path: root/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_u8qa_mmla_4x4VL/generic.cpp

/*
 * Copyright (c) 2021, 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_COMPUTE_ENABLE_SVE

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

#include <cassert>
#include <cstdint>

namespace arm_gemm {

void sve_hybrid_u8qa_mmla_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 43f\n"
      "cmp %x[M], #0x2\n"
      "bgt 29f\n"
      "beq 15f\n"
      "mov x10, %x[col_bias]\n"
      "mov z11.s, #0x0\n"
      "mov z15.b, #0x1\n"
      "bic %x[flags], %x[flags], #0x80000000\n"
      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
      "mov x27, %x[output_ptr]\n"
      "2:"  // Height 1: Column loop
      "mov x20, #0x0\n"
      "whilelt p1.b, x20, x9\n"
      "mov z16.s, #0x0\n"
      "mov z17.s, #0x0\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"
      "3:"  // Height 1: setup done
      "mov x26, #0x0\n"
      "4:"  // Height 1: String loop
      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
      "ldr w25, [x20, x26, LSL #0x2]\n"
      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "tbz %x[flags], #3, 5f\n"
      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
      "add x20, x20, x21, LSL #3\n"
      "ldr x24, [x20, #0x0]\n"
      "cbnz x26, 6f\n"
      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
      "add x24, x24, x20\n"
      "b 6f\n"
      "5:"  // Height 1: setup direct input
      "mov x24, %x[input_ptr]\n"
      "6:"  // Height 1: input setup done
      "cmp x25, #0x10\n"
      "ble 9f\n"
      "7:"  // Height 1: Multiply loop: Main loop head
      "whilelt p0.b, XZR, x25\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "trn1 z0.d, z1.d, z26.d\n"
      "ld1b { z24.b }, p2/Z, [x28]\n"
      ".inst 0x45d89810  // ummla z16.s, z0.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #1, MUL VL]\n"
      "trn2 z1.d, z1.d, z26.d\n"
      "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z26.b }, p2/Z, [x28, #3, MUL VL]\n"
      ".inst 0x45d99814  // ummla z20.s, z0.b, z25.b\n"
      ".inst 0x45d89811  // ummla z17.s, z0.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45da9815  // ummla z21.s, z0.b, z26.b\n"
      ".inst 0x45d99812  // ummla z18.s, z0.b, z25.b\n"
      "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
      "addvl x28, x28, #16\n"
      ".inst 0x45d89816  // ummla z22.s, z0.b, z24.b\n"
      "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
      ".inst 0x45da9813  // ummla z19.s, z0.b, z26.b\n"
      ".inst 0x45d99817  // ummla z23.s, z0.b, z25.b\n"
      "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n"
      ".inst 0x45d89830  // ummla z16.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n"
      ".inst 0x45da9834  // ummla z20.s, z1.b, z26.b\n"
      "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
      "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
      ".inst 0x45d99831  // ummla z17.s, z1.b, z25.b\n"
      ".inst 0x45d89835  // ummla z21.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
      ".inst 0x45db9832  // ummla z18.s, z1.b, z27.b\n"
      ".inst 0x45da9836  // ummla z22.s, z1.b, z26.b\n"
      ".inst 0x45d99833  // ummla z19.s, z1.b, z25.b\n"
      ".inst 0x45d89837  // ummla z23.s, z1.b, z24.b\n"
      "add x24, x24, #0x10\n"
      "tbnz %x[flags], #31, 8f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z11.s, z1.b, z15.b\n"
      "8:"  // Height 1: Multiply loop: unique 1: skip row sum
      "sub x25, x25, #0x10\n"
      "cmp x25, #0x10\n"
      "bgt 7b\n"
      "9:"  // Height 1: Multiply loop: Single iteration only
      "whilelt p0.b, XZR, x25\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "trn1 z0.d, z1.d, z27.d\n"
      "ld1b { z24.b }, p2/Z, [x28]\n"
      ".inst 0x45d89810  // ummla z16.s, z0.b, z24.b\n"
      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
      "subs x25, x25, #0x8\n"
      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
      "trn2 z1.d, z1.d, z27.d\n"
      ".inst 0x45da9814  // ummla z20.s, z0.b, z26.b\n"
      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45d99811  // ummla z17.s, z0.b, z25.b\n"
      ".inst 0x45d89815  // ummla z21.s, z0.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
      ".inst 0x45db9812  // ummla z18.s, z0.b, z27.b\n"
      ".inst 0x45da9816  // ummla z22.s, z0.b, z26.b\n"
      ".inst 0x45d99813  // ummla z19.s, z0.b, z25.b\n"
      ".inst 0x45d89817  // ummla z23.s, z0.b, z24.b\n"
      "addvl x28, x28, #8\n"
      "ble 10f\n"
      "ld1b { z24.b }, p2/Z, [x28]\n"
      ".inst 0x45d89830  // ummla z16.s, z1.b, z24.b\n"
      "ld1b { z24.b }, p2/Z, [x28, #1, MUL VL]\n"
      ".inst 0x45d89834  // ummla z20.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
      ".inst 0x45d99831  // ummla z17.s, z1.b, z25.b\n"
      ".inst 0x45d89835  // ummla z21.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45d99832  // ummla z18.s, z1.b, z25.b\n"
      ".inst 0x45d89836  // ummla z22.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
      ".inst 0x45d99833  // ummla z19.s, z1.b, z25.b\n"
      ".inst 0x45d89837  // ummla z23.s, z1.b, z24.b\n"
      "addvl x28, x28, #8\n"
      "10:"  // Height 1: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 11f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z11.s, z1.b, z15.b\n"
      "11:"  // Height 1: Multiply loop: unique 2: skip row sum
      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
      "add x26, x26, #0x1\n"
      "cmp x26, x20\n"
      "bne 4b\n"
      "uzp1 z16.d, z16.d, z20.d\n"
      "uzp1 z17.d, z17.d, z21.d\n"
      "uzp1 z18.d, z18.d, z22.d\n"
      "uzp1 z19.d, z19.d, z23.d\n"
      "mov z23.d, z16.d\n"
      "tbnz %x[flags], #31, 12f\n"
      "add x20, %x[qp], %[b_offset]\n"
      "ld1rw { z16.s }, p2/Z, [x20]\n"
      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
      "neg z16.s, p2/M, z16.s\n"
      "mov z11.s, z11.s[0]\n"
      "mul z11.s, p2/M, z11.s, z16.s\n"
      "12:"  // Height 1: skip row sum fixup
      "add z23.s, z23.s, z11.s\n"
      "add z17.s, z17.s, z11.s\n"
      "ld1w { z22.s }, p2/Z, [x10]\n"
      "ld1w { z21.s }, p2/Z, [x10, #1, MUL VL]\n"
      "add z18.s, z18.s, z11.s\n"
      "add z19.s, z19.s, z11.s\n"
      "ld1w { z20.s }, p2/Z, [x10, #2, MUL VL]\n"
      "ld1w { z16.s }, p2/Z, [x10, #3, MUL VL]\n"
      "add x20, %x[qp], %[per_layer_mul]\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "add z23.s, z23.s, z22.s\n"
      "add z17.s, z17.s, z21.s\n"
