/*
 * Copyright (c) 2022-2023 Arm Limited.
 *
 * SPDX-License-Identifier: MIT
 *
 * Permission is hereby granted, free of charge, to any person obtaining a copy
 * of this software and associated documentation files (the "Software"), to
 * deal in the Software without restriction, including without limitation the
 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
 * sell copies of the Software, and to permit persons to whom the Software is
 * furnished to do so, subject to the following conditions:
 *
 * The above copyright notice and this permission notice shall be included in all
 * copies or substantial portions of the Software.
 *
 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
 * SOFTWARE.
 */
#ifdef __ARM_FEATURE_SVE
#ifdef ARM_COMPUTE_ENABLE_SME2

#include "arm_gemm.hpp"
#include "../../utils.hpp"

#include <cassert>
#include <cstdint>

namespace arm_gemm {

void sme2_gemv_s8qa_dot_16VL (
    const int8_t *A_ptr, const int8_t *B_ptr, int8_t *output_ptr,
    size_t N, size_t K,
    const Requantize32 *qp, const int32_t *col_bias, unsigned int col_base
)
{
    ARM_COMPUTE_UNUSED(col_base);

    struct KernelArgs {
        const int8_t *B_ptr = {};
        size_t output_offset = {};
        unsigned int input_initial_col = {};
    } ka;

    unsigned long flags=0;
    ka.B_ptr = B_ptr;
    if (qp->c_offset > qp->minval) {
        flags |= 0x20;
    }
    __asm__ __volatile__(
      "ptrue p2.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"
      "mov x25, %x[output_ptr]\n"
      "ptrue p2.b\n"
      ".inst 0x25207810  // ptrue pn8.b\n"
      "mul x22, x22, %x[K]\n"
      "mov x21, #0x1\n"
      "1:"  // RHS size check loop
      "cmp x22, #0x200000\n"
      "blt 2f\n"
      "tbnz x22, #0, 3f\n"
      "lsr x22, x22, #0x1\n"
      "lsl x21, x21, #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"
      "3:"  // RHS prefetch exit
      "mov x24, %x[col_bias]\n"
      "mov z26.s, #0x0\n"
      "mov z24.b, #0x1\n"
      "bic %x[flags], %x[flags], #0x80000000\n"
      "4:"  // Column loop
      "cmp x27, #0x4\n"
      "bge 34f\n"
      "cmp x27, #0x2\n"
      "bgt 24f\n"
      "beq 14f\n"
      "mov x23, %x[A_ptr]\n"
      "mov x22, %x[K]\n"
      "mov x20, %x[N]\n"
      "mov x21, %x[K]\n"
      ".inst 0xf8b64af8  // rprfm pldmany, x22, [x23]\n"
      "whilelt p1.b, XZR, x20\n"
      "cbz x24, 5f\n"
      ".inst 0xa040c304  // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n"
      ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
      "b 6f\n"
      "5:"  // Width 1: no bias
      ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
      "6:"  // Width 1: setup done
      "cmp x21, #0x10\n"
      "ble 9f\n"
      "7:"  // Width 1: Multiply loop: Main loop head
      "whilelt p0.b, XZR, x21\n"
      "ld1rqb { z3.b }, p0/Z, [x23]\n"
      "add x23, x23, #0x10\n"
      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n"
      "addvl x26, x26, #16\n"
      "tbnz %x[flags], #31, 8f\n"
      "sdot z26.s, z3.b, z24.b\n"
      "8:"  // Width 1: Multiply loop: unique 1: skip row sum
      "sub x21, x21, #0x10\n"
      "cmp x21, #0x10\n"
      "bgt 7b\n"
      "9:"  // Width 1: Multiply loop: Single iteration only
      "whilelt p0.b, XZR, x21\n"
      "ld1rqb { z3.b }, p0/Z, [x23]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
      "add x23, x23, #0x10\n"
      ".inst 0xc153b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n"
      "addvl x26, x26, #16\n"
      "ble 10f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xc153b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n"
      "addvl x26, x26, #16\n"
      "ble 10f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xc153baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n"
      "addvl x26, x26, #16\n"
      "ble 10f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n"
