// Auto-generated file. Do not edit!
//   Template: src/f32-velu/avx-rr2-p6.c.in
//   Generator: tools/xngen
//
// Copyright 2020 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>


static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0};

void xnn_f32_velu_ukernel__avx_rr2_p6_x40(
    size_t n,
    const float* x,
    float* y,
    const union xnn_f32_elu_params params[restrict XNN_MIN_ELEMENTS(1)])
{
  assert(n % sizeof(float) == 0);

  const __m256 vprescale = _mm256_broadcast_ps((const __m128*) params->sse.prescale);
  const __m256 valpha = _mm256_broadcast_ps((const __m128*) params->sse.alpha);
  const __m256 vbeta = _mm256_broadcast_ps((const __m128*) params->sse.beta);

  const __m256 vsat_cutoff = _mm256_set1_ps(-0x1.154246p+4f);
  const __m256 vmagic_bias = _mm256_set1_ps(0x1.8000FEp23f);
  const __m256 vlog2e = _mm256_set1_ps(0x1.715476p+0f);
  const __m256 vminus_ln2_hi = _mm256_set1_ps(-0x1.62E440p-1f);
  const __m256 vminus_ln2_lo = _mm256_set1_ps(0x1.0105C6p-21f);
  const __m256 vc6 = _mm256_set1_ps(0x1.6b7338p-10f);
  const __m256 vc5 = _mm256_set1_ps(0x1.12278Ep-7f);
  const __m256 vc4 = _mm256_set1_ps(0x1.555716p-5f);
  const __m256 vc3 = _mm256_set1_ps(0x1.5554B0p-3f);
  const __m256 vc2 = _mm256_set1_ps(0x1.FFFFFEp-2f);
  const __m256 vone = _mm256_set1_ps(1.0f);

  for (; n >= 40 * sizeof(float); n -= 40 * sizeof(float)) {
    __m256 vx0 = _mm256_loadu_ps(x);
    __m256 vx1 = _mm256_loadu_ps(x + 8);
    __m256 vx2 = _mm256_loadu_ps(x + 16);
    __m256 vx3 = _mm256_loadu_ps(x + 24);
    __m256 vx4 = _mm256_loadu_ps(x + 32);
    x += 40;

    const __m256 vz0 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx0, vprescale));
    const __m256 vz1 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx1, vprescale));
    const __m256 vz2 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx2, vprescale));
    const __m256 vz3 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx3, vprescale));
    const __m256 vz4 = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx4, vprescale));

    __m256 vn0 = _mm256_add_ps(_mm256_mul_ps(vz0, vlog2e), vmagic_bias);
    __m256 vn1 = _mm256_add_ps(_mm256_mul_ps(vz1, vlog2e), vmagic_bias);
    __m256 vn2 = _mm256_add_ps(_mm256_mul_ps(vz2, vlog2e), vmagic_bias);
    __m256 vn3 = _mm256_add_ps(_mm256_mul_ps(vz3, vlog2e), vmagic_bias);
    __m256 vn4 = _mm256_add_ps(_mm256_mul_ps(vz4, vlog2e), vmagic_bias);

    const __m128 vs0_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn0)), 23));
    const __m128 vs0_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn0, 1)), 23));
    vn0 = _mm256_sub_ps(vn0, vmagic_bias);
    const __m128 vs1_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn1)), 23));
    const __m128 vs1_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn1, 1)), 23));
    vn1 = _mm256_sub_ps(vn1, vmagic_bias);
    const __m128 vs2_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn2)), 23));
    const __m128 vs2_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn2, 1)), 23));
    vn2 = _mm256_sub_ps(vn2, vmagic_bias);
    const __m128 vs3_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn3)), 23));
    const __m128 vs3_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn3, 1)), 23));
    vn3 = _mm256_sub_ps(vn3, vmagic_bias);
    const __m128 vs4_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn4)), 23));
    const __m128 vs4_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn4, 1)), 23));
    vn4 = _mm256_sub_ps(vn4, vmagic_bias);