      "add z18.s, z18.s, z20.s\n"
      "add z19.s, z19.s, z16.s\n"
      "ld1rw { z16.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      ".inst 0x04b076f7  // sqrdmulh z23.s, z23.s, z16.s\n"
      ".inst 0x04b07631  // sqrdmulh z17.s, z17.s, z16.s\n"
      "addvl x10, x10, #4\n"
      ".inst 0x04b07652  // sqrdmulh z18.s, z18.s, z16.s\n"
      ".inst 0x04b07673  // sqrdmulh z19.s, z19.s, z16.s\n"
      "tbz %x[flags], #5, 13f\n"
      "and z22.d, z23.d, z0.d\n"
      "and z21.d, z17.d, z0.d\n"
      "and z20.d, z18.d, z0.d\n"
      "and z16.d, z19.d, z0.d\n"
      "asr z22.s, z22.s, #0x1f\n"
      "asr z21.s, z21.s, #0x1f\n"
      "asr z20.s, z20.s, #0x1f\n"
      "asr z16.s, z16.s, #0x1f\n"
      "sqadd z23.s, z23.s, z22.s\n"
      "sqadd z17.s, z17.s, z21.s\n"
      "sqadd z18.s, z18.s, z20.s\n"
      "sqadd z19.s, z19.s, z16.s\n"
      "13:"  // Height 1: no shift correction
      "add x20, %x[qp], %[c_offset]\n"
      "ld1rw { z16.s }, p2/Z, [x20]\n"
      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
      "add z23.s, z23.s, z16.s\n"
      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
      "add z17.s, z17.s, z16.s\n"
      "add z18.s, z18.s, z16.s\n"
      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
      "add x20, %x[qp], %[maxval]\n"
      "ld1rw { z20.s }, p2/Z, [x20]\n"
      "add z19.s, z19.s, z16.s\n"
      "add x20, %x[qp], %[minval]\n"
      "ld1rw { z16.s }, p2/Z, [x20]\n"
      "smin z23.s, p2/M, z23.s, z20.s\n"
      "smin z17.s, p2/M, z17.s, z20.s\n"
      "smin z18.s, p2/M, z18.s, z20.s\n"
      "smin z19.s, p2/M, z19.s, z20.s\n"
      "smax z23.s, p2/M, z23.s, z16.s\n"
      "smax z17.s, p2/M, z17.s, z16.s\n"
      "smax z18.s, p2/M, z18.s, z16.s\n"
      "uzp1 z23.h, z23.h, z17.h\n"
      "smax z19.s, p2/M, z19.s, z16.s\n"
      "uzp1 z16.h, z18.h, z19.h\n"
      "uzp1 z23.b, z23.b, z16.b\n"
      "st1b { z23.b }, p1, [x27]\n"
      "addvl x27, x27, #1\n"
      "14:"  // Height 1: Writeback done
      "decw x9, ALL, MUL #4\n"
      "cmp x9, XZR\n"
      "bgt 2b\n"
      "b 58f\n"
      "15:"  // Height 2
      "mov x10, %x[col_bias]\n"
      "mov z11.s, #0x0\n"
      "mov z12.s, #0x0\n"
      "bic %x[flags], %x[flags], #0x80000000\n"
      "mov z15.b, #0x1\n"
      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
      "mov x27, %x[output_ptr]\n"
      "16:"  // Height 2: Column loop
      "mov x20, #0x0\n"
      "whilelt p1.b, x20, x9\n"
      "mov z16.s, #0x0\n"
      "mov z17.s, #0x0\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"
      "17:"  // Height 2: setup done
      "mov x26, #0x0\n"
      "18:"  // Height 2: String loop
      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
      "ldr w25, [x20, x26, LSL #0x2]\n"
      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "tbz %x[flags], #3, 19f\n"
      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
      "add x20, x20, x21, LSL #3\n"
      "ldr x24, [x20, #0x0]\n"
      "ldr x23, [x20, #0x8]\n"
      "cbnz x26, 20f\n"
      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
      "add x24, x24, x20\n"
      "add x23, x23, x20\n"
      "b 20f\n"
      "19:"  // Height 2: setup direct input
      "mov x24, %x[input_ptr]\n"
      "add x23, x24, x21\n"
      "20:"  // Height 2: input setup done
      "cmp x25, #0x10\n"
      "ble 23f\n"
      "21:"  // Height 2: Multiply loop: Main loop head
      "whilelt p0.b, XZR, x25\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "ld1rqb { z26.b }, p0/Z, [x23]\n"
      "trn1 z0.d, z1.d, z26.d\n"
      "ld1b { z24.b }, p2/Z, [x28]\n"
      ".inst 0x45d89810  // ummla z16.s, z0.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #1, MUL VL]\n"
      "trn2 z1.d, z1.d, z26.d\n"
      "ld1b { z24.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z26.b }, p2/Z, [x28, #3, MUL VL]\n"
      ".inst 0x45d99814  // ummla z20.s, z0.b, z25.b\n"
      ".inst 0x45d89811  // ummla z17.s, z0.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45da9815  // ummla z21.s, z0.b, z26.b\n"
      ".inst 0x45d99812  // ummla z18.s, z0.b, z25.b\n"
      "ld1b { z26.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z25.b }, p2/Z, [x28, #7, MUL VL]\n"
      "addvl x28, x28, #16\n"
      ".inst 0x45d89816  // ummla z22.s, z0.b, z24.b\n"
      "ld1b { z24.b }, p2/Z, [x28, #-8, MUL VL]\n"
      ".inst 0x45da9813  // ummla z19.s, z0.b, z26.b\n"
      ".inst 0x45d99817  // ummla z23.s, z0.b, z25.b\n"
      "ld1b { z26.b }, p2/Z, [x28, #-7, MUL VL]\n"
      ".inst 0x45d89830  // ummla z16.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #-6, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #-5, MUL VL]\n"
      ".inst 0x45da9834  // ummla z20.s, z1.b, z26.b\n"
      "ld1b { z27.b }, p2/Z, [x28, #-4, MUL VL]\n"
      "ld1b { z26.b }, p2/Z, [x28, #-3, MUL VL]\n"
      ".inst 0x45d99831  // ummla z17.s, z1.b, z25.b\n"
      ".inst 0x45d89835  // ummla z21.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #-2, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #-1, MUL VL]\n"
      ".inst 0x45db9832  // ummla z18.s, z1.b, z27.b\n"
      ".inst 0x45da9836  // ummla z22.s, z1.b, z26.b\n"
      ".inst 0x45d99833  // ummla z19.s, z1.b, z25.b\n"
      ".inst 0x45d89837  // ummla z23.s, z1.b, z24.b\n"
      "add x24, x24, #0x10\n"
      "add x23, x23, #0x10\n"
      "tbnz %x[flags], #31, 22f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z11.s, z1.b, z15.b\n"
      "22:"  // Height 2: Multiply loop: unique 3: skip row sum
      "sub x25, x25, #0x10\n"
      "cmp x25, #0x10\n"
      "bgt 21b\n"
      "23:"  // Height 2: Multiply loop: Single iteration only
      "whilelt p0.b, XZR, x25\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "ld1rqb { z27.b }, p0/Z, [x23]\n"
      "trn1 z0.d, z1.d, z27.d\n"
      "ld1b { z24.b }, p2/Z, [x28]\n"
      ".inst 0x45d89810  // ummla z16.s, z0.b, z24.b\n"
      "ld1b { z26.b }, p2/Z, [x28, #1, MUL VL]\n"
      "subs x25, x25, #0x8\n"
      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
      "trn2 z1.d, z1.d, z27.d\n"
      ".inst 0x45da9814  // ummla z20.s, z0.b, z26.b\n"
      "ld1b { z27.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z26.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45d99811  // ummla z17.s, z0.b, z25.b\n"
      ".inst 0x45d89815  // ummla z21.s, z0.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
      ".inst 0x45db9812  // ummla z18.s, z0.b, z27.b\n"
      ".inst 0x45da9816  // ummla z22.s, z0.b, z26.b\n"
      ".inst 0x45d99813  // ummla z19.s, z0.b, z25.b\n"
