/*
 * Copyright (c) 2019-2020 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

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

#include <cassert>
#include <cstdint>

namespace arm_gemm {

void sve_hybrid_u8qa_dot_4x4VL (
    unsigned int num_strings, const unsigned int *string_lengths, IndirectInputArg<uint8_t> A_arg,
    size_t M, size_t N, const uint8_t *B_ptr, IndirectOutputArg<uint8_t> output_arg,
    const Requantize32 *qp, const int32_t *col_bias, unsigned int
)
{
    struct KernelArgs {
        unsigned int num_strings = {};
        const unsigned int *string_lengths = {};
        size_t N = {};
        const uint8_t *B_ptr = {};
        size_t output_offset = {};
        size_t input_initial_col = {};
        size_t input_offset = {};
    } ka;

    unsigned long flags=0;
    void *output_ptr;
    void *input_ptr;

    if (output_arg.is_indirect) {
        output_ptr=(void *)(output_arg.indirect.ptr);
        ka.output_offset=output_arg.indirect.offset;
        flags |= 0x4;
    } else {
        output_ptr=(void *)(output_arg.direct.base);
        ka.output_offset=output_arg.direct.stride;
    }

    if (A_arg.is_indirect) {
        input_ptr=(void *)(A_arg.indirect.ptr);
        ka.input_offset=A_arg.indirect.start_row;
        ka.input_initial_col=A_arg.indirect.start_col;
        flags |= 0x8;
    } else {
        assert(num_strings==1);
        input_ptr=(void *)(A_arg.direct.base);
        ka.input_offset=A_arg.direct.stride;
    }
    ka.num_strings = num_strings;
    ka.string_lengths = string_lengths;
    ka.N = N;
    ka.B_ptr = B_ptr;
    if (qp->c_offset > qp->minval) {
        flags |= 0x20;
    }
    __asm__ __volatile__(
      "ptrue p2.b\n"
      "1:"  // Row loop
      "cmp %x[M], #0x4\n"
      "bge 46f\n"
      "cmp %x[M], #0x2\n"
      "bgt 31f\n"
      "beq 16f\n"
      "mov z11.s, #0x0\n"
      "ldr x12, [%x[args_ptr], %[offsetof_N]]\n"
      "mov z12.s, #0x0\n"
      "ldr x11, [%x[args_ptr], %[offsetof_B_ptr]]\n"
      "mov x10, %x[col_bias]\n"
      "mov z13.s, #0x0\n"
      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
      "bic %x[flags], %x[flags], #0x80000000\n"
      "mov z14.s, #0x0\n"
      "mov z15.b, #0x1\n"
      "tbz %x[flags], #2, 2f\n"
      "ldr x9, [%x[output_ptr], #0x0]\n"
      "add x9, x9, x19\n"
      "b 3f\n"
      "2:"  // Height 1: setup direct output
      "mov x9, %x[output_ptr]\n"
      "3:"  // Height 1: Column loop
      "mov z16.s, #0x0\n"
      "mov x19, #0x0\n"
      "mov z17.s, #0x0\n"
      "whilelt p1.b, x19, x12\n"
      "mov z18.s, #0x0\n"
      "mov z19.s, #0x0\n"
      "4:"  // Height 1: setup done
      "mov x28, #0x0\n"
      "5:"  // Height 1: String loop
      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "ldr w27, [x20, x28, LSL #0x2]\n"
      "tbz %x[flags], #3, 6f\n"
      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
      "add x20, x20, x19, LSL #3\n"
      "ldr x26, [x20, #0x0]\n"
      "cbnz x28, 7f\n"
      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
      "add x26, x26, x19\n"
      "b 7f\n"
      "6:"  // Height 1: setup direct input
      "mov x26, %x[input_ptr]\n"
      "7:"  // Height 1: input setup done
      "cmp x27, #0x10\n"
      "ble 10f\n"
      "8:"  // Height 1: Multiply loop: Main loop head
      "ld1b { z4.b }, p2/Z, [x11]\n"
      "whilelt p0.b, XZR, x27\n"
      "ld1b { z5.b }, p2/Z, [x11, #1, MUL VL]\n"
      "ld1rqb { z0.b }, p0/Z, [x26]\n"
      "udot z16.s, z4.b, z0.b[0]\n"
      "ld1b { z6.b }, p2/Z, [x11, #2, MUL VL]\n"
      "add x26, x26, #0x10\n"
      "udot z17.s, z5.b, z0.b[0]\n"
      "ld1b { z7.b }, p2/Z, [x11, #3, MUL VL]\n"
      "ld1b { z8.b }, p2/Z, [x11, #4, MUL VL]\n"
      "udot z18.s, z6.b, z0.b[0]\n"
      "ld1b { z9.b }, p2/Z, [x11, #5, MUL VL]\n"
      "ld1b { z10.b }, p2/Z, [x11, #6, MUL VL]\n"
      "udot z19.s, z7.b, z0.b[0]\n"
      "udot z16.s, z8.b, z0.b[1]\n"
      "ld1b { z4.b }, p2/Z, [x11, #7, MUL VL]\n"
      "addvl x11, x11, #16\n"
      "udot z17.s, z9.b, z0.b[1]\n"
      "ld1b { z5.b }, p2/Z, [x11, #-8, MUL VL]\n"
      "ld1b { z6.b }, p2/Z, [x11, #-7, MUL VL]\n"
      "udot z18.s, z10.b, z0.b[1]\n"
      "ld1b { z7.b }, p2/Z, [x11, #-6, MUL VL]\n"
      "udot z19.s, z4.b, z0.b[1]\n"
      "ld1b { z8.b }, p2/Z, [x11, #-5, MUL VL]\n"
      "udot z16.s, z5.b, z0.b[2]\n"
      "ld1b { z9.b }, p2/Z, [x11, #-4, MUL VL]\n"
      "udot z17.s, z6.b, z0.b[2]\n"
      "ld1b { z10.b }, p2/Z, [x11, #-3, MUL VL]\n"
      "udot z18.s, z7.b, z0.b[2]\n"
      "ld1b { z4.b }, p2/Z, [x11, #-2, MUL VL]\n"
      "udot z19.s, z8.b, z0.b[2]\n"
      "ld1b { z5.b }, p2/Z, [x11, #-1, MUL VL]\n"
      "udot z16.s, z9.b, z0.b[3]\n"
      "udot z17.s, z10.b, z0.b[3]\n"
      "udot z18.s, z4.b, z0.b[3]\n"
      "udot z19.s, z5.b, z0.b[3]\n"
      "tbnz %x[flags], #31, 9f\n"
      "udot z11.s, z0.b, z15.b\n"
      "9:"  // Height 1: Multiply loop: unique 1: skip row sum
      "prfm pldl1keep, [x26, #0x80]\n"
      "sub x27, x27, #0x10\n"
      "cmp x27, #0x10\n"
      "bgt 8b\n"
      "10:"  // Height 1: Multiply loop: Single iteration only
      "ld1b { z6.b }, p2/Z, [x11]\n"
      "whilelt p0.b, XZR, x27\n"
      "ld1b { z7.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "ld1rqb { z0.b }, p0/Z, [x26]\n"
      "udot z16.s, z6.b, z0.b[0]\n"
      "ld1b { z8.b }, p2/Z, [x11, #2, MUL VL]\n"
      "add x26, x26, #0x10\n"
      "udot z17.s, z7.b, z0.b[0]\n"
      "ld1b { z9.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z18.s, z8.b, z0.b[0]\n"
      "udot z19.s, z9.b, z0.b[0]\n"
      "ble 11f\n"
      "ld1b { z10.b }, p2/Z, [x11]\n"
      "udot z16.s, z10.b, z0.b[1]\n"
      "ld1b { z4.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "udot z17.s, z4.b, z0.b[1]\n"
      "ld1b { z5.b }, p2/Z, [x11, #2, MUL VL]\n"
      "ld1b { z6.b }, p2/Z, [x11, #3, MUL VL]\n"
      "udot z18.s, z5.b, z0.b[1]\n"
      "addvl x11, x11, #4\n"
      "udot z19.s, z6.b, z0.b[1]\n"
      "ble 11f\n"
      "ld1b { z7.b }, p2/Z, [x11]\n"
      "udot z16.s, z7.b, z0.b[2]\n"
      "ld1b { z8.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "udot z17.s, z8.b, z0.b[2]\n"
      "ld1b { z9.b }, p2/Z, [x11, #2, MUL VL]\n"
      "ld1b { z10.b }, p2/Z, [x11, #3, MUL VL]\n"
      "udot z18.s, z9.b, z0.b[2]\n"
      "addvl x11, x11, #4\n"
      "udot z19.s, z10.b, z0.b[2]\n"
      "ble 11f\n"
      "ld1b { z4.b }, p2/Z, [x11]\n"
      "udot z16.s, z4.b, z0.b[3]\n"
      "ld1b { z5.b }, p2/Z, [x11, #1, MUL VL]\n"
      "ld1b { z6.b }, p2/Z, [x11, #2, MUL VL]\n"
      "udot z17.s, z5.b, z0.b[3]\n"
      "ld1b { z7.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z18.s, z6.b, z0.b[3]\n"
      "udot z19.s, z7.b, z0.b[3]\n"
      "11:"  // Height 1: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 12f\n"
      "udot z11.s, z0.b, z15.b\n"
      "12:"  // Height 1: Multiply loop: unique 2: skip row sum
      "prfm pldl1keep, [x26, #0x80]\n"
      "add x28, x28, #0x1\n"
      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
      "cmp x28, x19\n"
      "bne 5b\n"
      "prfm pstl1keep, [x9, #0x0]\n"
      "tbnz %x[flags], #31, 13f\n"
      "add x19, %x[qp], %[b_offset]\n"
      "ld1rw { z1.s }, p2/Z, [x19]\n"
      "neg z1.s, p2/M, z1.s\n"
      "mov x19, #0x4\n"
      "whilelt p0.s, XZR, x19\n"
