za.s[x8, 0], { z20.s-z23.s }, z8.s[2]\n" "ble 7f\n" - ".inst 0xa04045ed // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa040460d // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x16]\n" ".inst 0xa04046cf // ldnt1w { z14.s-z15.s }, pn9.b/Z, [x22]\n" - "addvl x22, x22, #2\n" ".inst 0xc1588d80 // fmla za.s[x8, 0], { z12.s-z15.s }, z8.s[3]\n" "7:" // Width 1: Multiply loop: multiply skip "tbz %x[flags], #1, 8f\n" @@ -217,34 +213,32 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "ld1rw { z21.s }, p1/Z, [x21]\n" "ld1rw { z29.s }, p1/Z, [x20]\n" ".inst 0xc1bdcaa8 // fclamp { z8.s-z11.s }, z21.s, z29.s\n" - ".inst 0xa060c1a8 // st1w { z8.s-z11.s }, p8, [x13]\n" - "addvl x13, x13, #4\n" + ".inst 0xa060c1c8 // st1w { z8.s-z11.s }, p8, [x14]\n" "b 9f\n" "8:" // Width 1: No activation ".inst 0xc0060c08 // mova { z8.d-z11.d }, za.d[x8, #0]\n" - ".inst 0xa060c1a8 // st1w { z8.s-z11.s }, p8, [x13]\n" - "addvl x13, x13, #4\n" + ".inst 0xa060c1c8 // st1w { z8.s-z11.s }, p8, [x14]\n" "9:" // Width 1: Output done "b 28f\n" "10:" // Width 2 - "add x24, x15, x12, LSL #1\n" + "add x24, x16, x12, LSL #1\n" "cntw x20, ALL, MUL #6\n" - ".inst 0xa04045e4 // ld1w { z4.s-z5.s }, pn9.b/Z, [x15]\n" + ".inst 0xa0404604 // ld1w { z4.s-z5.s }, pn9.b/Z, [x16]\n" "add x23, x24, x12\n" "cmp %x[N], x20\n" ".inst 0xa0404700 // ld1w { z0.s-z1.s }, pn9.b/Z, [x24]\n" - "add x22, x15, x12\n" - "csel x23, x23, x15, GT\n" + "add x22, x16, x12\n" + "csel x23, x23, x16, GT\n" ".inst 0xa04046c6 // ld1w { z6.s-z7.s }, pn9.b/Z, [x22]\n" "mov x11, %x[K]\n" - "sub x21, %x[N], x16\n" + "sub x21, %x[N], x15\n" ".inst 0xa04046e2 // ld1w { z2.s-z3.s }, pn9.b/Z, [x23]\n" "mov x10, %x[A_ptr]\n" "lsl x20, %x[K], #0x2\n" ".inst 0x25b567f0 // whilelt p8.s, XZR, x21, VLx4\n" "cmp x11, #0x4\n" ".inst 0xf8b44958 // rprfm pldmany, x20, [x10]\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xc0040c80 // mova za.d[x8, #0], { z4.d-z7.d }\n" "addvl x22, x22, #2\n" "addvl x24, x24, #2\n" @@ -253,8 +247,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "ble 12f\n" "11:" // Width 2: Multiply loop: Main loop head "whilelt p0.s, XZR, x11\n" - ".inst 0xa04045e5 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404605 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" "ld1rqw { z0.s }, p0/Z, [x10]\n" "sub x11, x11, #0x4\n" "add x10, x10, #0x10\n" @@ -266,8 +260,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xa04046f7 // ldnt1w { z22.s-z23.s }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" ".inst 0xc1508080 // fmla za.s[x8, 0], { z4.s-z7.s }, z0.s[0]\n" - ".inst 0xa04045f9 // ldnt1w { z24.s-z25.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404619 // ldnt1w { z24.s-z25.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa04046db // ldnt1w { z26.s-z27.