// Copyright (c) Facebook, Inc. and its affiliates.
// All rights reserved.
//
// Copyright 2019 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/common.h>
#include <xnnpack/igemm.h>


void xnn_qu8_igemm_minmax_ukernel_4x8__neon(
    size_t mr,
    size_t nc,
    size_t kc,
    size_t ks,
    const uint8_t** restrict a,
    const void* restrict w,
    uint8_t* restrict c,
    size_t cm_stride,
    size_t cn_stride,
    size_t a_offset,
    const uint8_t* zero,
    const union xnn_qu8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
{
  assert(mr != 0);
  assert(mr <= 4);
  assert(nc != 0);
  assert(kc != 0);
  assert(ks != 0);
  assert(ks % (4 * sizeof(void*)) == 0);

  uint8_t* c0 = c;
  uint8_t* c1 = (uint8_t*) ((uintptr_t) c0 + cm_stride);
  if XNN_UNPREDICTABLE(mr < 2) {
    c1 = c0;
  }
  uint8_t* c2 = (uint8_t*) ((uintptr_t) c1 + cm_stride);
  if XNN_UNPREDICTABLE(mr <= 2) {
    c2 = c1;
  }
  uint8_t* c3 = (uint8_t*) ((uintptr_t) c2 + cm_stride);
  if (mr != 4) {
    c3 = c2;
  }

  const uint8x8_t vb_zero_point = vld1_dup_u8((const uint8_t*) &params->neon.kernel_zero_point);

  do {
    int32x4_t vacc0x0123 = vld1q_s32(w); w = (void*) ((uintptr_t) w + sizeof(int32x4_t));
    int32x4_t vacc0x4567 = vld1q_s32(w); w = (void*) ((uintptr_t) w + sizeof(int32x4_t));
    int32x4_t vacc1x0123 = vacc0x0123;
    int32x4_t vacc1x4567 = vacc0x4567;
    int32x4_t vacc2x0123 = vacc0x0123;
    int32x4_t vacc2x4567 = vacc0x4567;
    int32x4_t vacc3x0123 = vacc0x0123;
    int32x4_t vacc3x4567 = vacc0x4567;

    size_t p = ks;
    do {
      const uint8_t* restrict a0 = a[0];
      if XNN_UNPREDICTABLE(a0 != zero) {
        a0 = (const uint8_t*) ((uintptr_t) a0 + a_offset);
      }
      const uint8_t* restrict a1 = a[1];
      if XNN_UNPREDICTABLE(a1 != zero) {
        a1 = (const uint8_t*) ((uintptr_t) a1 + a_offset);
      }
      const uint8_t* restrict a2 = a[2];
      if XNN_UNPREDICTABLE(a2 != zero) {
        a2 = (const uint8_t*) ((uintptr_t) a2 + a_offset);
      }
      const uint8_t* restrict a3 = a[3];
      if XNN_UNPREDICTABLE(a3 != zero) {
        a3 = (const uint8_t*) ((uintptr_t) a3 + a_offset);
      }
      a += 4;

      size_t k = kc;
      while (k >= 8 * sizeof(uint8_t)) {
        const uint8x8_t va0 = vld1_u8(a0); a0 += 8;
        const uint8x8_t va1 = vld1_u8(a1); a1 += 8;
        const uint8x8_t va2 = vld1_u8(a2); a2 += 8;
        const uint8x8_t va3 = vld1_u8(a3); a3 += 8;
        const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
        const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
        const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
        const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));

        {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 0);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 0);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 0);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 0);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 0);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 0);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 0);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 0);
        }

        {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 1);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 1);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 1);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 1);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 1);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 1);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 1);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 1);
        }

        {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 2);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 2);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 2);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 2);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 2);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 2);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 2);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 2);
        }

        {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 3);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 3);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 3);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 3);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 3);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 3);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 3);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 3);
        }