      ".inst 0x45d89817  // ummla z23.s, z0.b, z24.b\n"
      "addvl x28, x28, #8\n"
      "ble 24f\n"
      "ld1b { z24.b }, p2/Z, [x28]\n"
      ".inst 0x45d89830  // ummla z16.s, z1.b, z24.b\n"
      "ld1b { z24.b }, p2/Z, [x28, #1, MUL VL]\n"
      ".inst 0x45d89834  // ummla z20.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #3, MUL VL]\n"
      ".inst 0x45d99831  // ummla z17.s, z1.b, z25.b\n"
      ".inst 0x45d89835  // ummla z21.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45d99832  // ummla z18.s, z1.b, z25.b\n"
      ".inst 0x45d89836  // ummla z22.s, z1.b, z24.b\n"
      "ld1b { z25.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z24.b }, p2/Z, [x28, #7, MUL VL]\n"
      ".inst 0x45d99833  // ummla z19.s, z1.b, z25.b\n"
      ".inst 0x45d89837  // ummla z23.s, z1.b, z24.b\n"
      "addvl x28, x28, #8\n"
      "24:"  // Height 2: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 25f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z11.s, z1.b, z15.b\n"
      "25:"  // Height 2: Multiply loop: unique 4: skip row sum
      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
      "add x26, x26, #0x1\n"
      "cmp x26, x20\n"
      "bne 18b\n"
      "uzp1 z24.d, z16.d, z20.d\n"
      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
      "uzp2 z16.d, z16.d, z20.d\n"
      "add x23, x27, x20\n"
      "uzp1 z20.d, z17.d, z21.d\n"
      "uzp2 z17.d, z17.d, z21.d\n"
      "uzp1 z21.d, z18.d, z22.d\n"
      "uzp2 z18.d, z18.d, z22.d\n"
      "uzp1 z22.d, z19.d, z23.d\n"
      "uzp2 z19.d, z19.d, z23.d\n"
      "mov z23.d, z24.d\n"
      "tbnz %x[flags], #31, 26f\n"
      "add x20, %x[qp], %[b_offset]\n"
      "ld1rw { z24.s }, p2/Z, [x20]\n"
      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
      "neg z24.s, p2/M, z24.s\n"
      "mov z12.s, z11.s[3]\n"
      "mov z11.s, z11.s[0]\n"
      "mul z11.s, p2/M, z11.s, z24.s\n"
      "mul z12.s, p2/M, z12.s, z24.s\n"
      "26:"  // Height 2: skip row sum fixup
      "add z23.s, z23.s, z11.s\n"
      "add z20.s, z20.s, z11.s\n"
      "ld1w { z28.s }, p2/Z, [x10]\n"
      "ld1w { z27.s }, p2/Z, [x10, #1, MUL VL]\n"
      "add z21.s, z21.s, z11.s\n"
      "add z22.s, z22.s, z11.s\n"
      "ld1w { z26.s }, p2/Z, [x10, #2, MUL VL]\n"
      "ld1w { z25.s }, p2/Z, [x10, #3, MUL VL]\n"
      "add z16.s, z16.s, z12.s\n"
      "add z17.s, z17.s, z12.s\n"
      "add x20, %x[qp], %[per_layer_mul]\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "add z18.s, z18.s, z12.s\n"
      "add z19.s, z19.s, z12.s\n"
      "ld1rw { z24.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      "add z23.s, z23.s, z28.s\n"
      "add z20.s, z20.s, z27.s\n"
      "addvl x10, x10, #4\n"
      "add z21.s, z21.s, z26.s\n"
      "add z22.s, z22.s, z25.s\n"
      "add z16.s, z16.s, z28.s\n"
      "add z17.s, z17.s, z27.s\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      "add z18.s, z18.s, z26.s\n"
      "add z19.s, z19.s, z25.s\n"
      ".inst 0x04b876f7  // sqrdmulh z23.s, z23.s, z24.s\n"
      ".inst 0x04b87694  // sqrdmulh z20.s, z20.s, z24.s\n"
      ".inst 0x04b876b5  // sqrdmulh z21.s, z21.s, z24.s\n"
      ".inst 0x04b876d6  // sqrdmulh z22.s, z22.s, z24.s\n"
      ".inst 0x04b87610  // sqrdmulh z16.s, z16.s, z24.s\n"
      ".inst 0x04b87631  // sqrdmulh z17.s, z17.s, z24.s\n"
      ".inst 0x04b87652  // sqrdmulh z18.s, z18.s, z24.s\n"
      ".inst 0x04b87673  // sqrdmulh z19.s, z19.s, z24.s\n"
      "tbz %x[flags], #5, 27f\n"
      "and z24.d, z23.d, z0.d\n"
      "asr z24.s, z24.s, #0x1f\n"
      "sqadd z23.s, z23.s, z24.s\n"
      "and z30.d, z20.d, z0.d\n"
      "and z29.d, z21.d, z0.d\n"
      "and z28.d, z22.d, z0.d\n"
      "and z27.d, z16.d, z0.d\n"
      "and z26.d, z17.d, z0.d\n"
      "and z25.d, z18.d, z0.d\n"
      "and z24.d, z19.d, z0.d\n"
      "asr z30.s, z30.s, #0x1f\n"
      "asr z29.s, z29.s, #0x1f\n"
      "asr z28.s, z28.s, #0x1f\n"
      "asr z27.s, z27.s, #0x1f\n"
      "asr z26.s, z26.s, #0x1f\n"
      "asr z25.s, z25.s, #0x1f\n"
      "asr z24.s, z24.s, #0x1f\n"
      "sqadd z20.s, z20.s, z30.s\n"
      "sqadd z21.s, z21.s, z29.s\n"
      "sqadd z22.s, z22.s, z28.s\n"
      "sqadd z16.s, z16.s, z27.s\n"
      "sqadd z17.s, z17.s, z26.s\n"
      "sqadd z18.s, z18.s, z25.s\n"
      "sqadd z19.s, z19.s, z24.s\n"
      "27:"  // Height 2: no shift correction
      "add x20, %x[qp], %[c_offset]\n"
      "ld1rw { z24.s }, p2/Z, [x20]\n"
      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
      "add z23.s, z23.s, z24.s\n"
      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
      "add z20.s, z20.s, z24.s\n"
      "add z21.s, z21.s, z24.s\n"
      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
      "add z22.s, z22.s, z24.s\n"
      "add z16.s, z16.s, z24.s\n"
      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
      "add z17.s, z17.s, z24.s\n"
      "add z18.s, z18.s, z24.s\n"
      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
      "add x20, %x[qp], %[maxval]\n"
      "ld1rw { z25.s }, p2/Z, [x20]\n"
      "add z19.s, z19.s, z24.s\n"
      "add x20, %x[qp], %[minval]\n"
      "ld1rw { z24.s }, p2/Z, [x20]\n"
      "smin z23.s, p2/M, z23.s, z25.s\n"
      "smin z20.s, p2/M, z20.s, z25.s\n"
      "smin z21.s, p2/M, z21.s, z25.s\n"
      "smin z22.s, p2/M, z22.s, z25.s\n"
      "smin z16.s, p2/M, z16.s, z25.s\n"
      "smin z17.s, p2/M, z17.s, z25.s\n"
      "smin z18.s, p2/M, z18.s, z25.s\n"
      "smin z19.s, p2/M, z19.s, z25.s\n"
      "smax z23.s, p2/M, z23.s, z24.s\n"
      "smax z20.s, p2/M, z20.s, z24.s\n"
      "smax z21.s, p2/M, z21.s, z24.s\n"
      "uzp1 z23.h, z23.h, z20.h\n"
      "smax z22.s, p2/M, z22.s, z24.s\n"
      "smax z16.s, p2/M, z16.s, z24.s\n"
      "uzp1 z20.h, z21.h, z22.h\n"
      "uzp1 z23.b, z23.b, z20.b\n"
      "smax z17.s, p2/M, z17.s, z24.s\n"
      "smax z18.s, p2/M, z18.s, z24.s\n"
      "uzp1 z16.h, z16.h, z17.h\n"
      "st1b { z23.b }, p1, [x27]\n"
      "smax z19.s, p2/M, z19.s, z24.s\n"
      "uzp1 z17.h, z18.h, z19.h\n"
      "uzp1 z16.b, z16.b, z17.b\n"
      "st1b { z16.b }, p1, [x23]\n"
      "addvl x27, x27, #1\n"
      "28:"  // Height 2: Writeback done
      "decw x9, ALL, MUL #4\n"
      "cmp x9, XZR\n"
      "bgt 16b\n"
      "b 58f\n"
      "29:"  // Height 3
      "mov x10, %x[col_bias]\n"
      "mov z11.s, #0x0\n"
      "mov z12.s, #0x0\n"
      "bic %x[flags], %x[flags], #0x80000000\n"
      "mov z13.s, #0x0\n"
      "mov z15.b, #0x1\n"
      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
      "mov x27, %x[output_ptr]\n"