      "addvl x26, x26, #16\n"
      "10:"  // Width 1: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 11f\n"
      "sdot z26.s, z3.b, z24.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 { z10.s }, p2/Z, [x21]\n"
      "neg z10.s, p2/M, z10.s\n"
      "whilelt p0.s, XZR, x20\n"
      "saddv d26, p0, z26.s\n"
      "mov z26.s, z26.s[0]\n"
      "mul z26.s, p2/M, z26.s, z10.s\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "12:"  // Width 1: skip row sum fixup
      ".inst 0xc0904b40  // addha za0.s, p2/M, p2/M, z26.s\n"
      "add x20, %x[qp], %[per_layer_mul]\n"
      "ld1rw { z5.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      ".inst 0xc0904b41  // addha za1.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z4.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[c_offset]\n"
      "add x21, %x[qp], %[minval]\n"
      ".inst 0xc0904b42  // addha za2.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z6.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[maxval]\n"
      ".inst 0xc0904b43  // addha za3.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z21.s }, p2/Z, [x21]\n"
      ".inst 0xc0062c1c  // mova { z28.d-z31.d }, za.d[x9, #0]\n"
      ".inst 0xc1a5ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z5.s\n"
      ".inst 0xc1a4aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z4.s\n"
      "ld1rw { z16.s }, p2/Z, [x20]\n"
      ".inst 0xc1a6ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z6.s\n"
      ".inst 0xc1b0cebc  // sclamp { z28.s-z31.s }, z21.s, z16.s\n"
      "uzp1 z28.h, z28.h, z29.h\n"
      "uzp1 z29.h, z30.h, z31.h\n"
      "uzp1 z28.b, z28.b, z29.b\n"
      "st1b { z28.b }, p1, [x25]\n"
      "addvl x25, x25, #1\n"
      "13:"  // Width 1: Output done
      "b 44f\n"
      "14:"  // Width 2
      "mov x23, %x[A_ptr]\n"
      "mov x22, %x[K]\n"
      "sub x20, %x[N], x28\n"
      "mov x21, %x[K]\n"
      ".inst 0xf8b64af8  // rprfm pldmany, x22, [x23]\n"
      "whilelt p1.b, XZR, x20\n"
      "cbz x24, 15f\n"
      ".inst 0xa040c304  // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n"
      ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
      ".inst 0xa041c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
      ".inst 0xc0042e81  // mova za.d[x9, #1], { z20.d-z23.d }\n"
      "b 16f\n"
      "15:"  // Width 2: no bias
      ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
      "16:"  // Width 2: setup done
      "cmp x21, #0x10\n"
      "ble 19f\n"
      "17:"  // Width 2: Multiply loop: Main loop head
      "whilelt p0.b, XZR, x21\n"
      "ld1rqb { z3.b }, p0/Z, [x23]\n"
      "add x23, x23, #0x10\n"
      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b0a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n"
      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b521  // sdot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b8a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0418355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153bea1  // sdot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n"
      "addvl x26, x26, #16\n"
      "tbnz %x[flags], #31, 18f\n"
      "sdot z26.s, z3.b, z24.b\n"
      "18:"  // Width 2: Multiply loop: unique 3: skip row sum
      "sub x21, x21, #0x10\n"
      "cmp x21, #0x10\n"
      "bgt 17b\n"
      "19:"  // Width 2: Multiply loop: Single iteration only
      "whilelt p0.b, XZR, x21\n"
      "ld1rqb { z3.b }, p0/Z, [x23]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
      "add x23, x23, #0x10\n"
      ".inst 0xc153b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b0a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n"
      "addvl x26, x26, #16\n"
      "ble 20f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xc153b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n"
