diff options
Diffstat (limited to 'src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp')
-rw-r--r-- | src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp | 770 |
1 files changed, 385 insertions, 385 deletions
diff --git a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp index 093feee6ce..1cbaf00052 100644 --- a/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp +++ b/src/core/NEON/kernels/arm_gemm/kernels/sme2_gemv_u8qa_dot_16VL/generic.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2022-2023 Arm Limited. + * Copyright (c) 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -52,20 +52,20 @@ void sme2_gemv_u8qa_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" "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 p2.b\n" + "udiv x26, x26, x28\n" ".inst 0x25207810 // ptrue pn8.b\n" - "mul x22, x22, %x[K]\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" "1:" // RHS size check loop "cmp x22, #0x200000\n" "blt 2f\n" @@ -79,16 +79,16 @@ void sme2_gemv_u8qa_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[col_bias]\n" - "mov z28.s, #0x0\n" - "mov z29.b, #0x1\n" + "mov z11.s, #0x0\n" + "mov z8.b, #0x1\n" "bic %x[flags], %x[flags], #0x80000000\n" "4:" // Column loop - "cmp x27, #0x4\n" + "cmp x26, #0x4\n" "bge 34f\n" - "cmp x27, #0x2\n" + "cmp x26, #0x2\n" "bgt 24f\n" "beq 14f\n" "mov x23, %x[A_ptr]\n" @@ -98,8 +98,8 @@ void sme2_gemv_u8qa_dot_16VL ( ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" "whilelt p1.b, XZR, x20\n" "cbz x24, 5f\n" - ".inst 0xa040c300 // ld1w { z0.s-z3.s }, pn8.b/Z, [x24]\n" - ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n" + ".inst 0xa040c30c // ld1w { z12.s-z15.s }, pn8.b/Z, [x24]\n" + ".inst 0xc0042d80 // mova za.d[x9, #0], { z12.d-z15.d }\n" "b 6f\n" "5:" // Width 1: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -108,82 +108,82 @@ void sme2_gemv_u8qa_dot_16VL ( "ble 9f\n" "7:" // Width 1: Multiply loop: Main loop head "whilelt p0.b, XZR, x22\n" - "ld1rqb { z1.b }, p0/Z, [x23]\n" + ".inst 0xa0408361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + "ld1rqb { z9.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b2b0 // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[0]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b5b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[1]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b9b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151bdb0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0408379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + ".inst 0xa0408375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b030 // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n" + ".inst 0xa040837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b730 // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[1]\n" + ".inst 0xc159bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[2]\n" + ".inst 0xc159bfb0 // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[3]\n" "tbnz %x[flags], #31, 8f\n" - "udot z28.s, z1.b, z29.b\n" + "udot z11.s, z9.b, z8.b\n" "8:" // Width 1: Multiply loop: unique 1: skip row sum "sub x22, x22, #0x10\n" "cmp x22, #0x10\n" "bgt 7b\n" "9:" // Width 1: Multiply loop: Single iteration only "whilelt p0.b, XZR, x22\n" - "ld1rqb { z1.b }, p0/Z, [x23]\n" + ".inst 0xa0408379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + "ld1rqb { z9.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xc151b1b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b330 // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[0]\n" "ble 10f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" + ".inst 0xa0408361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xc151b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[1]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b430 // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[1]\n" "ble 10f\n" - ".inst 0xa0408349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n" + ".inst 0xa040837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xc151b930 // udot za.s[x9, 0], { z8.b-z11.b }, z1.b[2]\n" - "addvl x26, x26, #16\n" + "addvl x27, x27, #16\n" + ".inst 0xc159bbb0 // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n" "ble 10f\n" - ".inst 0xa0408349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151bd30 // udot za.s[x9, 0], { z8.b-z11.b }, z1.b[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0408371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159be30 // udot za.s[x9, 0], { z16.b-z19.b }, z9.b[3]\n" "10:" // Width 1: Multiply loop: multiply skip "tbnz %x[flags], #31, 11f\n" - "udot z28.s, z1.b, z29.b\n" + "udot z11.s, z9.b, z8.b\n" "11:" // Width 1: Multiply loop: unique 2: skip row sum "tbnz %x[flags], #31, 12f\n" "add x21, %x[qp], %[b_offset]\n" "mov x20, #0x4\n" - "ld1rw { z26.s }, p2/Z, [x21]\n" - "neg z26.s, p2/M, z26.s\n" + "ld1rw { z17.s }, p2/Z, [x21]\n" "whilelt p0.s, XZR, x20\n" - "uaddv d28, p0, z28.s\n" - "mov z28.s, z28.s[0]\n" - "mul z28.s, p2/M, z28.s, z26.s\n" "orr %x[flags], %x[flags], #0x80000000\n" + "uaddv d11, p0, z11.s\n" + "mov z11.s, z11.s[0]\n" + "neg z17.s, p2/M, z17.s\n" + "mul z11.s, p2/M, z11.s, z17.s\n" "12:" // Width 1: skip row sum fixup - ".inst 0xc0904b80 // addha za0.s, p2/M, p2/M, z28.s\n" - "add x20, %x[qp], %[per_layer_mul]\n" - "ld1rw { z1.s }, p2/Z, [x20]\n" + ".inst 0xc0904960 // addha za0.s, p2/M, p2/M, z11.s\n" + "add x21, %x[qp], %[per_layer_mul]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" - ".inst 0xc0904b81 // addha za1.s, p2/M, p2/M, z28.s\n" - "ld1rw { z7.s }, p2/Z, [x20]\n" - "add x20, %x[qp], %[c_offset]\n" + ".inst 0xc0904961 // addha za1.s, p2/M, p2/M, z11.s\n" + "ld1rw { z10.s }, p2/Z, [x21]\n" + "add x22, %x[qp], %[c_offset]\n" "add x21, %x[qp], %[minval]\n" - ".inst 0xc0904b82 // addha za2.s, p2/M, p2/M, z28.s\n" - "ld1rw { z2.s }, p2/Z, [x20]\n" + ".inst 0xc0904962 // addha za2.s, p2/M, p2/M, z11.s\n" + "ld1rw { z6.