      "30:"  // Height 3: Column loop
      "mov x20, #0x0\n"
      "whilelt p1.b, x20, x9\n"
      "mov z16.s, #0x0\n"
      "mov z17.s, #0x0\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"
      "31:"  // Height 3: setup done
      "mov x26, #0x0\n"
      "32:"  // Height 3: String loop
      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
      "ldr w25, [x20, x26, LSL #0x2]\n"
      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "tbz %x[flags], #3, 33f\n"
      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
      "add x20, x20, x21, LSL #3\n"
      "ldr x24, [x20, #0x0]\n"
      "ldr x23, [x20, #0x8]\n"
      "ldr x22, [x20, #0x10]\n"
      "cbnz x26, 34f\n"
      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
      "add x24, x24, x20\n"
      "add x23, x23, x20\n"
      "add x22, x22, x20\n"
      "b 34f\n"
      "33:"  // Height 3: setup direct input
      "mov x24, %x[input_ptr]\n"
      "add x23, x24, x21\n"
      "add x22, x23, x21\n"
      "34:"  // Height 3: input setup done
      "cmp x25, #0x10\n"
      "ble 37f\n"
      "35:"  // Height 3: Multiply loop: Main loop head
      "whilelt p0.b, XZR, x25\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "ld1rqb { z2.b }, p0/Z, [x23]\n"
      "ld1rqb { z3.b }, p0/Z, [x22]\n"
      "trn1 z0.d, z1.d, z2.d\n"
      "trn2 z1.d, z1.d, z2.d\n"
      "ld1b { z4.b }, p2/Z, [x28]\n"
      "trn1 z2.d, z3.d, z5.d\n"
      "trn2 z3.d, z3.d, z5.d\n"
      ".inst 0x45c49810  // ummla z16.s, z0.b, z4.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
      ".inst 0x45c49858  // ummla z24.s, z2.b, z4.b\n"
      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z9.b }, p2/Z, [x28, #3, MUL VL]\n"
      ".inst 0x45c59814  // ummla z20.s, z0.b, z5.b\n"
      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45c5985c  // ummla z28.s, z2.b, z5.b\n"
      ".inst 0x45c49811  // ummla z17.s, z0.b, z4.b\n"
      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
      "addvl x28, x28, #16\n"
      ".inst 0x45c49859  // ummla z25.s, z2.b, z4.b\n"
      "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
      ".inst 0x45c99815  // ummla z21.s, z0.b, z9.b\n"
      ".inst 0x45c9985d  // ummla z29.s, z2.b, z9.b\n"
      "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
      ".inst 0x45c89812  // ummla z18.s, z0.b, z8.b\n"
      ".inst 0x45c8985a  // ummla z26.s, z2.b, z8.b\n"
      "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n"
      "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n"
      ".inst 0x45c79816  // ummla z22.s, z0.b, z7.b\n"
      ".inst 0x45c7985e  // ummla z30.s, z2.b, z7.b\n"
      "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n"
      "add x24, x24, #0x10\n"
      ".inst 0x45c69813  // ummla z19.s, z0.b, z6.b\n"
      ".inst 0x45c6985b  // ummla z27.s, z2.b, z6.b\n"
      "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
      "add x23, x23, #0x10\n"
      ".inst 0x45c59817  // ummla z23.s, z0.b, z5.b\n"
      ".inst 0x45c5985f  // ummla z31.s, z2.b, z5.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n"
      "add x22, x22, #0x10\n"
      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
      ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
      "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n"
      ".inst 0x45ca9834  // ummla z20.s, z1.b, z10.b\n"
      ".inst 0x45ca987c  // ummla z28.s, z3.b, z10.b\n"
      ".inst 0x45c99831  // ummla z17.s, z1.b, z9.b\n"
      ".inst 0x45c99879  // ummla z25.s, z3.b, z9.b\n"
      ".inst 0x45c89835  // ummla z21.s, z1.b, z8.b\n"
      ".inst 0x45c8987d  // ummla z29.s, z3.b, z8.b\n"
      ".inst 0x45c79832  // ummla z18.s, z1.b, z7.b\n"
      ".inst 0x45c7987a  // ummla z26.s, z3.b, z7.b\n"
      ".inst 0x45c69836  // ummla z22.s, z1.b, z6.b\n"
      ".inst 0x45c6987e  // ummla z30.s, z3.b, z6.b\n"
      ".inst 0x45c59833  // ummla z19.s, z1.b, z5.b\n"
      ".inst 0x45c5987b  // ummla z27.s, z3.b, z5.b\n"
      ".inst 0x45c49837  // ummla z23.s, z1.b, z4.b\n"
      ".inst 0x45c4987f  // ummla z31.s, z3.b, z4.b\n"
      "tbnz %x[flags], #31, 36f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z13.s, z2.b, z15.b\n"
      "udot z11.s, z1.b, z15.b\n"
      "udot z13.s, z3.b, z15.b\n"
      "36:"  // Height 3: Multiply loop: unique 5: skip row sum
      "sub x25, x25, #0x10\n"
      "cmp x25, #0x10\n"
      "bgt 35b\n"
      "37:"  // Height 3: Multiply loop: Single iteration only
      "whilelt p0.b, XZR, x25\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "ld1rqb { z2.b }, p0/Z, [x23]\n"
      "ld1rqb { z3.b }, p0/Z, [x22]\n"
      "trn1 z0.d, z1.d, z2.d\n"
      "trn2 z1.d, z1.d, z2.d\n"
      "ld1b { z5.b }, p2/Z, [x28]\n"
      "trn1 z2.d, z3.d, z4.d\n"
      "trn2 z3.d, z3.d, z4.d\n"
      ".inst 0x45c59810  // ummla z16.s, z0.b, z5.b\n"
      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
      ".inst 0x45c59858  // ummla z24.s, z2.b, z5.b\n"
      "ld1b { z9.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
      "subs x25, x25, #0x8\n"
      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45c49814  // ummla z20.s, z0.b, z4.b\n"
      ".inst 0x45c4985c  // ummla z28.s, z2.b, z4.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
      ".inst 0x45c99811  // ummla z17.s, z0.b, z9.b\n"
      ".inst 0x45c99859  // ummla z25.s, z2.b, z9.b\n"
      ".inst 0x45c89815  // ummla z21.s, z0.b, z8.b\n"
      ".inst 0x45c8985d  // ummla z29.s, z2.b, z8.b\n"
      "addvl x28, x28, #8\n"
      ".inst 0x45c79812  // ummla z18.s, z0.b, z7.b\n"
      ".inst 0x45c7985a  // ummla z26.s, z2.b, z7.b\n"
      ".inst 0x45c69816  // ummla z22.s, z0.b, z6.b\n"
      ".inst 0x45c6985e  // ummla z30.s, z2.b, z6.b\n"
      ".inst 0x45c59813  // ummla z19.s, z0.b, z5.b\n"
      ".inst 0x45c5985b  // ummla z27.s, z2.b, z5.b\n"
      ".inst 0x45c49817  // ummla z23.s, z0.b, z4.b\n"
      ".inst 0x45c4985f  // ummla z31.s, z2.b, z4.b\n"
      "ble 38f\n"
      "ld1b { z4.b }, p2/Z, [x28]\n"
      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
      ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
      ".inst 0x45c59834  // ummla z20.s, z1.b, z5.b\n"
      ".inst 0x45c5987c  // ummla z28.s, z3.b, z5.b\n"
      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45c49831  // ummla z17.s, z1.b, z4.b\n"
      ".inst 0x45c49879  // ummla z25.s, z3.b, z4.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
      ".inst 0x45c89835  // ummla z21.s, z1.b, z8.b\n"
      ".inst 0x45c8987d  // ummla z29.s, z3.b, z8.b\n"