      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b521  // sdot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n"
      "addvl x26, x26, #16\n"
      "ble 20f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xc153baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b8a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n"
      "addvl x26, x26, #16\n"
      "ble 20f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0418355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153bea1  // sdot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n"
      "addvl x26, x26, #16\n"
      "20:"  // Width 2: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 21f\n"
      "sdot z26.s, z3.b, z24.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 { z10.s }, p2/Z, [x21]\n"
      "neg z10.s, p2/M, z10.s\n"
      "whilelt p0.s, XZR, x20\n"
      "saddv d26, p0, z26.s\n"
      "mov z26.s, z26.s[0]\n"
      "mul z26.s, p2/M, z26.s, z10.s\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "22:"  // Width 2: skip row sum fixup
      ".inst 0xc0904b40  // addha za0.s, p2/M, p2/M, z26.s\n"
      "add x20, %x[qp], %[per_layer_mul]\n"
      "ld1rw { z5.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      ".inst 0xc0904b41  // addha za1.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z4.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[c_offset]\n"
      "add x21, %x[qp], %[minval]\n"
      ".inst 0xc0904b42  // addha za2.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z6.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[maxval]\n"
      ".inst 0xc0904b43  // addha za3.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z21.s }, p2/Z, [x21]\n"
      ".inst 0xc0062c1c  // mova { z28.d-z31.d }, za.d[x9, #0]\n"
      ".inst 0xc1a5ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z5.s\n"
      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
      ".inst 0xc1a5ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z5.s\n"
      ".inst 0xc1a4aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z4.s\n"
      "ld1rw { z16.s }, p2/Z, [x20]\n"
      ".inst 0xc1a4aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z4.s\n"
      ".inst 0xc1a6ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z6.s\n"
      ".inst 0xc1a6ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n"
      ".inst 0xc1b0cebc  // sclamp { z28.s-z31.s }, z21.s, z16.s\n"
      ".inst 0xc1b0ceac  // sclamp { z12.s-z15.s }, z21.s, z16.s\n"
      "uzp1 z28.h, z28.h, z29.h\n"
      "uzp1 z29.h, z30.h, z31.h\n"
      "uzp1 z12.h, z12.h, z13.h\n"
      "uzp1 z13.h, z14.h, z15.h\n"
      "uzp1 z28.b, z28.b, z29.b\n"
      "st1b { z28.b }, p2, [x25]\n"
      "uzp1 z12.b, z12.b, z13.b\n"
      "st1b { z12.b }, p1, [x25, #1, MUL VL]\n"
      "addvl x25, x25, #2\n"
      "23:"  // Width 2: Output done
      "b 44f\n"
      "24:"  // Width 3
      "mov x20, #0x2\n"
      "mov x23, %x[A_ptr]\n"
      "mov x22, %x[K]\n"
      "msub x20, x28, x20, %x[N]\n"
      "mov x21, %x[K]\n"
      ".inst 0xf8b64af8  // rprfm pldmany, x22, [x23]\n"
      "whilelt p1.b, XZR, x20\n"
      "cbz x24, 25f\n"
      ".inst 0xa040c304  // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n"
      ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
      ".inst 0xa041c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
      ".inst 0xc0042e81  // mova za.d[x9, #1], { z20.d-z23.d }\n"
      ".inst 0xa042c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n"
      ".inst 0xc0042e82  // mova za.d[x9, #2], { z20.d-z23.d }\n"
      "b 26f\n"
      "25:"  // Width 3: no bias
      ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
      "26:"  // Width 3: setup done
      "cmp x21, #0x10\n"
      "ble 29f\n"
      "27:"  // Width 3: Multiply loop: Main loop head
      "whilelt p0.b, XZR, x21\n"
      "ld1rqb { z3.b }, p0/Z, [x23]\n"
      "add x23, x23, #0x10\n"