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[maxval]\n" - ".inst 0xc0904b83 // addha za3.s, p2/M, p2/M, z28.s\n" - "ld1rw { z16.s }, p2/Z, [x21]\n" + ".inst 0xc0904963 // addha za3.s, p2/M, p2/M, z11.s\n" + "ld1rw { z7.s }, p2/Z, [x22]\n" + "ld1rw { z30.s }, p2/Z, [x21]\n" + "ld1rw { z23.s }, p2/Z, [x20]\n" ".inst 0xc0062c0c // mova { z12.d-z15.d }, za.d[x9, #0]\n" - ".inst 0xc1a1ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n" - ".inst 0xc1a7aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n" - "ld1rw { z30.s }, p2/Z, [x20]\n" - ".inst 0xc1a2ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n" - ".inst 0xc1bece0c // sclamp { z12.s-z15.s }, z16.s, z30.s\n" + ".inst 0xc1aaac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z10.s\n" + ".inst 0xc1a6aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n" + ".inst 0xc1a7ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n" + ".inst 0xc1b7cfcc // sclamp { z12.s-z15.s }, z30.s, z23.s\n" "uzp1 z12.h, z12.h, z13.h\n" "uzp1 z19.h, z14.h, z15.h\n" "uzp1 z12.b, z12.b, z19.b\n" @@ -199,10 +199,10 @@ void sme2_gemv_u8qa_dot_16VL ( ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" "whilelt p1.b, XZR, x20\n" "cbz x24, 15f\n" - ".inst 0xa040c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24]\n" - ".inst 0xc0042e00 // mova za.d[x9, #0], { z16.d-z19.d }\n" - ".inst 0xa041c318 // ld1w { z24.s-z27.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0042f01 // mova za.d[x9, #1], { z24.d-z27.d }\n" + ".inst 0xa040c304 // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n" + ".inst 0xa041c300 // ld1w { z0.s-z3.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xc0042c80 // mova za.d[x9, #0], { z4.d-z7.d }\n" + ".inst 0xc0042c01 // mova za.d[x9, #1], { z0.d-z3.d }\n" "b 16f\n" "15:" // Width 2: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -211,111 +211,111 @@ void sme2_gemv_u8qa_dot_16VL ( "ble 19f\n" "17:" // Width 2: Multiply loop: Main loop head "whilelt p0.b, XZR, x22\n" - "ld1rqb { z1.b }, p0/Z, [x23]\n" + ".inst 0xa0408375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n" + "ld1rqb { z9.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b1b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n" - ".inst 0xa0418359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b331 // udot za.s[x9, 1], { z24.b-z27.b }, z1.b[0]\n" - "addvl x26, x26, #16\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b630 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n" - ".inst 0xa0418355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b6b1 // udot za.s[x9, 1], { z20.b-z23.b }, z1.b[1]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b9b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n" - ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b9b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n" - "addvl x26, x26, #16\n" - ".inst 0xa0408345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151bcb0 // udot za.s[x9, 0], { z4.b-z7.b }, z1.b[3]\n" - ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151bd31 // udot za.s[x9, 1], { z8.b-z11.b }, z1.b[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xa0408371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27]\n" + ".inst 0xc159b2b0 // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[0]\n" + ".inst 0xa0418375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b331 // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[0]\n" + ".inst 0xa040837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n" + ".inst 0xa0418361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xa040836d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27]\n" + ".inst 0xc159b630 // udot za.s[x9, 0], { z16.b-z19.b }, z9.b[1]\n" + ".inst 0xa0418371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b6b1 // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[1]\n" + ".inst 0xc159bbb0 // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n" + ".inst 0xc159b831 // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n" + ".inst 0xc159bdb0 // udot za.s[x9, 0], { z12.b-z15.b }, z9.b[3]\n" + ".inst 0xc159be31 // udot za.s[x9, 1], { z16.b-z19.b }, z9.b[3]\n" "tbnz %x[flags], #31, 18f\n" - "udot z28.s, z1.b, z29.b\n" + "udot z11.s, z9.b, z8.b\n" "18:" // Width 2: Multiply loop: unique 3: skip row sum "sub x22, x22, #0x10\n" "cmp x22, #0x10\n" "bgt 17b\n" "19:" // Width 2: Multiply loop: Single iteration only "whilelt p0.b, XZR, x22\n" - "ld1rqb { z1.b }, p0/Z, [x23]\n" + ".inst 0xa0408361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xa0408359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n" + "ld1rqb { z9.