      ".inst 0x45c79832  // ummla z18.s, z1.b, z7.b\n"
      ".inst 0x45c7987a  // ummla z26.s, z3.b, z7.b\n"
      "addvl x28, x28, #8\n"
      ".inst 0x45c69836  // ummla z22.s, z1.b, z6.b\n"
      ".inst 0x45c6987e  // ummla z30.s, z3.b, z6.b\n"
      ".inst 0x45c59833  // ummla z19.s, z1.b, z5.b\n"
      ".inst 0x45c5987b  // ummla z27.s, z3.b, z5.b\n"
      ".inst 0x45c49837  // ummla z23.s, z1.b, z4.b\n"
      ".inst 0x45c4987f  // ummla z31.s, z3.b, z4.b\n"
      "38:"  // Height 3: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 39f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z13.s, z2.b, z15.b\n"
      "udot z11.s, z1.b, z15.b\n"
      "udot z13.s, z3.b, z15.b\n"
      "39:"  // Height 3: Multiply loop: unique 6: skip row sum
      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
      "add x26, x26, #0x1\n"
      "cmp x26, x20\n"
      "bne 32b\n"
      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
      "uzp1 z0.d, z16.d, z20.d\n"
      "add x23, x27, x20\n"
      "uzp2 z16.d, z16.d, z20.d\n"
      "uzp1 z20.d, z17.d, z21.d\n"
      "uzp2 z17.d, z17.d, z21.d\n"
      "add x22, x23, x20\n"
      "uzp1 z21.d, z18.d, z22.d\n"
      "uzp2 z18.d, z18.d, z22.d\n"
      "uzp1 z22.d, z19.d, z23.d\n"
      "uzp2 z19.d, z19.d, z23.d\n"
      "uzp1 z24.d, z24.d, z28.d\n"
      "uzp1 z25.d, z25.d, z29.d\n"
      "uzp1 z26.d, z26.d, z30.d\n"
      "uzp1 z27.d, z27.d, z31.d\n"
      "mov z31.d, z0.d\n"
      "tbnz %x[flags], #31, 40f\n"
      "add x20, %x[qp], %[b_offset]\n"
      "ld1rw { z23.s }, p2/Z, [x20]\n"
      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
      ".inst 0x4491a9ad  // addp z13.s, p2/m, z13.s, z13.s\n"
      "neg z23.s, p2/M, z23.s\n"
      "mov z12.s, z11.s[3]\n"
      "mov z11.s, z11.s[0]\n"
      "mul z11.s, p2/M, z11.s, z23.s\n"
      "mov z13.s, z13.s[0]\n"
      "mul z12.s, p2/M, z12.s, z23.s\n"
      "mul z13.s, p2/M, z13.s, z23.s\n"
      "40:"  // Height 3: skip row sum fixup
      "add z31.s, z31.s, z11.s\n"
      "add z20.s, z20.s, z11.s\n"
      "ld1w { z0.s }, p2/Z, [x10]\n"
      "ld1w { z30.s }, p2/Z, [x10, #1, MUL VL]\n"
      "add z21.s, z21.s, z11.s\n"
      "add z22.s, z22.s, z11.s\n"
      "ld1w { z29.s }, p2/Z, [x10, #2, MUL VL]\n"
      "ld1w { z28.s }, p2/Z, [x10, #3, MUL VL]\n"
      "add z16.s, z16.s, z12.s\n"
      "add z17.s, z17.s, z12.s\n"
      "add x20, %x[qp], %[per_layer_mul]\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "add z18.s, z18.s, z12.s\n"
      "add z19.s, z19.s, z12.s\n"
      "ld1rw { z23.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      "add z24.s, z24.s, z13.s\n"
      "add z25.s, z25.s, z13.s\n"
      "addvl x10, x10, #4\n"
      "add z26.s, z26.s, z13.s\n"
      "add z27.s, z27.s, z13.s\n"
      "add z31.s, z31.s, z0.s\n"
      "add z20.s, z20.s, z30.s\n"
      "add z21.s, z21.s, z29.s\n"
      "add z22.s, z22.s, z28.s\n"
      "add z16.s, z16.s, z0.s\n"
      "add z17.s, z17.s, z30.s\n"
      "add z18.s, z18.s, z29.s\n"
      "add z19.s, z19.s, z28.s\n"
      "add z24.s, z24.s, z0.s\n"
      "add z25.s, z25.s, z30.s\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      "add z26.s, z26.s, z29.s\n"
      "add z27.s, z27.s, z28.s\n"
      ".inst 0x04b777ff  // sqrdmulh z31.s, z31.s, z23.s\n"
      ".inst 0x04b77694  // sqrdmulh z20.s, z20.s, z23.s\n"
      ".inst 0x04b776b5  // sqrdmulh z21.s, z21.s, z23.s\n"
      ".inst 0x04b776d6  // sqrdmulh z22.s, z22.s, z23.s\n"
      ".inst 0x04b77610  // sqrdmulh z16.s, z16.s, z23.s\n"
      ".inst 0x04b77631  // sqrdmulh z17.s, z17.s, z23.s\n"
      ".inst 0x04b77652  // sqrdmulh z18.s, z18.s, z23.s\n"
      ".inst 0x04b77673  // sqrdmulh z19.s, z19.s, z23.s\n"
      ".inst 0x04b77718  // sqrdmulh z24.s, z24.s, z23.s\n"
      ".inst 0x04b77739  // sqrdmulh z25.s, z25.s, z23.s\n"
      ".inst 0x04b7775a  // sqrdmulh z26.s, z26.s, z23.s\n"
      ".inst 0x04b7777b  // sqrdmulh z27.s, z27.s, z23.s\n"
      "tbz %x[flags], #5, 41f\n"
      "and z1.d, z31.d, z0.d\n"
      "and z30.d, z20.d, z0.d\n"
      "and z29.d, z21.d, z0.d\n"
      "and z28.d, z22.d, z0.d\n"
      "and z23.d, z16.d, z0.d\n"
      "asr z1.s, z1.s, #0x1f\n"
      "asr z30.s, z30.s, #0x1f\n"
      "asr z29.s, z29.s, #0x1f\n"
      "asr z28.s, z28.s, #0x1f\n"
      "asr z23.s, z23.s, #0x1f\n"
      "sqadd z31.s, z31.s, z1.s\n"
      "sqadd z20.s, z20.s, z30.s\n"
      "sqadd z21.s, z21.s, z29.s\n"
      "sqadd z22.s, z22.s, z28.s\n"
      "sqadd z16.s, z16.s, z23.s\n"
      "and z3.d, z17.d, z0.d\n"
      "and z2.d, z18.d, z0.d\n"
      "and z1.d, z19.d, z0.d\n"
      "and z30.d, z24.d, z0.d\n"
      "and z29.d, z25.d, z0.d\n"
      "and z28.d, z26.d, z0.d\n"
      "and z23.d, z27.d, z0.d\n"
      "asr z3.s, z3.s, #0x1f\n"
      "asr z2.s, z2.s, #0x1f\n"
      "asr z1.s, z1.s, #0x1f\n"
      "asr z30.s, z30.s, #0x1f\n"
      "asr z29.s, z29.s, #0x1f\n"
      "asr z28.s, z28.s, #0x1f\n"
      "asr z23.s, z23.s, #0x1f\n"
      "sqadd z17.s, z17.s, z3.s\n"
      "sqadd z18.s, z18.s, z2.s\n"
      "sqadd z19.s, z19.s, z1.s\n"
      "sqadd z24.s, z24.s, z30.s\n"
      "sqadd z25.s, z25.s, z29.s\n"
      "sqadd z26.s, z26.s, z28.s\n"
      "sqadd z27.s, z27.s, z23.s\n"
      "41:"  // Height 3: no shift correction
      "add x20, %x[qp], %[c_offset]\n"
      "ld1rw { z23.s }, p2/Z, [x20]\n"
      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
      "add z31.s, z31.s, z23.s\n"
      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
      "add z20.s, z20.s, z23.s\n"
      "add z21.s, z21.s, z23.s\n"
      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
      "add z22.s, z22.s, z23.s\n"
      "add z16.s, z16.s, z23.s\n"
      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
      "add z17.s, z17.s, z23.s\n"
      "add z18.s, z18.s, z23.s\n"
      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
      ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
      "add z19.s, z19.s, z23.s\n"
      "add z24.s, z24.s, z23.s\n"
      ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
      ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
      "add z25.s, z25.s, z23.s\n"
      "add z26.s, z26.s, z23.s\n"
      ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
      "add x20, %x[qp], %[maxval]\n"
      "ld1rw { z28.s }, p2/Z, [x20]\n"
      "add z27.s, z27.s, z23.s\n"
      "add x20, %x[qp], %[minval]\n"
      "ld1rw { z23.s }, p2/Z, [x20]\n"
      "smin z31.s, p2/M, z31.s, z28.s\n"
      "smin z20.s, p2/M, z20.s, z28.s\n"
      "smin z21.s, p2/M, z21.s, z28.s\n"