      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b0a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n"
      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b1a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z3.b[0]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n"
      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b521  // sdot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n"
      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b622  // sdot za.s[x9, 2], { z16.b-z19.b }, z3.b[1]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b8a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n"
      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b9a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z3.b[2]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0418355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153bea1  // sdot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153be22  // sdot za.s[x9, 2], { z16.b-z19.b }, z3.b[3]\n"
      "addvl x26, x26, #16\n"
      "tbnz %x[flags], #31, 28f\n"
      "sdot z26.s, z3.b, z24.b\n"
      "28:"  // Width 3: Multiply loop: unique 5: skip row sum
      "sub x21, x21, #0x10\n"
      "cmp x21, #0x10\n"
      "bgt 27b\n"
      "29:"  // Width 3: Multiply loop: Single iteration only
      "whilelt p0.b, XZR, x21\n"
      "ld1rqb { z3.b }, p0/Z, [x23]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
      "add x23, x23, #0x10\n"
      ".inst 0xc153b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b0a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n"
      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b1a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z3.b[0]\n"
      "addvl x26, x26, #16\n"
      "ble 30f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xc153b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n"
      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b521  // sdot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n"
      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b622  // sdot za.s[x9, 2], { z16.b-z19.b }, z3.b[1]\n"
      "addvl x26, x26, #16\n"
      "ble 30f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xc153baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b8a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n"
      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b9a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z3.b[2]\n"
      "addvl x26, x26, #16\n"
      "ble 30f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0418355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153bea1  // sdot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153be22  // sdot za.s[x9, 2], { z16.b-z19.b }, z3.b[3]\n"
      "addvl x26, x26, #16\n"
      "30:"  // Width 3: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 31f\n"
      "sdot z26.s, z3.b, z24.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 { z10.s }, p2/Z, [x21]\n"
      "neg z10.s, p2/M, z10.s\n"
      "whilelt p0.s, XZR, x20\n"
      "saddv d26, p0, z26.s\n"
      "mov z26.s, z26.s[0]\n"
      "mul z26.s, p2/M, z26.s, z10.s\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "32:"  // Width 3: skip row sum fixup
      ".inst 0xc0904b40  // addha za0.s, p2/M, p2/M, z26.s\n"
      "add x20, %x[qp], %[per_layer_mul]\n"
      "ld1rw { z5.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      ".inst 0xc0904b41  // addha za1.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z4.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[c_offset]\n"
      "add x21, %x[qp], %[minval]\n"
      ".inst 0xc0904b42  // addha za2.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z6.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[maxval]\n"
      ".inst 0xc0904b43  // addha za3.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z21.s }, p2/Z, [x21]\n"