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xc151b330 // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[0]\n" - ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b231 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b030 // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n" + ".inst 0xc159b331 // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[0]\n" "ble 20f\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" + ".inst 0xa0408371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xc151b6b0 // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[1]\n" - ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b5b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b630 // udot za.s[x9, 0], { z16.b-z19.b }, z9.b[1]\n" + ".inst 0xc159b731 // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[1]\n" "ble 20f\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + ".inst 0xa040837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xc151b9b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[2]\n" - ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b9b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159bbb0 // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n" + ".inst 0xc159b831 // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n" "ble 20f\n" - ".inst 0xa0408359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151bf30 // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[3]\n" - ".inst 0xa0418349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151bd31 // udot za.s[x9, 1], { z8.b-z11.b }, z1.b[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0408375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n" + ".inst 0xa0418379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[3]\n" + ".inst 0xc159bf31 // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[3]\n" "20:" // Width 2: Multiply loop: multiply skip "tbnz %x[flags], #31, 21f\n" - "udot z28.s, z1.b, z29.b\n" + "udot z11.s, z9.b, z8.b\n" "21:" // Width 2: Multiply loop: unique 4: skip row sum "tbnz %x[flags], #31, 22f\n" "add x21, %x[qp], %[b_offset]\n" "mov x20, #0x4\n" - "ld1rw { z16.s }, p2/Z, [x21]\n" - "neg z16.s, p2/M, z16.s\n" + "ld1rw { z1.s }, p2/Z, [x21]\n" "whilelt p0.s, XZR, x20\n" - "uaddv d28, p0, z28.s\n" - "mov z28.s, z28.s[0]\n" - "mul z28.s, p2/M, z28.s, z16.s\n" "orr %x[flags], %x[flags], #0x80000000\n" + "uaddv d11, p0, z11.s\n" + "mov z11.s, z11.s[0]\n" + "neg z1.s, p2/M, z1.s\n" + "mul z11.s, p2/M, z11.s, z1.s\n" "22:" // Width 2: skip row sum fixup - ".inst 0xc0904b80 // addha za0.s, p2/M, p2/M, z28.s\n" - "add x20, %x[qp], %[per_layer_mul]\n" - "ld1rw { z6.s }, p2/Z, [x20]\n" + ".inst 0xc0904960 // addha za0.s, p2/M, p2/M, z11.s\n" + "add x21, %x[qp], %[per_layer_mul]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" - ".inst 0xc0904b81 // addha za1.s, p2/M, p2/M, z28.s\n" - "ld1rw { z5.s }, p2/Z, [x20]\n" - "add x20, %x[qp], %[c_offset]\n" + ".inst 0xc0904961 // addha za1.s, p2/M, p2/M, z11.s\n" + "ld1rw { z2.s }, p2/Z, [x21]\n" + "add x22, %x[qp], %[c_offset]\n" "add x21, %x[qp], %[minval]\n" - ".inst 0xc0904b82 // addha za2.s, p2/M, p2/M, z28.s\n" - "ld1rw { z9.s }, p2/Z, [x20]\n" + ".inst 0xc0904962 // addha za2.s, p2/M, p2/M, z11.s\n" + "ld1rw { z3.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[maxval]\n" - ".inst 0xc0904b83 // addha za3.s, p2/M, p2/M, z28.s\n" - "ld1rw { z16.s }, p2/Z, [x21]\n" - ".inst 0xc0062c18 // mova { z24.d-z27.d }, za.d[x9, #0]\n" - ".inst 0xc1a6ac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n" - ".inst 0xc0062c20 // mova { z0.d-z3.d }, za.d[x9, #1]\n" - ".inst 0xc1a6ac00 // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n" - ".inst 0xc1a5aa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z5.s\n" - "ld1rw { z21.s }, p2/Z, [x20]\n" - ".inst 0xc1a5aa20 // srshl { z0.s-z3.s }, { z0.s-z3.s }, z5.s\n" - ".inst 0xc1a9ab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z9.s\n" - ".inst 0xc1a9ab00 // add { z0.s-z3.s }, { z0.s-z3.s }, z9.s\n" - ".inst 0xc1b5ce18 // sclamp { z24.s-z27.s }, z16.s, z21.s\n" - ".inst 0xc1b5ce00 // sclamp { z0.s-z3.s }, z16.s, z21.s\n" - "uzp1 z24.h, z24.h, z25.h\n" - "uzp1 z9.h, z26.h, z27.h\n" - "uzp1 z0.h, z0.h, z1.h\n" - "uzp1 z26.h, z2.h, z3.h\n" - "uzp1 z24.b, z24.b, z9.b\n" - "st1b { z24.b }, p2, [x25]\n" - "uzp1 z0.b, z0.b, z26.b\n" - "st1b { z0.b }, p1, [x25, #1, MUL VL]\n" + ".inst 0xc0904963 // addha za3.s, p2/M, p2/M, z11.s\n" + "ld1rw { z9.s }, p2/Z, [x22]\n" + "ld1rw { z6.s }, p2/Z, [x21]\n" + "ld1rw { z29.s }, p2/Z, [x20]\n" + ".inst 0xc0062c14 // mova { z20.d-z23.d }, za.d[x9, #0]\n" + ".inst 0xc0062c2c // mova { z12.d-z15.d }, za.d[x9, #1]\n" + ".inst 0xc1a2ac14 // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z2.s\n" + ".inst 0xc1a2ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n" + ".inst 0xc1a3aa34 // srshl { z20.s-z23.s }, { z20.s-z23.s }, z3.s\n" + ".inst 0xc1a3aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z3.s\n" + ".inst 0xc1a9ab14 // add { z20.s-z23.s }, { z20.s-z23.s }, z9.s\n" + ".inst 0xc1a9ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z9.s\n" + ".inst 0xc1bdccd4 // sclamp { z20.s-z23.s }, z6.s, z29.s\n" + ".inst 0xc1bdcccc // sclamp { z12.s-z15.s }, z6.s, z29.s\n" + "uzp1 z20.h, z20.h, z21.h\n" + "uzp1 z16.h, z22.h, z23.h\n" + "uzp1 z12.h, z12.h, z13.h\n" + "uzp1 z24.h, z14.h, z15.h\n" + "uzp1 z20.b, z20.b, z16.b\n" + "uzp1 z12.b, z12.b, z24.b\n" + "st1b { z20.b }, p2, [x25]\n" + "st1b { z12.b }, p1, [x25, #1, MUL VL]\n" "addvl x25, x25, #2\n" "23:" // Width 2: Output done "b 44f\n" @@ -328,12 +328,12 @@ void sme2_gemv_u8qa_dot_16VL ( ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" "whilelt p1.b, XZR, x20\n" "cbz x24, 25f\n" - ".inst 0xa040c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24]\n" - ".inst 0xc0042e00 // mova za.d[x9, #0], { z16.d-z19.d }\n" - ".inst 0xa041c30c // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0042d81 // mova za.d[x9, #1], { z12.d-z15.d }\n" - ".inst 0xa042c318 // ld1w { z24.s-z27.