      "smin z22.s, p2/M, z22.s, z28.s\n"
      "smin z16.s, p2/M, z16.s, z28.s\n"
      "smin z17.s, p2/M, z17.s, z28.s\n"
      "smin z18.s, p2/M, z18.s, z28.s\n"
      "smin z19.s, p2/M, z19.s, z28.s\n"
      "smin z24.s, p2/M, z24.s, z28.s\n"
      "smin z25.s, p2/M, z25.s, z28.s\n"
      "smin z26.s, p2/M, z26.s, z28.s\n"
      "smin z27.s, p2/M, z27.s, z28.s\n"
      "smax z31.s, p2/M, z31.s, z23.s\n"
      "smax z20.s, p2/M, z20.s, z23.s\n"
      "smax z21.s, p2/M, z21.s, z23.s\n"
      "uzp1 z31.h, z31.h, z20.h\n"
      "smax z22.s, p2/M, z22.s, z23.s\n"
      "smax z16.s, p2/M, z16.s, z23.s\n"
      "uzp1 z20.h, z21.h, z22.h\n"
      "uzp1 z31.b, z31.b, z20.b\n"
      "smax z17.s, p2/M, z17.s, z23.s\n"
      "smax z18.s, p2/M, z18.s, z23.s\n"
      "uzp1 z16.h, z16.h, z17.h\n"
      "st1b { z31.b }, p1, [x27]\n"
      "smax z19.s, p2/M, z19.s, z23.s\n"
      "smax z24.s, p2/M, z24.s, z23.s\n"
      "uzp1 z17.h, z18.h, z19.h\n"
      "uzp1 z16.b, z16.b, z17.b\n"
      "smax z25.s, p2/M, z25.s, z23.s\n"
      "smax z26.s, p2/M, z26.s, z23.s\n"
      "uzp1 z24.h, z24.h, z25.h\n"
      "st1b { z16.b }, p1, [x23]\n"
      "smax z27.s, p2/M, z27.s, z23.s\n"
      "uzp1 z16.h, z26.h, z27.h\n"
      "uzp1 z24.b, z24.b, z16.b\n"
      "st1b { z24.b }, p1, [x22]\n"
      "addvl x27, x27, #1\n"
      "42:"  // Height 3: Writeback done
      "decw x9, ALL, MUL #4\n"
      "cmp x9, XZR\n"
      "bgt 30b\n"
      "b 58f\n"
      "43:"  // Height 4
      "ldr x21, [%x[args_ptr], %[offsetof_output_offset]]\n"
      "mov x20, #0x4\n"
      "mov x10, %x[col_bias]\n"
      "mov z11.s, #0x0\n"
      "mov z12.s, #0x0\n"
      "mov z13.s, #0x0\n"
      "bic %x[flags], %x[flags], #0x80000000\n"
      "ldr x9, [%x[args_ptr], %[offsetof_N]]\n"
      "mov z14.s, #0x0\n"
      "mov z15.b, #0x1\n"
      "ldr x28, [%x[args_ptr], %[offsetof_B_ptr]]\n"
      "mov x27, %x[output_ptr]\n"
      "madd %x[output_ptr], x21, x20, %x[output_ptr]\n"
      "44:"  // Height 4: Column loop
      "mov x20, #0x0\n"
      "whilelt p1.b, x20, x9\n"
      "mov z16.s, #0x0\n"
      "mov z17.s, #0x0\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"
      "45:"  // Height 4: setup done
      "mov x26, #0x0\n"
      "46:"  // Height 4: String loop
      "ldr x20, [%x[args_ptr], %[offsetof_string_lengths]]\n"
      "ldr w25, [x20, x26, LSL #0x2]\n"
      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "tbz %x[flags], #3, 47f\n"
      "ldr x20, [%x[input_ptr], x26, LSL #0x3]\n"
      "add x20, x20, x21, LSL #3\n"
      "ldr x24, [x20, #0x0]\n"
      "ldr x23, [x20, #0x8]\n"
      "ldr x22, [x20, #0x10]\n"
      "ldr x21, [x20, #0x18]\n"
      "cbnz x26, 48f\n"
      "ldr x20, [%x[args_ptr], %[offsetof_input_initial_col]]\n"
      "add x24, x24, x20\n"
      "add x23, x23, x20\n"
      "add x22, x22, x20\n"
      "add x21, x21, x20\n"
      "b 48f\n"
      "47:"  // Height 4: setup direct input
      "mov x24, %x[input_ptr]\n"
      "add x23, x24, x21\n"
      "add x22, x23, x21\n"
      "add x21, x22, x21\n"
      "48:"  // Height 4: input setup done
      "cmp x25, #0x10\n"
      "ble 51f\n"
      "49:"  // Height 4: Multiply loop: Main loop head
      "whilelt p0.b, XZR, x25\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "ld1rqb { z2.b }, p0/Z, [x23]\n"
      "trn1 z0.d, z1.d, z2.d\n"
      "ld1rqb { z3.b }, p0/Z, [x22]\n"
      "ld1rqb { z5.b }, p0/Z, [x21]\n"
      "trn2 z1.d, z1.d, z2.d\n"
      "trn1 z2.d, z3.d, z5.d\n"
      "ld1b { z4.b }, p2/Z, [x28]\n"
      "trn2 z3.d, z3.d, z5.d\n"
      ".inst 0x45c49810  // ummla z16.s, z0.b, z4.b\n"
      ".inst 0x45c49858  // ummla z24.s, z2.b, z4.b\n"
      "ld1b { z4.b }, p2/Z, [x28, #1, MUL VL]\n"
      "ld1b { z5.b }, p2/Z, [x28, #2, MUL VL]\n"
      ".inst 0x45c49814  // ummla z20.s, z0.b, z4.b\n"
      ".inst 0x45c4985c  // ummla z28.s, z2.b, z4.b\n"
      "ld1b { z4.b }, p2/Z, [x28, #3, MUL VL]\n"
      "ld1b { z8.b }, p2/Z, [x28, #4, MUL VL]\n"
      ".inst 0x45c59811  // ummla z17.s, z0.b, z5.b\n"
      ".inst 0x45c59859  // ummla z25.s, z2.b, z5.b\n"
      "ld1b { z7.b }, p2/Z, [x28, #5, MUL VL]\n"
      "ld1b { z6.b }, p2/Z, [x28, #6, MUL VL]\n"
      ".inst 0x45c49815  // ummla z21.s, z0.b, z4.b\n"
      ".inst 0x45c4985d  // ummla z29.s, z2.b, z4.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #7, MUL VL]\n"
      "addvl x28, x28, #16\n"
      "ld1b { z4.b }, p2/Z, [x28, #-8, MUL VL]\n"
      ".inst 0x45c89812  // ummla z18.s, z0.b, z8.b\n"
      ".inst 0x45c8985a  // ummla z26.s, z2.b, z8.b\n"
      ".inst 0x45c79816  // ummla z22.s, z0.b, z7.b\n"
      "ld1b { z10.b }, p2/Z, [x28, #-7, MUL VL]\n"
      "ld1b { z9.b }, p2/Z, [x28, #-6, MUL VL]\n"
      ".inst 0x45c7985e  // ummla z30.s, z2.b, z7.b\n"
      ".inst 0x45c69813  // ummla z19.s, z0.b, z6.b\n"
      "ld1b { z8.b }, p2/Z, [x28, #-5, MUL VL]\n"
      "ld1b { z7.b }, p2/Z, [x28, #-4, MUL VL]\n"
      ".inst 0x45c6985b  // ummla z27.s, z2.b, z6.b\n"
      ".inst 0x45c59817  // ummla z23.s, z0.b, z5.b\n"
      "ld1b { z6.b }, p2/Z, [x28, #-3, MUL VL]\n"
      "add x24, x24, #0x10\n"
      ".inst 0x45c5985f  // ummla z31.s, z2.b, z5.b\n"
      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #-2, MUL VL]\n"
      "add x23, x23, #0x10\n"
      ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
      "ld1b { z4.b }, p2/Z, [x28, #-1, MUL VL]\n"
      ".inst 0x45ca9834  // ummla z20.s, z1.b, z10.b\n"
      "add x22, x22, #0x10\n"
      ".inst 0x45ca987c  // ummla z28.s, z3.b, z10.b\n"
      ".inst 0x45c99831  // ummla z17.s, z1.b, z9.b\n"
      "add x21, x21, #0x10\n"
      ".inst 0x45c99879  // ummla z25.s, z3.b, z9.b\n"
      ".inst 0x45c89835  // ummla z21.s, z1.b, z8.b\n"
      ".inst 0x45c8987d  // ummla z29.s, z3.b, z8.b\n"
      ".inst 0x45c79832  // ummla z18.s, z1.b, z7.b\n"
      ".inst 0x45c7987a  // ummla z26.s, z3.b, z7.b\n"
      ".inst 0x45c69836  // ummla z22.s, z1.b, z6.b\n"
      ".inst 0x45c6987e  // ummla z30.s, z3.b, z6.b\n"
      ".inst 0x45c59833  // ummla z19.s, z1.b, z5.b\n"
      ".inst 0x45c5987b  // ummla z27.s, z3.b, z5.b\n"
      ".inst 0x45c49837  // ummla z23.s, z1.b, z4.b\n"
      ".inst 0x45c4987f  // ummla z31.s, z3.b, z4.b\n"
      "tbnz %x[flags], #31, 50f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z13.s, z2.b, z15.b\n"
      "udot z11.s, z1.b, z15.b\n"
      "udot z13.s, z3.b, z15.b\n"
      "50:"  // Height 4: Multiply loop: unique 7: skip row sum
      "sub x25, x25, #0x10\n"
      "cmp x25, #0x10\n"
      "bgt 49b\n"