      ".inst 0xc0062c1c  // mova { z28.d-z31.d }, za.d[x9, #0]\n"
      ".inst 0xc1a5ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z5.s\n"
      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
      ".inst 0xc1a5ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z5.s\n"
      ".inst 0xc0062c40  // mova { z0.d-z3.d }, za.d[x9, #2]\n"
      ".inst 0xc1a5ac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z5.s\n"
      ".inst 0xc1a4aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z4.s\n"
      "ld1rw { z16.s }, p2/Z, [x20]\n"
      ".inst 0xc1a4aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z4.s\n"
      ".inst 0xc1a4aa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z4.s\n"
      ".inst 0xc1a6ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z6.s\n"
      ".inst 0xc1a6ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n"
      ".inst 0xc1a6ab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n"
      ".inst 0xc1b0cebc  // sclamp { z28.s-z31.s }, z21.s, z16.s\n"
      ".inst 0xc1b0ceac  // sclamp { z12.s-z15.s }, z21.s, z16.s\n"
      "uzp1 z28.h, z28.h, z29.h\n"
      ".inst 0xc1b0cea0  // sclamp { z0.s-z3.s }, z21.s, z16.s\n"
      "uzp1 z29.h, z30.h, z31.h\n"
      "uzp1 z12.h, z12.h, z13.h\n"
      "uzp1 z13.h, z14.h, z15.h\n"
      "uzp1 z0.h, z0.h, z1.h\n"
      "uzp1 z1.h, z2.h, z3.h\n"
      "uzp1 z28.b, z28.b, z29.b\n"
      "st1b { z28.b }, p2, [x25]\n"
      "uzp1 z12.b, z12.b, z13.b\n"
      "st1b { z12.b }, p2, [x25, #1, MUL VL]\n"
      "uzp1 z0.b, z0.b, z1.b\n"
      "st1b { z0.b }, p1, [x25, #2, MUL VL]\n"
      "addvl x25, x25, #3\n"
      "33:"  // Width 3: Output done
      "b 44f\n"
      "34:"  // Width 4
      "mov x20, #0x3\n"
      "mov x23, %x[A_ptr]\n"
      "mov x22, %x[K]\n"
      "msub x20, x28, x20, %x[N]\n"
      "mov x21, %x[K]\n"
      ".inst 0xf8b64af8  // rprfm pldmany, x22, [x23]\n"
      "whilelt p1.b, XZR, x20\n"
      "cbz x24, 35f\n"
      ".inst 0xa040c304  // ld1w { z4.s-z7.s }, pn8.b/Z, [x24]\n"
      ".inst 0xc0042c80  // mova za.d[x9, #0], { z4.d-z7.d }\n"
      ".inst 0xa041c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x4, MUL VL]\n"
      ".inst 0xc0042e81  // mova za.d[x9, #1], { z20.d-z23.d }\n"
      ".inst 0xa042c314  // ld1w { z20.s-z23.s }, pn8.b/Z, [x24, #0x8, MUL VL]\n"
      ".inst 0xc0042e82  // mova za.d[x9, #2], { z20.d-z23.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"
      "addvl x24, x24, #16\n"
      "b 36f\n"
      "35:"  // Width 4: no bias
      ".inst 0xc00800ff  // zero { zad0, zad1, zad2, zad3, zad4, zad5, zad6, zad7 }\n"
      "36:"  // Width 4: setup done
      "cmp x21, #0x10\n"
      "ble 39f\n"
      "37:"  // Width 4: Multiply loop: Main loop head
      "whilelt p0.b, XZR, x21\n"
      "ld1rqb { z3.b }, p0/Z, [x23]\n"
      "add x23, x23, #0x10\n"
      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b0a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n"
      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b1a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z3.b[0]\n"
      ".inst 0xa043834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
      ".inst 0xc153b1a3  // sdot za.s[x9, 3], { z12.b-z15.b }, z3.b[0]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n"
      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b521  // sdot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n"
      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b622  // sdot za.s[x9, 2], { z16.b-z19.b }, z3.b[1]\n"
      ".inst 0xa043834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
      ".inst 0xc153b5a3  // sdot za.s[x9, 3], { z12.b-z15.b }, z3.b[1]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b8a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n"
