// 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_f32_avgpool_minmax_ukernel_9x__wasm_c1(
    size_t output_pixels,
    size_t kernel_elements,
    size_t channels,
    const float** input,
    size_t input_offset,
    const float* zero,
    float* output,
    size_t input_increment,
    size_t output_increment,
    const union xnn_f32_scaleminmax_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(output_pixels != 0);
  assert(kernel_elements != 0);
  assert(kernel_elements <= 9);
  assert(channels != 0);

  const float vscale = params->scalar.scale;
  const float vmin = params->scalar.min;
  const float vmax = params->scalar.max;

  do {
    const float* i0 = input[0];
    assert(i0 != NULL);
    const float* i1 = input[1];
    const float* i2 = input[2];
    const float* i3 = input[3];
    const float* i4 = input[4];
    const float* i5 = input[5];
    const float* i6 = input[6];
    const float* i7 = input[7];
    const float* i8 = input[8];
    input = (const float**) ((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 float*) ((uintptr_t) i0 + input_offset);
    }
    if XNN_UNPREDICTABLE(i1 != zero) {
      i1 = (const float*) ((uintptr_t) i1 + input_offset);
    }
    if XNN_UNPREDICTABLE(i2 != zero) {
      i2 = (const float*) ((uintptr_t) i2 + input_offset);
    }
    if XNN_UNPREDICTABLE(i3 != zero) {
      i3 = (const float*) ((uintptr_t) i3 + input_offset);
    }
    if XNN_UNPREDICTABLE(i4 != zero) {
      i4 = (const float*) ((uintptr_t) i4 + input_offset);
    }
    if XNN_UNPREDICTABLE(i5 != zero) {
      i5 = (const float*) ((uintptr_t) i5 + input_offset);
    }
    if XNN_UNPREDICTABLE(i6 != zero) {
      i6 = (const float*) ((uintptr_t) i6 + input_offset);
    }
    if XNN_UNPREDICTABLE(i7 != zero) {
      i7 = (const float*) ((uintptr_t) i7 + input_offset);
    }
    if XNN_UNPREDICTABLE(i8 != zero) {
      i8 = (const float*) ((uintptr_t) i8 + input_offset);
    }

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

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

      float vout = vsum * vscale;
      vout = __builtin_wasm_max_f32(vout, vmin);
      vout = __builtin_wasm_min_f32(vout, vmax);

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