Triangle* OctreeLeaf::Query(const Ray& ray, float& t) const
{
float tBox = std::numeric_limits<float>::min();
if (!Intersects(ray, bb, tBox) || tBox > t)
return nullptr;
const __m256 rayDirX = _mm256_set1_ps(ray.Direction.X);
const __m256 rayDirY = _mm256_set1_ps(ray.Direction.Y);
const __m256 rayDirZ = _mm256_set1_ps(ray.Direction.Z);
const __m256 rayPosX = _mm256_set1_ps(ray.Origin.X);
const __m256 rayPosY = _mm256_set1_ps(ray.Origin.Y);
const __m256 rayPosZ = _mm256_set1_ps(ray.Origin.Z);
union { float dists[MAXSIZE]; __m256 distances[MAXSIZE / NROFLANES]; };
for (int i = 0; i < count; i++)
{
// Vector3F e1 = triangle.Vertices[1].Position - triangle.Vertices[0].Position;
const __m256 e1X = edge1X8[i];
const __m256 e1Y = edge1Y8[i];
const __m256 e1Z = edge1Z8[i];
// Vector3F e2 = triangle.Vertices[2].Position - triangle.Vertices[0].Position;
const __m256 e2X = edge2X8[i];
const __m256 e2Y = edge2Y8[i];
const __m256 e2Z = edge2Z8[i];
// Vector3F p = ray.Direction.Cross(e2);
const __m256 pX = _mm256_sub_ps(_mm256_mul_ps(rayDirY, e2Z), _mm256_mul_ps(rayDirZ, e2Y));
const __m256 pY = _mm256_sub_ps(_mm256_mul_ps(rayDirZ, e2X), _mm256_mul_ps(rayDirX, e2Z));
const __m256 pZ = _mm256_sub_ps(_mm256_mul_ps(rayDirX, e2Y), _mm256_mul_ps(rayDirY, e2X));
// float det = e1.Dot(p);
const __m256 det = _mm256_add_ps(_mm256_mul_ps(e1X, pX), _mm256_add_ps(_mm256_mul_ps(e1Y, pY), _mm256_mul_ps(e1Z, pZ)));
// if (det > -EPSILON && det < EPSILON)
// return false;
__m256 mask = _mm256_or_ps(_mm256_cmp_ps(det, _mm256_set1_ps(-EPSILON), _CMP_LE_OS), _mm256_cmp_ps(det, _mm256_set1_ps(EPSILON), _CMP_GE_OS));
// float invDet = 1 / det;
const __m256 invDet = _mm256_div_ps(_mm256_set1_ps(1.0f), det);
// Vector3F r = ray.Origin - triangle.Vertices[0].Position;
const __m256 rX = _mm256_sub_ps(rayPosX, vert0X8[i]);
const __m256 rY = _mm256_sub_ps(rayPosY, vert0Y8[i]);
const __m256 rZ = _mm256_sub_ps(rayPosZ, vert0Z8[i]);
// float u = r.Dot(p) * invDet;
const __m256 u = _mm256_mul_ps(invDet, _mm256_add_ps(_mm256_mul_ps(rX, pX), _mm256_add_ps(_mm256_mul_ps(rY, pY), _mm256_mul_ps(rZ, pZ))));
// if (u < 0 || u > 1)
// return false;
mask = _mm256_and_ps(mask, _mm256_cmp_ps(u, _mm256_setzero_ps(), _CMP_GE_OS));
// Vector3F q = r.Cross(e1);
const __m256 qX = _mm256_sub_ps(_mm256_mul_ps(rY, e1Z), _mm256_mul_ps(rZ, e1Y));
const __m256 qY = _mm256_sub_ps(_mm256_mul_ps(rZ, e1X), _mm256_mul_ps(rX, e1Z));
const __m256 qZ = _mm256_sub_ps(_mm256_mul_ps(rX, e1Y), _mm256_mul_ps(rY, e1X));
// float v = ray.Direction.Dot(q) * invDet;
const __m256 v = _mm256_mul_ps(invDet, _mm256_add_ps(_mm256_mul_ps(rayDirX, qX), _mm256_add_ps(_mm256_mul_ps(rayDirY, qY), _mm256_mul_ps(rayDirZ, qZ))));
// if (v < 0 || u + v > 1)
// return false;
mask = _mm256_and_ps(mask, _mm256_and_ps(_mm256_cmp_ps(v, _mm256_setzero_ps(), _CMP_GE_OS), _mm256_cmp_ps(_mm256_add_ps(u, v), _mm256_set1_ps(1.0f), _CMP_LE_OS)));
// float tt = e2.Dot(q) * invDet;
const __m256 tt = _mm256_mul_ps(invDet, _mm256_add_ps(_mm256_mul_ps(e2X, qX), _mm256_add_ps(_mm256_mul_ps(e2Y, qY), _mm256_mul_ps(e2Z, qZ))));
// if (tt > EPSILON)
// {
// t = tt;
// return true;
// }
//
// return false;
distances[i] = _mm256_and_ps(tt, mask);
}
Triangle* triangle = nullptr;
for (int i = 0; i < count * NROFLANES; i++)
if (dists[i] < t && dists[i] > EPSILON)
{
t = dists[i];
triangle = triangles[i];
}
return triangle;
}
inline void newsincos_ps_dual(avx_m256_t x1, avx_m256_t x2, avx_m256_t *s1, avx_m256_t *s2,
avx_m256_t *c1, avx_m256_t *c2) {
avx_m256_t tempa = _ps_sign_mask;
avx_m256_t tempb = _ps_inv_sign_mask;
avx_m256_t sign_bit1 = _mm256_and_ps(x1, tempa);
avx_m256_t sign_bit2 = _mm256_and_ps(x2, tempa);
x1 = _mm256_and_ps(x1, tempb);
x2 = _mm256_and_ps(x2, tempb);
tempa = _ps_cephes_FOPI;
avx_m256_t y1 = _mm256_mul_ps(x1, tempa);