      "uaddv d11, p0, z11.s\n"
      "mov z11.s, z11.s[0]\n"
      "mul z11.s, p2/M, z11.s, z1.s\n"
      "13:"  // Height 1: skip row sum fixup
      "add z16.s, z16.s, z11.s\n"
      "ld1w { z0.s }, p2/Z, [x10]\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "add z17.s, z17.s, z11.s\n"
      "ld1w { z1.s }, p2/Z, [x10, #1, MUL VL]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      "add z18.s, z18.s, z11.s\n"
      "ld1w { z2.s }, p2/Z, [x10, #2, MUL VL]\n"
      "add x19, %x[qp], %[per_layer_mul]\n"
      "add z19.s, z19.s, z11.s\n"
      "ld1w { z3.s }, p2/Z, [x10, #3, MUL VL]\n"
      "addvl x10, x10, #4\n"
      "add z16.s, z16.s, z0.s\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      "add z17.s, z17.s, z1.s\n"
      "ld1rw { z4.s }, p2/Z, [x19]\n"
      "add z18.s, z18.s, z2.s\n"
      "add z19.s, z19.s, z3.s\n"
      ".inst 0x04a47610  // sqrdmulh z16.s, z16.s, z4.s\n"
      ".inst 0x04a47631  // sqrdmulh z17.s, z17.s, z4.s\n"
      ".inst 0x04a47652  // sqrdmulh z18.s, z18.s, z4.s\n"
      ".inst 0x04a47673  // sqrdmulh z19.s, z19.s, z4.s\n"
      "tbz %x[flags], #5, 14f\n"
      "and z4.d, z16.d, z0.d\n"
      "asr z4.s, z4.s, #0x1f\n"
      "and z5.d, z17.d, z0.d\n"
      "and z6.d, z18.d, z0.d\n"
      "asr z5.s, z5.s, #0x1f\n"
      "and z7.d, z19.d, z0.d\n"
      "asr z6.s, z6.s, #0x1f\n"
      "sqadd z16.s, z16.s, z4.s\n"
      "asr z7.s, z7.s, #0x1f\n"
      "sqadd z17.s, z17.s, z5.s\n"
      "sqadd z18.s, z18.s, z6.s\n"
      "sqadd z19.s, z19.s, z7.s\n"
      "14:"  // Height 1: no shift correction
      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
      "add x19, %x[qp], %[c_offset]\n"
      "ld1rw { z4.s }, p2/Z, [x19]\n"
      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
      "add x19, %x[qp], %[minval]\n"
      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
      "ld1rw { z5.s }, p2/Z, [x19]\n"
      "add x19, %x[qp], %[maxval]\n"
      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
      "ld1rw { z6.s }, p2/Z, [x19]\n"
      "add z16.s, z16.s, z4.s\n"
      "add z17.s, z17.s, z4.s\n"
      "add z18.s, z18.s, z4.s\n"
      "add z19.s, z19.s, z4.s\n"
      "smin z16.s, p2/M, z16.s, z6.s\n"
      "smin z17.s, p2/M, z17.s, z6.s\n"
      "smin z18.s, p2/M, z18.s, z6.s\n"
      "smin z19.s, p2/M, z19.s, z6.s\n"
      "smax z16.s, p2/M, z16.s, z5.s\n"
      "smax z17.s, p2/M, z17.s, z5.s\n"
      "smax z18.s, p2/M, z18.s, z5.s\n"
      "smax z19.s, p2/M, z19.s, z5.s\n"
      "uzp1 z16.h, z16.h, z17.h\n"
      "uzp1 z17.h, z18.h, z19.h\n"
      "uzp1 z16.b, z16.b, z17.b\n"
      "st1b { z16.b }, p1, [x9]\n"
      "addvl x9, x9, #1\n"
      "15:"  // Height 1: Writeback done
      "mov x19, #0x0\n"
      "incw x19, ALL, MUL #4\n"
      "subs x12, x12, x19\n"
      "bgt 3b\n"
      "b 62f\n"
      "16:"  // Height 2
      "mov z11.s, #0x0\n"
      "ldr x12, [%x[args_ptr], %[offsetof_N]]\n"
      "mov x10, %x[col_bias]\n"
      "mov z12.s, #0x0\n"
      "ldr x11, [%x[args_ptr], %[offsetof_B_ptr]]\n"
      "bic %x[flags], %x[flags], #0x80000000\n"
      "mov z13.s, #0x0\n"
      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
      "mov z14.s, #0x0\n"
      "mov z15.b, #0x1\n"
      "tbz %x[flags], #2, 17f\n"
      "ldr x9, [%x[output_ptr], #0x0]\n"
      "ldr x25, [%x[output_ptr], #0x8]\n"
      "add x9, x9, x19\n"
      "add x25, x25, x19\n"
      "b 18f\n"
      "17:"  // Height 2: setup direct output
      "mov x9, %x[output_ptr]\n"
      "add x25, x9, x19\n"
      "18:"  // Height 2: Column loop
      "mov z16.s, #0x0\n"
      "mov x19, #0x0\n"
      "mov z17.s, #0x0\n"
      "whilelt p1.b, x19, x12\n"
      "mov z18.s, #0x0\n"
      "mov z19.s, #0x0\n"
      "mov z20.s, #0x0\n"
      "mov z21.s, #0x0\n"
      "mov z22.s, #0x0\n"
      "mov z23.s, #0x0\n"
      "19:"  // Height 2: setup done
      "mov x28, #0x0\n"
      "20:"  // Height 2: String loop
      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "ldr w27, [x20, x28, LSL #0x2]\n"
      "tbz %x[flags], #3, 21f\n"
      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
      "add x20, x20, x19, LSL #3\n"
      "ldr x26, [x20, #0x0]\n"
      "ldr x24, [x20, #0x8]\n"
      "cbnz x28, 22f\n"
      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
      "add x26, x26, x19\n"
      "add x24, x24, x19\n"
      "b 22f\n"
      "21:"  // Height 2: setup direct input
      "mov x26, %x[input_ptr]\n"
      "add x24, x26, x19\n"
      "22:"  // Height 2: input setup done
      "cmp x27, #0x10\n"
      "ble 25f\n"
      "23:"  // Height 2: Multiply loop: Main loop head
      "ld1b { z4.b }, p2/Z, [x11]\n"
      "whilelt p0.b, XZR, x27\n"
      "ld1b { z5.b }, p2/Z, [x11, #1, MUL VL]\n"
      "ld1rqb { z0.b }, p0/Z, [x26]\n"
      "udot z16.s, z4.b, z0.b[0]\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "add x26, x26, #0x10\n"
      "udot z17.s, z5.b, z0.b[0]\n"
      "ld1b { z6.b }, p2/Z, [x11, #2, MUL VL]\n"
      "add x24, x24, #0x10\n"
      "udot z20.s, z4.b, z1.b[0]\n"
      "ld1b { z7.b }, p2/Z, [x11, #3, MUL VL]\n"
      "udot z21.s, z5.b, z1.b[0]\n"
      "ld1b { z8.b }, p2/Z, [x11, #4, MUL VL]\n"
      "udot z18.s, z6.b, z0.b[0]\n"
      "ld1b { z9.b }, p2/Z, [x11, #5, MUL VL]\n"
      "udot z22.s, z6.b, z1.b[0]\n"
      "ld1b { z10.b }, p2/Z, [x11, #6, MUL VL]\n"
      "udot z19.s, z7.b, z0.b[0]\n"
      "ld1b { z4.b }, p2/Z, [x11, #7, MUL VL]\n"
      "addvl x11, x11, #16\n"
      "udot z23.s, z7.b, z1.b[0]\n"
      "ld1b { z5.b }, p2/Z, [x11, #-8, MUL VL]\n"
      "udot z16.s, z8.b, z0.b[1]\n"
      "ld1b { z6.b }, p2/Z, [x11, #-7, MUL VL]\n"
      "udot z20.s, z8.b, z1.b[1]\n"
      "ld1b { z7.b }, p2/Z, [x11, #-6, MUL VL]\n"
      "udot z17.s, z9.b, z0.b[1]\n"
      "ld1b { z8.b }, p2/Z, [x11, #-5, MUL VL]\n"
      "udot z21.s, z9.b, z1.b[1]\n"
      "ld1b { z9.b }, p2/Z, [x11, #-4, MUL VL]\n"
      "udot z18.s, z10.b, z0.b[1]\n"
      "udot z22.s, z10.b, z1.b[1]\n"
      "ld1b { z10.b }, p2/Z, [x11, #-3, MUL VL]\n"
      "udot z19.s, z4.b, z0.b[1]\n"
      "udot z23.s, z4.b, z1.b[1]\n"
      "ld1b { z4.b }, p2/Z, [x11, #-2, MUL VL]\n"
      "udot z16.s, z5.b, z0.b[2]\n"
      "udot z20.s, z5.b, z1.b[2]\n"
      "ld1b { z5.b }, p2/Z, [x11, #-1, MUL VL]\n"
      "udot z17.s, z6.b, z0.b[2]\n"
      "udot z21.s, z6.b, z1.b[2]\n"
      "udot z18.s, z7.b, z0.b[2]\n"
      "udot z22.s, z7.b, z1.b[2]\n"
      "udot z19.s, z8.b, z0.b[2]\n"
      "udot z23.s, z8.b, z1.b[2]\n"
      "udot z16.s, z9.b, z0.b[3]\n"
      "udot z20.s, z9.b, z1.b[3]\n"
      "udot z17.s, z10.b, z0.b[3]\n"
      "udot z21.s, z10.b, z1.b[3]\n"
      "udot z18.s, z4.b, z0.b[3]\n"
      "udot z22.s, z4.b, z1.b[3]\n"
      "udot z19.s, z5.b, z0.b[3]\n"
      "udot z23.s, z5.b, z1.b[3]\n"
      "tbnz %x[flags], #31, 24f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z12.s, z1.b, z15.b\n"
      "24:"  // Height 2: Multiply loop: unique 3: skip row sum
      "prfm pldl1keep, [x26, #0x80]\n"
      "sub x27, x27, #0x10\n"
      "prfm pldl1keep, [x24, #0x80]\n"
      "cmp x27, #0x10\n"
      "bgt 23b\n"
      "25:"  // Height 2: Multiply loop: Single iteration only
      "ld1b { z6.b }, p2/Z, [x11]\n"
      "whilelt p0.b, XZR, x27\n"
      "ld1b { z7.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "ld1rqb { z0.b }, p0/Z, [x26]\n"