    __m256 vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_hi), vz0);
    __m256 vs0 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs0_lo), vs0_hi, 1);
    __m256 vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_hi), vz1);
    __m256 vs1 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs1_lo), vs1_hi, 1);
    __m256 vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_hi), vz2);
    __m256 vs2 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs2_lo), vs2_hi, 1);
    __m256 vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_hi), vz3);
    __m256 vs3 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs3_lo), vs3_hi, 1);
    __m256 vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_hi), vz4);
    __m256 vs4 = _mm256_insertf128_ps(_mm256_castps128_ps256(vs4_lo), vs4_hi, 1);

    vt0 = _mm256_add_ps(_mm256_mul_ps(vn0, vminus_ln2_lo), vt0);
    vt1 = _mm256_add_ps(_mm256_mul_ps(vn1, vminus_ln2_lo), vt1);
    vt2 = _mm256_add_ps(_mm256_mul_ps(vn2, vminus_ln2_lo), vt2);
    vt3 = _mm256_add_ps(_mm256_mul_ps(vn3, vminus_ln2_lo), vt3);
    vt4 = _mm256_add_ps(_mm256_mul_ps(vn4, vminus_ln2_lo), vt4);

    __m256 vp0 = _mm256_add_ps(_mm256_mul_ps(vc6, vt0), vc5);
    __m256 vp1 = _mm256_add_ps(_mm256_mul_ps(vc6, vt1), vc5);
    __m256 vp2 = _mm256_add_ps(_mm256_mul_ps(vc6, vt2), vc5);
    __m256 vp3 = _mm256_add_ps(_mm256_mul_ps(vc6, vt3), vc5);
    __m256 vp4 = _mm256_add_ps(_mm256_mul_ps(vc6, vt4), vc5);

    vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc4);
    vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc4);
    vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc4);
    vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc4);
    vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc4);

    vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc3);
    vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc3);
    vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc3);
    vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc3);
    vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc3);

    vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vc2);
    vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vc2);
    vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vc2);
    vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vc2);
    vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vc2);

    vp0 = _mm256_mul_ps(vp0, vt0);
    vp1 = _mm256_mul_ps(vp1, vt1);
    vp2 = _mm256_mul_ps(vp2, vt2);
    vp3 = _mm256_mul_ps(vp3, vt3);
    vp4 = _mm256_mul_ps(vp4, vt4);

    vt0 = _mm256_mul_ps(vt0, vs0);
    vs0 = _mm256_sub_ps(vs0, vone);
    vt1 = _mm256_mul_ps(vt1, vs1);
    vs1 = _mm256_sub_ps(vs1, vone);
    vt2 = _mm256_mul_ps(vt2, vs2);
    vs2 = _mm256_sub_ps(vs2, vone);
    vt3 = _mm256_mul_ps(vt3, vs3);
    vs3 = _mm256_sub_ps(vs3, vone);
    vt4 = _mm256_mul_ps(vt4, vs4);
    vs4 = _mm256_sub_ps(vs4, vone);

    vp0 = _mm256_add_ps(_mm256_mul_ps(vp0, vt0), vt0);
    vp1 = _mm256_add_ps(_mm256_mul_ps(vp1, vt1), vt1);
    vp2 = _mm256_add_ps(_mm256_mul_ps(vp2, vt2), vt2);
    vp3 = _mm256_add_ps(_mm256_mul_ps(vp3, vt3), vt3);
    vp4 = _mm256_add_ps(_mm256_mul_ps(vp4, vt4), vt4);

