// clang-format off

// Auto-generated file. Do not edit!

//   Template: src/f32-dwconv/unipass-avx.c.in

//   Generator: tools/xngen

//

// 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 <stddef.h>

#include <stdint.h>

#include <immintrin.h>

#include "src/xnnpack/common.h"

#include "src/xnnpack/microparams.h"

#include "src/xnnpack/dwconv.h"

void xnn_f32_dwconv_minmax_ukernel_3p16c__fma3(

    size_t channels,

    size_t output_width,

    const float** input,

    const float* weights,

    float* output,

    intptr_t input_stride,

    size_t output_increment,

    size_t input_offset,

    size_t input_pixel_stride,

    const float* zero,

    const struct xnn_f32_minmax_params* restrict params)

{

  assert(channels != 0);

  assert(output_width != 0);

  static const int32_t mask_table[16] = {-1, -1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0, 0};

  const __m256 vmin = _mm256_set1_ps(params->scalar.min);

  const __m256 vmax = _mm256_set1_ps(params->scalar.max);

  XNN_FORCE_REALIZATION(vmin);

  XNN_FORCE_REALIZATION(vmax);

  do {

    const float* i0 = input[0];

    assert(i0 != NULL);

    if XNN_UNPREDICTABLE(i0 != zero) {

      i0 = (const float*) ((uintptr_t) i0 + input_offset);

    }

    const float* i1 = input[1];

    assert(i1 != NULL);

    if XNN_UNPREDICTABLE(i1 != zero) {

      i1 = (const float*) ((uintptr_t) i1 + input_offset);

    }

    const float* i2 = input[2];

    assert(i2 != NULL);

    if XNN_UNPREDICTABLE(i2 != zero) {

      i2 = (const float*) ((uintptr_t) i2 + input_offset);

    }

    input = (const float**) ((uintptr_t) input + input_stride);

    size_t c = channels;

    const float* w = weights;

    for (; c >= 16; c -= 16) {

      __m256 vacc01234567p0 = _mm256_load_ps(w);

      __m256 vacc89ABCDEFp0 = _mm256_load_ps(w + 8);

      const __m256 vi0x01234567 = _mm256_loadu_ps(i0);

      const __m256 vi0x89ABCDEF = _mm256_loadu_ps(i0 + 8);

      i0 += 16;

      const __m256 vk0x01234567 = _mm256_load_ps(w + 16);

      const __m256 vk0x89ABCDEF = _mm256_load_ps(w + 24);

      vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);

      vacc89ABCDEFp0 = _mm256_fmadd_ps(vi0x89ABCDEF, vk0x89ABCDEF, vacc89ABCDEFp0);

      const __m256 vi1x01234567 = _mm256_loadu_ps(i1);

      const __m256 vi1x89ABCDEF = _mm256_loadu_ps(i1 + 8);

      i1 += 16;

      const __m256 vk1x01234567 = _mm256_load_ps(w + 32);

      const __m256 vk1x89ABCDEF = _mm256_load_ps(w + 40);

      vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);

      vacc89ABCDEFp0 = _mm256_fmadd_ps(vi1x89ABCDEF, vk1x89ABCDEF, vacc89ABCDEFp0);

      const __m256 vi2x01234567 = _mm256_loadu_ps(i2);

      const __m256 vi2x89ABCDEF = _mm256_loadu_ps(i2 + 8);

      i2 += 16;

      const __m256 vk2x01234567 = _mm256_load_ps(w + 48);

      const __m256 vk2x89ABCDEF = _mm256_load_ps(w + 56);

      vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);

      vacc89ABCDEFp0 = _mm256_fmadd_ps(vi2x89ABCDEF, vk2x89ABCDEF, vacc89ABCDEFp0);

      w += 64;

      __m256 vacc01234567 = _mm256_max_ps(vmin, vacc01234567p0);

      __m256 vacc89ABCDEF = _mm256_max_ps(vmin, vacc89ABCDEFp0);

      vacc01234567 = _mm256_min_ps(vmax, vacc01234567);

      vacc89ABCDEF = _mm256_min_ps(vmax, vacc89ABCDEF);

      _mm256_storeu_ps(output, vacc01234567);

      _mm256_storeu_ps(output + 8, vacc89ABCDEF);

      output += 16;

    }

    for (; c >= 8; c -= 8) {

      __m256 vacc01234567p0 = _mm256_load_ps(w);

      const __m256 vi0x01234567 = _mm256_loadu_ps(i0);

      i0 += 8;

      const __m256 vk0x01234567 = _mm256_load_ps(w + 16);

      vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);

      const __m256 vi1x01234567 = _mm256_loadu_ps(i1);

      i1 += 8;

      const __m256 vk1x01234567 = _mm256_load_ps(w + 32);

      vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);

      const __m256 vi2x01234567 = _mm256_loadu_ps(i2);

      i2 += 8;

      const __m256 vk2x01234567 = _mm256_load_ps(w + 48);

      vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);

      w += 8;

      __m256 vacc01234567 = _mm256_max_ps(vmin, vacc01234567p0);

      vacc01234567 = _mm256_min_ps(vmax, vacc01234567);

      _mm256_storeu_ps(output, vacc01234567);

      output += 8;

    }

    if XNN_UNLIKELY(c != 0) {

      assert(c >= 1);

      assert(c <= 7);

      const __m256i vmask = _mm256_loadu_si256((const __m256i*) &mask_table[8 - c]);

      __m256 vacc01234567p0 = _mm256_load_ps(w);

      const __m256 vi0x01234567 = _mm256_maskload_ps(i0, vmask);

      const __m256 vk0x01234567 = _mm256_load_ps(w + 16);

      vacc01234567p0 = _mm256_fmadd_ps(vi0x01234567, vk0x01234567, vacc01234567p0);

      const __m256 vi1x01234567 = _mm256_maskload_ps(i1, vmask);

      const __m256 vk1x01234567 = _mm256_load_ps(w + 32);

      vacc01234567p0 = _mm256_fmadd_ps(vi1x01234567, vk1x01234567, vacc01234567p0);

      const __m256 vi2x01234567 = _mm256_maskload_ps(i2, vmask);

      const __m256 vk2x01234567 = _mm256_load_ps(w + 48);

      vacc01234567p0 = _mm256_fmadd_ps(vi2x01234567, vk2x01234567, vacc01234567p0);

      __m256 vacc01234567 = _mm256_max_ps(vmin, vacc01234567p0);

      vacc01234567 = _mm256_min_ps(vmax, vacc01234567);

      _mm256_maskstore_ps(output, vmask, vacc01234567);

      output += c;

    }

    input_offset += input_pixel_stride;

    output = (float*) ((uintptr_t) output + output_increment);

  } while (--output_width != 0);

}