      "udot z16.s, z6.b, z0.b[0]\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "add x26, x26, #0x10\n"
      "udot z17.s, z7.b, z0.b[0]\n"
      "ld1b { z8.b }, p2/Z, [x11, #2, MUL VL]\n"
      "add x24, x24, #0x10\n"
      "udot z20.s, z6.b, z1.b[0]\n"
      "ld1b { z9.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z21.s, z7.b, z1.b[0]\n"
      "udot z18.s, z8.b, z0.b[0]\n"
      "udot z22.s, z8.b, z1.b[0]\n"
      "udot z19.s, z9.b, z0.b[0]\n"
      "udot z23.s, z9.b, z1.b[0]\n"
      "ble 26f\n"
      "ld1b { z10.b }, p2/Z, [x11]\n"
      "udot z16.s, z10.b, z0.b[1]\n"
      "ld1b { z4.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "udot z20.s, z10.b, z1.b[1]\n"
      "ld1b { z5.b }, p2/Z, [x11, #2, MUL VL]\n"
      "ld1b { z6.b }, p2/Z, [x11, #3, MUL VL]\n"
      "udot z17.s, z4.b, z0.b[1]\n"
      "addvl x11, x11, #4\n"
      "udot z21.s, z4.b, z1.b[1]\n"
      "udot z18.s, z5.b, z0.b[1]\n"
      "udot z22.s, z5.b, z1.b[1]\n"
      "udot z19.s, z6.b, z0.b[1]\n"
      "udot z23.s, z6.b, z1.b[1]\n"
      "ble 26f\n"
      "ld1b { z7.b }, p2/Z, [x11]\n"
      "udot z16.s, z7.b, z0.b[2]\n"
      "ld1b { z8.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "udot z20.s, z7.b, z1.b[2]\n"
      "ld1b { z9.b }, p2/Z, [x11, #2, MUL VL]\n"
      "ld1b { z10.b }, p2/Z, [x11, #3, MUL VL]\n"
      "udot z17.s, z8.b, z0.b[2]\n"
      "addvl x11, x11, #4\n"
      "udot z21.s, z8.b, z1.b[2]\n"
      "udot z18.s, z9.b, z0.b[2]\n"
      "udot z22.s, z9.b, z1.b[2]\n"
      "udot z19.s, z10.b, z0.b[2]\n"
      "udot z23.s, z10.b, z1.b[2]\n"
      "ble 26f\n"
      "ld1b { z4.b }, p2/Z, [x11]\n"
      "udot z16.s, z4.b, z0.b[3]\n"
      "ld1b { z5.b }, p2/Z, [x11, #1, MUL VL]\n"
      "udot z20.s, z4.b, z1.b[3]\n"
      "ld1b { z6.b }, p2/Z, [x11, #2, MUL VL]\n"
      "ld1b { z7.b }, p2/Z, [x11, #3, MUL VL]\n"
      "udot z17.s, z5.b, z0.b[3]\n"
      "addvl x11, x11, #4\n"
      "udot z21.s, z5.b, z1.b[3]\n"
      "udot z18.s, z6.b, z0.b[3]\n"
      "udot z22.s, z6.b, z1.b[3]\n"
      "udot z19.s, z7.b, z0.b[3]\n"
      "udot z23.s, z7.b, z1.b[3]\n"
      "26:"  // Height 2: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 27f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z12.s, z1.b, z15.b\n"
      "27:"  // Height 2: Multiply loop: unique 4: skip row sum
      "prfm pldl1keep, [x26, #0x80]\n"
      "add x28, x28, #0x1\n"
      "prfm pldl1keep, [x24, #0x80]\n"
      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
      "cmp x28, x19\n"
      "bne 20b\n"
      "prfm pstl1keep, [x9, #0x0]\n"
      "prfm pstl1keep, [x25, #0x0]\n"
      "tbnz %x[flags], #31, 28f\n"
      "add x19, %x[qp], %[b_offset]\n"
      "ld1rw { z2.s }, p2/Z, [x19]\n"
      "neg z2.s, p2/M, z2.s\n"
      "mov x20, #0x4\n"
      "mov x19, #0x4\n"
      "whilelt p0.s, XZR, x20\n"
      "uaddv d11, p0, z11.s\n"
      "whilelt p0.s, XZR, x19\n"
      "uaddv d12, p0, z12.s\n"
      "mov z11.s, z11.s[0]\n"
      "mov z12.s, z12.s[0]\n"
      "mul z11.s, p2/M, z11.s, z2.s\n"
      "mul z12.s, p2/M, z12.s, z2.s\n"
      "28:"  // Height 2: skip row sum fixup
      "add z16.s, z16.s, z11.s\n"
      "ld1w { z0.s }, p2/Z, [x10]\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "add z17.s, z17.s, z11.s\n"
      "ld1w { z1.s }, p2/Z, [x10, #1, MUL VL]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      "add z18.s, z18.s, z11.s\n"
      "ld1w { z2.s }, p2/Z, [x10, #2, MUL VL]\n"
      "add x19, %x[qp], %[per_layer_mul]\n"
      "add z19.s, z19.s, z11.s\n"
      "ld1w { z3.s }, p2/Z, [x10, #3, MUL VL]\n"
      "addvl x10, x10, #4\n"
      "add z20.s, z20.s, z12.s\n"
      "ld1rw { z4.s }, p2/Z, [x19]\n"
      "add z21.s, z21.s, z12.s\n"
      "add z22.s, z22.s, z12.s\n"
      "add z23.s, z23.s, z12.s\n"
      "add z16.s, z16.s, z0.s\n"
      "add z17.s, z17.s, z1.s\n"
      "add z18.s, z18.s, z2.s\n"
      "add z19.s, z19.s, z3.s\n"
      "add z20.s, z20.s, z0.s\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      "add z21.s, z21.s, z1.s\n"
      "add z22.s, z22.s, z2.s\n"
      "add z23.s, z23.s, z3.s\n"
      ".inst 0x04a47610  // sqrdmulh z16.s, z16.s, z4.s\n"
      ".inst 0x04a47631  // sqrdmulh z17.s, z17.s, z4.s\n"
      ".inst 0x04a47652  // sqrdmulh z18.s, z18.s, z4.s\n"
      ".inst 0x04a47673  // sqrdmulh z19.s, z19.s, z4.s\n"
      ".inst 0x04a47694  // sqrdmulh z20.s, z20.s, z4.s\n"
      ".inst 0x04a476b5  // sqrdmulh z21.s, z21.s, z4.s\n"
      ".inst 0x04a476d6  // sqrdmulh z22.s, z22.s, z4.s\n"
      ".inst 0x04a476f7  // sqrdmulh z23.s, z23.s, z4.s\n"
      "tbz %x[flags], #5, 29f\n"
      "and z4.d, z16.d, z0.d\n"
      "asr z4.s, z4.s, #0x1f\n"
      "and z5.d, z17.d, z0.d\n"
      "and z6.d, z18.d, z0.d\n"
      "asr z5.s, z5.s, #0x1f\n"
      "and z7.d, z19.d, z0.d\n"
      "and z8.d, z20.d, z0.d\n"
      "asr z6.s, z6.s, #0x1f\n"
      "and z9.d, z21.d, z0.d\n"
      "asr z7.s, z7.s, #0x1f\n"
      "sqadd z16.s, z16.s, z4.s\n"
      "and z10.d, z22.d, z0.d\n"
      "asr z8.s, z8.s, #0x1f\n"
      "and z4.d, z23.d, z0.d\n"
      "asr z9.s, z9.s, #0x1f\n"
      "sqadd z17.s, z17.s, z5.s\n"
      "asr z10.s, z10.s, #0x1f\n"
      "sqadd z18.s, z18.s, z6.s\n"
      "asr z4.s, z4.s, #0x1f\n"
      "sqadd z19.s, z19.s, z7.s\n"
      "sqadd z20.s, z20.s, z8.s\n"
      "sqadd z21.s, z21.s, z9.s\n"
      "sqadd z22.s, z22.s, z10.s\n"
      "sqadd z23.s, z23.s, z4.s\n"
      "29:"  // Height 2: no shift correction
      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
      "add x19, %x[qp], %[c_offset]\n"
      "ld1rw { z4.s }, p2/Z, [x19]\n"
      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
      "add x19, %x[qp], %[minval]\n"
      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
      "ld1rw { z5.s }, p2/Z, [x19]\n"
      "add x19, %x[qp], %[maxval]\n"
      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
      "ld1rw { z6.s }, p2/Z, [x19]\n"
      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
      "add z16.s, z16.s, z4.s\n"
      "add z17.s, z17.s, z4.s\n"
      "add z18.s, z18.s, z4.s\n"
      "add z19.s, z19.s, z4.s\n"
      "add z20.s, z20.s, z4.s\n"
      "smin z16.s, p2/M, z16.s, z6.s\n"
      "smin z17.s, p2/M, z17.s, z6.s\n"
      "smin z18.s, p2/M, z18.s, z6.s\n"
      "smin z19.s, p2/M, z19.s, z6.s\n"
      "smax z16.s, p2/M, z16.s, z5.s\n"
      "smax z17.s, p2/M, z17.s, z5.s\n"
      "smax z18.s, p2/M, z18.s, z5.s\n"
      "smax z19.s, p2/M, z19.s, z5.s\n"
      "smin z20.s, p2/M, z20.s, z6.s\n"
      "uzp1 z16.h, z16.h, z17.h\n"
      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
      "uzp1 z17.h, z18.h, z19.h\n"
      "smax z20.s, p2/M, z20.s, z5.s\n"
      "uzp1 z16.b, z16.b, z17.b\n"
      "st1b { z16.b }, p1, [x9]\n"
      "add z21.s, z21.s, z4.s\n"
      "addvl x9, x9, #1\n"
      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
      "smin z21.s, p2/M, z21.s, z6.s\n"
      "add z22.s, z22.s, z4.s\n"
      "add z23.s, z23.s, z4.s\n"
      "smax z21.s, p2/M, z21.s, z5.s\n"
      "smin z22.s, p2/M, z22.s, z6.s\n"
      "smin z23.s, p2/M, z23.s, z6.s\n"
      "uzp1 z20.h, z20.h, z21.h\n"
      "smax z22.s, p2/M, z22.s, z5.s\n"
      "smax z23.s, p2/M, z23.s, z5.s\n"
      "uzp1 z21.h, z22.h, z23.h\n"
      "uzp1 z20.b, z20.b, z21.b\n"
      "st1b { z20.b }, p1, [x25]\n"