avx_m256_t y2 = _mm256_mul_ps(x2, tempa);
//avx_m256i_t emm21 = _mm256_cvttps_epi32(y1);
//avx_m256i_t emm22 = _mm256_cvttps_epi32(y2);
//emm21 = _mm256_add_epi32(emm21, _pi32_1);
//emm22 = _mm256_add_epi32(emm22, _pi32_1);
avx_m256i_t emm21 = _mm256_cvttps_epi32(_mm256_add_ps(y1, _ps_1));
avx_m256i_t emm22 = _mm256_cvttps_epi32(_mm256_add_ps(y2, _ps_1));
//emm21 = _mm256_and_si256(emm21, _pi32_inv1);
//emm22 = _mm256_and_si256(emm22, _pi32_inv1);
emm21 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm21), _mm256_castsi256_ps(_pi32_inv1)));
emm22 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm22), _mm256_castsi256_ps(_pi32_inv1)));
y1 = _mm256_cvtepi32_ps(emm21);
y2 = _mm256_cvtepi32_ps(emm22);
//avx_m256i_t tempia = _pi32_2;
//avx_m256i_t cos_emm21 = _mm256_sub_epi32(emm21, tempia);
//avx_m256i_t cos_emm22 = _mm256_sub_epi32(emm22, tempia);
avx_m256i_t cos_emm21 = _mm256_cvtps_epi32(_mm256_sub_ps(_mm256_cvtepi32_ps(emm21), _ps_2));
avx_m256i_t cos_emm22 = _mm256_cvtps_epi32(_mm256_sub_ps(_mm256_cvtepi32_ps(emm22), _ps_2));
//avx_m256i_t tempib = _pi32_4;
//avx_m256i_t emm01 = _mm256_and_si256(emm21, tempib);
//avx_m256i_t emm02 = _mm256_and_si256(emm22, tempib);
avx_m256i_t emm01 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm21),
_mm256_castsi256_ps(_pi32_4)));
avx_m256i_t emm02 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm22),
_mm256_castsi256_ps(_pi32_4)));
//avx_m256i_t cos_emm01 = _mm256_andnot_si256(cos_emm21, tempib);
//avx_m256i_t cos_emm02 = _mm256_andnot_si256(cos_emm22, tempib);
avx_m256i_t cos_emm01 = _mm256_castps_si256(_mm256_andnot_ps(_mm256_castsi256_ps(cos_emm21),
_mm256_castsi256_ps(_pi32_4)));
avx_m256i_t cos_emm02 = _mm256_castps_si256(_mm256_andnot_ps(_mm256_castsi256_ps(cos_emm22),
_mm256_castsi256_ps(_pi32_4)));
//emm01 = _mm256_slli_epi32(emm01, 29);
__m128i emm0hi1 = _mm256_extractf128_si256(emm01, 0);
__m128i emm0lo1 = _mm256_extractf128_si256(emm01, 1);
emm0hi1 = _mm_slli_epi32(emm0hi1, 29);
emm0lo1 = _mm_slli_epi32(emm0lo1, 29);
emm01 = _mm256_insertf128_si256(emm01, emm0hi1, 0);
emm01 = _mm256_insertf128_si256(emm01, emm0lo1, 1);
//emm02 = _mm256_slli_epi32(emm02, 29);
__m128i emm0hi2 = _mm256_extractf128_si256(emm02, 0);
__m128i emm0lo2 = _mm256_extractf128_si256(emm02, 1);
emm0hi2 = _mm_slli_epi32(emm0hi2, 29);
emm0lo2 = _mm_slli_epi32(emm0lo2, 29);
emm02 = _mm256_insertf128_si256(emm02, emm0hi1, 0);
emm02 = _mm256_insertf128_si256(emm02, emm0lo1, 1);
//cos_emm01 = _mm256_slli_epi32(cos_emm01, 29);
__m128i cos_emm0hi1 = _mm256_extractf128_si256(cos_emm01, 0);
__m128i cos_emm0lo1 = _mm256_extractf128_si256(cos_emm01, 1);
cos_emm0hi1 = _mm_slli_epi32(cos_emm0hi1, 29);
cos_emm0lo1 = _mm_slli_epi32(cos_emm0lo1, 29);
cos_emm01 = _mm256_insertf128_si256(cos_emm01, cos_emm0hi1, 0);
cos_emm01 = _mm256_insertf128_si256(cos_emm01, cos_emm0lo1, 1);
//cos_emm02 = _mm256_slli_epi32(cos_emm02, 29);
__m128i cos_emm0hi2 = _mm256_extractf128_si256(cos_emm02, 0);
__m128i cos_emm0lo2 = _mm256_extractf128_si256(cos_emm02, 1);
cos_emm0hi2 = _mm_slli_epi32(cos_emm0hi2, 29);
cos_emm0lo2 = _mm_slli_epi32(cos_emm0lo2, 29);
cos_emm02 = _mm256_insertf128_si256(cos_emm02, cos_emm0hi2, 0);
cos_emm02 = _mm256_insertf128_si256(cos_emm02, cos_emm0lo2, 1);
//tempia = _pi32_2;
//tempib = _mm256_setzero_si256();
//emm21 = _mm256_and_si256(emm21, tempia);
//emm22 = _mm256_and_si256(emm22, tempia);
emm21 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm21),
_mm256_castsi256_ps(_pi32_2)));
emm22 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm22),
_mm256_castsi256_ps(_pi32_2)));
//cos_emm21 = _mm256_and_si256(cos_emm21, tempia);
//cos_emm22 = _mm256_and_si256(cos_emm22, tempia);
cos_emm21 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(cos_emm21),