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xc1508281 // fmla za.s[x8, 1], { z20.s-z23.s }, z0.s[0]\n" @@ -276,8 +270,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xa04046eb // ldnt1w { z10.s-z11.s }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" ".inst 0xc1508700 // fmla za.s[x8, 0], { z24.s-z27.s }, z0.s[1]\n" - ".inst 0xa04045fd // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa040461d // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa04046df // ldnt1w { z30.s-z31.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xc1508501 // fmla za.s[x8, 1], { z8.s-z11.s }, z0.s[1]\n" @@ -286,8 +280,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xa04046eb // ldnt1w { z10.s-z11.s }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" ".inst 0xc1508b80 // fmla za.s[x8, 0], { z28.s-z31.s }, z0.s[2]\n" - ".inst 0xa04045f9 // ldnt1w { z24.s-z25.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404619 // ldnt1w { z24.s-z25.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa04046db // ldnt1w { z26.s-z27.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xc1508901 // fmla za.s[x8, 1], { z8.s-z11.s }, z0.s[2]\n" @@ -300,11 +294,10 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "bgt 11b\n" "12:" // Width 2: Multiply loop: Single iteration only "whilelt p0.s, XZR, x11\n" - ".inst 0xa04045e5 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x15]\n" + ".inst 0xa0404605 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" "ld1rqw { z8.s }, p0/Z, [x10]\n" - "add x10, x10, #0x10\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa04046c7 // ldnt1w { z6.s-z7.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xa0404715 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x24]\n" @@ -314,9 +307,9 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1588080 // fmla za.s[x8, 0], { z4.s-z7.s }, z8.s[0]\n" ".inst 0xc1588281 // fmla za.s[x8, 1], { z20.s-z23.s }, z8.s[0]\n" "ble 13f\n" - ".inst 0xa04045ed // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x15]\n" + ".inst 0xa040460d // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa04046cf // ldnt1w { z14.s-z15.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xa040471d // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x24]\n" @@ -326,9 +319,9 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1588580 // fmla za.s[x8, 0], { z12.s-z15.s }, z8.s[1]\n" ".inst 0xc1588781 // fmla za.s[x8, 1], { z28.s-z31.s }, z8.s[1]\n" "ble 13f\n" - ".inst 0xa04045fd // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x15]\n" + ".inst 0xa040461d // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa04046df // ldnt1w { z30.s-z31.s }, pn9.b/Z, [x22]\n" "addvl x22, x22, #2\n" ".inst 0xa0404701 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x24]\n" @@ -338,14 +331,10 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1588b80 // fmla za.s[x8, 0], { z28.s-z31.s }, z8.s[2]\n" ".inst 0xc1588801 // fmla za.s[x8, 1], { z0.s-z3.s }, z8.s[2]\n" "ble 13f\n" - ".inst 0xa04045f5 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404615 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x16]\n" ".inst 0xa04046d7 // ldnt1w { z22.s-z23.s }, pn9.b/Z, [x22]\n" - "addvl