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n" - ".inst 0xc0042f02 // mova za.d[x9, #2], { z24.d-z27.d }\n" + ".inst 0xa040c314 // ld1w { z20.s-z23.s }, pn8.b/Z, [x24]\n" + ".inst 0xa041c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" + ".inst 0xa042c30c // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n" + ".inst 0xc0042e80 // mova za.d[x9, #0], { z20.d-z23.d }\n" + ".inst 0xc0042e01 // mova za.d[x9, #1], { z16.d-z19.d }\n" + ".inst 0xc0042d82 // mova za.d[x9, #2], { z12.d-z15.d }\n" "b 26f\n" "25:" // Width 3: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -342,136 +342,136 @@ void sme2_gemv_u8qa_dot_16VL ( "ble 29f\n" "27:" // Width 3: Multiply loop: Main loop head "whilelt p0.b, XZR, x22\n" - "ld1rqb { z1.b }, p0/Z, [x23]\n" + ".inst 0xa040837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n" + "ld1rqb { z9.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b230 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[0]\n" - ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b231 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151b1b2 // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[0]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b5b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[1]\n" - ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b5b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n" - ".inst 0xa0428355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151b6b2 // udot za.s[x9, 2], { z20.b-z23.b }, z1.b[1]\n" - "addvl x26, x26, #16\n" - ".inst 0xa0408349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b930 // udot za.s[x9, 0], { z8.b-z11.b }, z1.b[2]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b8b1 // udot za.s[x9, 1], { z4.b-z7.b }, z1.b[2]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151ba32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n" - "addvl x26, x26, #16\n" - ".inst 0xa0408359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151bf30 // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[3]\n" - ".inst 0xa0418345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151bcb1 // udot za.s[x9, 1], { z4.b-z7.b }, z1.b[3]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151be32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042836d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b3b0 // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[0]\n" + ".inst 0xa0408379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n" + ".inst 0xc159b231 // udot za.s[x9, 1], { z16.b-z19.b }, z9.b[0]\n" + ".inst 0xa041837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc159b1b2 // udot za.s[x9, 2], { z12.b-z15.b }, z9.b[0]\n" + ".inst 0xa0428371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xa040836d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27]\n" + ".inst 0xc159b730 // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[1]\n" + ".inst 0xa0418361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc159b7b1 // udot za.s[x9, 1], { z28.b-z31.b }, z9.b[1]\n" + ".inst 0xa042837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b632 // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[1]\n" + ".inst 0xa0408375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n" + ".inst 0xa0418379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc159b9b0 // udot za.s[x9, 0], { z12.b-z15.b }, z9.b[2]\n" + ".inst 0xa0428371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b831 // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n" + ".inst 0xc159bbb2 // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[2]\n" + ".inst 0xc159beb0 // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[3]\n" + ".inst 0xc159bf31 // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[3]\n" + ".inst 0xc159be32 // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[3]\n" "tbnz %x[flags], #31, 28f\n" - "udot z28.s, z1.b, z29.b\n" + "udot z11.s, z9.b, z8.b\n" "28:" // Width 3: Multiply loop: unique 5: skip row sum "sub x22, x22, #0x10\n" "cmp x22, #0x10\n" "bgt 27b\n" "29:" // Width 3: Multiply loop: Single iteration only "whilelt p0.b, XZR, x22\n" - "ld1rqb { z1.b }, p0/Z, [x23]\n" + ".inst 0xa040836d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xa0408355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n" + "ld1rqb { z9.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xc151b2b0 // udot za.s[x9, 0], { z20.b-z23.b }, z1.b[0]\n" - ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b231 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151b232 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[0]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b1b0 // udot za.s[x9, 0], { z12.b-z15.b }, z9.b[0]\n" + ".inst 0xc159b031 // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[0]\n" + ".inst 0xc159b3b2 // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[0]\n" "ble 30f\n" - ".inst 0xa0408359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26]\n" + ".inst 0xa0408365 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xc151b730 // udot za.s[x9, 0], { z24.b-z27.b }, z1.b[1]\n" - ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b5b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151b632 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[1]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa0428371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b4b0 // udot za.s[x9, 0], { z4.b-z7.b }, z9.b[1]\n" + ".inst 0xc159b731 // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[1]\n" + ".inst 0xc159b632 // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[1]\n" "ble 