_mm256_castsi256_ps(_pi32_2)));
cos_emm22 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(cos_emm22),
_mm256_castsi256_ps(_pi32_2)));
//emm21 = _mm256_cmpeq_epi32(emm21, tempib);
//emm22 = _mm256_cmpeq_epi32(emm22, tempib);
emm21 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(emm21), _mm256_setzero_ps(), _CMP_EQ_UQ));
emm22 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(emm22), _mm256_setzero_ps(), _CMP_EQ_UQ));
//cos_emm21 = _mm256_cmpeq_epi32(cos_emm21, tempib);
//cos_emm22 = _mm256_cmpeq_epi32(cos_emm22, tempib);
cos_emm21 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(cos_emm21), _mm256_setzero_ps(), _CMP_EQ_UQ));
cos_emm22 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(cos_emm22), _mm256_setzero_ps(), _CMP_EQ_UQ));
avx_m256_t emm0f1 = _mm256_castsi256_ps(emm01);
avx_m256_t emm0f2 = _mm256_castsi256_ps(emm02);
avx_m256_t emm2f1 = _mm256_castsi256_ps(emm21);
avx_m256_t emm2f2 = _mm256_castsi256_ps(emm22);
avx_m256_t cos_emm0f1 = _mm256_castsi256_ps(cos_emm01);
avx_m256_t cos_emm0f2 = _mm256_castsi256_ps(cos_emm02);
avx_m256_t cos_emm2f1 = _mm256_castsi256_ps(cos_emm21);
avx_m256_t cos_emm2f2 = _mm256_castsi256_ps(cos_emm22);
sign_bit1 = _mm256_xor_ps(sign_bit1, emm0f1);
sign_bit2 = _mm256_xor_ps(sign_bit2, emm0f2);
tempa = _ps_minus_cephes_DP123;
tempb = _mm256_mul_ps(y2, tempa);
tempa = _mm256_mul_ps(y1, tempa);
x2 = _mm256_add_ps(x2, tempb);
x1 = _mm256_add_ps(x1, tempa);
avx_m256_t x21 = _mm256_mul_ps(x1, x1);
avx_m256_t x22 = _mm256_mul_ps(x2, x2);
avx_m256_t x31 = _mm256_mul_ps(x21, x1);
avx_m256_t x32 = _mm256_mul_ps(x22, x2);
avx_m256_t x41 = _mm256_mul_ps(x21, x21);
avx_m256_t x42 = _mm256_mul_ps(x22, x22);
tempa = _ps_coscof_p0;
tempb = _ps_sincof_p0;
y1 = _mm256_mul_ps(x21, tempa);
y2 = _mm256_mul_ps(x22, tempa);
avx_m256_t y21 = _mm256_mul_ps(x21, tempb);
avx_m256_t y22 = _mm256_mul_ps(x22, tempb);
tempa = _ps_coscof_p1;
tempb = _ps_sincof_p1;
y1 = _mm256_add_ps(y1, tempa);
y2 = _mm256_add_ps(y2, tempa);
y21 = _mm256_add_ps(y21, tempb);
y22 = _mm256_add_ps(y22, tempb);
y1 = _mm256_mul_ps(y1, x21);
y2 = _mm256_mul_ps(y2, x22);
y21 = _mm256_mul_ps(y21, x21);
y22 = _mm256_mul_ps(y22, x22);
tempa = _ps_coscof_p2;
tempb = _ps_sincof_p2;
y1 = _mm256_add_ps(y1, tempa);
y2 = _mm256_add_ps(y2, tempa);
y21 = _mm256_add_ps(y21, tempb);
y22 = _mm256_add_ps(y22, tempb);
y1 = _mm256_mul_ps(y1, x41);
y2 = _mm256_mul_ps(y2, x42);
y21 = _mm256_mul_ps(y21, x31);
y22 = _mm256_mul_ps(y22, x32);
tempa = _ps_0p5;
tempb = _ps_1;
avx_m256_t temp_21 = _mm256_mul_ps(x21, tempa);
avx_m256_t temp_22 = _mm256_mul_ps(x22, tempa);
y21 = _mm256_add_ps(y21, x1);
y22 = _mm256_add_ps(y22, x2);
temp_21 = _mm256_sub_ps(temp_21, tempb);
temp_22 = _mm256_sub_ps(temp_22, tempb);
y1 = _mm256_sub_ps(y1, temp_21);
y2 = _mm256_sub_ps(y2, temp_22);
avx_m256_t cos_y1 = y1;
avx_m256_t cos_y2 = y2;
avx_m256_t cos_y21 = y21;
avx_m256_t cos_y22 = y22;
y1 = _mm256_andnot_ps(emm2f1, y1);
y2 = _mm256_andnot_ps(emm2f2, y2);
cos_y1 = _mm256_andnot_ps(cos_emm2f1, cos_y1);
cos_y2 = _mm256_andnot_ps(cos_emm2f2, cos_y2);
y21 = _mm256_and_ps(emm2f1, y21);
y22 = _mm256_and_ps(emm2f2, y22);
cos_y21 = _mm256_and_ps(cos_emm2f1, cos_y21);
cos_y22 = _mm256_and_ps(cos_emm2f2, cos_y22);
y1 = _mm256_add_ps(y1, y21);
y2 = _mm256_add_ps(y2, y22);
cos_y1 = _mm256_add_ps(cos_y1, cos_y21);
cos_y2 = _mm256_add_ps(cos_y2, cos_y22);
*s1 = _mm256_xor_ps(y1, sign_bit1);
*s2 = _mm256_xor_ps(y2, sign_bit2);
*c1 = _mm256_xor_ps(cos_y1, cos_emm0f1);
*c2 = _mm256_xor_ps(cos_y2, cos_emm0f2);
} // newsincos_ps_dual()
inline avx_m256_t newexp_ps(avx_m256_t x) {
avx_m256_t one = _ps_1;
avx_m256_t zero = _ps_0;
x = _mm256_min_ps(x, _ps_exp_hi);
x = _mm256_max_ps(x, _ps_exp_lo);
avx_m256_t temp_2 = _mm256_mul_ps(x, _ps_cephes_LOG2EF);
