// clang-format off

// Auto-generated file. Do not edit!

//   Template: src/f32-igemm/avx-broadcast.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/igemm.h"

void xnn_f32_igemm_minmax_ukernel_5x8__avx_broadcast(

    size_t mr,

    size_t nc,

    size_t kc,

    size_t ks,

    const float** restrict a,

    const float* restrict w,

    float* restrict c,

    size_t cm_stride,

    size_t cn_stride,

    size_t a_offset,

    const float* zero,

    const struct xnn_f32_minmax_params* restrict params)

{

  assert(mr != 0);

  assert(mr <= 5);

  assert(nc != 0);

  assert(kc != 0);

  assert(kc % sizeof(float) == 0);

  assert(ks != 0);

  assert(ks % (5 * sizeof(void*)) == 0);

  assert(a_offset % sizeof(float) == 0);

  assert(a != NULL);

  assert(w != NULL);

  assert(c != NULL);

  float* c0 = c;

  float* c1 = (float*) ((uintptr_t) c0 + cm_stride);

  if XNN_UNPREDICTABLE(mr < 2) {

    c1 = c0;

  }

  float* c2 = (float*) ((uintptr_t) c1 + cm_stride);

  if XNN_UNPREDICTABLE(mr <= 2) {

    c2 = c1;

  }

  float* c3 = (float*) ((uintptr_t) c2 + cm_stride);

  if XNN_UNPREDICTABLE(mr < 4) {

    c3 = c2;

  }

  float* c4 = (float*) ((uintptr_t) c3 + cm_stride);

  if XNN_UNPREDICTABLE(mr <= 4) {

    c4 = c3;

  }

  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 {

    __m256 vacc0x01234567 = _mm256_load_ps(w);

    __m256 vacc1x01234567 = vacc0x01234567;

    __m256 vacc2x01234567 = vacc0x01234567;

    __m256 vacc3x01234567 = vacc0x01234567;

    __m256 vacc4x01234567 = vacc0x01234567;

    w += 8;

    size_t p = ks;

    do {

      const float* restrict a0 = a[0];

      assert(a0 != NULL);

      if XNN_UNPREDICTABLE(a0 != zero) {

        a0 = (const float*) ((uintptr_t) a0 + a_offset);

      }

      const float* restrict a1 = a[1];

      assert(a1 != NULL);

      if XNN_UNPREDICTABLE(a1 != zero) {

        a1 = (const float*) ((uintptr_t) a1 + a_offset);

      }

      const float* restrict a2 = a[2];

      assert(a2 != NULL);

      if XNN_UNPREDICTABLE(a2 != zero) {

        a2 = (const float*) ((uintptr_t) a2 + a_offset);

      }

      const float* restrict a3 = a[3];

      assert(a3 != NULL);

      if XNN_UNPREDICTABLE(a3 != zero) {

        a3 = (const float*) ((uintptr_t) a3 + a_offset);

      }

      const float* restrict a4 = a[4];

      assert(a4 != NULL);

      if XNN_UNPREDICTABLE(a4 != zero) {

        a4 = (const float*) ((uintptr_t) a4 + a_offset);

      }

      a += 5;

      size_t k = kc;

      do {

        const __m256 vb01234567 = _mm256_load_ps(w);

        w += 8;

        const __m256 va0 = _mm256_broadcast_ss(a0);

        a0 += 1;

        const __m256 va1 = _mm256_broadcast_ss(a1);

        a1 += 1;

        const __m256 va2 = _mm256_broadcast_ss(a2);

        a2 += 1;

        const __m256 va3 = _mm256_broadcast_ss(a3);

        a3 += 1;

        const __m256 va4 = _mm256_broadcast_ss(a4);

        a4 += 1;

        vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));

        vacc1x01234567 = _mm256_add_ps(vacc1x01234567, _mm256_mul_ps(va1, vb01234567));

        vacc2x01234567 = _mm256_add_ps(vacc2x01234567, _mm256_mul_ps(va2, vb01234567));

        vacc3x01234567 = _mm256_add_ps(vacc3x01234567, _mm256_mul_ps(va3, vb01234567));

        vacc4x01234567 = _mm256_add_ps(vacc4x01234567, _mm256_mul_ps(va4, vb01234567));

        k -= sizeof(float);

