xref: /aosp_15_r20/external/XNNPACK/src/qu8-avgpool/9x-minmax-scalar-c1.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)

// 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 <xnnpack/avgpool.h>
#include <xnnpack/math.h>


void xnn_qu8_avgpool_minmax_ukernel_9x__scalar_c1(
    size_t output_pixels,
    size_t kernel_elements,
    size_t channels,
    const uint8_t** input,
    size_t input_offset,
    const uint8_t* zero,
    uint8_t* output,
    size_t input_increment,
    size_t output_increment,
    const union xnn_qu8_avgpool_minmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(output_pixels != 0);
  assert(kernel_elements != 0);
  assert(kernel_elements <= 9);
  assert(channels != 0);

  const int32_t vbias = params->scalar.bias;
  const int32_t vmultiplier = params->scalar.multiplier;
  const int64_t vrounding = params->scalar.rounding;
  const uint32_t vshift = params->scalar.right_shift;
  const int32_t voutput_min = params->scalar.output_min_less_zero_point;
  const int32_t voutput_max = params->scalar.output_max_less_zero_point;
  const int32_t voutput_zero_point = params->scalar.output_zero_point;
  do {
    const uint8_t* i0 = input[0];
    assert(i0 != NULL);
    const uint8_t* i1 = input[1];
    const uint8_t* i2 = input[2];
    const uint8_t* i3 = input[3];
    const uint8_t* i4 = input[4];
    const uint8_t* i5 = input[5];
    const uint8_t* i6 = input[6];
    const uint8_t* i7 = input[7];
    const uint8_t* i8 = input[8];
    input = (const uint8_t**) ((uintptr_t) input + input_increment);
    if (kernel_elements < 2) {
      i1 = zero;
    }
    assert(i1 != NULL);
    if (kernel_elements <= 2) {
      i2 = zero;
    }
    assert(i2 != NULL);
    if (kernel_elements < 4) {
      i3 = zero;
    }
    assert(i3 != NULL);
    if (kernel_elements <= 4) {
      i4 = zero;
    }
    assert(i4 != NULL);
    if (kernel_elements < 6) {
      i5 = zero;
    }
    assert(i5 != NULL);
    if (kernel_elements <= 6) {
      i6 = zero;
    }
    assert(i6 != NULL);
    if (kernel_elements < 8) {
      i7 = zero;
    }
    assert(i7 != NULL);
    if (kernel_elements <= 8) {
      i8 = zero;
    }
    assert(i8 != NULL);
    if XNN_UNPREDICTABLE(i0 != zero) {
      i0 = (const uint8_t*) ((uintptr_t) i0 + input_offset);
    }
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const uint8_t*) ((uintptr_t) i1 + input_offset);
    }
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const uint8_t*) ((uintptr_t) i2 + input_offset);
    }
    if XNN_UNPREDICTABLE(i3 != zero) {
      i3 = (const uint8_t*) ((uintptr_t) i3 + input_offset);
    }
    if XNN_UNPREDICTABLE(i4 != zero) {
      i4 = (const uint8_t*) ((uintptr_t) i4 + input_offset);
    }
    if XNN_UNPREDICTABLE(i5 != zero) {
      i5 = (const uint8_t*) ((uintptr_t) i5 + input_offset);
    }
    if XNN_UNPREDICTABLE(i6 != zero) {
      i6 = (const uint8_t*) ((uintptr_t) i6 + input_offset);
    }
    if XNN_UNPREDICTABLE(i7 != zero) {
      i7 = (const uint8_t*) ((uintptr_t) i7 + input_offset);
    }
    if XNN_UNPREDICTABLE(i8 != zero) {
      i8 = (const uint8_t*) ((uintptr_t) i8 + input_offset);
    }

    size_t c = channels;
    do {
      const uint32_t vi0 = (uint32_t) *i0++;
      const uint32_t vi1 = (uint32_t) *i1++;
      const uint32_t vi2 = (uint32_t) *i2++;
      const uint32_t vi3 = (uint32_t) *i3++;
      const uint32_t vi4 = (uint32_t) *i4++;
      const uint32_t vi5 = (uint32_t) *i5++;
      const uint32_t vi6 = (uint32_t) *i6++;
      const uint32_t vi7 = (uint32_t) *i7++;
      const uint32_t vi8 = (uint32_t) *i8++;

      const uint32_t vsum01 = vi0 + vi1;
      const uint32_t vsum23 = vi2 + vi3;
      const uint32_t vsum45 = vi4 + vi5;
      const uint32_t vsum67 = vi6 + vi7;
      const uint32_t vsum018 = vsum01 + vi8;
      const uint32_t vsum2345 = vsum23 + vsum45;
      const uint32_t vsum01678 = vsum018 + vsum67;
      const uint32_t vsum = vsum2345 + vsum01678;

      const int32_t vacc = vbias + (int32_t) vsum;

      const int64_t vproduct = (int64_t) vacc * (int64_t) vmultiplier;
      const int64_t vadjusted_product = vproduct - (int64_t) (vacc < 0);
      int32_t vout = (int32_t) math_asr_s64(vadjusted_product + vrounding, vshift);
      vout = vout < voutput_min ? voutput_min : vout;
      vout = vout > voutput_max ? voutput_max : vout;
      vout += voutput_zero_point;

      *output++ = (uint8_t) vout;
    } while (--c != 0);
    output = (uint8_t*) ((uintptr_t) output + output_increment);
  } while (--output_pixels != 0);
}