temp_2 = _mm256_add_ps(temp_2, _ps_0p5);
avx_m256i_t emm0 = _mm256_cvttps_epi32(temp_2);
avx_m256_t temp_1 = _mm256_cvtepi32_ps(emm0);
avx_m256_t temp_3 = _mm256_sub_ps(temp_1, temp_2);
avx_m256_t mask = _mm256_cmp_ps(temp_3, zero, _CMP_GT_OQ);
mask = _mm256_and_ps(mask, one);
temp_2 = _mm256_sub_ps(temp_1, mask);
emm0 = _mm256_cvttps_epi32(temp_2);
temp_1 = _mm256_mul_ps(temp_2, _ps_cephes_exp_C12);
x = _mm256_sub_ps(x, temp_1);
avx_m256_t x2 = _mm256_mul_ps(x, x);
avx_m256_t x3 = _mm256_mul_ps(x2, x);
avx_m256_t x4 = _mm256_mul_ps(x2, x2);
temp_1 = _mm256_add_ps(x, one);
temp_2 = _mm256_mul_ps(x2, _ps_cephes_exp_p5);
temp_3 = _mm256_mul_ps(x3, _ps_cephes_exp_p4);
temp_1 = _mm256_add_ps(temp_1, temp_2);
temp_2 = _mm256_mul_ps(x3, _ps_cephes_exp_p0);
temp_1 = _mm256_add_ps(temp_1, temp_3);
avx_m256_t temp_4 = _mm256_mul_ps(x, _ps_cephes_exp_p2);
temp_3 = _mm256_mul_ps(x2, _ps_cephes_exp_p1);
emm0 = _mm256_castps_si256(_mm256_add_ps(_mm256_castsi256_ps(emm0), _mm256_castsi256_ps(_pi32_0x7f)));
temp_2 = _mm256_add_ps(temp_2, temp_3);
temp_3 = _mm256_add_ps(temp_3, temp_4);
//emm0 = _mm256_slli_epi32(emm0, 23);
// convert emm0 into two 128-bit integer vectors
// perform shift on both vectors
// combine both vectors into 256-bit emm0
__m128i emm0hi = _mm256_extractf128_si256(emm0, 0);
__m128i emm0lo = _mm256_extractf128_si256(emm0, 1);
emm0hi = _mm_slli_epi32(emm0hi, 23);
emm0lo = _mm_slli_epi32(emm0lo, 23);
emm0 = _mm256_insertf128_si256(emm0, emm0hi, 0);
emm0 = _mm256_insertf128_si256(emm0, emm0lo, 1);
avx_m256_t pow2n = _mm256_castsi256_ps(emm0);
temp_2 = _mm256_add_ps(temp_2, temp_3);
temp_2 = _mm256_mul_ps(temp_2, x4);
avx_m256_t y = _mm256_add_ps(temp_1, temp_2);
y = _mm256_mul_ps(y, pow2n);
return y;
} // newexp_ps()
inline void newexp_ps_dual(avx_m256_t x1, avx_m256_t x2, avx_m256_t* exp1, avx_m256_t* exp2) {
avx_m256_t one = _ps_1;
avx_m256_t zero = _ps_0;
x1 = _mm256_min_ps(x1, _ps_exp_hi);
x2 = _mm256_min_ps(x2, _ps_exp_hi);
x1 = _mm256_max_ps(x1, _ps_exp_lo);
x2 = _mm256_max_ps(x2, _ps_exp_lo);
avx_m256_t temp_21 = _mm256_mul_ps(x1, _ps_cephes_LOG2EF);
avx_m256_t temp_22 = _mm256_mul_ps(x2, _ps_cephes_LOG2EF);
temp_21 = _mm256_add_ps(temp_21, _ps_0p5);
temp_22 = _mm256_add_ps(temp_22, _ps_0p5);
avx_m256i_t emm01 = _mm256_cvttps_epi32(temp_21);
avx_m256i_t emm02 = _mm256_cvttps_epi32(temp_22);
avx_m256_t temp_11 = _mm256_cvtepi32_ps(emm01);
avx_m256_t temp_12 = _mm256_cvtepi32_ps(emm02);
avx_m256_t temp_31 = _mm256_sub_ps(temp_11, temp_21);
avx_m256_t temp_32 = _mm256_sub_ps(temp_12, temp_22);
avx_m256_t mask1 = _mm256_cmp_ps(temp_31, zero, _CMP_GT_OQ);
avx_m256_t mask2 = _mm256_cmp_ps(temp_32, zero, _CMP_GT_OQ);
mask1 = _mm256_and_ps(mask1, one);
mask2 = _mm256_and_ps(mask2, one);
temp_21 = _mm256_sub_ps(temp_11, mask1);
temp_22 = _mm256_sub_ps(temp_12, mask2);
emm01 = _mm256_cvttps_epi32(temp_21);
emm02 = _mm256_cvttps_epi32(temp_22);
temp_11 = _mm256_mul_ps(temp_21, _ps_cephes_exp_C12);
temp_12 = _mm256_mul_ps(temp_22, _ps_cephes_exp_C12);
x1 = _mm256_sub_ps(x1, temp_11);
x2 = _mm256_sub_ps(x2, temp_12);
avx_m256_t x21 = _mm256_mul_ps(x1, x1);
avx_m256_t x22 = _mm256_mul_ps(x2, x2);
avx_m256_t x31 = _mm256_mul_ps(x21, x1);
avx_m256_t x32 = _mm256_mul_ps(x22, x2);
avx_m256_t x41 = _mm256_mul_ps(x21, x21);
avx_m256_t x42 = _mm256_mul_ps(x22, x22);
temp_11 = _mm256_add_ps(x1, one);
temp_12 = _mm256_add_ps(x2, one);
temp_21 = _mm256_mul_ps(x21, _ps_cephes_exp_p5);
temp_22 = _mm256_mul_ps(x22, _ps_cephes_exp_p5);
temp_31 = _mm256_mul_ps(x31, _ps_cephes_exp_p4);
temp_32 = _mm256_mul_ps(x32, _ps_cephes_exp_p4);
temp_11 = _mm256_add_ps(temp_11, temp_21);
temp_12 = _mm256_add_ps(temp_12, temp_22);
temp_21 = _mm256_mul_ps(x31, _ps_cephes_exp_p0);
temp_22 = _mm256_mul_ps(x32, _ps_cephes_exp_p0);