      "addvl x25, x25, #1\n"
      "30:"  // Height 2: Writeback done
      "mov x19, #0x0\n"
      "incw x19, ALL, MUL #4\n"
      "subs x12, x12, x19\n"
      "bgt 18b\n"
      "b 62f\n"
      "31:"  // Height 3
      "mov z11.s, #0x0\n"
      "ldr x12, [%x[args_ptr], %[offsetof_N]]\n"
      "mov x10, %x[col_bias]\n"
      "mov z12.s, #0x0\n"
      "ldr x11, [%x[args_ptr], %[offsetof_B_ptr]]\n"
      "bic %x[flags], %x[flags], #0x80000000\n"
      "mov z13.s, #0x0\n"
      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
      "mov z14.s, #0x0\n"
      "mov z15.b, #0x1\n"
      "tbz %x[flags], #2, 32f\n"
      "ldr x9, [%x[output_ptr], #0x0]\n"
      "ldr x25, [%x[output_ptr], #0x8]\n"
      "add x9, x9, x19\n"
      "ldr x23, [%x[output_ptr], #0x10]\n"
      "add x25, x25, x19\n"
      "add x23, x23, x19\n"
      "b 33f\n"
      "32:"  // Height 3: setup direct output
      "mov x9, %x[output_ptr]\n"
      "add x25, x9, x19\n"
      "add x23, x25, x19\n"
      "33:"  // Height 3: Column loop
      "mov z16.s, #0x0\n"
      "mov x19, #0x0\n"
      "mov z17.s, #0x0\n"
      "whilelt p1.b, x19, x12\n"
      "mov z18.s, #0x0\n"
      "mov z19.s, #0x0\n"
      "mov z20.s, #0x0\n"
      "mov z21.s, #0x0\n"
      "mov z22.s, #0x0\n"
      "mov z23.s, #0x0\n"
      "mov z24.s, #0x0\n"
      "mov z25.s, #0x0\n"
      "mov z26.s, #0x0\n"
      "mov z27.s, #0x0\n"
      "34:"  // Height 3: setup done
      "mov x28, #0x0\n"
      "35:"  // Height 3: String loop
      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "ldr w27, [x20, x28, LSL #0x2]\n"
      "tbz %x[flags], #3, 36f\n"
      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
      "add x20, x20, x19, LSL #3\n"
      "ldr x26, [x20, #0x0]\n"
      "ldr x24, [x20, #0x8]\n"
      "ldr x22, [x20, #0x10]\n"
      "cbnz x28, 37f\n"
      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
      "add x26, x26, x19\n"
      "add x24, x24, x19\n"
      "add x22, x22, x19\n"
      "b 37f\n"
      "36:"  // Height 3: setup direct input
      "mov x26, %x[input_ptr]\n"
      "add x24, x26, x19\n"
      "add x22, x24, x19\n"
      "37:"  // Height 3: input setup done
      "cmp x27, #0x10\n"
      "ble 40f\n"
      "38:"  // Height 3: Multiply loop: Main loop head
      "ld1b { z4.b }, p2/Z, [x11]\n"
      "whilelt p0.b, XZR, x27\n"
      "ld1b { z5.b }, p2/Z, [x11, #1, MUL VL]\n"
      "ld1rqb { z0.b }, p0/Z, [x26]\n"
      "udot z16.s, z4.b, z0.b[0]\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "add x26, x26, #0x10\n"
      "udot z17.s, z5.b, z0.b[0]\n"
      "ld1rqb { z2.b }, p0/Z, [x22]\n"
      "add x24, x24, #0x10\n"
      "udot z20.s, z4.b, z1.b[0]\n"
      "ld1b { z6.b }, p2/Z, [x11, #2, MUL VL]\n"
      "add x22, x22, #0x10\n"
      "udot z24.s, z4.b, z2.b[0]\n"
      "ld1b { z7.b }, p2/Z, [x11, #3, MUL VL]\n"
      "udot z21.s, z5.b, z1.b[0]\n"
      "ld1b { z8.b }, p2/Z, [x11, #4, MUL VL]\n"
      "udot z25.s, z5.b, z2.b[0]\n"
      "ld1b { z9.b }, p2/Z, [x11, #5, MUL VL]\n"
      "udot z18.s, z6.b, z0.b[0]\n"
      "ld1b { z10.b }, p2/Z, [x11, #6, MUL VL]\n"
      "udot z22.s, z6.b, z1.b[0]\n"
      "ld1b { z4.b }, p2/Z, [x11, #7, MUL VL]\n"
      "addvl x11, x11, #16\n"
      "udot z26.s, z6.b, z2.b[0]\n"
      "ld1b { z5.b }, p2/Z, [x11, #-8, MUL VL]\n"
      "udot z19.s, z7.b, z0.b[0]\n"
      "ld1b { z6.b }, p2/Z, [x11, #-7, MUL VL]\n"
      "udot z23.s, z7.b, z1.b[0]\n"
      "udot z27.s, z7.b, z2.b[0]\n"
      "ld1b { z7.b }, p2/Z, [x11, #-6, MUL VL]\n"
      "udot z16.s, z8.b, z0.b[1]\n"
      "udot z20.s, z8.b, z1.b[1]\n"
      "udot z24.s, z8.b, z2.b[1]\n"
      "ld1b { z8.b }, p2/Z, [x11, #-5, MUL VL]\n"
      "udot z17.s, z9.b, z0.b[1]\n"
      "udot z21.s, z9.b, z1.b[1]\n"
      "udot z25.s, z9.b, z2.b[1]\n"
      "ld1b { z9.b }, p2/Z, [x11, #-4, MUL VL]\n"
      "udot z18.s, z10.b, z0.b[1]\n"
      "udot z22.s, z10.b, z1.b[1]\n"
      "udot z26.s, z10.b, z2.b[1]\n"
      "ld1b { z10.b }, p2/Z, [x11, #-3, MUL VL]\n"
      "udot z19.s, z4.b, z0.b[1]\n"
      "udot z23.s, z4.b, z1.b[1]\n"
      "udot z27.s, z4.b, z2.b[1]\n"
      "ld1b { z4.b }, p2/Z, [x11, #-2, MUL VL]\n"
      "udot z16.s, z5.b, z0.b[2]\n"
      "udot z20.s, z5.b, z1.b[2]\n"
      "udot z24.s, z5.b, z2.b[2]\n"
      "ld1b { z5.b }, p2/Z, [x11, #-1, MUL VL]\n"
      "udot z17.s, z6.b, z0.b[2]\n"
      "udot z21.s, z6.b, z1.b[2]\n"
      "udot z25.s, z6.b, z2.b[2]\n"
      "udot z18.s, z7.b, z0.b[2]\n"
      "udot z22.s, z7.b, z1.b[2]\n"
      "udot z26.s, z7.b, z2.b[2]\n"
      "udot z19.s, z8.b, z0.b[2]\n"
      "udot z23.s, z8.b, z1.b[2]\n"
      "udot z27.s, z8.b, z2.b[2]\n"
      "udot z16.s, z9.b, z0.b[3]\n"
      "udot z20.s, z9.b, z1.b[3]\n"
      "udot z24.s, z9.b, z2.b[3]\n"
      "udot z17.s, z10.b, z0.b[3]\n"
      "udot z21.s, z10.b, z1.b[3]\n"
      "udot z25.s, z10.b, z2.b[3]\n"
      "udot z18.s, z4.b, z0.b[3]\n"
      "udot z22.s, z4.b, z1.b[3]\n"
      "udot z26.s, z4.b, z2.b[3]\n"
      "udot z19.s, z5.b, z0.b[3]\n"
      "udot z23.s, z5.b, z1.b[3]\n"
      "udot z27.s, z5.b, z2.b[3]\n"
      "tbnz %x[flags], #31, 39f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z12.s, z1.b, z15.b\n"
      "udot z13.s, z2.b, z15.b\n"
      "39:"  // Height 3: Multiply loop: unique 5: skip row sum
      "prfm pldl1keep, [x26, #0x80]\n"
      "sub x27, x27, #0x10\n"
      "prfm pldl1keep, [x24, #0x80]\n"
      "cmp x27, #0x10\n"
      "prfm pldl1keep, [x22, #0x80]\n"
      "bgt 38b\n"
      "40:"  // Height 3: Multiply loop: Single iteration only
      "ld1b { z6.b }, p2/Z, [x11]\n"
      "whilelt p0.b, XZR, x27\n"
      "ld1b { z7.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "ld1rqb { z0.b }, p0/Z, [x26]\n"
      "udot z16.s, z6.b, z0.b[0]\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "add x26, x26, #0x10\n"
      "udot z17.s, z7.b, z0.b[0]\n"
      "ld1rqb { z2.b }, p0/Z, [x22]\n"
      "add x24, x24, #0x10\n"
      "udot z20.s, z6.b, z1.b[0]\n"
      "ld1b { z8.b }, p2/Z, [x11, #2, MUL VL]\n"
      "add x22, x22, #0x10\n"
      "udot z24.s, z6.b, z2.b[0]\n"
      "ld1b { z9.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z21.s, z7.b, z1.b[0]\n"
      "udot z25.s, z7.b, z2.b[0]\n"
      "udot z18.s, z8.b, z0.b[0]\n"
      "udot z22.s, z8.b, z1.b[0]\n"
      "udot z26.s, z8.b, z2.b[0]\n"
      "udot z19.s, z9.b, z0.b[0]\n"
      "udot z23.s, z9.b, z1.b[0]\n"
      "udot z27.s, z9.b, z2.b[0]\n"
      "ble 41f\n"
      "ld1b { z10.b }, p2/Z, [x11]\n"
      "udot z16.s, z10.b, z0.b[1]\n"
      "ld1b { z4.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "udot z20.s, z10.b, z1.b[1]\n"
      "ld1b { z5.b }, p2/Z, [x11, #2, MUL VL]\n"
      "udot z24.s, z10.b, z2.b[1]\n"
      "ld1b { z6.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z17.s, z4.b, z0.b[1]\n"
      "udot z21.s, z4.b, z1.b[1]\n"
      "udot z25.s, z4.b, z2.b[1]\n"
      "udot z18.s, z5.b, z0.b[1]\n"
      "udot z22.s, z5.b, z1.b[1]\n"
      "udot z26.s, z5.b, z2.b[1]\n"
      "udot z19.s, z6.b, z0.b[1]\n"
      "udot z23.s, z6.b, z1.b[1]\n"
      "udot z27.s, z6.b, z2.b[1]\n"
      "ble 41f\n"
      "ld1b { z7.b }, p2/Z, [x11]\n"
      "udot z16.s, z7.b, z0.b[2]\n"
      "ld1b { z8.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "udot z20.s, z7.b, z1.b[2]\n"
      "ld1b { z9.b }, p2/Z, [x11, #2, MUL VL]\n"