30f\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + ".inst 0xa040837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xc151ba30 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n" - ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151ba31 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[2]\n" - ".inst 0xa0428355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151bab2 // udot za.s[x9, 2], { z20.b-z23.b }, z1.b[2]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa0428379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159bbb0 // udot za.s[x9, 0], { z28.b-z31.b }, z9.b[2]\n" + ".inst 0xc159b831 // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n" + ".inst 0xc159bb32 // udot za.s[x9, 2], { z24.b-z27.b }, z9.b[2]\n" "ble 30f\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151bdb0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n" - ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151be31 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[3]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151bdb2 // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0408361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n" + ".inst 0xa0418375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa0428371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159bc30 // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[3]\n" + ".inst 0xc159beb1 // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[3]\n" + ".inst 0xc159be32 // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[3]\n" "30:" // Width 3: Multiply loop: multiply skip "tbnz %x[flags], #31, 31f\n" - "udot z28.s, z1.b, z29.b\n" + "udot z11.s, z9.b, z8.b\n" "31:" // Width 3: Multiply loop: unique 6: skip row sum "tbnz %x[flags], #31, 32f\n" "add x21, %x[qp], %[b_offset]\n" "mov x20, #0x4\n" "ld1rw { z16.s }, p2/Z, [x21]\n" - "neg z16.s, p2/M, z16.s\n" "whilelt p0.s, XZR, x20\n" - "uaddv d28, p0, z28.s\n" - "mov z28.s, z28.s[0]\n" - "mul z28.s, p2/M, z28.s, z16.s\n" "orr %x[flags], %x[flags], #0x80000000\n" + "uaddv d11, p0, z11.s\n" + "mov z11.s, z11.s[0]\n" + "neg z16.s, p2/M, z16.s\n" + "mul z11.s, p2/M, z11.s, z16.s\n" "32:" // Width 3: skip row sum fixup - ".inst 0xc0904b80 // addha za0.s, p2/M, p2/M, z28.s\n" - "add x20, %x[qp], %[per_layer_mul]\n" - "ld1rw { z2.s }, p2/Z, [x20]\n" + ".inst 0xc0904960 // addha za0.s, p2/M, p2/M, z11.s\n" + "add x21, %x[qp], %[per_layer_mul]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" - ".inst 0xc0904b81 // addha za1.s, p2/M, p2/M, z28.s\n" - "ld1rw { z1.s }, p2/Z, [x20]\n" - "add x20, %x[qp], %[c_offset]\n" + ".inst 0xc0904961 // addha za1.s, p2/M, p2/M, z11.s\n" + "ld1rw { z2.s }, p2/Z, [x21]\n" + "add x22, %x[qp], %[c_offset]\n" "add x21, %x[qp], %[minval]\n" - ".inst 0xc0904b82 // addha za2.s, p2/M, p2/M, z28.s\n" - "ld1rw { z3.s }, p2/Z, [x20]\n" + ".inst 0xc0904962 // addha za2.s, p2/M, p2/M, z11.s\n" + "ld1rw { z1.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[maxval]\n" - ".inst 0xc0904b83 // addha za3.s, p2/M, p2/M, z28.s\n" - "ld1rw { z16.s }, p2/Z, [x21]\n" - ".inst 0xc0062c08 // mova { z8.d-z11.d }, za.d[x9, #0]\n" - ".inst 0xc1a2ac08 // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z2.s\n" - ".inst 0xc0062c24 // mova { z4.d-z7.d }, za.d[x9, #1]\n" - ".inst 0xc1a2ac04 // sqdmulh { z4.s-z7.s }, { z4.s-z7.s }, z2.s\n" - ".inst 0xc0062c4c // mova { z12.d-z15.d }, za.d[x9, #2]\n" + ".inst 0xc0904963 // addha za3.s, p2/M, p2/M, z11.s\n" + "ld1rw { z0.s }, p2/Z, [x22]\n" + "ld1rw { z21.s }, p2/Z, [x21]\n" + "ld1rw { z20.s }, p2/Z, [x20]\n" + ".inst 0xc0062c1c // mova { z28.d-z31.d }, za.d[x9, #0]\n" + ".inst 0xc0062c2c // mova { z12.d-z15.d }, za.d[x9, #1]\n" + ".inst 0xc0062c50 // mova { z16.d-z19.d }, za.d[x9, #2]\n" + ".inst 0xc1a2ac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z2.s\n" ".inst 0xc1a2ac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z2.s\n" - ".inst 0xc1a1aa28 // srshl { z8.s-z11.s }, { z8.s-z11.s }, z1.s\n" - "ld1rw { z0.s }, p2/Z, [x20]\n" - ".inst 0xc1a1aa24 // srshl { z4.s-z7.s }, { z4.s-z7.s }, z1.s\n" + ".inst 0xc1a2ac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z2.s\n" + ".inst 0xc1a1aa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z1.s\n" ".inst 0xc1a1aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z1.s\n" - ".inst 0xc1a3ab08 // add { z8.s-z11.s }, { z8.s-z11.s }, z3.s\n" - ".inst 0xc1a3ab04 // add { z4.s-z7.s }, { z4.s-z7.s }, z3.s\n" - ".inst 0xc1a3ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z3.s\n" - ".inst 0xc1a0ce08 // sclamp { z8.s-z11.s }, z16.s, z0.s\n" - ".inst 0xc1a0ce04 // sclamp { z4.s-z7.s }, z16.s, z0.s\n" - "uzp1 z8.h, z8.h, z9.h\n" - ".inst 0xc1a0ce0c // sclamp { z12.s-z15.s }, z16.s, z0.s\n" - "uzp1 z18.h, z10.h, z11.h\n" - "uzp1 z4.h, z4.h, z5.h\n" - "uzp1 z17.h, z6.h, z7.h\n" + ".inst 0xc1a1aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z1.s\n" + ".inst 0xc1a0ab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z0.s\n" + ".inst 0xc1a0ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z0.s\n" + ".inst 0xc1a0ab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z0.s\n" + ".inst 0xc1b4cebc // sclamp { z28.s-z31.s }, z21.s, z20.s\n" + ".inst 0xc1b4ceac // sclamp { z12.s-z15.s }, z21.s, z20.s\n" + ".inst 0xc1b4ceb0 // sclamp { z16.s-z19.s }, z21.s, z20.s\n" + "uzp1 z28.h, z28.h, z29.h\n" + "uzp1 z21.h, z30.h, z31.h\n" "uzp1 z12.h, z12.h, z13.h\n" - "uzp1 z16.h, z14.h, z15.h\n" - "uzp1 z8.b, z8.b, z18.b\n" - "st1b { z8.b }, p2, [x25]\n" - "uzp1 z4.b, z4.b, z17.b\n" - "st1b { z4.b }, p2, [x25, #1, MUL