      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b9a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z3.b[2]\n"
      ".inst 0xa043835d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
      ".inst 0xc153bba3  // sdot za.s[x9, 3], { z28.b-z31.b }, z3.b[2]\n"
      "addvl x26, x26, #16\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0418355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153bea1  // sdot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153be22  // sdot za.s[x9, 2], { z16.b-z19.b }, z3.b[3]\n"
      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
      ".inst 0xc153be23  // sdot za.s[x9, 3], { z16.b-z19.b }, z3.b[3]\n"
      "addvl x26, x26, #16\n"
      "tbnz %x[flags], #31, 38f\n"
      "sdot z26.s, z3.b, z24.b\n"
      "38:"  // Width 4: Multiply loop: unique 7: skip row sum
      "sub x21, x21, #0x10\n"
      "cmp x21, #0x10\n"
      "bgt 37b\n"
      "39:"  // Width 4: Multiply loop: Single iteration only
      "whilelt p0.b, XZR, x21\n"
      "ld1rqb { z3.b }, p0/Z, [x23]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xa0408351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26]\n"
      "add x23, x23, #0x10\n"
      ".inst 0xc153b220  // sdot za.s[x9, 0], { z16.b-z19.b }, z3.b[0]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b0a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[0]\n"
      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b1a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z3.b[0]\n"
      ".inst 0xa043834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
      ".inst 0xc153b1a3  // sdot za.s[x9, 3], { z12.b-z15.b }, z3.b[0]\n"
      "addvl x26, x26, #16\n"
      "ble 40f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xc153b6a0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[1]\n"
      ".inst 0xa0418349  // ldnt1b { z8.b-z11.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b521  // sdot za.s[x9, 1], { z8.b-z11.b }, z3.b[1]\n"
      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b622  // sdot za.s[x9, 2], { z16.b-z19.b }, z3.b[1]\n"
      ".inst 0xa043834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
      ".inst 0xc153b5a3  // sdot za.s[x9, 3], { z12.b-z15.b }, z3.b[1]\n"
      "addvl x26, x26, #16\n"
      "ble 40f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      "subs x21, x21, #0x4\n"
      ".inst 0xc153baa0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[2]\n"
      ".inst 0xa0418345  // ldnt1b { z4.b-z7.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153b8a1  // sdot za.s[x9, 1], { z4.b-z7.b }, z3.b[2]\n"
      ".inst 0xa042834d  // ldnt1b { z12.b-z15.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153b9a2  // sdot za.s[x9, 2], { z12.b-z15.b }, z3.b[2]\n"
      ".inst 0xa043835d  // ldnt1b { z28.b-z31.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
      ".inst 0xc153bba3  // sdot za.s[x9, 3], { z28.b-z31.b }, z3.b[2]\n"
      "addvl x26, x26, #16\n"
      "ble 40f\n"
      ".inst 0xa0408355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26]\n"
      ".inst 0xc153bea0  // sdot za.s[x9, 0], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0418355  // ldnt1b { z20.b-z23.b }, pn8.b/Z, [x26, #0x4, MUL VL]\n"
      ".inst 0xc153bea1  // sdot za.s[x9, 1], { z20.b-z23.b }, z3.b[3]\n"
      ".inst 0xa0428351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0x8, MUL VL]\n"
      ".inst 0xc153be22  // sdot za.s[x9, 2], { z16.b-z19.b }, z3.b[3]\n"
      ".inst 0xa0438351  // ldnt1b { z16.b-z19.b }, pn8.b/Z, [x26, #0xc, MUL VL]\n"
      ".inst 0xc153be23  // sdot za.s[x9, 3], { z16.b-z19.b }, z3.b[3]\n"
      "addvl x26, x26, #16\n"
      "40:"  // Width 4: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 41f\n"
      "sdot z26.s, z3.b, z24.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 { z10.s }, p2/Z, [x21]\n"
      "neg z10.s, p2/M, z10.s\n"