x22, x22, #2\n" ".inst 0xa0404701 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x24]\n" - "addvl x24, x24, #2\n" ".inst 0xa04046e3 // ldnt1w { z2.s-z3.s }, pn9.b/Z, [x23]\n" - "addvl x23, x23, #2\n" ".inst 0xc1588e80 // fmla za.s[x8, 0], { z20.s-z23.s }, z8.s[3]\n" ".inst 0xc1588c01 // fmla za.s[x8, 1], { z0.s-z3.s }, z8.s[3]\n" "13:" // Width 2: Multiply loop: multiply skip @@ -358,36 +347,34 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "ld1rw { z9.s }, p1/Z, [x20]\n" ".inst 0xc1a9ca3c // fclamp { z28.s-z31.s }, z17.s, z9.s\n" ".inst 0xc1a9ca24 // fclamp { z4.s-z7.s }, z17.s, z9.s\n" - ".inst 0xa060c5bc // st1w { z28.s-z31.s }, pn9.b, [x13]\n" - ".inst 0xa061c1a4 // st1w { z4.s-z7.s }, p8, [x13, #0x4, MUL VL]\n" - "addvl x13, x13, #8\n" + ".inst 0xa060c5dc // st1w { z28.s-z31.s }, pn9.b, [x14]\n" + ".inst 0xa061c1c4 // st1w { z4.s-z7.s }, p8, [x14, #0x4, MUL VL]\n" "b 15f\n" "14:" // Width 2: No activation ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" ".inst 0xc0060c3c // mova { z28.d-z31.d }, za.d[x8, #1]\n" - ".inst 0xa060c5a4 // st1w { z4.s-z7.s }, pn9.b, [x13]\n" - ".inst 0xa061c1bc // st1w { z28.s-z31.s }, p8, [x13, #0x4, MUL VL]\n" - "addvl x13, x13, #8\n" + ".inst 0xa060c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14]\n" + ".inst 0xa061c1dc // st1w { z28.s-z31.s }, p8, [x14, #0x4, MUL VL]\n" "15:" // Width 2: Output done "b 28f\n" "16:" // Width 3 - "add x26, x15, x12, LSL #2\n" + "add x26, x16, x12, LSL #2\n" "cntw x20, ALL, MUL #10\n" - ".inst 0xa04045f4 // ld1w { z20.s-z21.s }, pn9.b/Z, [x15]\n" - "add x25, x15, x12, LSL #1\n" + ".inst 0xa0404614 // ld1w { z20.s-z21.s }, pn9.b/Z, [x16]\n" + "add x25, x16, x12, LSL #1\n" "add x24, x26, x12\n" ".inst 0xa0404740 // ld1w { z0.s-z1.s }, pn9.b/Z, [x26]\n" "cmp %x[N], x20\n" - "add x23, x15, x12\n" + "add x23, x16, x12\n" ".inst 0xa0404730 // ld1w { z16.s-z17.s }, pn9.b/Z, [x25]\n" "add x22, x25, x12\n" - "csel x24, x24, x15, GT\n" + "csel x24, x24, x16, GT\n" ".inst 0xa04046f6 // ld1w { z22.s-z23.s }, pn9.b/Z, [x23]\n" "mov x20, #0x2\n" ".inst 0xa04046d2 // ld1w { z18.s-z19.s }, pn9.b/Z, [x22]\n" "mov x11, %x[K]\n" ".inst 0xa0404702 // ld1w { z2.s-z3.s }, pn9.b/Z, [x24]\n" - "msub x21, x16, x20, %x[N]\n" + "msub x21, x15, x20, %x[N]\n" "mov x10, %x[A_ptr]\n" "lsl x20, %x[K], #0x2\n" ".inst 0x25b567f0 // whilelt p8.s, XZR, x21, VLx4\n" @@ -395,7 +382,7 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "cmp x11, #0x4\n" ".inst 0xf8b44958 // rprfm pldmany, x20, [x10]\n" ".inst 0xc0040e01 // mova za.d[x8, #1], { z16.d-z19.d }\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" "addvl x23, x23, #2\n" ".inst 0xc0040c02 // mova za.d[x8, #2], { z0.d-z3.d }\n" "addvl x25, x25, #2\n" @@ -405,8 +392,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "ble 18f\n" "17:" // Width 3: Multiply loop: Main loop head "whilelt p0.s, XZR, x11\n" - ".inst 0xa04045ed // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa040460d // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" "ld1rqw { z3.s }, p0/Z, [x10]\n" "sub x11, x11, #0x4\n" "add x10, x10, #0x10\n" @@ -423,8 +410,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xa0404713 // ldnt1w { z18.s-z19.s }, pn9.b/Z, [x24]\n" "addvl x24, x24, #2\n" ".inst 0xc1538101 // fmla za.s[x8, 1], { z8.s-z11.s }, z3.s[0]\n" - ".inst 0xa04045e9 // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404609 // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa04046eb // ldnt1w { z10.s-z11.s }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" ".inst 0xc1538202 // fmla za.s[x8, 2], { z16.s-z19.s }, z3.s[0]\n" @@ -438,8 +425,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xa0404707 // ldnt1w { z6.s-z7.s }, pn9.b/Z, [x24]\n" "addvl x24, x24, #2\n" ".inst 0xc1538601 // fmla za.s[x8, 1], { z16.s-z19.s }, z3.s[1]\n" - ".inst 0xa04045e9 // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404609 // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa04046eb // ldnt1w { z10.s-z11.s }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" ".inst 0xc1538482 // fmla za.s[x8, 2], { z4.s-z7.s }, z3.s[1]\n" @@ -453,8 +440,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xa0404707 // ldnt1w { z6.s-z7.s }, pn9.b/Z, [x24]\n" "addvl x24, x24, #2\n" ".inst 0xc1538a01 // fmla za.s[x8, 1], { z16.s-z19.s }, z3.s[2]\n" - ".inst 0xa04045f5 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404615 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa04046f7 // ldnt1w { z22.s-z23.s }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" ".inst 0xc1538882 // fmla za.s[x8, 2], { z4.s-z7.s }, z3.s[2]\n" @@ -472,11 +459,10 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "bgt 17b\n" "18:" // Width 3: Multiply loop: Single iteration only "whilelt p0.s, XZR, x11\n" - ".inst 0xa04045e5 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x15]\n" + ".inst 0xa0404605 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" "ld1rqw { z8.s }, p0/Z, [x10]\n" - "add x10, x10, #0x10\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa04046e7 // ldnt1w { z6.s-z7.s }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" ".inst 0xa040473d // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x25]\n" @@ -491,9 +477,9 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1588381 // fmla za.s[x8, 1], { z28.s-z31.s }, z8.s[0]\n" ".inst 0xc1588282 // fmla za.s[x8, 2], { z20.s-z23.s }, z8.s[0]\n" "ble 19f\n" - ".inst 0xa04045ed // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x15]\n" + ".inst 0xa040460d // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa04046ef // ldnt1w { z14.s-z15.s }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" ".inst 0xa0404725 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x25]\n" @@ -508,9 +494,9 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1588481 // fmla za.s[x8, 1], { z4.s-z7.s }, z8.s[1]\n" ".inst 0xc1588602 // fmla za.s[x8, 2], { z16.s-z19.s }, z8.s[1]\n" "ble 19f\n" - ".inst 0xa04045e1 