      "udot z24.s, z7.b, z2.b[2]\n"
      "ld1b { z10.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z17.s, z8.b, z0.b[2]\n"
      "udot z21.s, z8.b, z1.b[2]\n"
      "udot z25.s, z8.b, z2.b[2]\n"
      "udot z18.s, z9.b, z0.b[2]\n"
      "udot z22.s, z9.b, z1.b[2]\n"
      "udot z26.s, z9.b, z2.b[2]\n"
      "udot z19.s, z10.b, z0.b[2]\n"
      "udot z23.s, z10.b, z1.b[2]\n"
      "udot z27.s, z10.b, z2.b[2]\n"
      "ble 41f\n"
      "ld1b { z4.b }, p2/Z, [x11]\n"
      "udot z16.s, z4.b, z0.b[3]\n"
      "ld1b { z5.b }, p2/Z, [x11, #1, MUL VL]\n"
      "udot z20.s, z4.b, z1.b[3]\n"
      "ld1b { z6.b }, p2/Z, [x11, #2, MUL VL]\n"
      "udot z24.s, z4.b, z2.b[3]\n"
      "ld1b { z7.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z17.s, z5.b, z0.b[3]\n"
      "udot z21.s, z5.b, z1.b[3]\n"
      "udot z25.s, z5.b, z2.b[3]\n"
      "udot z18.s, z6.b, z0.b[3]\n"
      "udot z22.s, z6.b, z1.b[3]\n"
      "udot z26.s, z6.b, z2.b[3]\n"
      "udot z19.s, z7.b, z0.b[3]\n"
      "udot z23.s, z7.b, z1.b[3]\n"
      "udot z27.s, z7.b, z2.b[3]\n"
      "41:"  // Height 3: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 42f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z12.s, z1.b, z15.b\n"
      "udot z13.s, z2.b, z15.b\n"
      "42:"  // Height 3: Multiply loop: unique 6: skip row sum
      "prfm pldl1keep, [x26, #0x80]\n"
      "add x28, x28, #0x1\n"
      "prfm pldl1keep, [x24, #0x80]\n"
      "prfm pldl1keep, [x22, #0x80]\n"
      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
      "cmp x28, x19\n"
      "bne 35b\n"
      "prfm pstl1keep, [x9, #0x0]\n"
      "prfm pstl1keep, [x25, #0x0]\n"
      "prfm pstl1keep, [x23, #0x0]\n"
      "tbnz %x[flags], #31, 43f\n"
      "add x19, %x[qp], %[b_offset]\n"
      "ld1rw { z3.s }, p2/Z, [x19]\n"
      "neg z3.s, p2/M, z3.s\n"
      "mov x20, #0x4\n"
      "mov x19, #0x4\n"
      "whilelt p0.s, XZR, x20\n"
      "uaddv d11, p0, z11.s\n"
      "whilelt p0.s, XZR, x19\n"
      "uaddv d12, p0, z12.s\n"
      "mov x19, #0x4\n"
      "mov z11.s, z11.s[0]\n"
      "whilelt p0.s, XZR, x19\n"
      "mov z12.s, z12.s[0]\n"
      "uaddv d13, p0, z13.s\n"
      "mul z11.s, p2/M, z11.s, z3.s\n"
      "mul z12.s, p2/M, z12.s, z3.s\n"
      "mov z13.s, z13.s[0]\n"
      "mul z13.s, p2/M, z13.s, z3.s\n"
      "43:"  // Height 3: skip row sum fixup
      "add z16.s, z16.s, z11.s\n"
      "ld1w { z0.s }, p2/Z, [x10]\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "add z17.s, z17.s, z11.s\n"
      "ld1w { z1.s }, p2/Z, [x10, #1, MUL VL]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      "add z18.s, z18.s, z11.s\n"
      "ld1w { z2.s }, p2/Z, [x10, #2, MUL VL]\n"
      "add x19, %x[qp], %[per_layer_mul]\n"
      "add z19.s, z19.s, z11.s\n"
      "ld1w { z3.s }, p2/Z, [x10, #3, MUL VL]\n"
      "addvl x10, x10, #4\n"
      "add z20.s, z20.s, z12.s\n"
      "ld1rw { z4.s }, p2/Z, [x19]\n"
      "add z21.s, z21.s, z12.s\n"
      "add z22.s, z22.s, z12.s\n"
      "add z23.s, z23.s, z12.s\n"
      "add z24.s, z24.s, z13.s\n"
      "add z25.s, z25.s, z13.s\n"
      "add z26.s, z26.s, z13.s\n"
      "add z27.s, z27.s, z13.s\n"
      "add z16.s, z16.s, z0.s\n"
      "add z17.s, z17.s, z1.s\n"
      "add z18.s, z18.s, z2.s\n"
      "add z19.s, z19.s, z3.s\n"
      "add z20.s, z20.s, z0.s\n"
      "add z21.s, z21.s, z1.s\n"
      "add z22.s, z22.s, z2.s\n"
      "add z23.s, z23.s, z3.s\n"
      "add z24.s, z24.s, z0.s\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      "add z25.s, z25.s, z1.s\n"
      "add z26.s, z26.s, z2.s\n"
      "add z27.s, z27.s, z3.s\n"
      ".inst 0x04a47610  // sqrdmulh z16.s, z16.s, z4.s\n"
      ".inst 0x04a47631  // sqrdmulh z17.s, z17.s, z4.s\n"
      ".inst 0x04a47652  // sqrdmulh z18.s, z18.s, z4.s\n"
      ".inst 0x04a47673  // sqrdmulh z19.s, z19.s, z4.s\n"
      ".inst 0x04a47694  // sqrdmulh z20.s, z20.s, z4.s\n"
      ".inst 0x04a476b5  // sqrdmulh z21.s, z21.s, z4.s\n"
      ".inst 0x04a476d6  // sqrdmulh z22.s, z22.s, z4.s\n"
      ".inst 0x04a476f7  // sqrdmulh z23.s, z23.s, z4.s\n"
      ".inst 0x04a47718  // sqrdmulh z24.s, z24.s, z4.s\n"
      ".inst 0x04a47739  // sqrdmulh z25.s, z25.s, z4.s\n"
      ".inst 0x04a4775a  // sqrdmulh z26.s, z26.s, z4.s\n"
      ".inst 0x04a4777b  // sqrdmulh z27.s, z27.s, z4.s\n"
      "tbz %x[flags], #5, 44f\n"
      "and z4.d, z16.d, z0.d\n"
      "asr z4.s, z4.s, #0x1f\n"
      "and z5.d, z17.d, z0.d\n"
      "and z6.d, z18.d, z0.d\n"
      "asr z5.s, z5.s, #0x1f\n"
      "and z7.d, z19.d, z0.d\n"
      "and z8.d, z20.d, z0.d\n"
      "asr z6.s, z6.s, #0x1f\n"
      "and z9.d, z21.d, z0.d\n"
      "asr z7.s, z7.s, #0x1f\n"
      "sqadd z16.s, z16.s, z4.s\n"
      "and z10.d, z22.d, z0.d\n"
      "asr z8.s, z8.s, #0x1f\n"
      "and z4.d, z23.d, z0.d\n"
      "asr z9.s, z9.s, #0x1f\n"
      "sqadd z17.s, z17.s, z5.s\n"
      "asr z10.s, z10.s, #0x1f\n"
      "sqadd z18.s, z18.s, z6.s\n"
      "asr z4.s, z4.s, #0x1f\n"
      "and z5.d, z24.d, z0.d\n"
      "asr z5.s, z5.s, #0x1f\n"
      "sqadd z19.s, z19.s, z7.s\n"
      "sqadd z20.s, z20.s, z8.s\n"
      "sqadd z21.s, z21.s, z9.s\n"
      "sqadd z22.s, z22.s, z10.s\n"
      "sqadd z23.s, z23.s, z4.s\n"
      "and z6.d, z25.d, z0.d\n"
      "asr z6.s, z6.s, #0x1f\n"
      "sqadd z24.s, z24.s, z5.s\n"
      "and z7.d, z26.d, z0.d\n"
      "asr z7.s, z7.s, #0x1f\n"
      "and z8.d, z27.d, z0.d\n"
      "sqadd z25.s, z25.s, z6.s\n"
      "asr z8.s, z8.s, #0x1f\n"
      "sqadd z26.s, z26.s, z7.s\n"
      "sqadd z27.s, z27.s, z8.s\n"
      "44:"  // Height 3: no shift correction
      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
      "add x19, %x[qp], %[c_offset]\n"
      "ld1rw { z4.s }, p2/Z, [x19]\n"
      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
      "add x19, %x[qp], %[minval]\n"
      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
      "ld1rw { z5.s }, p2/Z, [x19]\n"
      "add x19, %x[qp], %[maxval]\n"
      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
      "ld1rw { z6.s }, p2/Z, [x19]\n"
      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
      "add z16.s, z16.s, z4.s\n"
      "add z17.s, z17.s, z4.s\n"
      "add z18.s, z18.s, z4.s\n"
      "add z19.s, z19.s, z4.s\n"
      "add z20.s, z20.s, z4.s\n"
      "smin z16.s, p2/M, z16.s, z6.s\n"
      "smin z17.s, p2/M, z17.s, z6.s\n"
      "smin z18.s, p2/M, z18.s, z6.s\n"
      "smin z19.s, p2/M, z19.s, z6.s\n"
      "smax z16.s, p2/M, z16.s, z5.s\n"
      "smax z17.s, p2/M, z17.s, z5.s\n"
      "smax z18.s, p2/M, z18.s, z5.s\n"
      "smax z19.s, p2/M, z19.s, z5.s\n"
      "smin z20.s, p2/M, z20.s, z6.s\n"
      "uzp1 z16.h, z16.h, z17.h\n"
      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
      "uzp1 z17.h, z18.h, z19.h\n"
      "smax z20.s, p2/M, z20.s, z5.s\n"
      "uzp1 z16.b, z16.b, z17.b\n"
      "st1b { z16.b }, p1, [x9]\n"
      "add z21.s, z21.s, z4.s\n"
      "addvl x9, x9, #1\n"
      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
      ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
      "smin z21.s, p2/M, z21.s, z6.s\n"
      ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
      "add z22.s, z22.s, z4.s\n"
      "add z23.s, z23.s, z4.s\n"
      "add z24.s, z24.s, z4.s\n"
      "add z25.s, z25.s, z4.s\n"