      "51:"  // Height 4: Multiply loop: Single iteration only
      "whilelt p0.b, XZR, x25\n"
      "ld1rqb { z1.b }, p0/Z, [x24]\n"
      "ld1rqb { z2.b }, p0/Z, [x23]\n"
      "trn1 z0.d, z1.d, z2.d\n"
      "ld1rqb { z3.b }, p0/Z, [x22]\n"
      "ld1rqb { z5.b }, p0/Z, [x21]\n"
      "trn2 z1.d, z1.d, z2.d\n"
      "trn1 z2.d, z3.d, z5.d\n"
      "ld1b { z4.b }, p2/Z, [x28]\n"
      "trn2 z3.d, z3.d, z5.d\n"
      ".inst 0x45c49810  // ummla z16.s, z0.b, z4.b\n"
      ".inst 0x45c49858  // ummla z24.s, z2.b, z4.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
      "subs x25, x25, #0x8\n"
      ".inst 0x45c59814  // ummla z20.s, z0.b, z5.b\n"
      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
      ".inst 0x45c5985c  // ummla z28.s, z2.b, z5.b\n"
      ".inst 0x45c49811  // ummla z17.s, z0.b, z4.b\n"
      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
      "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
      ".inst 0x45c49859  // ummla z25.s, z2.b, z4.b\n"
      ".inst 0x45c89815  // ummla z21.s, z0.b, z8.b\n"
      "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
      ".inst 0x45c8985d  // ummla z29.s, z2.b, z8.b\n"
      ".inst 0x45c79812  // ummla z18.s, z0.b, z7.b\n"
      "addvl x28, x28, #8\n"
      ".inst 0x45c7985a  // ummla z26.s, z2.b, z7.b\n"
      ".inst 0x45c69816  // ummla z22.s, z0.b, z6.b\n"
      ".inst 0x45c6985e  // ummla z30.s, z2.b, z6.b\n"
      ".inst 0x45c59813  // ummla z19.s, z0.b, z5.b\n"
      ".inst 0x45c5985b  // ummla z27.s, z2.b, z5.b\n"
      ".inst 0x45c49817  // ummla z23.s, z0.b, z4.b\n"
      ".inst 0x45c4985f  // ummla z31.s, z2.b, z4.b\n"
      "ble 52f\n"
      "ld1b { z4.b }, p2/Z, [x28]\n"
      ".inst 0x45c49830  // ummla z16.s, z1.b, z4.b\n"
      ".inst 0x45c49878  // ummla z24.s, z3.b, z4.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #1, MUL VL]\n"
      "ld1b { z4.b }, p2/Z, [x28, #2, MUL VL]\n"
      "ld1b { z8.b }, p2/Z, [x28, #3, MUL VL]\n"
      ".inst 0x45c59834  // ummla z20.s, z1.b, z5.b\n"
      ".inst 0x45c5987c  // ummla z28.s, z3.b, z5.b\n"
      "ld1b { z7.b }, p2/Z, [x28, #4, MUL VL]\n"
      "ld1b { z6.b }, p2/Z, [x28, #5, MUL VL]\n"
      ".inst 0x45c49831  // ummla z17.s, z1.b, z4.b\n"
      ".inst 0x45c49879  // ummla z25.s, z3.b, z4.b\n"
      "ld1b { z5.b }, p2/Z, [x28, #6, MUL VL]\n"
      "ld1b { z4.b }, p2/Z, [x28, #7, MUL VL]\n"
      ".inst 0x45c89835  // ummla z21.s, z1.b, z8.b\n"
      ".inst 0x45c8987d  // ummla z29.s, z3.b, z8.b\n"
      ".inst 0x45c79832  // ummla z18.s, z1.b, z7.b\n"
      ".inst 0x45c7987a  // ummla z26.s, z3.b, z7.b\n"
      "addvl x28, x28, #8\n"
      ".inst 0x45c69836  // ummla z22.s, z1.b, z6.b\n"
      ".inst 0x45c6987e  // ummla z30.s, z3.b, z6.b\n"
      ".inst 0x45c59833  // ummla z19.s, z1.b, z5.b\n"
      ".inst 0x45c5987b  // ummla z27.s, z3.b, z5.b\n"
      ".inst 0x45c49837  // ummla z23.s, z1.b, z4.b\n"
      ".inst 0x45c4987f  // ummla z31.s, z3.b, z4.b\n"
      "52:"  // Height 4: Multiply loop: multiply skip
      "tbnz %x[flags], #31, 53f\n"
      "udot z11.s, z0.b, z15.b\n"
      "udot z13.s, z2.b, z15.b\n"
      "udot z11.s, z1.b, z15.b\n"
      "udot z13.s, z3.b, z15.b\n"
      "53:"  // Height 4: Multiply loop: unique 8: skip row sum
      "ldr w20, [%x[args_ptr], %[offsetof_num_strings]]\n"
      "add x26, x26, #0x1\n"
      "cmp x26, x20\n"
      "bne 46b\n"
      "ldr x20, [%x[args_ptr], %[offsetof_output_offset]]\n"
      "uzp1 z0.d, z16.d, z20.d\n"
      "add x23, x27, x20\n"
      "add x22, x23, x20\n"
      "uzp2 z16.d, z16.d, z20.d\n"
      "uzp1 z20.d, z17.d, z21.d\n"
      "add x21, x22, x20\n"
      "uzp2 z17.d, z17.d, z21.d\n"
      "uzp1 z21.d, z18.d, z22.d\n"
      "uzp2 z18.d, z18.d, z22.d\n"
      "uzp1 z22.d, z19.d, z23.d\n"
      "uzp2 z19.d, z19.d, z23.d\n"
      "uzp1 z23.d, z24.d, z28.d\n"
      "uzp2 z24.d, z24.d, z28.d\n"
      "uzp1 z28.d, z25.d, z29.d\n"
      "uzp2 z25.d, z25.d, z29.d\n"
      "uzp1 z29.d, z26.d, z30.d\n"
      "uzp2 z26.d, z26.d, z30.d\n"
      "uzp1 z30.d, z27.d, z31.d\n"
      "uzp2 z27.d, z27.d, z31.d\n"
      "mov z31.d, z0.d\n"
      "tbnz %x[flags], #31, 54f\n"
      "add x20, %x[qp], %[b_offset]\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      ".inst 0x4491a96b  // addp z11.s, p2/m, z11.s, z11.s\n"
      ".inst 0x4491a9ad  // addp z13.s, p2/m, z13.s, z13.s\n"
      "neg z0.s, p2/M, z0.s\n"
      "mov z12.s, z11.s[3]\n"
      "mov z11.s, z11.s[0]\n"
      "mul z11.s, p2/M, z11.s, z0.s\n"
      "mov z14.s, z13.s[3]\n"
      "mov z13.s, z13.s[0]\n"
      "mul z12.s, p2/M, z12.s, z0.s\n"
      "mul z13.s, p2/M, z13.s, z0.s\n"
      "mul z14.s, p2/M, z14.s, z0.s\n"
      "54:"  // Height 4: skip row sum fixup
      "add z31.s, z31.s, z11.s\n"
      "add z20.s, z20.s, z11.s\n"
      "ld1w { z4.s }, p2/Z, [x10]\n"
      "ld1w { z0.s }, p2/Z, [x10, #1, MUL VL]\n"
      "add z21.s, z21.s, z11.s\n"
      "add z22.s, z22.s, z11.s\n"
      "ld1w { z3.s }, p2/Z, [x10, #2, MUL VL]\n"
      "ld1w { z2.s }, p2/Z, [x10, #3, MUL VL]\n"
      "add z16.s, z16.s, z12.s\n"
      "add z17.s, z17.s, z12.s\n"
      "add x20, %x[qp], %[per_layer_mul]\n"
      "orr %x[flags], %x[flags], #0x80000000\n"
      "add z18.s, z18.s, z12.s\n"
      "add z19.s, z19.s, z12.s\n"
      "ld1rw { z1.s }, p2/Z, [x20]\n"
      "add x20, %x[qp], %[per_layer_right_shift]\n"
      "add z23.s, z23.s, z13.s\n"
      "add z28.s, z28.s, z13.s\n"
      "addvl x10, x10, #4\n"
      "add z29.s, z29.s, z13.s\n"
      "add z30.s, z30.s, z13.s\n"
      "add z24.s, z24.s, z14.s\n"
      "add z25.s, z25.s, z14.s\n"
      "add z26.s, z26.s, z14.s\n"
      "add z27.s, z27.s, z14.s\n"
      "add z31.s, z31.s, z4.s\n"
      "add z20.s, z20.s, z0.s\n"
      "add z21.s, z21.s, z3.s\n"
      "add z22.s, z22.s, z2.s\n"
      "add z16.s, z16.s, z4.s\n"
      "add z17.s, z17.s, z0.s\n"
      "add z18.s, z18.s, z3.s\n"
      "add z19.s, z19.s, z2.s\n"
      "add z23.s, z23.s, z4.s\n"
      "add z28.s, z28.s, z0.s\n"
      "add z29.s, z29.s, z3.s\n"
      "add z30.s, z30.s, z2.s\n"
      "add z24.s, z24.s, z4.s\n"
      "add z25.s, z25.s, z0.s\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      "add z26.s, z26.s, z3.s\n"
      "add z27.s, z27.s, z2.s\n"
      ".inst 0x04a177ff  // sqrdmulh z31.s, z31.s, z1.s\n"
      ".inst 0x04a17694  // sqrdmulh z20.s, z20.s, z1.s\n"
      ".inst 0x04a176b5  // sqrdmulh z21.s, z21.s, z1.s\n"
      ".inst 0x04a176d6  // sqrdmulh z22.s, z22.s, z1.s\n"
      ".inst 0x04a17610  // sqrdmulh z16.s, z16.s, z1.s\n"