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x15]\n" + ".inst 0xa0404601 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa04046e3 // ldnt1w { z2.s-z3.s }, pn9.b/Z, [x23]\n" "addvl x23, x23, #2\n" ".inst 0xa040472d // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x25]\n" @@ -525,19 +511,13 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1588981 // fmla za.s[x8, 1], { z12.s-z15.s }, z8.s[2]\n" ".inst 0xc1588a02 // fmla za.s[x8, 2], { z16.s-z19.s }, z8.s[2]\n" "ble 19f\n" - ".inst 0xa04045e5 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404605 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x16]\n" ".inst 0xa04046e7 // ldnt1w { z6.s-z7.s }, pn9.b/Z, [x23]\n" - "addvl x23, x23, #2\n" ".inst 0xa040472d // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x25]\n" - "addvl x25, x25, #2\n" ".inst 0xa04046cf // ldnt1w { z14.s-z15.s }, pn9.b/Z, [x22]\n" - "addvl x22, x22, #2\n" ".inst 0xa0404755 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x26]\n" ".inst 0xc1588c80 // fmla za.s[x8, 0], { z4.s-z7.s }, z8.s[3]\n" - "addvl x26, x26, #2\n" ".inst 0xa0404717 // ldnt1w { z22.s-z23.s }, pn9.b/Z, [x24]\n" - "addvl x24, x24, #2\n" ".inst 0xc1588d81 // fmla za.s[x8, 1], { z12.s-z15.s }, z8.s[3]\n" ".inst 0xc1588e82 // fmla za.s[x8, 2], { z20.s-z23.s }, z8.s[3]\n" "19:" // Width 3: Multiply loop: multiply skip @@ -552,42 +532,40 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".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 0xa060c5a8 // st1w { z8.s-z11.s }, pn9.b, [x13]\n" - ".inst 0xa061c5ac // st1w { z12.s-z15.s }, pn9.b, [x13, #0x4, MUL VL]\n" - ".inst 0xa062c1b0 // st1w { z16.s-z19.s }, p8, [x13, #0x8, MUL VL]\n" - "addvl x13, x13, #12\n" + ".inst 0xa060c5c8 // st1w { z8.s-z11.s }, pn9.b, [x14]\n" + ".inst 0xa061c5cc // st1w { z12.s-z15.s }, pn9.b, [x14, #0x4, MUL VL]\n" + ".inst 0xa062c1d0 // st1w { z16.s-z19.s }, p8, [x14, #0x8, MUL VL]\n" "b 21f\n" "20:" // Width 3: No activation ".inst 0xc0060c04 // mova { z4.d-z7.d }, za.d[x8, #0]\n" ".inst 0xc0060c20 // mova { z0.d-z3.d }, za.d[x8, #1]\n" ".inst 0xc0060c50 // mova { z16.d-z19.d }, za.d[x8, #2]\n" - ".inst 0xa060c5a4 // st1w { z4.s-z7.s }, pn9.b, [x13]\n" - ".inst 0xa061c5a0 // st1w { z0.s-z3.s }, pn9.b, [x13, #0x4, MUL VL]\n" - ".inst 0xa062c1b0 // st1w { z16.s-z19.s }, p8, [x13, #0x8, MUL VL]\n" - "addvl x13, x13, #12\n" + ".inst 0xa060c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14]\n" + ".inst 0xa061c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n" + ".inst 0xa062c1d0 // st1w { z16.s-z19.s }, p8, [x14, #0x8, MUL VL]\n" "21:" // Width 3: Output done "b 28f\n" "22:" // Width 4 - "add x9, x15, x12, LSL #2\n" + "add x9, x16, x12, LSL #2\n" "cntw x20, ALL, MUL #14\n" - ".inst 0xa04045ec // ld1w { z12.s-z13.s }, pn9.b/Z, [x15]\n" + ".inst 0xa040460c // ld1w { z12.s-z13.s }, pn9.b/Z, [x16]\n" "add x28, x9, x12, LSL #1\n" - "add x27, x15, x12, LSL #1\n" + "add x27, x16, x12, LSL #1\n" ".inst 0xa0404528 // ld1w { z8.s-z9.s }, pn9.b/Z, [x9]\n" "add