// 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 <immintrin.h>

#include <xnnpack/igemm.h>


void xnn_f32_igemm_ukernel_1x8__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 union xnn_f32_output_params params[restrict static 1])
{
  assert(mr != 0);
  assert(mr <= 1);
  assert(nc != 0);
  assert(kc != 0);
  assert(kc % sizeof(float) == 0);
  assert(ks != 0);
  assert(ks % (1 * sizeof(void*)) == 0);
  assert(a_offset % sizeof(float) == 0);
  assert(a != NULL);
  assert(w != NULL);
  assert(c != NULL);

  float* c0 = c;

  do {
    __m256 vacc0x01234567 = _mm256_load_ps(w);
    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);
      }
      a += 1;

      size_t k = kc;
      do {
        const __m256 vb01234567 = _mm256_load_ps(w);
        w += 8;

        const __m256 va0 = _mm256_broadcast_ss(a0);
        a0 += 1;

        vacc0x01234567 = _mm256_add_ps(vacc0x01234567, _mm256_mul_ps(va0, vb01234567));
        k -= sizeof(float);
      } while (k != 0);
      p -= 1 * sizeof(void*);
    } while (p != 0);

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

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

    if XNN_LIKELY(nc >= 8) {
      _mm256_storeu_ps(c0, vacc0x01234567);
      c0 = (float*) ((uintptr_t) c0 + cn_stride);

      a = (const float**restrict) ((uintptr_t) a - ks);
      nc -= 8;
    } else {
      __m128 vacc0x0123 = _mm256_castps256_ps128(vacc0x01234567);
      if (nc & 4) {
        _mm_storeu_ps(c0, vacc0x0123);

        vacc0x0123 = _mm256_extractf128_ps(vacc0x01234567, 1);

        c0 += 4;
      }
      if (nc & 2) {
        _mm_storel_pi((__m64*) c0, vacc0x0123);

        vacc0x0123 = _mm_movehl_ps(vacc0x0123, vacc0x0123);

        c0 += 2;
      }
      if (nc & 1) {
        _mm_store_ss(c0, vacc0x0123);
      }

      nc = 0;
    }
  } while (nc != 0);
}