VL]\n" - "uzp1 z12.b, z12.b, z16.b\n" - "st1b { z12.b }, p1, [x25, #2, MUL VL]\n" + "uzp1 z20.h, z14.h, z15.h\n" + "uzp1 z16.h, z16.h, z17.h\n" + "uzp1 z17.h, z18.h, z19.h\n" + "uzp1 z28.b, z28.b, z21.b\n" + "uzp1 z12.b, z12.b, z20.b\n" + "uzp1 z16.b, z16.b, z17.b\n" + "st1b { z28.b }, p2, [x25]\n" + "st1b { z12.b }, p2, [x25, #1, MUL VL]\n" + "st1b { z16.b }, p1, [x25, #2, MUL VL]\n" "addvl x25, x25, #3\n" "33:" // Width 3: Output done "b 44f\n" @@ -484,15 +484,15 @@ void sme2_gemv_u8qa_dot_16VL ( ".inst 0xf8b54af8 // rprfm pldmany, x21, [x23]\n" "whilelt p1.b, XZR, x20\n" "cbz x24, 35f\n" - ".inst 0xa040c314 // ld1w { z20.s-z23.s }, pn8.b/Z, [x24]\n" - ".inst 0xc0042e80 // mova za.d[x9, #0], { z20.d-z23.d }\n" - ".inst 0xa041c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" - ".inst 0xc0042e01 // mova za.d[x9, #1], { z16.d-z19.d }\n" + ".inst 0xa040c300 // ld1w { z0.s-z3.s }, pn8.b/Z, [x24]\n" + ".inst 0xa041c30c // ld1w { z12.s-z15.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n" ".inst 0xa042c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n" - ".inst 0xc0042e02 // mova za.d[x9, #2], { z16.d-z19.d }\n" - ".inst 0xa043c310 // ld1w { z16.s-z19.s }, pn8.b/Z, [x24, #0xc, MUL VL]\n" - ".inst 0xc0042e03 // mova za.d[x9, #3], { z16.d-z19.d }\n" + ".inst 0xa043c314 // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0xc, MUL VL]\n" + ".inst 0xc0042c00 // mova za.d[x9, #0], { z0.d-z3.d }\n" "addvl x24, x24, #16\n" + ".inst 0xc0042d81 // mova za.d[x9, #1], { z12.d-z15.d }\n" + ".inst 0xc0042e02 // mova za.d[x9, #2], { z16.d-z19.d }\n" + ".inst 0xc0042e83 // mova za.d[x9, #3], { z20.d-z23.d }\n" "b 36f\n" "35:" // Width 4: no bias ".inst 0xc00800ff // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n" @@ -501,164 +501,164 @@ void sme2_gemv_u8qa_dot_16VL ( "ble 39f\n" "37:" // Width 4: Multiply loop: Main loop head "whilelt p0.b, XZR, x22\n" - "ld1rqb { z1.b }, p0/Z, [x23]\n" + ".inst 0xa0408361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n" + "ld1rqb { z9.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b230 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[0]\n" - ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b231 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[0]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151b1b2 // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[0]\n" - ".inst 0xa043834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc151b1b3 // udot za.s[x9, 3], { z12.b-z15.b }, z1.b[0]\n" - "addvl x26, x26, #16\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151b630 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n" - ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b5b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[1]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151b632 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[1]\n" - ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc151b633 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[1]\n" - "addvl x26, x26, #16\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151ba30 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n" - ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b9b1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[2]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151ba32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n" - ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc151ba33 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[2]\n" - "addvl x26, x26, #16\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151bdb0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[3]\n" - ".inst 0xa041834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151bdb1 // udot za.s[x9, 1], { z12.b-z15.b }, z1.b[3]\n" - ".inst 0xa0428359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151bf32 // udot za.s[x9, 2], { z24.b-z27.b }, z1.b[3]\n" - ".inst 0xa0438345 // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc151bcb3 // udot za.s[x9, 3], { z4.b-z7.b }, z1.b[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa0438379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc159b030 // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b2b1 // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[0]\n" + ".inst 0xa0408361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n" + ".inst 0xc159b3b2 // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[0]\n" + ".inst 0xa041837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc159b333 // udot za.s[x9, 3], { z24.b-z27.b }, z9.b[0]\n" + ".inst 0xa0428371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043836d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc159b430 // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[1]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b7b1 // udot za.s[x9, 1], { z28.b-z31.b }, z9.b[1]\n" + ".inst 0xa0408375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27]\n" + ".inst 0xc159b632 // udot za.s[x9, 2], { z16.b-z19.b }, z9.b[1]\n" + ".inst 0xa041837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc159b5b3 // udot za.s[x9, 3], { z12.b-z15.b }, z9.b[1]\n" + ".inst 0xa0428361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043836d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc159bab0 // udot za.s[x9, 0], { z20.b-z23.b }, z9.b[2]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159bbb1 // udot za.s[x9, 1], { z28.b-z31.b }, z9.b[2]\n" + ".inst 0xa0408379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n" + ".inst 0xc159b832 // udot za.s[x9, 