x26, x28, x12\n" "cmp %x[N], x20\n" ".inst 0xa0404760 // ld1w { z0.s-z1.s }, pn9.b/Z, [x27]\n" - "add x25, x15, x12\n" + "add x25, x16, x12\n" "add x24, x27, x12\n" ".inst 0xa0404790 // ld1w { z16.s-z17.s }, pn9.b/Z, [x28]\n" "add x23, x9, x12\n" - "csel x26, x26, x15, GT\n" + "csel x26, x26, x16, GT\n" ".inst 0xa040472e // ld1w { z14.s-z15.s }, pn9.b/Z, [x25]\n" "mov x20, #0x3\n" ".inst 0xa0404702 // ld1w { z2.s-z3.s }, pn9.b/Z, [x24]\n" "mov x11, %x[K]\n" ".inst 0xa04046ea // ld1w { z10.s-z11.s }, pn9.b/Z, [x23]\n" - "msub x21, x16, x20, %x[N]\n" + "msub x21, x15, x20, %x[N]\n" "mov x10, %x[A_ptr]\n" ".inst 0xa0404752 // ld1w { z18.s-z19.s }, pn9.b/Z, [x26]\n" "lsl x20, %x[K], #0x2\n" @@ -596,8 +574,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "cmp x11, #0x4\n" ".inst 0xf8b44958 // rprfm pldmany, x20, [x10]\n" ".inst 0xc0040c01 // mova za.d[x8, #1], { z0.d-z3.d }\n" - "add x22, x15, x12, LSL #3\n" - "addvl x15, x15, #2\n" + "add x22, x16, x12, LSL #3\n" + "addvl x16, x16, #2\n" ".inst 0xc0040d02 // mova za.d[x8, #2], { z8.d-z11.d }\n" "addvl x25, x25, #2\n" "addvl x27, x27, #2\n" @@ -610,8 +588,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "ble 24f\n" "23:" // Width 4: Multiply loop: Main loop head "whilelt p0.s, XZR, x11\n" - ".inst 0xa04045e9 // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404609 // ldnt1w { z8.s-z9.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" "ld1rqw { z13.s }, p0/Z, [x10]\n" "sub x11, x11, #0x4\n" "add x10, x10, #0x10\n" @@ -633,8 +611,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xa0404743 // ldnt1w { z2.s-z3.s }, pn9.b/Z, [x26]\n" "addvl x26, x26, #2\n" ".inst 0xc15d8202 // fmla za.s[x8, 2], { z16.s-z19.s }, z13.s[0]\n" - ".inst 0xa04045fd // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa040461d // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa040473f // ldnt1w { z30.s-z31.s }, pn9.b/Z, [x25]\n" "addvl x25, x25, #2\n" ".inst 0xc15d8003 // fmla za.s[x8, 3], { z0.s-z3.s }, z13.s[0]\n" @@ -653,8 +631,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xa0404753 // ldnt1w { z18.s-z19.s }, pn9.b/Z, [x26]\n" "addvl x26, x26, #2\n" ".inst 0xc15d8502 // fmla za.s[x8, 2], { z8.s-z11.s }, z13.s[1]\n" - ".inst 0xa04045e5 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404605 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa0404727 // ldnt1w { z6.s-z7.s }, pn9.b/Z, [x25]\n" "addvl x25, x25, #2\n" ".inst 0xc15d8603 // fmla za.s[x8, 3], { z16.s-z19.s }, z13.s[1]\n" @@ -673,8 +651,8 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xa0404753 // ldnt1w { z18.s-z19.s }, pn9.b/Z, [x26]\n" "addvl x26, x26, #2\n" ".inst 0xc15d8902 // fmla za.s[x8, 2], { z8.s-z11.s }, z13.s[2]\n" - ".inst 0xa04045f5 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404615 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x16]\n" + "addvl x16, x16, #2\n" ".inst 0xa0404737 // ldnt1w { z22.s-z23.s }, pn9.b/Z, [x25]\n" "addvl x25, x25, #2\n" ".inst 0xc15d8a03 // fmla za.s[x8, 3], { z16.s-z19.s }, z13.s[2]\n" @@ -697,11 +675,10 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( "bgt 23b\n" "24:" // Width 4: Multiply loop: Single iteration only "whilelt p0.s, XZR, x11\n" - ".inst 0xa04045e5 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x15]\n" + ".inst 0xa0404605 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" "ld1rqw { z8.s }, p0/Z, [x10]\n" - "add x10, x10, #0x10\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa0404727 // ldnt1w { z6.s-z7.s }, pn9.b/Z, [x25]\n" "addvl x25, x25, #2\n" ".inst 0xa0404761 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x27]\n" @@ -721,9 +698,9 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1588182 // fmla za.s[x8, 2], { z12.s-z15.s }, z8.s[0]\n" ".inst 0xc1588203 // fmla za.s[x8, 3], { z16.s-z19.s }, z8.s[0]\n" "ble 25f\n" - ".inst 0xa04045fd // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x15]\n" + ".inst 0xa040461d // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa040473f // ldnt1w { z30.s-z31.s }, pn9.b/Z, [x25]\n" "addvl x25, x25, #2\n" ".inst 0xa0404761 // ldnt1w { z0.s-z1.s }, pn9.b/Z, [x27]\n" @@ -743,9 +720,9 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1588482 // fmla za.s[x8, 2], { z4.s-z7.s }, z8.s[1]\n" ".inst 0xc1588603 // fmla za.s[x8, 3], { z16.s-z19.s }, z8.s[1]\n" "ble 25f\n" - ".inst 0xa04045fd // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x15]\n" + ".inst 0xa040461d // ldnt1w { z28.s-z29.s }, pn9.b/Z, [x16]\n" "subs x11, x11, #0x1\n" - "addvl x15, x15, #2\n" + "addvl x16, x16, #2\n" ".inst 0xa040473f // ldnt1w { z30.s-z31.s }, pn9.b/Z, [x25]\n" "addvl x25, x25, #2\n" ".inst 0xa040476d // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x27]\n" @@ -765,24 +742,16 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1588802 // fmla za.s[x8, 2], { z0.s-z3.s }, z8.s[2]\n" ".inst 0xc1588a03 // fmla za.s[x8, 3], { z16.s-z19.s }, z8.s[2]\n" "ble 25f\n" - ".inst 0xa04045e5 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x15]\n" - "addvl x15, x15, #2\n" + ".inst 0xa0404605 // ldnt1w { z4.s-z5.s }, pn9.b/Z, [x16]\n" ".inst 0xa0404727 // ldnt1w { z6.s-z7.s }, pn9.b/Z, [x25]\n" - "addvl x25, x25, #2\n" ".inst 0xa040476d // ldnt1w { z12.s-z13.s }, pn9.b/Z, [x27]\n" - "addvl x27, x27, #2\n" ".inst 0xa040470f // ldnt1w { z14.s-z15.s }, pn9.b/Z, [x24]\n" - "addvl x24, x24, #2\n" ".inst 0xa0404535 // ldnt1w { z20.s-z21.s }, pn9.b/Z, [x9]\n" ".inst 0xc1588c80 // fmla za.s[x8, 0], { z4.s-z7.s }, z8.s[3]\n" - "addvl x9, x9, #2\n" ".inst 0xa04046f7 // ldnt1w { z22.s-z23.s }, pn9.b/Z, [x23]\n" - "addvl x23, x23, #2\n" ".inst 0xa0404791 // ldnt1w { z16.s-z17.s }, pn9.b/Z, [x28]\n" ".inst 0xc1588d81 // fmla za.s[x8, 1], { z12.s-z15.s }, z8.s[3]\n" - "addvl x28, x28, #2\n" ".inst 0xa0404753 // ldnt1w { z18.s-z19.s }, pn9.b/Z, [x26]\n" - "addvl x26, x26, #2\n" ".inst 0xc1588e82 // fmla za.s[x8, 2], { z20.s-z23.s }, z8.s[3]\n" ".inst 0xc1588e03 // fmla za.s[x8, 3], { z16.s-z19.s }, z8.s[3]\n" "25:" // Width 4: Multiply loop: multiply skip @@ -799,39 +768,38 @@ void kai_run_matmul_clamp_f32_f32_f32p2vlx1b_1x16vl_sme2_mla( ".inst 0xc1b4caa0 // fclamp { z0.s-z3.