      "whilelt p0.s, XZR, x20\n"
      "saddv d26, p0, z26.s\n"
      "mov z26.s, z26.s[0]\n"
      "mul z26.s, p2/M, z26.s, z10.s\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "42:"  // Width 4: skip row sum fixup
      ".inst 0xc0904b40  // addha za0.s, p2/M, p2/M, z26.s\n"
      "add x20, %x[qp], %[per_layer_mul]\n"
      "ld1rw { z5.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      ".inst 0xc0904b41  // addha za1.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z4.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[c_offset]\n"
      "add x21, %x[qp], %[minval]\n"
      ".inst 0xc0904b42  // addha za2.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z6.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[maxval]\n"
      ".inst 0xc0904b43  // addha za3.s, p2/M, p2/M, z26.s\n"
      "ld1rw { z21.s }, p2/Z, [x21]\n"
      ".inst 0xc0062c1c  // mova { z28.d-z31.d }, za.d[x9, #0]\n"
      ".inst 0xc1a5ac1c  // sqdmulh { z28.s-z31.s }, { z28.s-z31.s }, z5.s\n"
      ".inst 0xc0062c2c  // mova { z12.d-z15.d }, za.d[x9, #1]\n"
      ".inst 0xc1a5ac0c  // sqdmulh { z12.s-z15.s }, { z12.s-z15.s }, z5.s\n"
      ".inst 0xc0062c40  // mova { z0.d-z3.d }, za.d[x9, #2]\n"
      ".inst 0xc1a5ac00  // sqdmulh { z0.s-z3.s }, { z0.s-z3.s }, z5.s\n"
      ".inst 0xc0062c68  // mova { z8.d-z11.d }, za.d[x9, #3]\n"
      ".inst 0xc1a5ac08  // sqdmulh { z8.s-z11.s }, { z8.s-z11.s }, z5.s\n"
      ".inst 0xc1a4aa3c  // srshl { z28.s-z31.s }, { z28.s-z31.s }, z4.s\n"
      "ld1rw { z16.s }, p2/Z, [x20]\n"
      ".inst 0xc1a4aa2c  // srshl { z12.s-z15.s }, { z12.s-z15.s }, z4.s\n"
      ".inst 0xc1a4aa20  // srshl { z0.s-z3.s }, { z0.s-z3.s }, z4.s\n"
      ".inst 0xc1a4aa28  // srshl { z8.s-z11.s }, { z8.s-z11.s }, z4.s\n"
      ".inst 0xc1a6ab1c  // add { z28.s-z31.s }, { z28.s-z31.s }, z6.s\n"
      ".inst 0xc1a6ab0c  // add { z12.s-z15.s }, { z12.s-z15.s }, z6.s\n"
      ".inst 0xc1a6ab00  // add { z0.s-z3.s }, { z0.s-z3.s }, z6.s\n"
      ".inst 0xc1a6ab08  // add { z8.s-z11.s }, { z8.s-z11.s }, z6.s\n"
      ".inst 0xc1b0cebc  // sclamp { z28.s-z31.s }, z21.s, z16.s\n"
      ".inst 0xc1b0ceac  // sclamp { z12.s-z15.s }, z21.s, z16.s\n"
      "uzp1 z28.h, z28.h, z29.h\n"
      ".inst 0xc1b0cea0  // sclamp { z0.s-z3.s }, z21.s, z16.s\n"
      ".inst 0xc1b0cea8  // sclamp { z8.s-z11.s }, z21.s, z16.s\n"
      "uzp1 z29.h, z30.h, z31.h\n"
      "uzp1 z12.h, z12.h, z13.h\n"
      "uzp1 z13.h, z14.h, z15.h\n"
      "uzp1 z0.h, z0.h, z1.h\n"
      "uzp1 z1.h, z2.h, z3.h\n"
      "uzp1 z8.h, z8.h, z9.h\n"
      "uzp1 z9.h, z10.h, z11.h\n"
      "uzp1 z28.b, z28.b, z29.b\n"
      "st1b { z28.b }, p2, [x25]\n"
      "uzp1 z12.b, z12.b, z13.b\n"
      "st1b { z12.b }, p2, [x25, #1, MUL VL]\n"
      "uzp1 z0.b, z0.b, z1.b\n"
      "uzp1 z8.b, z8.b, z9.b\n"
      "st1b { z0.b }, p2, [x25, #2, MUL VL]\n"
      "st1b { z8.b }, p1, [x25, #3, MUL VL]\n"
      "addvl x25, x25, #4\n"
      "43:"  // Width 4: Output done
      "subs x27, x27, #0x4\n"
      "sub %x[N], %x[N], x28, LSL #2\n"
      "bgt 4b\n"
      "44:"  // Exit
      ".inst 0xd503467f  // SMSTOP\n"
      "ptrue p2.b\n"
      : [N] "+&r" (N), [flags] "+&r" (flags)
      : [A_ptr] "r" (A_ptr), [B_ptr] "r" (B_ptr), [K] "r" (K), [b_offset] "I" (offsetof(Requantize32, b_offset)), [c_offset] "I" (offsetof(Requantize32, c_offset)), [col_bias] "r" (col_bias), [maxval] "I" (offsetof(Requantize32, maxval)), [minval] "I" (offsetof(Requantize32, minval)), [output_ptr] "r" (output_ptr), [per_layer_mul] "I" (offsetof(Requantize32, per_layer_mul)), [per_layer_right_shift] "I" (offsetof(Requantize32, per_layer_right_shift)), [qp] "r" (qp)
      : "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", "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"
    );
}

} // namespace arm_gemm

#endif // ARM_COMPUTE_ENABLE_SME2
#endif