2], { z0.b-z3.b }, z9.b[2]\n" + ".inst 0xa0418375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xc159b9b3 // udot za.s[x9, 3], { z12.b-z15.b }, z9.b[2]\n" + ".inst 0xa0428361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa0438371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc159bf30 // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[3]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159beb1 // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[3]\n" + ".inst 0xc159bc32 // udot za.s[x9, 2], { z0.b-z3.b }, z9.b[3]\n" + ".inst 0xc159be33 // udot za.s[x9, 3], { z16.b-z19.b }, z9.b[3]\n" "tbnz %x[flags], #31, 38f\n" - "udot z28.s, z1.b, z29.b\n" + "udot z11.s, z9.b, z8.b\n" "38:" // Width 4: Multiply loop: unique 7: skip row sum "sub x22, x22, #0x10\n" "cmp x22, #0x10\n" "bgt 37b\n" "39:" // Width 4: Multiply loop: Single iteration only "whilelt p0.b, XZR, x22\n" - "ld1rqb { z1.b }, p0/Z, [x23]\n" + ".inst 0xa0408361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xa040834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26]\n" + "ld1rqb { z9.b }, p0/Z, [x23]\n" "add x23, x23, #0x10\n" - ".inst 0xc151b1b0 // udot za.s[x9, 0], { z12.b-z15.b }, z1.b[0]\n" - ".inst 0xa0418359 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b331 // udot za.s[x9, 1], { z24.b-z27.b }, z1.b[0]\n" - ".inst 0xa0428349 // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151b132 // udot za.s[x9, 2], { z8.b-z11.b }, z1.b[0]\n" - ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc151b233 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[0]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa0438371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc159b030 // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[0]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b2b1 // udot za.s[x9, 1], { z20.b-z23.b }, z9.b[0]\n" + ".inst 0xc159b3b2 // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[0]\n" + ".inst 0xc159b233 // udot za.s[x9, 3], { z16.b-z19.b }, z9.b[0]\n" "ble 40f\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + ".inst 0xa0408361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xc151b630 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[1]\n" - ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151b631 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[1]\n" - ".inst 0xa042834d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151b5b2 // udot za.s[x9, 2], { z12.b-z15.b }, z1.b[1]\n" - ".inst 0xa0438355 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc151b6b3 // udot za.s[x9, 3], { z20.b-z23.b }, z1.b[1]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa042837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa0438371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc159b430 // udot za.s[x9, 0], { z0.b-z3.b }, z9.b[1]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b731 // udot za.s[x9, 1], { z24.b-z27.b }, z9.b[1]\n" + ".inst 0xc159b7b2 // udot za.s[x9, 2], { z28.b-z31.b }, z9.b[1]\n" + ".inst 0xc159b633 // udot za.s[x9, 3], { z16.b-z19.b }, z9.b[1]\n" "ble 40f\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" + ".inst 0xa0408379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n" "subs x22, x22, #0x4\n" - ".inst 0xc151ba30 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[2]\n" - ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151ba31 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[2]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151ba32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[2]\n" - ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc151ba33 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[2]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0418361 // ldnt1b { z0.b-z3.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa0428375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa043836d // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc159bb30 // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[2]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159b831 // udot za.s[x9, 1], { z0.b-z3.b }, z9.b[2]\n" + ".inst 0xc159bab2 // udot za.s[x9, 2], { z20.b-z23.b }, z9.b[2]\n" + ".inst 0xc159b9b3 // udot za.s[x9, 3], { z12.b-z15.b }, z9.b[2]\n" "ble 40f\n" - ".inst 0xa0408351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n" - ".inst 0xc151be30 // udot za.s[x9, 0], { z16.b-z19.b }, z1.b[3]\n" - ".inst 0xa0418351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n" - ".inst 0xc151be31 // udot za.s[x9, 1], { z16.b-z19.b }, z1.b[3]\n" - ".inst 0xa0428351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n" - ".inst 0xc151be32 // udot za.s[x9, 2], { z16.b-z19.b }, z1.b[3]\n" - ".inst 0xa0438351 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n" - ".inst 0xc151be33 // udot za.s[x9, 3], { z16.b-z19.b }, z1.b[3]\n" - "addvl x26, x26, #16\n" + ".inst 0xa0408379 // ldnt1b { z24.b-z27.b }, pn8.b/Z, [x27]\n" + ".inst 0xa041837d // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x27, #0x4, MUL VL]\n" + ".inst 0xa0428375 // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x27, #0x8, MUL VL]\n" + ".inst 0xa0438371 // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x27, #0xc, MUL VL]\n" + ".inst 0xc159bf30 // udot za.s[x9, 0], { z24.b-z27.b }, z9.b[3]\n" + "addvl x27, x27, #16\n" + ".inst 0xc159bfb1 // udot za.s[x9, 1], { z28.b-z31.b }, z9.b[3]\n" + ".inst 0xc159beb2 // udot za.s[x9, 2], { z20.b-z23.b }, z9.b[3]\n" + ".inst 0xc159be33 // udot za.s[x9, 3], { z16.b-z19.b }, z9.b[3]\n" "40:" // Width 4: Multiply