      ".inst 0x04a17631  // sqrdmulh z17.s, z17.s, z1.s\n"
      ".inst 0x04a17652  // sqrdmulh z18.s, z18.s, z1.s\n"
      ".inst 0x04a17673  // sqrdmulh z19.s, z19.s, z1.s\n"
      ".inst 0x04a176f7  // sqrdmulh z23.s, z23.s, z1.s\n"
      ".inst 0x04a1779c  // sqrdmulh z28.s, z28.s, z1.s\n"
      ".inst 0x04a177bd  // sqrdmulh z29.s, z29.s, z1.s\n"
      ".inst 0x04a177de  // sqrdmulh z30.s, z30.s, z1.s\n"
      ".inst 0x04a17718  // sqrdmulh z24.s, z24.s, z1.s\n"
      ".inst 0x04a17739  // sqrdmulh z25.s, z25.s, z1.s\n"
      ".inst 0x04a1775a  // sqrdmulh z26.s, z26.s, z1.s\n"
      ".inst 0x04a1777b  // sqrdmulh z27.s, z27.s, z1.s\n"
      "tbz %x[flags], #5, 55f\n"
      "and z2.d, z31.d, z0.d\n"
      "and z1.d, z20.d, z0.d\n"
      "asr z2.s, z2.s, #0x1f\n"
      "asr z1.s, z1.s, #0x1f\n"
      "sqadd z31.s, z31.s, z2.s\n"
      "sqadd z20.s, z20.s, z1.s\n"
      "and z7.d, z21.d, z0.d\n"
      "and z6.d, z22.d, z0.d\n"
      "and z5.d, z16.d, z0.d\n"
      "and z4.d, z17.d, z0.d\n"
      "and z3.d, z18.d, z0.d\n"
      "and z2.d, z19.d, z0.d\n"
      "and z1.d, z23.d, z0.d\n"
      "asr z7.s, z7.s, #0x1f\n"
      "asr z6.s, z6.s, #0x1f\n"
      "asr z5.s, z5.s, #0x1f\n"
      "asr z4.s, z4.s, #0x1f\n"
      "asr z3.s, z3.s, #0x1f\n"
      "asr z2.s, z2.s, #0x1f\n"
      "asr z1.s, z1.s, #0x1f\n"
      "sqadd z21.s, z21.s, z7.s\n"
      "sqadd z22.s, z22.s, z6.s\n"
      "sqadd z16.s, z16.s, z5.s\n"
      "sqadd z17.s, z17.s, z4.s\n"
      "sqadd z18.s, z18.s, z3.s\n"
      "sqadd z19.s, z19.s, z2.s\n"
      "sqadd z23.s, z23.s, z1.s\n"
      "and z7.d, z28.d, z0.d\n"
      "and z6.d, z29.d, z0.d\n"
      "and z5.d, z30.d, z0.d\n"
      "and z4.d, z24.d, z0.d\n"
      "and z3.d, z25.d, z0.d\n"
      "and z2.d, z26.d, z0.d\n"
      "and z1.d, z27.d, z0.d\n"
      "asr z7.s, z7.s, #0x1f\n"
      "asr z6.s, z6.s, #0x1f\n"
      "asr z5.s, z5.s, #0x1f\n"
      "asr z4.s, z4.s, #0x1f\n"
      "asr z3.s, z3.s, #0x1f\n"
      "asr z2.s, z2.s, #0x1f\n"
      "asr z1.s, z1.s, #0x1f\n"
      "sqadd z28.s, z28.s, z7.s\n"
      "sqadd z29.s, z29.s, z6.s\n"
      "sqadd z30.s, z30.s, z5.s\n"
      "sqadd z24.s, z24.s, z4.s\n"
      "sqadd z25.s, z25.s, z3.s\n"
      "sqadd z26.s, z26.s, z2.s\n"
      "sqadd z27.s, z27.s, z1.s\n"
      "55:"  // Height 4: no shift correction
      "add x20, %x[qp], %[c_offset]\n"
      "ld1rw { z2.s }, p2/Z, [x20]\n"
      ".inst 0x4482881f  // srshl z31.s, p2/M, z31.s, z0.s\n"
      "add z31.s, z31.s, z2.s\n"
      ".inst 0x44828814  // srshl z20.s, p2/M, z20.s, z0.s\n"
      ".inst 0x44828815  // srshl z21.s, p2/M, z21.s, z0.s\n"
      "add z20.s, z20.s, z2.s\n"
      "add z21.s, z21.s, z2.s\n"
      ".inst 0x44828816  // srshl z22.s, p2/M, z22.s, z0.s\n"
      ".inst 0x44828810  // srshl z16.s, p2/M, z16.s, z0.s\n"
      "add z22.s, z22.s, z2.s\n"
      "add z16.s, z16.s, z2.s\n"
      ".inst 0x44828811  // srshl z17.s, p2/M, z17.s, z0.s\n"
      ".inst 0x44828812  // srshl z18.s, p2/M, z18.s, z0.s\n"
      "add z17.s, z17.s, z2.s\n"
      "add z18.s, z18.s, z2.s\n"
      ".inst 0x44828813  // srshl z19.s, p2/M, z19.s, z0.s\n"
      ".inst 0x44828817  // srshl z23.s, p2/M, z23.s, z0.s\n"
      "add z19.s, z19.s, z2.s\n"
      "add z23.s, z23.s, z2.s\n"
      ".inst 0x4482881c  // srshl z28.s, p2/M, z28.s, z0.s\n"
      ".inst 0x4482881d  // srshl z29.s, p2/M, z29.s, z0.s\n"
      "add z28.s, z28.s, z2.s\n"
      "add z29.s, z29.s, z2.s\n"
      ".inst 0x4482881e  // srshl z30.s, p2/M, z30.s, z0.s\n"
      ".inst 0x44828818  // srshl z24.s, p2/M, z24.s, z0.s\n"
      "add z30.s, z30.s, z2.s\n"
      "add z24.s, z24.s, z2.s\n"
      ".inst 0x44828819  // srshl z25.s, p2/M, z25.s, z0.s\n"
      ".inst 0x4482881a  // srshl z26.s, p2/M, z26.s, z0.s\n"
      "add z25.s, z25.s, z2.s\n"
      "add z26.s, z26.s, z2.s\n"
      ".inst 0x4482881b  // srshl z27.s, p2/M, z27.s, z0.s\n"
      "add x20, %x[qp], %[maxval]\n"
      "ld1rw { z1.s }, p2/Z, [x20]\n"
      "add z27.s, z27.s, z2.s\n"
      "add x20, %x[qp], %[minval]\n"
      "ld1rw { z0.s }, p2/Z, [x20]\n"
      "smin z31.s, p2/M, z31.s, z1.s\n"
      "smin z20.s, p2/M, z20.s, z1.s\n"
      "smin z21.s, p2/M, z21.s, z1.s\n"
      "smin z22.s, p2/M, z22.s, z1.s\n"
      "smin z16.s, p2/M, z16.s, z1.s\n"
      "smin z17.s, p2/M, z17.s, z1.s\n"
      "smin z18.s, p2/M, z18.s, z1.s\n"
      "smin z19.s, p2/M, z19.s, z1.s\n"
      "smin z23.s, p2/M, z23.s, z1.s\n"
      "smin z28.s, p2/M, z28.s, z1.s\n"
      "smin z29.s, p2/M, z29.s, z1.s\n"
      "smin z30.s, p2/M, z30.s, z1.s\n"
      "smin z24.s, p2/M, z24.s, z1.s\n"
      "smin z25.s, p2/M, z25.s, z1.s\n"
      "smin z26.s, p2/M, z26.s, z1.s\n"
      "smin z27.s, p2/M, z27.s, z1.s\n"
      "smax z31.s, p2/M, z31.s, z0.s\n"
      "smax z20.s, p2/M, z20.s, z0.s\n"
      "smax z21.s, p2/M, z21.s, z0.s\n"
      "uzp1 z31.h, z31.h, z20.h\n"
      "smax z22.s, p2/M, z22.s, z0.s\n"
      "smax z16.s, p2/M, z16.s, z0.s\n"
      "uzp1 z20.h, z21.h, z22.h\n"
      "uzp1 z31.b, z31.b, z20.b\n"
      "smax z17.s, p2/M, z17.s, z0.s\n"
      "smax z18.s, p2/M, z18.s, z0.s\n"
      "uzp1 z16.h, z16.h, z17.h\n"
      "st1b { z31.b }, p1, [x27]\n"
      "smax z19.s, p2/M, z19.s, z0.s\n"
      "smax z23.s, p2/M, z23.s, z0.s\n"
      "uzp1 z17.h, z18.h, z19.h\n"
      "uzp1 z16.b, z16.b, z17.b\n"
      "smax z28.s, p2/M, z28.s, z0.s\n"
      "smax z29.s, p2/M, z29.s, z0.s\n"
      "uzp1 z23.h, z23.h, z28.h\n"
      "st1b { z16.b }, p1, [x23]\n"
      "smax z30.s, p2/M, z30.s, z0.s\n"
      "smax z24.s, p2/M, z24.s, z0.s\n"
      "uzp1 z16.h, z29.h, z30.h\n"
      "uzp1 z23.b, z23.b, z16.b\n"
      "smax z25.s, p2/M, z25.s, z0.s\n"
      "smax z26.s, p2/M, z26.s, z0.s\n"
      "uzp1 z24.h, z24.h, z25.h\n"
      "st1b { z23.b }, p1, [x22]\n"
      "smax z27.s, p2/M, z27.s, z0.s\n"
      "uzp1 z16.h, z26.h, z27.h\n"
      "uzp1 z24.b, z24.b, z16.b\n"
      "st1b { z24.b }, p1, [x21]\n"
      "addvl x27, x27, #1\n"
      "56:"  // Height 4: Writeback done
      "decw x9, ALL, MUL #4\n"
      "cmp x9, XZR\n"
      "bgt 44b\n"
      "subs %x[M], %x[M], #0x4\n"
      "beq 58f\n"
      "ldr x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "tbz %x[flags], #3, 57f\n"
      "add x21, x21, #0x4\n"
      "str x21, [%x[args_ptr], %[offsetof_input_offset]]\n"
      "b 1b\n"
      "57:"  // Update direct input
      "mov x20, #0x4\n"
      "madd %x[input_ptr], x20, x21, %x[input_ptr]\n"
      "b 1b\n"
      "58:"  // 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", "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_SVE