// Auto-generated file. Do not edit!
//   Template: src/f32-gemm/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 <immintrin.h>

#include <xnnpack/gemm.h>


void xnn_f32_gemm_ukernel_4x8__avx_broadcast(
    size_t mr,
    size_t nc,
    size_t kc,
    const float*restrict a,
    size_t a_stride,
    const float*restrict w,
    float*restrict c,
    size_t cm_stride,
    size_t cn_stride,
    const union xnn_f32_minmax_params params[restrict static 1])
{
  assert(mr != 0);
  assert(mr <= 4);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  const float* a0 = a;
  float* c0 = c;
  const float* a1 = (const float*) ((uintptr_t) a0 + a_stride);
  float* c1 = (float*) ((uintptr_t) c0 + cm_stride);
  if XNN_UNPREDICTABLE(mr < 2) {
    a1 = a0;
    c1 = c0;
  }
  const float* a2 = (const float*) ((uintptr_t) a1 + a_stride);
  float* c2 = (float*) ((uintptr_t) c1 + cm_stride);
  if XNN_UNPREDICTABLE(mr <= 2) {
    a2 = a1;
    c2 = c1;
  }
  const float* a3 = (const float*) ((uintptr_t) a2 + a_stride);
  float* c3 = (float*) ((uintptr_t) c2 + cm_stride);
  if XNN_UNPREDICTABLE(mr != 4) {
    a3 = a2;
    c3 = c2;
  }

  do {
    __m256 vacc0x01234567 = _mm256_load_ps(w + 0);
    __m256 vacc1x01234567 = vacc0x01234567;
    __m256 vacc2x01234567 = vacc0x01234567;
    __m256 vacc3x01234567 = vacc0x01234567;
    w += 8;

    size_t k = kc;
    do {
      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 vb01234567 = _mm256_load_ps(w);
      w += 8;

      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));

      k -= sizeof(float);
    } while (k != 0);

    const __m256 vmax = _mm256_broadcast_ps((const __m128*) params->sse.max);
    vacc0x01234567 = _mm256_min_ps(vacc0x01234567, vmax);
    vacc1x01234567 = _mm256_min_ps(vacc1x01234567, vmax);
    vacc2x01234567 = _mm256_min_ps(vacc2x01234567, vmax);
    vacc3x01234567 = _mm256_min_ps(vacc3x01234567, vmax);

    const __m256 vmin = _mm256_broadcast_ps((const __m128*) params->sse.min);
    vacc0x01234567 = _mm256_max_ps(vacc0x01234567, vmin);
    vacc1x01234567 = _mm256_max_ps(vacc1x01234567, vmin);
    vacc2x01234567 = _mm256_max_ps(vacc2x01234567, vmin);
    vacc3x01234567 = _mm256_max_ps(vacc3x01234567, vmin);

    if XNN_LIKELY(nc >= 8) {
      _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);

      a3 = (const float*) ((uintptr_t) a3 - kc);
      a2 = (const float*) ((uintptr_t) a2 - kc);
      a1 = (const float*) ((uintptr_t) a1 - kc);
      a0 = (const float*) ((uintptr_t) a0 - kc);

      nc -= 8;
    } else {
      __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(c3, vacc3x0123);
        _mm_storeu_ps(c2, vacc2x0123);
        _mm_storeu_ps(c1, vacc1x0123);
        _mm_storeu_ps(c0, vacc0x0123);

        vacc3x0123 = _mm256_extractf128_ps(vacc3x01234567, 1);
        vacc2x0123 = _mm256_extractf128_ps(vacc2x01234567, 1);
        vacc1x0123 = _mm256_extractf128_ps(vacc1x01234567, 1);
        vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);

        c3 += 4;
        c2 += 4;
        c1 += 4;
        c0 += 4;
      }
      if (nc & 2) {
        _mm_storel_pi((__m64*) c3, vacc3x0123);
        _mm_storel_pi((__m64*) c2, vacc2x0123);
        _mm_storel_pi((__m64*) c1, vacc1x0123);
        _mm_storel_pi((__m64*) c0, vacc0x0123);

        vacc3x0123 = _mm_movehl_ps(vacc3x0123, vacc3x0123);
        vacc2x0123 = _mm_movehl_ps(vacc2x0123, vacc2x0123);
        vacc1x0123 = _mm_movehl_ps(vacc1x0123, vacc1x0123);
        vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);

        c3 += 2;
        c2 += 2;
        c1 += 2;
        c0 += 2;
      }
      if (nc & 1) {
        _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);
}