        {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 0);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 0);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 0);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 0);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 0);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 0);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 0);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 0);
        }

        {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 1);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 1);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 1);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 1);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 1);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 1);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 1);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 1);
        }

        {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 2);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 2);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 2);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 2);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 2);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 2);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 2);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 2);
        }

        {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 3);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 3);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 3);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 3);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 3);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 3);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 3);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 3);
        }

        k -= 8 * sizeof(uint8_t);
      }
      if (k != 0) {
        const uint8x8_t va0 = vld1_u8(a0);
        const uint8x8_t va1 = vld1_u8(a1);
        const uint8x8_t va2 = vld1_u8(a2);
        const uint8x8_t va3 = vld1_u8(a3);
        const int16x8_t vxa0 = vreinterpretq_s16_u16(vmovl_u8(va0));
        const int16x8_t vxa1 = vreinterpretq_s16_u16(vmovl_u8(va1));
        const int16x8_t vxa2 = vreinterpretq_s16_u16(vmovl_u8(va2));
        const int16x8_t vxa3 = vreinterpretq_s16_u16(vmovl_u8(va3));

        {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 0);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 0);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 0);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 0);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 0);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 0);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 0);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 0);
        }

        if (k >= 2 * sizeof(uint8_t)) {
          const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
          const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

          vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 1);
          vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 1);
          vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 1);
          vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 1);
          vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 1);
          vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 1);
          vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 1);
          vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 1);

          if (k > 2 * sizeof(uint8_t)) {
            const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
            const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

            vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 2);
            vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 2);
            vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 2);
            vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 2);
            vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 2);
            vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 2);
            vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 2);
            vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 2);

            if (k >= 4 * sizeof(uint8_t)) {
              const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
              const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

              vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa0), 3);
              vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa0), 3);
              vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa1), 3);
              vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa1), 3);
              vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa2), 3);
              vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa2), 3);
              vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_low_s16(vxa3), 3);
              vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_low_s16(vxa3), 3);

              if (k > 4 * sizeof(uint8_t)) {
                const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
                const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

                vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 0);
                vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 0);
                vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 0);
                vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 0);
                vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 0);
                vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 0);
                vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 0);
                vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 0);

                if (k >= 6 * sizeof(uint8_t)) {
                  const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
                  const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

                  vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 1);
                  vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 1);
                  vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 1);
                  vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 1);
                  vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 1);
                  vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 1);
                  vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 1);
                  vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 1);

                  if (k > 6 * sizeof(uint8_t)) {
                    const uint8x8_t vb01234567 = vld1_u8(w); w = (void*) ((uintptr_t) w + sizeof(uint8x8_t));
                    const int16x8_t vxb01234567 = vreinterpretq_s16_u16(vsubl_u8(vb01234567, vb_zero_point));

                    vacc0x0123 = vmlal_lane_s16(vacc0x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa0), 2);
                    vacc0x4567 = vmlal_lane_s16(vacc0x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa0), 2);
                    vacc1x0123 = vmlal_lane_s16(vacc1x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa1), 2);
                    vacc1x4567 = vmlal_lane_s16(vacc1x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa1), 2);
                    vacc2x0123 = vmlal_lane_s16(vacc2x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa2), 2);
                    vacc2x4567 = vmlal_lane_s16(vacc2x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa2), 2);
                    vacc3x0123 = vmlal_lane_s16(vacc3x0123, vget_low_s16(vxb01234567), vget_high_s16(vxa3), 2);
                    vacc3x4567 = vmlal_lane_s16(vacc3x4567, vget_high_s16(vxb01234567), vget_high_s16(vxa3), 2);
                  }
                }
              }
            }
          }
        }
      }
      p -= 4 * sizeof(void*);
    } while (p != 0);