      "smax z21.s, p2/M, z21.s, z5.s\n"
      "smin z22.s, p2/M, z22.s, z6.s\n"
      "smin z23.s, p2/M, z23.s, z6.s\n"
      "smin z24.s, p2/M, z24.s, z6.s\n"
      "uzp1 z20.h, z20.h, z21.h\n"
      "smax z22.s, p2/M, z22.s, z5.s\n"
      "smax z23.s, p2/M, z23.s, z5.s\n"
      "smax z24.s, p2/M, z24.s, z5.s\n"
      "smin z25.s, p2/M, z25.s, z6.s\n"
      ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
      "uzp1 z21.h, z22.h, z23.h\n"
      ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
      "uzp1 z20.b, z20.b, z21.b\n"
      "st1b { z20.b }, p1, [x25]\n"
      "add z26.s, z26.s, z4.s\n"
      "addvl x25, x25, #1\n"
      "add z27.s, z27.s, z4.s\n"
      "smax z25.s, p2/M, z25.s, z5.s\n"
      "smin z26.s, p2/M, z26.s, z6.s\n"
      "smin z27.s, p2/M, z27.s, z6.s\n"
      "uzp1 z24.h, z24.h, z25.h\n"
      "smax z26.s, p2/M, z26.s, z5.s\n"
      "smax z27.s, p2/M, z27.s, z5.s\n"
      "uzp1 z25.h, z26.h, z27.h\n"
      "uzp1 z24.b, z24.b, z25.b\n"
      "st1b { z24.b }, p1, [x23]\n"
      "addvl x23, x23, #1\n"
      "45:"  // Height 3: Writeback done
      "mov x19, #0x0\n"
      "incw x19, ALL, MUL #4\n"
      "subs x12, x12, x19\n"
      "bgt 33b\n"
      "b 62f\n"
      "46:"  // Height 4
      "mov z11.s, #0x0\n"
      "ldr x12, [%x[args_ptr], %[offsetof_N]]\n"
      "mov x10, %x[col_bias]\n"
      "mov z12.s, #0x0\n"
      "ldr x11, [%x[args_ptr], %[offsetof_B_ptr]]\n"
      "bic %x[flags], %x[flags], #0x80000000\n"
      "mov z13.s, #0x0\n"
      "ldr x19, [%x[args_ptr], %[offsetof_output_offset]]\n"
      "mov z14.s, #0x0\n"
      "mov z15.b, #0x1\n"
      "tbz %x[flags], #2, 47f\n"
      "ldr x9, [%x[output_ptr], #0x0]\n"
      "ldr x25, [%x[output_ptr], #0x8]\n"
      "add x9, x9, x19\n"
      "ldr x23, [%x[output_ptr], #0x10]\n"
      "ldr x21, [%x[output_ptr], #0x18]\n"
      "add x25, x25, x19\n"
      "add %x[output_ptr], %x[output_ptr], #0x20\n"
      "add x23, x23, x19\n"
      "add x21, x21, x19\n"
      "b 48f\n"
      "47:"  // Height 4: setup direct output
      "mov x9, %x[output_ptr]\n"
      "add x25, x9, x19\n"
      "add x23, x25, x19\n"
      "add x21, x23, x19\n"
      "add %x[output_ptr], x21, x19\n"
      "48:"  // Height 4: Column loop
      "mov z16.s, #0x0\n"
      "mov x19, #0x0\n"
      "mov z17.s, #0x0\n"
      "whilelt p1.b, x19, x12\n"
      "mov z18.s, #0x0\n"
      "mov z19.s, #0x0\n"
      "mov z20.s, #0x0\n"
      "mov z21.s, #0x0\n"
      "mov z22.s, #0x0\n"
      "mov z23.s, #0x0\n"
      "mov z24.s, #0x0\n"
      "mov z25.s, #0x0\n"
      "mov z26.s, #0x0\n"
      "mov z27.s, #0x0\n"
      "mov z28.s, #0x0\n"
      "mov z29.s, #0x0\n"
      "mov z30.s, #0x0\n"
      "mov z31.s, #0x0\n"
      "49:"  // Height 4: setup done
      "mov x28, #0x0\n"
      "50:"  // Height 4: String loop
      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
      "ldr x19, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "ldr w27, [x20, x28, LSL #0x2]\n"
      "tbz %x[flags], #3, 51f\n"
      "ldr x20, [%x[input_ptr], x28, LSL #0x3]\n"
      "add x20, x20, x19, LSL #3\n"
      "ldr x26, [x20, #0x0]\n"
      "ldr x24, [x20, #0x8]\n"
      "ldr x22, [x20, #0x10]\n"
      "ldr x20, [x20, #0x18]\n"
      "cbnz x28, 52f\n"
      "ldr x19, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
      "add x26, x26, x19\n"
      "add x24, x24, x19\n"
      "add x22, x22, x19\n"
      "add x20, x20, x19\n"
      "b 52f\n"
      "51:"  // Height 4: setup direct input
      "mov x26, %x[input_ptr]\n"
      "add x24, x26, x19\n"
      "add x22, x24, x19\n"
      "add x20, x22, x19\n"
      "52:"  // Height 4: input setup done
      "cmp x27, #0x10\n"
      "ble 55f\n"
      "53:"  // Height 4: Multiply loop: Main loop head
      "ld1b { z4.b }, p2/Z, [x11]\n"
      "whilelt p0.b, XZR, x27\n"
      "ld1b { z5.b }, p2/Z, [x11, #1, MUL VL]\n"
      "ld1rqb { z0.b }, p0/Z, [x26]\n"
      "udot z16.s, z4.b, z0.b[0]\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "add x26, x26, #0x10\n"
      "udot z17.s, z5.b, z0.b[0]\n"
      "ld1rqb { z2.b }, p0/Z, [x22]\n"
      "add x24, x24, #0x10\n"
      "udot z20.s, z4.b, z1.b[0]\n"
      "ld1rqb { z3.b }, p0/Z, [x20]\n"
      "add x22, x22, #0x10\n"
      "udot z24.s, z4.b, z2.b[0]\n"
      "ld1b { z6.b }, p2/Z, [x11, #2, MUL VL]\n"
      "add x20, x20, #0x10\n"
      "udot z21.s, z5.b, z1.b[0]\n"
      "ld1b { z7.b }, p2/Z, [x11, #3, MUL VL]\n"
      "udot z25.s, z5.b, z2.b[0]\n"
      "ld1b { z8.b }, p2/Z, [x11, #4, MUL VL]\n"
      "udot z28.s, z4.b, z3.b[0]\n"
      "ld1b { z9.b }, p2/Z, [x11, #5, MUL VL]\n"
      "udot z29.s, z5.b, z3.b[0]\n"
      "ld1b { z10.b }, p2/Z, [x11, #6, MUL VL]\n"
      "udot z18.s, z6.b, z0.b[0]\n"
      "ld1b { z4.b }, p2/Z, [x11, #7, MUL VL]\n"
      "addvl x11, x11, #16\n"
      "udot z22.s, z6.b, z1.b[0]\n"
      "ld1b { z5.b }, p2/Z, [x11, #-8, MUL VL]\n"
      "udot z26.s, z6.b, z2.b[0]\n"
      "udot z30.s, z6.b, z3.b[0]\n"
      "ld1b { z6.b }, p2/Z, [x11, #-7, MUL VL]\n"
      "udot z19.s, z7.b, z0.b[0]\n"
      "udot z23.s, z7.b, z1.b[0]\n"
      "udot z27.s, z7.b, z2.b[0]\n"
      "udot z31.s, z7.b, z3.b[0]\n"
      "ld1b { z7.b }, p2/Z, [x11, #-6, MUL VL]\n"
      "udot z16.s, z8.b, z0.b[1]\n"
      "udot z20.s, z8.b, z1.b[1]\n"
      "udot z24.s, z8.b, z2.b[1]\n"
      "udot z28.s, z8.b, z3.b[1]\n"
      "ld1b { z8.b }, p2/Z, [x11, #-5, MUL VL]\n"
      "udot z17.s, z9.b, z0.b[1]\n"
      "udot z21.s, z9.b, z1.b[1]\n"
      "udot z25.s, z9.b, z2.b[1]\n"
      "udot z29.s, z9.b, z3.b[1]\n"
      "ld1b { z9.b }, p2/Z, [x11, #-4, MUL VL]\n"
      "udot z18.s, z10.b, z0.b[1]\n"
      "udot z22.s, z10.b, z1.b[1]\n"
      "udot z26.s, z10.b, z2.b[1]\n"
      "udot z30.s, z10.b, z3.b[1]\n"
      "ld1b { z10.b }, p2/Z, [x11, #-3, MUL VL]\n"
      "udot z19.s, z4.b, z0.b[1]\n"
      "udot z23.s, z4.b, z1.b[1]\n"
      "udot z27.s, z4.b, z2.b[1]\n"
      "udot z31.s, z4.b, z3.b[1]\n"
      "ld1b { z4.b }, p2/Z, [x11, #-2, MUL VL]\n"
      "udot z16.s, z5.b, z0.b[2]\n"
      "udot z20.s, z5.b, z1.b[2]\n"
      "udot z24.s, z5.b, z2.b[2]\n"
      "udot z28.s, z5.b, z3.b[2]\n"
      "ld1b { z5.b }, p2/Z, [x11, #-1, MUL VL]\n"
      "udot z17.s, z6.b, z0.b[2]\n"
      "udot z21.s, z6.b, z1.b[2]\n"
      "udot z25.s, z6.b, z2.b[2]\n"
      "udot z29.s, z6.b, z3.b[2]\n"
      "udot z18.s, z7.b, z0.b[2]\n"
      "udot z22.s, z7.b, z1.b[2]\n"
      "udot z26.s, z7.b, z2.b[2]\n"
      "udot z30.s, z7.b, z3.b[2]\n"
      "udot z19.s, z8.b, z0.b[2]\n"
      "udot z23.s, z8.b, z1.b[2]\n"
      "udot z27.s, z8.b, z2.b[2]\n"
      "udot z31.s, z8.b, z3.b[2]\n"
      "udot z16.s, z9.b, z0.b[3]\n"
      "udot z20.s, z9.b, z1.b[3]\n"
      "udot z24.s, z9.b, z2.b[3]\n"
      "udot z28.s, z9.b, z3.b[3]\n"
      "udot z17.s, z10.b, z0.b[3]\n"
      "udot z21.s, z10.b, z1.b[3]\n"
      "udot z25.s, z10.b, z2.b[3]\n"
      "udot z29.s, z10.b, z3.b[3]\n"
      "udot z18.s, z4.b, z0.b[3]\n"
      "udot z22.s, z4.b, z1.b[3]\n"
      "udot z26.s, z4.b, z2.b[3]\n"
      "udot z30.s, z4.b, z3.b[3]\n"
      "udot z19.s, z5.b, z0.b[3]\n"
      "udot z23.s, z5.b, z1.b[3]\n"
      "udot z27.s, z5.b, z2.b[3]\n"
      "udot z31.s, z5.b, z3.b[3]\n"
      "tbnz %x[flags], #31, 54f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z12.s, z1.b, z15.b\n"
      "udot z13.s, z2.b, z15.b\n"