loop: multiply skip "tbnz %x[flags], #31, 41f\n" - "udot z28.s, z1.b, z29.b\n" + "udot z11.s, z9.b, z8.b\n" "41:" // Width 4: Multiply loop: unique 8: skip row sum "tbnz %x[flags], #31, 42f\n" "add x21, %x[qp], %[b_offset]\n" "mov x20, #0x4\n" "ld1rw { z16.s }, p2/Z, [x21]\n" - "neg z16.s, p2/M, z16.s\n" "whilelt p0.s, XZR, x20\n" - "uaddv d28, p0, z28.s\n" - "mov z28.s, z28.s[0]\n" - "mul z28.s, p2/M, z28.s, z16.s\n" "orr %x[flags], %x[flags], #0x80000000\n" + "uaddv d11, p0, z11.s\n" + "mov z11.s, z11.s[0]\n" + "neg z16.s, p2/M, z16.s\n" + "mul z11.s, p2/M, z11.s, z16.s\n" "42:" // Width 4: skip row sum fixup - ".inst 0xc0904b80 // addha za0.s, p2/M, p2/M, z28.s\n" - "add x20, %x[qp], %[per_layer_mul]\n" - "ld1rw { z11.s }, p2/Z, [x20]\n" + ".inst 0xc0904960 // addha za0.s, p2/M, p2/M, z11.s\n" + "add x21, %x[qp], %[per_layer_mul]\n" "add x20, %x[qp], %[per_layer_right_shift]\n" - ".inst 0xc0904b81 // addha za1.s, p2/M, p2/M, z28.s\n" - "ld1rw { z7.s }, p2/Z, [x20]\n" - "add x20, %x[qp], %[c_offset]\n" + ".inst 0xc0904961 // addha za1.s, p2/M, p2/M, z11.s\n" + "ld1rw { z10.s }, p2/Z, [x21]\n" + "add x22, %x[qp], %[c_offset]\n" "add x21, %x[qp], %[minval]\n" - ".inst 0xc0904b82 // addha za2.s, p2/M, p2/M, z28.s\n" - "ld1rw { z6.s }, p2/Z, [x20]\n" + ".inst 0xc0904962 // addha za2.s, p2/M, p2/M, z11.s\n" + "ld1rw { z12.s }, p2/Z, [x20]\n" "add x20, %x[qp], %[maxval]\n" - ".inst 0xc0904b83 // addha za3.s, p2/M, p2/M, z28.s\n" - "ld1rw { z3.s }, p2/Z, [x21]\n" - ".inst 0xc0062c18 // mova { z24.d-z27.d }, za.d[x9, #0]\n" - ".inst 0xc1abac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z11.s\n" - ".inst 0xc0062c30 // mova { z16.d-z19.d }, za.d[x9, #1]\n" - ".inst 0xc1abac10 // sqdmulh { z16.s-z19.s }, { z16.s-z19.s }, z11.s\n" - ".inst 0xc0062c54 // mova { z20.d-z23.d }, za.d[x9, #2]\n" - ".inst 0xc1abac14 // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z11.s\n" - ".inst 0xc0062c6c // mova { z12.d-z15.d }, za.d[x9, #3]\n" - ".inst 0xc1abac0c // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z11.s\n" - ".inst 0xc1a7aa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z7.s\n" - "ld1rw { z31.s }, p2/Z, [x20]\n" - ".inst 0xc1a7aa30 // srshl { z16.s-z19.s }, { z16.s-z19.s }, z7.s\n" - ".inst 0xc1a7aa34 // srshl { z20.s-z23.s }, { z20.s-z23.s }, z7.s\n" - ".inst 0xc1a7aa2c // srshl { z12.s-z15.s }, { z12.s-z15.s }, z7.s\n" - ".inst 0xc1a6ab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z6.s\n" - ".inst 0xc1a6ab10 // add { z16.s-z19.s }, { z16.s-z19.s }, z6.s\n" - ".inst 0xc1a6ab14 // add { z20.s-z23.s }, { z20.s-z23.s }, z6.s\n" - ".inst 0xc1a6ab0c // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n" - ".inst 0xc1bfcc78 // sclamp { z24.s-z27.s }, z3.s, z31.s\n" - ".inst 0xc1bfcc70 // sclamp { z16.s-z19.s }, z3.s, z31.s\n" + ".inst 0xc0904963 // addha za3.s, p2/M, p2/M, z11.s\n" + "ld1rw { z13.s }, p2/Z, [x22]\n" + "ld1rw { z17.s }, p2/Z, [x21]\n" + "ld1rw { z18.s }, p2/Z, [x20]\n" + ".inst 0xc0062c1c // mova { z28.d-z31.d }, za.d[x9, #0]\n" + ".inst 0xc0062c20 // mova { z0.d-z3.d }, za.d[x9, #1]\n" + ".inst 0xc0062c58 // mova { z24.d-z27.d }, za.d[x9, #2]\n" + ".inst 0xc0062c74 // mova { z20.d-z23.d }, za.d[x9, #3]\n" + ".inst 0xc1aaac1c // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z10.s\n" + ".inst 0xc1aaac00 // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z10.s\n" + ".inst 0xc1aaac18 // sqdmulh { z24.s-z27.s }, { z24.s-z27.s }, z10.s\n" + ".inst 0xc1aaac14 // sqdmulh { z20.s-z23.s }, { z20.s-z23.s }, z10.s\n" + ".inst 0xc1acaa3c // srshl { z28.s-z31.s }, { z28.s-z31.s }, z12.s\n" + ".inst 0xc1acaa20 // srshl { z0.s-z3.s }, { z0.s-z3.s }, z12.s\n" + ".inst 0xc1acaa38 // srshl { z24.s-z27.s }, { z24.s-z27.s }, z12.s\n" + ".inst 0xc1acaa34 // srshl { z20.s-z23.s }, { z20.s-z23.s }, z12.s\n" + ".inst 0xc1adab1c // add { z28.s-z31.s }, { z28.s-z31.s }, z13.s\n" + ".inst 0xc1adab00 // add { z0.s-z3.s }, { z0.s-z3.s }, z13.s\n" + ".inst 0xc1adab18 // add { z24.s-z27.s }, { z24.s-z27.s }, z13.s\n" + ".inst 0xc1adab14 // add { z20.s-z23.s }, { z20.s-z23.s }, z13.s\n" + ".inst 0xc1b2ce3c // sclamp { z28.s-z31.s }, z17.s, z18.s\n" + ".inst 0xc1b2ce20 // sclamp { z0.s-z3.s }, z17.s, z18.s\n" + ".inst 0xc1b2ce38 // sclamp { z24.s-z27.s }, z17.s, z18.s\n" + ".inst 0xc1b2ce34 // sclamp { z20.s-z23.s }, z17.s, z18.s\n" + "uzp1 z28.h, z28.h, z29.h\n" + "uzp1 z14.h, z30.h, z31.h\n" + "uzp1 z0.h, z0.h, z1.h\n" + "uzp1 z18.h, z2.h, z3.h\n" "uzp1 z24.h, z24.h, z25.h\n" - ".inst 0xc1bfcc74 // sclamp { z20.s-z23.s }, z3.s, z31.s\n" - ".inst 0xc1bfcc6c // sclamp { z12.s-z15.s }, z3.s, z31.s\n" - "uzp1 z25.h, z26.h, z27.h\n" - "uzp1 z16.h, z16.h, z17.h\n" - "uzp1 z18.h, z18.h, z19.h\n" + "uzp1 z17.h, z26.h, z27.h\n" "uzp1 z20.h, z20.h, z21.h\n" - "uzp1 z17.h, z22.h, z23.h\n" - "uzp1 z12.h, z12.h, z13.h\n" - "uzp1 z30.h, z14.h, z15.h\n" - "uzp1 z24.b, z24.b, z25.b\n" - "st1b { z24.b }, p2, [x25]\n" - "uzp1 z16.b, z16.b, z18.b\n" - "st1b { z16.b }, p2, [x25, #1, MUL VL]\n" - "uzp1 z20.b, z20.b, z17.b\n" - "uzp1 z12.b, z12.b, z30.b\n" - "st1b { z20.b }, p2, [x25, #2, MUL VL]\n" - "st1b { z12.b }, p1, [x25, #3, MUL VL]\n" + "uzp1 z16.h, z22.h, z23.h\n" + "uzp1 z28.b, z28.b, z14.b\n" + "uzp1 z0.b, z0.b, z18.b\n" + "uzp1 z24.b, z24.b, z17.b\n" + "uzp1 z20.b, z20.b, z16.b\n" + "st1b { z28.b }, p2, [x25]\n" + "st1b { z0.b }, p2, [x25, #1, MUL VL]\n" + "st1b { z24.b }, p2, [x25, #2, MUL VL]\n" + "st1b { z20.b }, p1, [x25, #3, MUL VL]\n" "addvl x25, x25, #4\n" "43:" // 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" "44:" // Exit |