// Auto-generated file. Do not edit!
//   Template: src/qs8-vaddc/neon.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/vadd.h>


void xnn_qu8_vaddc_minmax_ukernel__neon_ld64_x8(
    size_t n,
    const uint8_t* input_a,
    const uint8_t* input_b,
    uint8_t* output,
    const union xnn_qu8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
{
  const uint8x8_t va_zero_point = vld1_dup_u8(&params->neon.a_zero_point);
  const int32x4_t va_multiplier = vld1q_dup_s32(&params->neon.a_multiplier);
  const int32x4_t vright_shift = vld1q_dup_s32(&params->neon.right_shift);
  const int16x8_t voutput_zero_point = vld1q_dup_s16(&params->neon.output_zero_point);
  const uint8x8_t voutput_min = vld1_dup_u8(&params->neon.output_min);
  const uint8x8_t voutput_max = vld1_dup_u8(&params->neon.output_max);

  const int32_t vxb = (int32_t) *input_b - (int32_t) params->neon.b_zero_point;
  const int32_t vb = params->neon.b_multiplier;
  const int32x4_t vbias = vdupq_n_s32(vxb * vb);

  for (; n >= 8 * sizeof(uint8_t); n -= 8 * sizeof(uint8_t)) {
    const uint8x8_t va01234567 = vld1_u8(input_a); input_a += 8;

    const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));

    int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
    int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);

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

    const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);

    uint8x8_t vout01234567 = vqmovun_s16(vacc01234567);

    vout01234567 = vmax_u8(vout01234567, voutput_min);

    vout01234567 = vmin_u8(vout01234567, voutput_max);

    vst1_u8(output, vout01234567); output += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    {
      const uint8x8_t va01234567 = vld1_u8(input_a);

      const int16x8_t vxa01234567 = vreinterpretq_s16_u16(vsubl_u8(va01234567, va_zero_point));

      int32x4_t vacc0123 = vmlaq_s32(vbias, vmovl_s16(vget_low_s16(vxa01234567)), va_multiplier);
      int32x4_t vacc4567 = vmlaq_s32(vbias, vmovl_s16(vget_high_s16(vxa01234567)), va_multiplier);

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

      const int16x8_t vacc01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0123), vqmovn_s32(vacc4567)), voutput_zero_point);

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

      if (n & (4 * sizeof(uint8_t))) {
        vst1_lane_u32((void*) output, vreinterpret_u32_u8(vout01234567), 0); output += 4;
        vout01234567 = vext_u8(vout01234567, vout01234567, 4);
      }
      if (n & (2 * sizeof(uint8_t))) {
        vst1_lane_u16((void*) output, vreinterpret_u16_u8(vout01234567), 0); output += 2;
        vout01234567 = vext_u8(vout01234567, vout01234567, 2);
      }
      if (n & (1 * sizeof(uint8_t))) {
        vst1_lane_u8(output, vout01234567, 0);
      }
    }
  }
}