temp_11 = _mm256_add_ps(temp_11, temp_31);
temp_12 = _mm256_add_ps(temp_12, temp_32);
avx_m256_t temp_41 = _mm256_mul_ps(x1, _ps_cephes_exp_p2);
avx_m256_t temp_42 = _mm256_mul_ps(x2, _ps_cephes_exp_p2);
temp_31 = _mm256_mul_ps(x21, _ps_cephes_exp_p1);
temp_32 = _mm256_mul_ps(x22, _ps_cephes_exp_p1);
//emm01 = _mm256_add_epi32(emm01, _pi32_0x7f);
//emm02 = _mm256_add_epi32(emm02, _pi32_0x7f);
emm01 = _mm256_castps_si256(_mm256_add_ps(_mm256_castsi256_ps(emm01), _mm256_castsi256_ps(_pi32_0x7f)));
emm02 = _mm256_castps_si256(_mm256_add_ps(_mm256_castsi256_ps(emm02), _mm256_castsi256_ps(_pi32_0x7f)));
temp_21 = _mm256_add_ps(temp_21, temp_31);
temp_22 = _mm256_add_ps(temp_22, temp_32);
temp_31 = _mm256_add_ps(temp_31, temp_41);
temp_32 = _mm256_add_ps(temp_32, temp_42);
//emm01 = _mm256_slli_epi32(emm01, 23);
__m128i emm0hi1 = _mm256_extractf128_si256(emm01, 0);
__m128i emm0lo1 = _mm256_extractf128_si256(emm01, 1);
emm0hi1 = _mm_slli_epi32(emm0hi1, 23);
emm0lo1 = _mm_slli_epi32(emm0lo1, 23);
emm01 = _mm256_insertf128_si256(emm01, emm0hi1, 0);
emm01 = _mm256_insertf128_si256(emm01, emm0lo1, 1);
//emm02 = _mm256_slli_epi32(emm02, 23);
__m128i emm0hi2 = _mm256_extractf128_si256(emm02, 0);
__m128i emm0lo2 = _mm256_extractf128_si256(emm02, 1);
emm0hi2 = _mm_slli_epi32(emm0hi2, 23);
emm0lo2 = _mm_slli_epi32(emm0lo2, 23);
emm02 = _mm256_insertf128_si256(emm02, emm0hi2, 0);
emm02 = _mm256_insertf128_si256(emm02, emm0lo2, 1);
avx_m256_t pow2n1 = _mm256_castsi256_ps(emm01);
avx_m256_t pow2n2 = _mm256_castsi256_ps(emm02);
temp_21 = _mm256_add_ps(temp_21, temp_31);
temp_22 = _mm256_add_ps(temp_22, temp_32);
temp_21 = _mm256_mul_ps(temp_21, x41);
temp_22 = _mm256_mul_ps(temp_22, x42);
avx_m256_t y1 = _mm256_add_ps(temp_11, temp_21);
avx_m256_t y2 = _mm256_add_ps(temp_12, temp_22);
*exp1 = _mm256_mul_ps(y1, pow2n1);
*exp2 = _mm256_mul_ps(y2, pow2n2);
} // newexp_ps_dual()
inline void newsincos_ps(avx_m256_t x, avx_m256_t *s, avx_m256_t *c) {
avx_m256_t sign_bit = _mm256_and_ps(x, _ps_sign_mask);
x = _mm256_and_ps(x, _ps_inv_sign_mask);
avx_m256_t y = _mm256_mul_ps(x, _ps_cephes_FOPI);
//avx_m256i_t emm2 = _mm256_cvttps_epi32(y);
//emm2 = _mm256_add_epi32(emm2, _pi32_1);
avx_m256i_t emm2 = _mm256_cvttps_epi32(_mm256_add_ps(y, _ps_1));
//emm2 = _mm256_and_si256(emm2, _pi32_inv1);
emm2 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm2), _mm256_castsi256_ps(_pi32_inv1)));
y = _mm256_cvtepi32_ps(emm2);
//avx_m256i_t cos_emm2 = _mm256_sub_epi32(emm2, _pi32_2);
avx_m256i_t cos_emm2 = _mm256_cvtps_epi32(_mm256_sub_ps(_mm256_cvtepi32_ps(emm2), _ps_2));
//avx_m256i_t emm0 = _mm256_and_si256(emm2, _pi32_4);
avx_m256i_t emm0 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm2),
_mm256_castsi256_ps(_pi32_4)));
//avx_m256i_t cos_emm0 = _mm256_andnot_si256(cos_emm2, _pi32_4);
avx_m256i_t cos_emm0 = _mm256_castps_si256(_mm256_andnot_ps(_mm256_castsi256_ps(cos_emm2),
_mm256_castsi256_ps(_pi32_4)));
//emm0 = _mm256_slli_epi32(emm0, 29);
__m128i emm0hi = _mm256_extractf128_si256(emm0, 0);
__m128i emm0lo = _mm256_extractf128_si256(emm0, 1);
emm0hi = _mm_slli_epi32(emm0hi, 29);
emm0lo = _mm_slli_epi32(emm0lo, 29);
emm0 = _mm256_insertf128_si256(emm0, emm0hi, 0);
emm0 = _mm256_insertf128_si256(emm0, emm0lo, 1);
//cos_emm0 = _mm256_slli_epi32(cos_emm0, 29);
__m128i cos_emm0hi = _mm256_extractf128_si256(cos_emm0, 0);
__m128i cos_emm0lo = _mm256_extractf128_si256(cos_emm0, 1);
cos_emm0hi = _mm_slli_epi32(cos_emm0hi, 29);
cos_emm0lo = _mm_slli_epi32(cos_emm0lo, 29);
cos_emm0 = _mm256_insertf128_si256(cos_emm0, cos_emm0hi, 0);
cos_emm0 = _mm256_insertf128_si256(cos_emm0, cos_emm0lo, 1);
//emm2 = _mm256_and_si256(emm2, _pi32_2);
emm2 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(emm2),
_mm256_castsi256_ps(_pi32_2)));
//cos_emm2 = _mm256_and_si256(cos_emm2, _pi32_2);