      "udot z14.s, z3.b, z15.b\n"
      "54:"  // Height 4: Multiply loop: unique 7: skip row sum
      "prfm pldl1keep, [x26, #0x80]\n"
      "sub x27, x27, #0x10\n"
      "prfm pldl1keep, [x24, #0x80]\n"
      "cmp x27, #0x10\n"
      "prfm pldl1keep, [x22, #0x80]\n"
      "prfm pldl1keep, [x20, #0x80]\n"
      "bgt 53b\n"
      "55:"  // Height 4: Multiply loop: Single iteration only
      "ld1b { z6.b }, p2/Z, [x11]\n"
      "whilelt p0.b, XZR, x27\n"
      "ld1b { z7.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "ld1rqb { z0.b }, p0/Z, [x26]\n"
      "udot z16.s, z6.b, z0.b[0]\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "add x26, x26, #0x10\n"
      "udot z17.s, z7.b, z0.b[0]\n"
      "ld1rqb { z2.b }, p0/Z, [x22]\n"
      "add x24, x24, #0x10\n"
      "udot z20.s, z6.b, z1.b[0]\n"
      "ld1rqb { z3.b }, p0/Z, [x20]\n"
      "add x22, x22, #0x10\n"
      "udot z24.s, z6.b, z2.b[0]\n"
      "ld1b { z8.b }, p2/Z, [x11, #2, MUL VL]\n"
      "add x20, x20, #0x10\n"
      "udot z21.s, z7.b, z1.b[0]\n"
      "ld1b { z9.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z28.s, z6.b, z3.b[0]\n"
      "udot z25.s, z7.b, z2.b[0]\n"
      "udot z29.s, z7.b, z3.b[0]\n"
      "udot z18.s, z8.b, z0.b[0]\n"
      "udot z22.s, z8.b, z1.b[0]\n"
      "udot z26.s, z8.b, z2.b[0]\n"
      "udot z30.s, z8.b, z3.b[0]\n"
      "udot z19.s, z9.b, z0.b[0]\n"
      "udot z23.s, z9.b, z1.b[0]\n"
      "udot z27.s, z9.b, z2.b[0]\n"
      "udot z31.s, z9.b, z3.b[0]\n"
      "ble 56f\n"
      "ld1b { z10.b }, p2/Z, [x11]\n"
      "udot z16.s, z10.b, z0.b[1]\n"
      "ld1b { z4.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "udot z20.s, z10.b, z1.b[1]\n"
      "ld1b { z5.b }, p2/Z, [x11, #2, MUL VL]\n"
      "udot z24.s, z10.b, z2.b[1]\n"
      "ld1b { z6.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z28.s, z10.b, z3.b[1]\n"
      "udot z17.s, z4.b, z0.b[1]\n"
      "udot z21.s, z4.b, z1.b[1]\n"
      "udot z25.s, z4.b, z2.b[1]\n"
      "udot z29.s, z4.b, z3.b[1]\n"
      "udot z18.s, z5.b, z0.b[1]\n"
      "udot z22.s, z5.b, z1.b[1]\n"
      "udot z26.s, z5.b, z2.b[1]\n"
      "udot z30.s, z5.b, z3.b[1]\n"
      "udot z19.s, z6.b, z0.b[1]\n"
      "udot z23.s, z6.b, z1.b[1]\n"
      "udot z27.s, z6.b, z2.b[1]\n"
      "udot z31.s, z6.b, z3.b[1]\n"
      "ble 56f\n"
      "ld1b { z7.b }, p2/Z, [x11]\n"
      "udot z16.s, z7.b, z0.b[2]\n"
      "ld1b { z8.b }, p2/Z, [x11, #1, MUL VL]\n"
      "subs x27, x27, #0x4\n"
      "udot z20.s, z7.b, z1.b[2]\n"
      "ld1b { z9.b }, p2/Z, [x11, #2, MUL VL]\n"
      "udot z24.s, z7.b, z2.b[2]\n"
      "ld1b { z10.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z28.s, z7.b, z3.b[2]\n"
      "udot z17.s, z8.b, z0.b[2]\n"
      "udot z21.s, z8.b, z1.b[2]\n"
      "udot z25.s, z8.b, z2.b[2]\n"
      "udot z29.s, z8.b, z3.b[2]\n"
      "udot z18.s, z9.b, z0.b[2]\n"
      "udot z22.s, z9.b, z1.b[2]\n"
      "udot z26.s, z9.b, z2.b[2]\n"
      "udot z30.s, z9.b, z3.b[2]\n"
      "udot z19.s, z10.b, z0.b[2]\n"
      "udot z23.s, z10.b, z1.b[2]\n"
      "udot z27.s, z10.b, z2.b[2]\n"
      "udot z31.s, z10.b, z3.b[2]\n"
      "ble 56f\n"
      "ld1b { z4.b }, p2/Z, [x11]\n"
      "udot z16.s, z4.b, z0.b[3]\n"
      "ld1b { z5.b }, p2/Z, [x11, #1, MUL VL]\n"
      "udot z20.s, z4.b, z1.b[3]\n"
      "ld1b { z6.b }, p2/Z, [x11, #2, MUL VL]\n"
      "udot z24.s, z4.b, z2.b[3]\n"
      "ld1b { z7.b }, p2/Z, [x11, #3, MUL VL]\n"
      "addvl x11, x11, #4\n"
      "udot z28.s, z4.b, z3.b[3]\n"
      "udot z17.s, z5.b, z0.b[3]\n"
      "udot z21.s, z5.b, z1.b[3]\n"
      "udot z25.s, z5.b, z2.b[3]\n"
      "udot z29.s, z5.b, z3.b[3]\n"
      "udot z18.s, z6.b, z0.b[3]\n"
      "udot z22.s, z6.b, z1.b[3]\n"
      "udot z26.s, z6.b, z2.b[3]\n"
      "udot z30.s, z6.b, z3.b[3]\n"
      "udot z19.s, z7.b, z0.b[3]\n"
      "udot z23.s, z7.b, z1.b[3]\n"
      "udot z27.s, z7.b, z2.b[3]\n"
      "udot z31.s, z7.b, z3.b[3]\n"
      "56:"  // Height 4: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 57f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z12.s, z1.b, z15.b\n"
      "udot z13.s, z2.b, z15.b\n"
      "udot z14.s, z3.b, z15.b\n"
      "57:"  // Height 4: Multiply loop: unique 8: skip row sum
      "prfm pldl1keep, [x26, #0x80]\n"
      "add x28, x28, #0x1\n"
      "prfm pldl1keep, [x24, #0x80]\n"
      "prfm pldl1keep, [x22, #0x80]\n"
      "prfm pldl1keep, [x20, #0x80]\n"
      "ldr w19, [%x[args_ptr], %[offsetof_num_strings]]\n"
      "cmp x28, x19\n"
      "bne 50b\n"
      "prfm pstl1keep, [x9, #0x0]\n"
      "prfm pstl1keep, [x25, #0x0]\n"
      "prfm pstl1keep, [x23, #0x0]\n"
      "prfm pstl1keep, [x21, #0x0]\n"
      "tbnz %x[flags], #31, 58f\n"
      "add x19, %x[qp], %[b_offset]\n"
      "ld1rw { z4.s }, p2/Z, [x19]\n"
      "neg z4.s, p2/M, z4.s\n"
      "mov x20, #0x4\n"
      "mov x19, #0x4\n"
      "whilelt p0.s, XZR, x20\n"
      "uaddv d11, p0, z11.s\n"
      "whilelt p0.s, XZR, x19\n"
      "uaddv d12, p0, z12.s\n"
      "mov x19, #0x4\n"
      "mov z11.s, z11.s[0]\n"
      "whilelt p0.s, XZR, x19\n"
      "mov x19, #0x4\n"
      "mov z12.s, z12.s[0]\n"
      "uaddv d13, p0, z13.s\n"
      "whilelt p0.s, XZR, x19\n"
      "mul z11.s, p2/M, z11.s, z4.s\n"
      "uaddv d14, p0, z14.s\n"
      "mul z12.s, p2/M, z12.s, z4.s\n"
      "mov z13.s, z13.s[0]\n"
      "mul z13.s, p2/M, z13.s, z4.s\n"
      "mov z14.s, z14.s[0]\n"
      "mul z14.s, p2/M, z14.s, z4.s\n"
      "58:"  // Height 4: skip row sum fixup
      "add z16.s, z16.s, z11.s\n"
      "ld1w { z0.s }, p2/Z, [x10]\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "add z17.s, z17.s, z11.s\n"
      "ld1w { z1.s }, p2/Z, [x10, #1, MUL VL]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      "add z18.s, z18.s, z11.s\n"
      "ld1w { z2.s }, p2/Z, [x10, #2, MUL VL]\n"
      "add x19, %x[qp], %[per_layer_mul]\n"
      "add z19.s, z19.s, z11.s\n"
      "ld1w { z3.s }, p2/Z, [x10, #3, MUL VL]\n"
      "addvl x10, x10, #4\n"
      "add z20.s, z20.s, z12.s\n"
      "ld1rw { z4.s }, p2/Z, [x19]\n"
      "add z21.s, z21.s, z12.s\n"
      "add z22.s, z22.s, z12.s\n"
      "add z23.s, z23.s, z12.s\n"
      "add z24.s, z24.s, z13.s\n"
      "add z25.s, z25.s, z13.s\n"
      "add z26.s, z26.s, z13.s\n"
      "add z27.s, z27.s, z13.s\n"
      "add z28.s, z28.s, z14.s\n"
      "add z29.s, z29.s, z14.s\n"
      "add z30.s, z30.s, z14.s\n"
      "add z31.s, z31.s, z14.s\n"
      "add z16.s, z16.s, z0.s\n"
      "add z17.s, z17.s, z1.s\n"
      "add z18.s, z18.s, z2.s\n"
      "add z19.s, z19.s, z3.s\n"
      "add z20.s, z20.s, z0.s\n"
      "add z21.s, z21.s, z1.s\n"
      "add z22.s, z22.s, z2.s\n"
      "add z23.s, z23.s, z3.s\n"
      "add z24.s, z24.s, z0.s\n"
      "add z25.s, z25.s, z1.s\n"
      "add z26.s, z26.s, z2.s\n"
      "add z27.s, z27.s, z3.s\n"
      "add z28.s, z28.s, z0.s\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      "add z29.s, z29.s, z1.s\n"
      "add z30.s, z30.s, z2.s\n"
      "add z31.s, z31.s, z3.s\n"
      ".inst 0x04a47610  // sqrdmulh z16.s, z16.s, z4.s\n"
      ".inst 0x04a47631  // sqrdmulh z17.s, z17.s, z4.s\n"
      ".inst 0x04a47652  // sqrdmulh z18.s, z18.s, z4.s\n"
      ".inst 0x04a47673  // sqrdmulh z19.s, z19.s, z4.s\n"
      ".inst 0x04a47694  // sqrdmulh z20.s, z20.s, z4.s\n"