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 0xa060c5a4 // st1w { z4.s-z7.s }, pn9.b, [x13]\n" - ".inst 0xa061c5a0 // st1w { z0.s-z3.s }, pn9.b, [x13, #0x4, MUL VL]\n" - ".inst 0xa062c5ac // st1w { z12.s-z15.s }, pn9.b, [x13, #0x8, MUL VL]\n" - ".inst 0xa063c1b0 // st1w { z16.s-z19.s }, p8, [x13, #0xc, MUL VL]\n" - "addvl x13, x13, #16\n" + ".inst 0xa060c5c4 // st1w { z4.s-z7.s }, pn9.b, [x14]\n" + ".inst 0xa061c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n" + ".inst 0xa062c5cc // st1w { z12.s-z15.s }, pn9.b, [x14, #0x8, MUL VL]\n" + ".inst 0xa063c1d0 // st1w { z16.s-z19.s }, p8, [x14, #0xc, MUL VL]\n" + "addvl x14, x14, #16\n" "b 27f\n" "26:" // Width 4: No activation ".inst 0xc0060c0c // mova { z12.d-z15.d }, za.d[x8, #0]\n" ".inst 0xc0060c20 // mova { z0.d-z3.d }, za.d[x8, #1]\n" ".inst 0xc0060c50 // mova { z16.d-z19.d }, za.d[x8, #2]\n" ".inst 0xc0060c64 // mova { z4.d-z7.d }, za.d[x8, #3]\n" - ".inst 0xa060c5ac // st1w { z12.s-z15.s }, pn9.b, [x13]\n" - ".inst 0xa061c5a0 // st1w { z0.s-z3.s }, pn9.b, [x13, #0x4, MUL VL]\n" - ".inst 0xa062c5b0 // st1w { z16.s-z19.s }, pn9.b, [x13, #0x8, MUL VL]\n" - ".inst 0xa063c1a4 // st1w { z4.s-z7.s }, p8, [x13, #0xc, MUL VL]\n" - "addvl x13, x13, #16\n" + ".inst 0xa060c5cc // st1w { z12.s-z15.s }, pn9.b, [x14]\n" + ".inst 0xa061c5c0 // st1w { z0.s-z3.s }, pn9.b, [x14, #0x4, MUL VL]\n" + ".inst 0xa062c5d0 // st1w { z16.s-z19.s }, pn9.b, [x14, #0x8, MUL VL]\n" + ".inst 0xa063c1c4 // st1w { z4.s-z7.s }, p8, [x14, #0xc, MUL VL]\n" + "addvl x14, x14, #16\n" "27:" // Width 4: Output done - "subs x14, x14, #0x4\n" - "mov x15, x22\n" - "sub %x[N], %x[N], x16, LSL #2\n" + "subs x13, x13, #0x4\n" + "mov x16, x22\n" + "sub %x[N], %x[N], x15, LSL #2\n" "bgt 4b\n" "28:" // Exit ".inst 0xd503467f // SMSTOP\n" - "ptrue p8.b\n" : [N] "+&r"(N) : [A_ptr] "r"(A_ptr), [B_ptr] "r"(B_ptr), [K] "r"(K), [args_ptr] "r"(&ka), [flags] "r"(flags), [offset_max] "I"(offsetof(KernelArgs, maxval)), [offset_min] "I"(offsetof(KernelArgs, minval)), [output_ptr] "r"(output_ptr) - : "cc", "memory", "p0", "p1", "p2", "p3", "p4", "p5", "p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", - "p15", "x8", "x9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", - "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", - "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", - "z29", "z30", "z31"); + : "cc", "memory", "p0", "p1", "p10", "p11", "p12", "p13", "p14", "p15", "p2", "p3", "p4", "p5", "p6", "p7", + "p8", "p9", "x10", "x11", "x12", "x13", "x14", "x15", "x16", "x20", "x21", "x22", "x23", "x24", "x25", "x26", + "x27", "x28", "x8", "x9", "z0", "z1", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", + "z2", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z3", "z30", "z31", "z4", "z5", + "z6", "z7", "z8", "z9"); } #endif // Architectural features check. -- GitLab