xref: /aosp_15_r20/external/XNNPACK/src/qu8-dwconv/gen/up24x9-minmax-rndnu-neon-mul16.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// Auto-generated file. Do not edit!
//   Template: src/qs8-dwconv/unipass-neon-mul16.c.in
//   Generator: tools/xngen
//
// Copyright 2020 Google LLC
//
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.

#include <assert.h>

#include <arm_neon.h>

#include <xnnpack/dwconv.h>


void xnn_qu8_dwconv_minmax_rndnu_ukernel_up24x9__neon_mul16(
    size_t channels,
    size_t output_width,
    const uint8_t** input,
    const void* weights,
    uint8_t* output,
    size_t input_stride,
    size_t output_increment,
    size_t input_offset,
    const uint8_t* zero,
    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  assert(channels != 0);
  assert(output_width != 0);

  const uint8x8_t vkernel_zero_point = vld1_dup_u8(params->rndnu_neon.kernel_zero_point);
  const int32x4_t vright_pre_shift = vld1q_dup_s32(&params->rndnu_neon.right_pre_shift);
  const int32x4_t vmultiplier = vld1q_dup_s32(&params->rndnu_neon.multiplier);
  const int32x4_t vright_post_shift = vld1q_dup_s32(&params->rndnu_neon.right_post_shift);
  const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->rndnu_neon.output_zero_point);
  const uint8x16_t voutput_min = vld1q_dup_u8(&params->rndnu_neon.output_min);
  const uint8x16_t voutput_max = vld1q_dup_u8(&params->rndnu_neon.output_max);
  do {
    const uint8_t* i0 = input[0];
    assert(i0 != NULL);
    if XNN_UNPREDICTABLE(i0 != zero) {
      i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
    }
    const uint8_t* i1 = input[1];
    assert(i1 != NULL);
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
    }
    const uint8_t* i2 = input[2];
    assert(i2 != NULL);
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
    }
    const uint8_t* i3 = input[3];
    assert(i3 != NULL);
    if XNN_UNPREDICTABLE(i3 != zero) {
      i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
    }
    const uint8_t* i4 = input[4];
    assert(i4 != NULL);
    if XNN_UNPREDICTABLE(i4 != zero) {
      i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
    }
    const uint8_t* i5 = input[5];
    assert(i5 != NULL);
    if XNN_UNPREDICTABLE(i5 != zero) {
      i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
    }
    const uint8_t* i6 = input[6];
    assert(i6 != NULL);
    if XNN_UNPREDICTABLE(i6 != zero) {
      i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
    }
    const uint8_t* i7 = input[7];
    assert(i7 != NULL);
    if XNN_UNPREDICTABLE(i7 != zero) {
      i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
    }
    const uint8_t* i8 = input[8];
    assert(i8 != NULL);
    if XNN_UNPREDICTABLE(i8 != zero) {
      i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
    }
    input = (const uint8_t**) ((uintptr_t) input + input_stride);

    size_t c = channels;
    const void* w = weights;
    for (; c >= 24; c -= 24) {
      int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
      int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
      int32x4_t vacc89AB = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
      int32x4_t vaccCDEF = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
      int32x4_t vaccGHIJ = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
      int32x4_t vaccKLMN = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);


      const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
      const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi0x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
      const int16x8_t vk0x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi0xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
      const int16x8_t vk0xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);

      vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
      vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
      vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi0x89ABCDEF), vget_low_s16(vk0x89ABCDEF));
      vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi0x89ABCDEF), vget_high_s16(vk0x89ABCDEF));
      vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi0xGHIJKLMN), vget_low_s16(vk0xGHIJKLMN));
      vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi0xGHIJKLMN), vget_high_s16(vk0xGHIJKLMN));

      const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
      const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi1x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
      const int16x8_t vk1x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi1xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
      const int16x8_t vk1xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);

      vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
      vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
      vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi1x89ABCDEF), vget_low_s16(vk1x89ABCDEF));
      vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi1x89ABCDEF), vget_high_s16(vk1x89ABCDEF));
      vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi1xGHIJKLMN), vget_low_s16(vk1xGHIJKLMN));
      vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi1xGHIJKLMN), vget_high_s16(vk1xGHIJKLMN));

      const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
      const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi2x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
      const int16x8_t vk2x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi2xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
      const int16x8_t vk2xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);

      vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
      vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
      vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi2x89ABCDEF), vget_low_s16(vk2x89ABCDEF));
      vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi2x89ABCDEF), vget_high_s16(vk2x89ABCDEF));
      vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi2xGHIJKLMN), vget_low_s16(vk2xGHIJKLMN));
      vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi2xGHIJKLMN), vget_high_s16(vk2xGHIJKLMN));

      const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
      const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi3x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
      const int16x8_t vk3x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi3xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
      const int16x8_t vk3xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);

      vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
      vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
      vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi3x89ABCDEF), vget_low_s16(vk3x89ABCDEF));
      vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi3x89ABCDEF), vget_high_s16(vk3x89ABCDEF));
      vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi3xGHIJKLMN), vget_low_s16(vk3xGHIJKLMN));
      vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi3xGHIJKLMN), vget_high_s16(vk3xGHIJKLMN));

      const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
      const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi4x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
      const int16x8_t vk4x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi4xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
      const int16x8_t vk4xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);

      vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
      vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
      vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi4x89ABCDEF), vget_low_s16(vk4x89ABCDEF));
      vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi4x89ABCDEF), vget_high_s16(vk4x89ABCDEF));
      vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi4xGHIJKLMN), vget_low_s16(vk4xGHIJKLMN));
      vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi4xGHIJKLMN), vget_high_s16(vk4xGHIJKLMN));

      const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
      const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi5x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
      const int16x8_t vk5x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi5xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
      const int16x8_t vk5xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);

      vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
      vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
      vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi5x89ABCDEF), vget_low_s16(vk5x89ABCDEF));
      vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi5x89ABCDEF), vget_high_s16(vk5x89ABCDEF));
      vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi5xGHIJKLMN), vget_low_s16(vk5xGHIJKLMN));
      vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi5xGHIJKLMN), vget_high_s16(vk5xGHIJKLMN));

      const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
      const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi6x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
      const int16x8_t vk6x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi6xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
      const int16x8_t vk6xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);

      vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
      vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
      vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi6x89ABCDEF), vget_low_s16(vk6x89ABCDEF));
      vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi6x89ABCDEF), vget_high_s16(vk6x89ABCDEF));
      vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi6xGHIJKLMN), vget_low_s16(vk6xGHIJKLMN));
      vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi6xGHIJKLMN), vget_high_s16(vk6xGHIJKLMN));

      const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
      const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi7x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
      const int16x8_t vk7x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi7xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
      const int16x8_t vk7xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);

      vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
      vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
      vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi7x89ABCDEF), vget_low_s16(vk7x89ABCDEF));
      vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi7x89ABCDEF), vget_high_s16(vk7x89ABCDEF));
      vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi7xGHIJKLMN), vget_low_s16(vk7xGHIJKLMN));
      vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi7xGHIJKLMN), vget_high_s16(vk7xGHIJKLMN));

      const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
      const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi8x89ABCDEF = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
      const int16x8_t vk8x89ABCDEF = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);
      const int16x8_t vi8xGHIJKLMN = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
      const int16x8_t vk8xGHIJKLMN = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(w), vkernel_zero_point)); w = (const void*) ((const uint8_t*) w + 8);

      vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
      vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));
      vacc89AB = vmlal_s16(vacc89AB, vget_low_s16(vi8x89ABCDEF), vget_low_s16(vk8x89ABCDEF));
      vaccCDEF = vmlal_s16(vaccCDEF, vget_high_s16(vi8x89ABCDEF), vget_high_s16(vk8x89ABCDEF));
      vaccGHIJ = vmlal_s16(vaccGHIJ, vget_low_s16(vi8xGHIJKLMN), vget_low_s16(vk8xGHIJKLMN));
      vaccKLMN = vmlal_s16(vaccKLMN, vget_high_s16(vi8xGHIJKLMN), vget_high_s16(vk8xGHIJKLMN));

      vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
      vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);
      vacc89AB = vqshlq_s32(vacc89AB, vright_pre_shift);
      vaccCDEF = vqshlq_s32(vaccCDEF, vright_pre_shift);
      vaccGHIJ = vqshlq_s32(vaccGHIJ, vright_pre_shift);
      vaccKLMN = vqshlq_s32(vaccKLMN, vright_pre_shift);

      vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
      vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);
      vacc89AB = vqdmulhq_s32(vacc89AB, vmultiplier);
      vaccCDEF = vqdmulhq_s32(vaccCDEF, vmultiplier);
      vaccGHIJ = vqdmulhq_s32(vaccGHIJ, vmultiplier);
      vaccKLMN = vqdmulhq_s32(vaccKLMN, vmultiplier);

      vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
      vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);
      vacc89AB = vrshlq_s32(vacc89AB, vright_post_shift);
      vaccCDEF = vrshlq_s32(vaccCDEF, vright_post_shift);
      vaccGHIJ = vrshlq_s32(vaccGHIJ, vright_post_shift);
      vaccKLMN = vrshlq_s32(vaccKLMN, vright_post_shift);

#if XNN_ARCH_ARM64
      int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
      int16x8_t vacc89ABCDEF = vqmovn_high_s32(vqmovn_s32(vacc89AB), vaccCDEF);
      int16x8_t vaccGHIJKLMN = vqmovn_high_s32(vqmovn_s32(vaccGHIJ), vaccKLMN);

      vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
      vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
      vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);

      uint8x16_t vout0123456789ABCDEF = vqmovun_high_s16(vqmovun_s16(vacc01234567), vacc89ABCDEF);
      uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
#else  // !XNN_ARCH_ARM64
      int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
      int16x8_t vacc89ABCDEF = vcombine_s16(vqmovn_s32(vacc89AB), vqmovn_s32(vaccCDEF));
      int16x8_t vaccGHIJKLMN = vcombine_s16(vqmovn_s32(vaccGHIJ), vqmovn_s32(vaccKLMN));

      vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);
      vacc89ABCDEF = vqaddq_s16(vacc89ABCDEF, voutput_zero_point);
      vaccGHIJKLMN = vqaddq_s16(vaccGHIJKLMN, voutput_zero_point);

      uint8x16_t vout0123456789ABCDEF = vcombine_u8(vqmovun_s16(vacc01234567), vqmovun_s16(vacc89ABCDEF));
      uint8x8_t voutGHIJKLMN = vqmovun_s16(vaccGHIJKLMN);
#endif  // !XNN_ARCH_ARM64

      vout0123456789ABCDEF = vmaxq_u8(vout0123456789ABCDEF, voutput_min);
      voutGHIJKLMN = vmax_u8(voutGHIJKLMN, vget_low_u8(voutput_min));

      vout0123456789ABCDEF = vminq_u8(vout0123456789ABCDEF, voutput_max);
      voutGHIJKLMN = vmin_u8(voutGHIJKLMN, vget_low_u8(voutput_max));

      vst1q_u8(output, vout0123456789ABCDEF); output += 16;
      vst1_u8(output, voutGHIJKLMN); output += 8;
    }
    if XNN_UNLIKELY(c != 0) {
      const uint8_t* k = (const uint8_t*) ((const int32_t*) w + 24);
      do {
        int32x4_t vacc0123 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);
        int32x4_t vacc4567 = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);

        const int16x8_t vi0x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i0))); i0 += 8;
        const int16x8_t vk0x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8(k), vkernel_zero_point)); k += 8;

        vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi0x01234567), vget_low_s16(vk0x01234567));
        vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi0x01234567), vget_high_s16(vk0x01234567));
        const int16x8_t vi1x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i1))); i1 += 8;
        const int16x8_t vk1x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 16)), vkernel_zero_point));

        vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi1x01234567), vget_low_s16(vk1x01234567));
        vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi1x01234567), vget_high_s16(vk1x01234567));
        const int16x8_t vi2x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i2))); i2 += 8;
        const int16x8_t vk2x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 40)), vkernel_zero_point));

        vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi2x01234567), vget_low_s16(vk2x01234567));
        vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi2x01234567), vget_high_s16(vk2x01234567));
        const int16x8_t vi3x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i3))); i3 += 8;
        const int16x8_t vk3x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 64)), vkernel_zero_point));

        vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi3x01234567), vget_low_s16(vk3x01234567));
        vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi3x01234567), vget_high_s16(vk3x01234567));
        const int16x8_t vi4x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i4))); i4 += 8;
        const int16x8_t vk4x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 88)), vkernel_zero_point));

        vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi4x01234567), vget_low_s16(vk4x01234567));
        vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi4x01234567), vget_high_s16(vk4x01234567));
        const int16x8_t vi5x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i5))); i5 += 8;
        const int16x8_t vk5x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 112)), vkernel_zero_point));

        vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi5x01234567), vget_low_s16(vk5x01234567));
        vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi5x01234567), vget_high_s16(vk5x01234567));
        const int16x8_t vi6x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i6))); i6 += 8;
        const int16x8_t vk6x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 136)), vkernel_zero_point));

        vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi6x01234567), vget_low_s16(vk6x01234567));
        vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi6x01234567), vget_high_s16(vk6x01234567));
        const int16x8_t vi7x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i7))); i7 += 8;
        const int16x8_t vk7x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 160)), vkernel_zero_point));

        vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi7x01234567), vget_low_s16(vk7x01234567));
        vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi7x01234567), vget_high_s16(vk7x01234567));
        const int16x8_t vi8x01234567 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(i8))); i8 += 8;
        const int16x8_t vk8x01234567 = vreinterpretq_s16_u16(vsubl_u8(vld1_u8((const void*) (k + 184)), vkernel_zero_point));

        vacc0123 = vmlal_s16(vacc0123, vget_low_s16(vi8x01234567), vget_low_s16(vk8x01234567));
        vacc4567 = vmlal_s16(vacc4567, vget_high_s16(vi8x01234567), vget_high_s16(vk8x01234567));

        vacc0123 = vqshlq_s32(vacc0123, vright_pre_shift);
        vacc4567 = vqshlq_s32(vacc4567, vright_pre_shift);

        vacc0123 = vqdmulhq_s32(vacc0123, vmultiplier);
        vacc4567 = vqdmulhq_s32(vacc4567, vmultiplier);

        vacc0123 = vrshlq_s32(vacc0123, vright_post_shift);
        vacc4567 = vrshlq_s32(vacc4567, vright_post_shift);

#if XNN_ARCH_ARM64
        int16x8_t vacc01234567 = vqmovn_high_s32(vqmovn_s32(vacc0123), vacc4567);
#else
        int16x8_t vacc01234567 = vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567));
#endif
        vacc01234567 = vqaddq_s16(vacc01234567, voutput_zero_point);

        uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);
        vout01234567 = vmax_u8(vout01234567, vget_low_u8(voutput_min));
        vout01234567 = vmin_u8(vout01234567, vget_low_u8(voutput_max));

        if XNN_LIKELY(c >= 8) {
          vst1_u8(output, vout01234567); output += 8;
          c -= 8;
        } else {
          if (c & 4) {
            vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
            vout01234567 = vext_u8(vout01234567, vout01234567, 4);
          }
          if (c & 2) {
            vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
            vout01234567 = vext_u8(vout01234567, vout01234567, 2);
          }
          if (c & 1) {
            vst1_lane_u8(output, vout01234567, 0); output += 1;
          }
          c = 0;
        }
      } while (c != 0);
    }

    output = (uint8_t*) ((uintptr_t) output + output_increment);
  } while (--output_width != 0);
}