// Auto-generated file. Do not edit!
//   Template: src/f32-vsigmoid/avx2-rr1-p5.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/common.h>
#include <xnnpack/vunary.h>


void xnn_f32_vsigmoid_ukernel__avx2_rr1_p5_nr1fma_x8(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_sigmoid_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n % sizeof(float) == 0);

  const __m256 vsign_mask = _mm256_load_ps(params->avx2_rr1_p5.sign_mask);
  const __m256 vmagic_bias = _mm256_load_ps(params->avx2_rr1_p5.magic_bias);
  const __m256 vlog2e = _mm256_load_ps(params->avx2_rr1_p5.log2e);
  const __m256 vminus_ln2 = _mm256_load_ps(params->avx2_rr1_p5.minus_ln2);
  const __m256 vc5 = _mm256_load_ps(params->avx2_rr1_p5.c5);
  const __m256 vc4 = _mm256_load_ps(params->avx2_rr1_p5.c4);
  const __m256 vc3 = _mm256_load_ps(params->avx2_rr1_p5.c3);
  const __m256 vc2 = _mm256_load_ps(params->avx2_rr1_p5.c2);
  const __m256 vc1 = _mm256_load_ps(params->avx2_rr1_p5.c1);
  const __m256 vone = _mm256_load_ps(params->avx2_rr1_p5.one);
  const __m256 vdenorm_cutoff = _mm256_load_ps(params->avx2_rr1_p5.denorm_cutoff);

  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
    const __m256 vx = _mm256_loadu_ps(x);
    x += 8;

    const __m256 vz = _mm256_or_ps(vx, vsign_mask);

    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
    vn = _mm256_sub_ps(vn, vmagic_bias);

    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);

    __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
    vp = _mm256_fmadd_ps(vp, vt, vc3);
    vp = _mm256_fmadd_ps(vp, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vc1);

    vt = _mm256_mul_ps(vt, vs);
    const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);

    const __m256 vd = _mm256_add_ps(ve, vone);
    __m256 vr = _mm256_rcp_ps(vd);
    vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
    __m256 vf = _mm256_mul_ps(ve, vr);

    vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
    vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);

    _mm256_storeu_ps(y, vf);
    y += 8;
  }
  if XNN_UNLIKELY(n != 0) {
    assert(n >= 1 * sizeof(float));
    assert(n <= 7 * sizeof(float));
    const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &params->avx2_rr1_p5.mask_table[7] - n));

    const __m256 vx = _mm256_maskload_ps(x, vmask);

    const __m256 vz = _mm256_or_ps(vx, vsign_mask);

    __m256 vn = _mm256_fmadd_ps(vz, vlog2e, vmagic_bias);
    const __m256 vs = _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_castps_si256(vn), 23));
    vn = _mm256_sub_ps(vn, vmagic_bias);

    __m256 vt = _mm256_fmadd_ps(vn, vminus_ln2, vz);

    __m256 vp = _mm256_fmadd_ps(vc5, vt, vc4);
    vp = _mm256_fmadd_ps(vp, vt, vc3);
    vp = _mm256_fmadd_ps(vp, vt, vc2);
    vp = _mm256_fmadd_ps(vp, vt, vc1);

    vt = _mm256_mul_ps(vt, vs);
    const __m256 ve = _mm256_fmadd_ps(vt, vp, vs);

    const __m256 vd = _mm256_add_ps(ve, vone);
    __m256 vr = _mm256_rcp_ps(vd);
    vr = _mm256_fmadd_ps(_mm256_fnmadd_ps(vr, vd, vone), vr, vr);
    __m256 vf = _mm256_mul_ps(ve, vr);

    vf = _mm256_andnot_ps(_mm256_cmp_ps(vz, vdenorm_cutoff, _CMP_LT_OS), vf);
    vf = _mm256_blendv_ps(_mm256_sub_ps(vone, vf), vf, vx);

    __m128 vf_lo = _mm256_castps256_ps128(vf);
    if (n & (4 * sizeof(float))) {
      _mm_storeu_ps(y, vf_lo);
      vf_lo = _mm256_extractf128_ps(vf, 1);
      y += 4;
    }
    if (n & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) y, vf_lo);
      vf_lo = _mm_movehl_ps(vf_lo, vf_lo);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      _mm_store_ss(y, vf_lo);
    }
  }
}