cos_emm2 = _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(cos_emm2),
_mm256_castsi256_ps(_pi32_2)));
//emm2 = _mm256_cmpeq_epi32(emm2, _mm256_setzero_si256());
emm2 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(emm2), _mm256_setzero_ps(), _CMP_EQ_UQ));
//cos_emm2 = _mm256_cmpeq_epi32(cos_emm2, _mm256_setzero_si256());
cos_emm2 = _mm256_castps_si256(_mm256_cmp_ps(_mm256_castsi256_ps(cos_emm2), _mm256_setzero_ps(), _CMP_EQ_UQ));
avx_m256_t emm0f = _mm256_castsi256_ps(emm0);
avx_m256_t emm2f = _mm256_castsi256_ps(emm2);
avx_m256_t cos_emm0f = _mm256_castsi256_ps(cos_emm0);
avx_m256_t cos_emm2f = _mm256_castsi256_ps(cos_emm2);
sign_bit = _mm256_xor_ps(sign_bit, emm0f);
avx_m256_t temp_2 = _ps_minus_cephes_DP123;
temp_2 = _mm256_mul_ps(y, temp_2);
x = _mm256_add_ps(x, temp_2);
avx_m256_t x2 = _mm256_mul_ps(x, x);
avx_m256_t x3 = _mm256_mul_ps(x2, x);
avx_m256_t x4 = _mm256_mul_ps(x2, x2);
y = _ps_coscof_p0;
avx_m256_t y2 = _ps_sincof_p0;
y = _mm256_mul_ps(y, x2);
y2 = _mm256_mul_ps(y2, x2);
y = _mm256_add_ps(y, _ps_coscof_p1);
y2 = _mm256_add_ps(y2, _ps_sincof_p1);
y = _mm256_mul_ps(y, x2);
y2 = _mm256_mul_ps(y2, x2);
y = _mm256_add_ps(y, _ps_coscof_p2);
y2 = _mm256_add_ps(y2, _ps_sincof_p2);
y = _mm256_mul_ps(y, x4);
y2 = _mm256_mul_ps(y2, x3);
temp_2 = _mm256_mul_ps(x2, _ps_0p5);
y2 = _mm256_add_ps(y2, x);
temp_2 = _mm256_sub_ps(temp_2, _ps_1);
y = _mm256_sub_ps(y, temp_2);
avx_m256_t cos_y = y;
avx_m256_t cos_y2 = y2;
y = _mm256_andnot_ps(emm2f, y);
cos_y = _mm256_andnot_ps(cos_emm2f, cos_y);
y2 = _mm256_and_ps(emm2f, y2);
cos_y2 = _mm256_and_ps(cos_emm2f, cos_y2);
y = _mm256_add_ps(y, y2);
cos_y = _mm256_add_ps(cos_y, cos_y2);
*s = _mm256_xor_ps(y, sign_bit);
*c = _mm256_xor_ps(cos_y, cos_emm0f);
} // newsincos_ps()
__m256 mm256_cos_ps(__m256 x) {
__m256 xmm1, xmm2 = _mm256_setzero_ps(), xmm3, y;
__m256i emm0, emm2;
/* take the absolute value */
x = _mm256_and_ps(x, *(__m256*)m256_ps_inv_sign_mask);
/* scale by 4/Pi */
y = _mm256_mul_ps(x, *(__m256*)m256_ps_cephes_FOPI);
/* store the integer part of y in mm0 */
emm2 = _mm256_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
emm2 = _mm256_add_epi32(emm2, *(__m256i*)m256_pi32_1);
emm2 = _mm256_and_si256(emm2, *(__m256i*)m256_pi32_inv1);
y = _mm256_cvtepi32_ps(emm2);
emm2 = _mm256_sub_epi32(emm2, *(__m256i*)m256_pi32_2);
/* get the swap sign flag */
emm0 = _mm256_andnot_si256(emm2, *(__m256i*)m256_pi32_4);
emm0 = _mm256_slli_epi32(emm0, 29);
/* get the polynom selection mask */
emm2 = _mm256_and_si256(emm2, *(__m256i*)m256_pi32_2);
emm2 = _mm256_cmpeq_epi32(emm2, _mm256_setzero_si256());
__m256 sign_bit = _mm256_castsi256_ps(emm0);
__m256 poly_mask = _mm256_castsi256_ps(emm2);
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(__m256*)m256_ps_minus_cephes_DP1;
xmm2 = *(__m256*)m256_ps_minus_cephes_DP2;
xmm3 = *(__m256*)m256_ps_minus_cephes_DP3;
xmm1 = _mm256_mul_ps(y, xmm1);
xmm2 = _mm256_mul_ps(y, xmm2);
xmm3 = _mm256_mul_ps(y, xmm3);
x = _mm256_add_ps(x, xmm1);
x = _mm256_add_ps(x, xmm2);
x = _mm256_add_ps(x, xmm3);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
y = *(__m256*)m256_ps_coscof_p0;
__m256 z = _mm256_mul_ps(x,x);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(__m256*)m256_ps_coscof_p1);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(__m256*)m256_ps_coscof_p2);
y = _mm256_mul_ps(y, z);
y = _mm256_mul_ps(y, z);
__m256 tmp = _mm256_mul_ps(z, *(__m256*)m256_ps_0p5);
y = _mm256_sub_ps(y, tmp);
y = _mm256_add_ps(y, *(__m256*)m256_ps_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
__m256 y2 = *(__m256*)m256_ps_sincof_p0;
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(__m256*)m256_ps_sincof_p1);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(__m256*)m256_ps_sincof_p2);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_mul_ps(y2, x);