    const int32x4_t vmultiplier = vld1q_dup_s32(&params->neon.multiplier);
    vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
    vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
    vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
    vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
    vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
    vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
    vacc3x0123 = vqrdmulhq_s32(vacc3x0123, vmultiplier);
    vacc3x4567 = vqrdmulhq_s32(vacc3x4567, vmultiplier);

    const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
    const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
    vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
    vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
    vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
    vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
    vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
    vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
    vacc3x0123 = vsraq_n_s32(vacc3x0123, vbicq_s32(vacc3x0123, vzero_shift_mask), 31);
    vacc3x4567 = vsraq_n_s32(vacc3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31);

    vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
    vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
    vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
    vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
    vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
    vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
    vacc3x0123 = vrshlq_s32(vacc3x0123, vright_shift);
    vacc3x4567 = vrshlq_s32(vacc3x4567, vright_shift);

    const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
  #if XNN_ARCH_ARM64
    const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
    const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
    const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
    const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);

    uint8x16_t vout0x01234567_1x01234567 = vqmovun_high_s16(vqmovun_s16(vacc0x01234567), vacc1x01234567);
    uint8x16_t vout2x01234567_3x01234567 = vqmovun_high_s16(vqmovun_s16(vacc2x01234567), vacc3x01234567);
  #else
    const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
    const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
    const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
    const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);

    uint8x16_t vout0x01234567_1x01234567 = vcombine_u8(vqmovun_s16(vacc0x01234567), vqmovun_s16(vacc1x01234567));
    uint8x16_t vout2x01234567_3x01234567 = vcombine_u8(vqmovun_s16(vacc2x01234567), vqmovun_s16(vacc3x01234567));
  #endif
    const uint8x16_t voutput_min = vld1q_dup_u8(&params->neon.output_min);
    const uint8x16_t voutput_max = vld1q_dup_u8(&params->neon.output_max);

    vout0x01234567_1x01234567 = vmaxq_u8(vout0x01234567_1x01234567, voutput_min);
    vout2x01234567_3x01234567 = vmaxq_u8(vout2x01234567_3x01234567, voutput_min);
    vout0x01234567_1x01234567 = vminq_u8(vout0x01234567_1x01234567, voutput_max);
    vout2x01234567_3x01234567 = vminq_u8(vout2x01234567_3x01234567, voutput_max);

    if XNN_LIKELY(nc >= 8) {
      vst1_u8(c3, vget_high_u8(vout2x01234567_3x01234567)); c3 += cn_stride;
      vst1_u8(c2,  vget_low_u8(vout2x01234567_3x01234567)); c2 += cn_stride;
      vst1_u8(c1, vget_high_u8(vout0x01234567_1x01234567)); c1 += cn_stride;
      vst1_u8(c0,  vget_low_u8(vout0x01234567_1x01234567)); c0 += cn_stride;

      a = (const uint8_t**restrict) ((uintptr_t) a - ks);
      nc -= 8;
    } else {
      if (nc & 4) {
        vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 2); c3 += 4;
        vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_u8(vout2x01234567_3x01234567), 0); c2 += 4;
        vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 2); c1 += 4;
        vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_u8(vout0x01234567_1x01234567), 0); c0 += 4;
        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
      }
      if (nc & 2) {
        vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 4); c3 += 2;
        vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_u8(vout2x01234567_3x01234567), 0); c2 += 2;
        vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 4); c1 += 2;
        vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_u8(vout0x01234567_1x01234567), 0); c0 += 2;
        vout0x01234567_1x01234567 = vextq_u8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
        vout2x01234567_3x01234567 = vextq_u8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
      }
      if (nc & 1) {
        vst1q_lane_u8(c3, vout2x01234567_3x01234567, 8);
        vst1q_lane_u8(c2, vout2x01234567_3x01234567, 0);
        vst1q_lane_u8(c1, vout0x01234567_1x01234567, 8);
        vst1q_lane_u8(c0, vout0x01234567_1x01234567, 0);
      }

      nc = 0;
    }
  } while (nc != 0);
}