      ".inst 0x04a476b5  // sqrdmulh z21.s, z21.s, z4.s\n"
      ".inst 0x04a476d6  // sqrdmulh z22.s, z22.s, z4.s\n"
      ".inst 0x04a476f7  // sqrdmulh z23.s, z23.s, z4.s\n"
      ".inst 0x04a47718  // sqrdmulh z24.s, z24.s, z4.s\n"
      ".inst 0x04a47739  // sqrdmulh z25.s, z25.s, z4.s\n"
      ".inst 0x04a4775a  // sqrdmulh z26.s, z26.s, z4.s\n"
      ".inst 0x04a4777b  // sqrdmulh z27.s, z27.s, z4.s\n"
      ".inst 0x04a4779c  // sqrdmulh z28.s, z28.s, z4.s\n"
      ".inst 0x04a477bd  // sqrdmulh z29.s, z29.s, z4.s\n"
      ".inst 0x04a477de  // sqrdmulh z30.s, z30.s, z4.s\n"
      ".inst 0x04a477ff  // sqrdmulh z31.s, z31.s, z4.s\n"
      "tbz %x[flags], #5, 59f\n"
      "and z4.d, z16.d, z0.d\n"
      "asr z4.s, z4.s, #0x1f\n"
      "and z5.d, z17.d, z0.d\n"
      "and z6.d, z18.d, z0.d\n"
      "asr z5.s, z5.s, #0x1f\n"
      "and z7.d, z19.d, z0.d\n"
      "and z8.d, z20.d, z0.d\n"
      "asr z6.s, z6.s, #0x1f\n"
      "and z9.d, z21.d, z0.d\n"
      "asr z7.s, z7.s, #0x1f\n"
      "sqadd z16.s, z16.s, z4.s\n"
      "and z10.d, z22.d, z0.d\n"
      "asr z8.s, z8.s, #0x1f\n"
      "and z4.d, z23.d, z0.d\n"
      "asr z9.s, z9.s, #0x1f\n"
      "sqadd z17.s, z17.s, z5.s\n"
      "asr z10.s, z10.s, #0x1f\n"
      "sqadd z18.s, z18.s, z6.s\n"
      "asr z4.s, z4.s, #0x1f\n"
      "and z5.d, z24.d, z0.d\n"
      "asr z5.s, z5.s, #0x1f\n"
      "sqadd z19.s, z19.s, z7.s\n"
      "sqadd z20.s, z20.s, z8.s\n"
      "sqadd z21.s, z21.s, z9.s\n"
      "sqadd z22.s, z22.s, z10.s\n"
      "sqadd z23.s, z23.s, z4.s\n"
      "and z6.d, z25.d, z0.d\n"
      "asr z6.s, z6.s, #0x1f\n"
      "sqadd z24.s, z24.s, z5.s\n"
      "and z7.d, z26.d, z0.d\n"
      "asr z7.s, z7.s, #0x1f\n"
      "and z8.d, z27.d, z0.d\n"
      "and z9.d, z28.d, z0.d\n"
      "asr z8.s, z8.s, #0x1f\n"
      "sqadd z25.s, z25.s, z6.s\n"
      "and z10.d, z29.d, z0.d\n"
      "asr z9.s, z9.s, #0x1f\n"
      "and z4.d, z30.d, z0.d\n"
      "asr z10.s, z10.s, #0x1f\n"
      "sqadd z26.s, z26.s, z7.s\n"
      "and z5.d, z31.d, z0.d\n"
      "asr z4.s, z4.s, #0x1f\n"
      "sqadd z27.s, z27.s, z8.s\n"
      "asr z5.s, z5.s, #0x1f\n"
      "sqadd z28.s, z28.s, z9.s\n"
      "sqadd z29.s, z29.s, z10.s\n"
      "sqadd z30.s, z30.s, z4.s\n"
      "sqadd z31.s, z31.s, z5.s\n"
      "59:"  // Height 4: no shift correction
      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
      "add x19, %x[qp], %[c_offset]\n"
      "ld1rw { z4.s }, p2/Z, [x19]\n"
      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
      "add x19, %x[qp], %[minval]\n"
      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
      "ld1rw { z5.s }, p2/Z, [x19]\n"
      "add x19, %x[qp], %[maxval]\n"
      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
      "ld1rw { z6.s }, p2/Z, [x19]\n"
      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
      "add z16.s, z16.s, z4.s\n"
      "add z17.s, z17.s, z4.s\n"
      "add z18.s, z18.s, z4.s\n"
      "add z19.s, z19.s, z4.s\n"
      "add z20.s, z20.s, z4.s\n"
      "smin z16.s, p2/M, z16.s, z6.s\n"
      "smin z17.s, p2/M, z17.s, z6.s\n"
      "smin z18.s, p2/M, z18.s, z6.s\n"
      "smin z19.s, p2/M, z19.s, z6.s\n"
      "smax z16.s, p2/M, z16.s, z5.s\n"
      "smax z17.s, p2/M, z17.s, z5.s\n"
      "smax z18.s, p2/M, z18.s, z5.s\n"
      "smax z19.s, p2/M, z19.s, z5.s\n"
      "smin z20.s, p2/M, z20.s, z6.s\n"
      "uzp1 z16.h, z16.h, z17.h\n"
      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
      "uzp1 z17.h, z18.h, z19.h\n"
      "smax z20.s, p2/M, z20.s, z5.s\n"
      "uzp1 z16.b, z16.b, z17.b\n"
      "st1b { z16.b }, p1, [x9]\n"
      "add z21.s, z21.s, z4.s\n"
      "addvl x9, x9, #1\n"
      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
      ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
      "smin z21.s, p2/M, z21.s, z6.s\n"
      ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
      "add z22.s, z22.s, z4.s\n"
      "add z23.s, z23.s, z4.s\n"
      "add z24.s, z24.s, z4.s\n"
      "add z25.s, z25.s, z4.s\n"
      "smax z21.s, p2/M, z21.s, z5.s\n"
      "smin z22.s, p2/M, z22.s, z6.s\n"
      "smin z23.s, p2/M, z23.s, z6.s\n"
      "smin z24.s, p2/M, z24.s, z6.s\n"
      "uzp1 z20.h, z20.h, z21.h\n"
      "smax z22.s, p2/M, z22.s, z5.s\n"
      "smax z23.s, p2/M, z23.s, z5.s\n"
      "smax z24.s, p2/M, z24.s, z5.s\n"
      "smin z25.s, p2/M, z25.s, z6.s\n"
      ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
      "uzp1 z21.h, z22.h, z23.h\n"
      ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
      "uzp1 z20.b, z20.b, z21.b\n"
      "st1b { z20.b }, p1, [x25]\n"
      "add z26.s, z26.s, z4.s\n"
      "addvl x25, x25, #1\n"
      "add z27.s, z27.s, z4.s\n"
      "smax z25.s, p2/M, z25.s, z5.s\n"
      ".inst 0x4482881c  // srshl z28.s, p2/M, z28.s, z0.s\n"
      "smin z26.s, p2/M, z26.s, z6.s\n"
      "smin z27.s, p2/M, z27.s, z6.s\n"
      "uzp1 z24.h, z24.h, z25.h\n"
      "add z28.s, z28.s, z4.s\n"
      "smax z26.s, p2/M, z26.s, z5.s\n"
      "smax z27.s, p2/M, z27.s, z5.s\n"
      "smin z28.s, p2/M, z28.s, z6.s\n"
      ".inst 0x4482881d  // srshl z29.s, p2/M, z29.s, z0.s\n"
      ".inst 0x4482881e  // srshl z30.s, p2/M, z30.s, z0.s\n"
      "uzp1 z25.h, z26.h, z27.h\n"
      "smax z28.s, p2/M, z28.s, z5.s\n"
      "add z29.s, z29.s, z4.s\n"
      "add z30.s, z30.s, z4.s\n"
      "uzp1 z24.b, z24.b, z25.b\n"
      "st1b { z24.b }, p1, [x23]\n"
      "smin z29.s, p2/M, z29.s, z6.s\n"
      "addvl x23, x23, #1\n"
      "smin z30.s, p2/M, z30.s, z6.s\n"
      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
      "smax z29.s, p2/M, z29.s, z5.s\n"
      "add z31.s, z31.s, z4.s\n"
      "smax z30.s, p2/M, z30.s, z5.s\n"
      "uzp1 z28.h, z28.h, z29.h\n"
      "smin z31.s, p2/M, z31.s, z6.s\n"
      "smax z31.s, p2/M, z31.s, z5.s\n"
      "uzp1 z29.h, z30.h, z31.h\n"
      "uzp1 z28.b, z28.b, z29.b\n"
      "st1b { z28.b }, p1, [x21]\n"
      "addvl x21, x21, #1\n"
      "60:"  // Height 4: Writeback done
      "mov x19, #0x0\n"
      "incw x19, ALL, MUL #4\n"
      "subs x12, x12, x19\n"
      "bgt 48b\n"
      "subs %x[M], %x[M], #0x4\n"
      "beq 62f\n"
      "ldr x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "tbz %x[flags], #3, 61f\n"
      "add x20, x20, #0x4\n"
      "str x20, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "b 1b\n"
      "61:"  // Update direct input
      "mov x19, #0x4\n"
      "madd %x[input_ptr], x19, x20, %x[input_ptr]\n"
      "b 1b\n"
      "62:"  // Exit

      : [M] "+r" (M), [flags] "+r" (flags), [input_ptr] "+r" (input_ptr), [output_ptr] "+r" (output_ptr)
      : [args_ptr] "r" (&ka), [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)), [offsetof_B_ptr] "I" (offsetof(KernelArgs, B_ptr)), [offsetof_N] "I" (offsetof(KernelArgs, N)), [offsetof_input_initial_col] "I" (offsetof(KernelArgs, input_initial_col)), [offsetof_input_offset] "I" (offsetof(KernelArgs, input_offset)), [offsetof_num_strings] "I" (offsetof(KernelArgs, num_strings)), [offsetof_output_offset] "I" (offsetof(KernelArgs, output_offset)), [offsetof_string_lengths] "I" (offsetof(KernelArgs, string_lengths)), [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", "x9", "x10", "x11", "x12", "x19", "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_FEATURE_SVE