y2 = _mm256_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
y2 = _mm256_and_ps(xmm3, y2); //, xmm3);
y = _mm256_andnot_ps(xmm3, y);
y = _mm256_add_ps(y,y2);
/* update the sign */
y = _mm256_xor_ps(y, sign_bit);
_mm256_zeroupper();
return y;
}
INLINE const avxi operator &( const avxi& a, const avxi& b ) { return _mm256_castps_si256(_mm256_and_ps(_mm256_castsi256_ps(a), _mm256_castsi256_ps(b))); }
void Decoder::ADMMDecoder_deg_6_7_2_3_6()
{
int maxIter = maxIteration;
float mu = 5.5f;
float tableau[12] = { 0.0f };
if ((mBlocklength == 576) && (mNChecks == 288))
{
mu = 3.37309f;//penalty
tableau[2] = 0.00001f;
tableau[3] = 2.00928f;
tableau[6] = 4.69438f;
}
else if((mBlocklength == 2304) && (mNChecks == 1152) )
{
mu = 3.81398683f;//penalty
tableau[2] = 0.29669288f;
tableau[3] = 0.46964023f;
tableau[6] = 3.19548154f;
}
else
{
mu = 5.5;//penalty
tableau[2] = 0.8f;
tableau[3] = 0.8f;
tableau[6] = 0.8f;
}
const float rho = 1.9f; //over relaxation parameter;
const float un_m_rho = 1.0 - rho;
const auto _rho = _mm256_set1_ps( rho );
const auto _un_m_rho = _mm256_set1_ps( un_m_rho );
float tableaX[12];
//
// ON PRECALCULE LES CONSTANTES
//
#pragma unroll
for (int i = 0; i < 7; i++)
{
tableaX[i] = tableau[ i ] / mu;
}
const auto t_mu = _mm256_set1_ps ( mu );
const auto t2_amu = _mm256_set1_ps ( tableau[ 2 ] / mu );
const auto t3_amu = _mm256_set1_ps ( tableau[ 3 ] / mu );
const auto t6_amu = _mm256_set1_ps ( tableau[ 6 ] / mu );
const auto t2_2amu = _mm256_set1_ps ( 2.0f * tableau[ 2 ] / mu );
const auto t3_2amu = _mm256_set1_ps ( 2.0f * tableau[ 3 ] / mu );
const auto t6_2amu = _mm256_set1_ps ( 2.0f * tableau[ 6 ] / mu );
const auto t2_deg = _mm256_set1_ps ( 2.0f );
const auto t3_deg = _mm256_set1_ps ( 3.0f );
const auto t6_deg = _mm256_set1_ps ( 6.0f );
const auto zero = _mm256_set1_ps ( 0.0f );
const auto un = _mm256_set1_ps ( 1.0f );
const __m256 a = _mm256_set1_ps ( 0.0f );
const __m256 b = _mm256_set1_ps ( 0.5f );
//////////////////////////////////////////////////////////////////////////////////////
#pragma unroll
for( int j = 0; j < _mPCheckMapSize; j+=8 )
{
_mm256_store_ps(&Lambda [j], a);
_mm256_store_ps(&zReplica[j], b);
_mm256_store_ps(&latestProjVector[j], b);
}
//////////////////////////////////////////////////////////////////////////////////////
for(int i = 0; i < maxIter; i++)
{
int ptr = 0;
mIteration = i + 1;
//
// MEASURE OF THE VN EXECUTION TIME
//
#ifdef PROFILE_ON
const auto start = timer();
#endif
//
// VN processing kernel
//
#pragma unroll
for (int j = 0; j < _mBlocklength; j++)
{
const int degVn = VariableDegree[j];
float M[8] __attribute__((aligned(64)));
if( degVn == 2 ){
#if 1
const int dVN = 2;
for(int qq = 0; qq < 8; qq++)
{
M[qq] = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
#pragma unroll
for(int k = 1; k < dVN; k++)
{
M[qq] += (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
}
}
const auto m = _mm256_loadu_ps( M );
const auto llr = _mm256_loadu_ps( &_LogLikelihoodRatio[j] );
const auto t1 = _mm256_sub_ps(m, _mm256_div_ps(llr, t_mu));
const auto xx = _mm256_div_ps(_mm256_sub_ps(t1, t2_amu), _mm256_sub_ps(t2_deg, t2_2amu));
const auto vMin = _mm256_max_ps(_mm256_min_ps(xx, un) , zero);
_mm256_storeu_ps(&OutputFromDecoder[j], vMin);
j += 7;
#else
const int degVN = 2;
float temp = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]);
#pragma unroll
for(int k = 1; k < degVN; k++)
temp += (zReplica[ t_row[ptr + k] ] + Lambda[ t_row[ptr + k] ]);
ptr += degVN;
const float _amu_ = tableaX[ degVN ];
const float _2_amu_ = _amu_+ _amu_;
const float llr = _LogLikelihoodRatio[j];
const float t = temp - llr / mu;
const float xx = (t - _amu_)/(degVn - _2_amu_);
const float vMax = std::min(xx, 1.0f);