    const __m256 ve0 = _mm256_mul_ps(_mm256_add_ps(vp0, vs0), valpha);
    vx0 = _mm256_mul_ps(vx0, vbeta);
    const __m256 ve1 = _mm256_mul_ps(_mm256_add_ps(vp1, vs1), valpha);
    vx1 = _mm256_mul_ps(vx1, vbeta);
    const __m256 ve2 = _mm256_mul_ps(_mm256_add_ps(vp2, vs2), valpha);
    vx2 = _mm256_mul_ps(vx2, vbeta);
    const __m256 ve3 = _mm256_mul_ps(_mm256_add_ps(vp3, vs3), valpha);
    vx3 = _mm256_mul_ps(vx3, vbeta);
    const __m256 ve4 = _mm256_mul_ps(_mm256_add_ps(vp4, vs4), valpha);
    vx4 = _mm256_mul_ps(vx4, vbeta);

    const __m256 vy0 = _mm256_blendv_ps(vx0, ve0, vx0);
    const __m256 vy1 = _mm256_blendv_ps(vx1, ve1, vx1);
    const __m256 vy2 = _mm256_blendv_ps(vx2, ve2, vx2);
    const __m256 vy3 = _mm256_blendv_ps(vx3, ve3, vx3);
    const __m256 vy4 = _mm256_blendv_ps(vx4, ve4, vx4);

    _mm256_storeu_ps(y, vy0);
    _mm256_storeu_ps(y + 8, vy1);
    _mm256_storeu_ps(y + 16, vy2);
    _mm256_storeu_ps(y + 24, vy3);
    _mm256_storeu_ps(y + 32, vy4);
    y += 40;
  }
  for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) {
    __m256 vx = _mm256_loadu_ps(x);
    x += 8;

    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));

    __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
    const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
    const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
    vn = _mm256_sub_ps(vn, vmagic_bias);

    __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
    __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
    vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);

    __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
    vp = _mm256_mul_ps(vp, vt);

    vt = _mm256_mul_ps(vt, vs);
    vs = _mm256_sub_ps(vs, vone);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);

    const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
    vx = _mm256_mul_ps(vx, vbeta);
    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);

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

    __m256 vx = _mm256_maskload_ps(x, vmask);

    const __m256 vz = _mm256_max_ps(vsat_cutoff, _mm256_mul_ps(vx, vprescale));

    __m256 vn = _mm256_add_ps(_mm256_mul_ps(vz, vlog2e), vmagic_bias);
    const __m128 vs_lo = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_castps256_ps128(vn)), 23));
    const __m128 vs_hi = _mm_castsi128_ps(_mm_slli_epi32(_mm_castps_si128(_mm256_extractf128_ps(vn, 1)), 23));
    vn = _mm256_sub_ps(vn, vmagic_bias);

    __m256 vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_hi), vz);
    __m256 vs = _mm256_insertf128_ps(_mm256_castps128_ps256(vs_lo), vs_hi, 1);
    vt = _mm256_add_ps(_mm256_mul_ps(vn, vminus_ln2_lo), vt);

    __m256 vp = _mm256_add_ps(_mm256_mul_ps(vc6, vt), vc5);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc4);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc3);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vc2);
    vp = _mm256_mul_ps(vp, vt);

    vt = _mm256_mul_ps(vt, vs);
    vs = _mm256_sub_ps(vs, vone);
    vp = _mm256_add_ps(_mm256_mul_ps(vp, vt), vt);

    const __m256 ve = _mm256_mul_ps(_mm256_add_ps(vp, vs), valpha);
    vx = _mm256_mul_ps(vx, vbeta);
    const __m256 vy = _mm256_blendv_ps(vx, ve, vx);

    // _mm256_maskstore_ps(y, vmask, vf) could be used here, but triggers msan failures (probably an msan bug).
    __m128 vy_lo = _mm256_castps256_ps128(vy);
    if (n & (4 * sizeof(float))) {
      _mm_storeu_ps(y, vy_lo);
      vy_lo = _mm256_extractf128_ps(vy, 1);
      y += 4;
    }
    if (n & (2 * sizeof(float))) {
      _mm_storel_pi((__m64*) y, vy_lo);
      vy_lo = _mm_movehl_ps(vy_lo, vy_lo);
      y += 2;
    }
    if (n & (1 * sizeof(float))) {
      _mm_store_ss(y, vy_lo);
    }
  }
}