      } while (k != 0);

      p -= 5 * sizeof(void*);

    } while (p != 0);

    vacc0x01234567 = _mm256_max_ps(vmin, vacc0x01234567);

    vacc1x01234567 = _mm256_max_ps(vmin, vacc1x01234567);

    vacc2x01234567 = _mm256_max_ps(vmin, vacc2x01234567);

    vacc3x01234567 = _mm256_max_ps(vmin, vacc3x01234567);

    vacc4x01234567 = _mm256_max_ps(vmin, vacc4x01234567);

    vacc0x01234567 = _mm256_min_ps(vmax, vacc0x01234567);

    vacc1x01234567 = _mm256_min_ps(vmax, vacc1x01234567);

    vacc2x01234567 = _mm256_min_ps(vmax, vacc2x01234567);

    vacc3x01234567 = _mm256_min_ps(vmax, vacc3x01234567);

    vacc4x01234567 = _mm256_min_ps(vmax, vacc4x01234567);

    if XNN_LIKELY(nc >= 8) {

      _mm256_storeu_ps(c4, vacc4x01234567);

      c4 = (float*) ((uintptr_t) c4 + cn_stride);

      _mm256_storeu_ps(c3, vacc3x01234567);

      c3 = (float*) ((uintptr_t) c3 + cn_stride);

      _mm256_storeu_ps(c2, vacc2x01234567);

      c2 = (float*) ((uintptr_t) c2 + cn_stride);

      _mm256_storeu_ps(c1, vacc1x01234567);

      c1 = (float*) ((uintptr_t) c1 + cn_stride);

      _mm256_storeu_ps(c0, vacc0x01234567);

      c0 = (float*) ((uintptr_t) c0 + cn_stride);

      a = (const float**restrict) ((uintptr_t) a - ks);

      nc -= 8;

    } else {

      __m128 vacc4x0123 = _mm256_castps256_ps128(vacc4x01234567);

      __m128 vacc3x0123 = _mm256_castps256_ps128(vacc3x01234567);

      __m128 vacc2x0123 = _mm256_castps256_ps128(vacc2x01234567);

      __m128 vacc1x0123 = _mm256_castps256_ps128(vacc1x01234567);

      __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);

      if (nc & 4) {

        _mm_storeu_ps(c4, vacc4x0123);

        _mm_storeu_ps(c3, vacc3x0123);

        _mm_storeu_ps(c2, vacc2x0123);

        _mm_storeu_ps(c1, vacc1x0123);

        _mm_storeu_ps(c0, vacc0x0123);

        vacc4x0123 = _mm256_extractf128_ps(vacc4x01234567, 1);

        vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);

        vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);

        vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);

        vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);

        c4 += 4;

        c3 += 4;

        c2 += 4;

        c1 += 4;

        c0 += 4;

      }

      if (nc & 2) {

        _mm_storel_pi((__m64*) c4, vacc4x0123);

        _mm_storel_pi((__m64*) c3, vacc3x0123);

        _mm_storel_pi((__m64*) c2, vacc2x0123);

        _mm_storel_pi((__m64*) c1, vacc1x0123);

        _mm_storel_pi((__m64*) c0, vacc0x0123);

        vacc4x0123 = _mm_movehl_ps(vacc4x0123, vacc4x0123);

        vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);

        vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);

        vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);

        vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);

        c4 += 2;

        c3 += 2;

        c2 += 2;

        c1 += 2;

        c0 += 2;

      }

      if (nc & 1) {

        _mm_store_ss(c4, vacc4x0123);

        _mm_store_ss(c3, vacc3x0123);

        _mm_store_ss(c2, vacc2x0123);

        _mm_store_ss(c1, vacc1x0123);

        _mm_store_ss(c0, vacc0x0123);

      }

      nc = 0;

    }

  } while (nc != 0);

}