const float vMin = std::max(vMax, 0.0f);
OutputFromDecoder[j] = vMin;
#endif
}else if( degVn == 3 ){
#if 1
const int dVN = 3;
for(int qq = 0; qq < 8; qq++)
{
M[qq] = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
#pragma unroll
for(int k = 1; k < dVN; k++)
{
M[qq] += (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
}
}
const auto m = _mm256_loadu_ps( M );
const auto llr = _mm256_loadu_ps( &_LogLikelihoodRatio[j] );
const auto t1 = _mm256_sub_ps(m, _mm256_div_ps(llr, t_mu));
const auto xx = _mm256_div_ps(_mm256_sub_ps(t1, t3_amu), _mm256_sub_ps(t3_deg, t3_2amu));
const auto vMin = _mm256_max_ps(_mm256_min_ps(xx, un) , zero);
_mm256_storeu_ps(&OutputFromDecoder[j], vMin);
j += 7;
#else
const int degVN = 3;
float temp = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]);
#pragma unroll
for(int k = 1; k < degVN; k++)
temp += (zReplica[ t_row[ptr + k] ] + Lambda[ t_row[ptr + k] ]);
ptr += degVN;
const float _amu_ = tableaX[ degVN ];
const float _2_amu_ = _amu_+ _amu_;
const float llr = _LogLikelihoodRatio[j];
const float t = temp - llr / mu;
const float xx = (t - _amu_)/(degVn - _2_amu_);
const float vMax = std::min(xx, 1.0f);
const float vMin = std::max(vMax, 0.0f);
OutputFromDecoder[j] = vMin;
#endif
}else if( degVn == 6 ){
#if 1
const int dVN = 6;
for(int qq = 0; qq < 8; qq++)
{
M[qq] = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
#pragma unroll
for(int k = 1; k < dVN; k++)
{
M[qq] += (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]); ptr += 1;
}
}
const auto m = _mm256_loadu_ps( M );
const auto llr = _mm256_loadu_ps( &_LogLikelihoodRatio[j] );
const auto t1 = _mm256_sub_ps(m, _mm256_div_ps(llr, t_mu));
const auto xx = _mm256_div_ps(_mm256_sub_ps(t1, t6_amu), _mm256_sub_ps(t6_deg, t6_2amu));
const auto vMin = _mm256_max_ps(_mm256_min_ps(xx, un) , zero);
_mm256_storeu_ps(&OutputFromDecoder[j], vMin);
j += 7;
#else
const int degVN = 6;
float temp = (zReplica[ t_row[ptr] ] + Lambda[ t_row[ptr] ]);
#pragma unroll
for(int k = 1; k < degVN; k++)
temp += (zReplica[ t_row[ptr + k] ] + Lambda[ t_row[ptr + k] ]);
ptr += degVN;
const float _amu_ = tableaX[ degVN ];
const float _2_amu_ = _amu_+ _amu_;
const float llr = _LogLikelihoodRatio[j];
const float t = temp - llr / mu;
const float xx = (t - _amu_)/(degVn - _2_amu_);
const float vMax = std::min(xx, 1.0f);
const float vMin = std::max(vMax, 0.0f);
OutputFromDecoder[j] = vMin;
#endif
}
}
//
// MEASURE OF THE VN EXECUTION TIME
//
#ifdef PROFILE_ON
t_vn += (timer() - start);
#endif
//
// CN processing kernel
//
int CumSumCheckDegree = 0; // cumulative position of currect edge in factor graph
int allVerified = 0;
float vector_before_proj[8] __attribute__((aligned(64)));
const auto zero = _mm256_set1_ps ( 0.0f );
const auto mask_6 = _mm256_set_epi32(0x00000000, 0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
const auto mask_7 = _mm256_set_epi32(0x00000000, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF);
const auto dot5 = _mm256_set1_ps( 0.5f );
//
// MEASURE OF THE CN EXECUTION TIME
//
#ifdef PROFILE_ON
const auto starT = timer();
#endif
const auto seuilProj = _mm256_set1_ps( 1e-5f );
for(int j = 0; j < _mNChecks; j++)
{
if( CheckDegree[j] == 6 ){
const int cDeg6 = 0x3F;
const auto offsets = _mm256_loadu_si256 ((const __m256i*)&t_col1 [CumSumCheckDegree]);
const auto xpred = _mm256_mask_i32gather_ps (zero, OutputFromDecoder, offsets, _mm256_castsi256_ps(mask_6), 4);
const auto synd = _mm256_cmp_ps( xpred, dot5, _CMP_GT_OS );
int test = (_mm256_movemask_ps( synd ) & cDeg6); // deg 6
const auto syndrom = _mm_popcnt_u32( test );
const auto _Replica = _mm256_loadu_ps( &zReplica[CumSumCheckDegree]);
const auto _ambda = _mm256_loadu_ps( &Lambda [CumSumCheckDegree]);
const auto v1 = _mm256_mul_ps (xpred, _rho );
const auto v2 = _mm256_mul_ps ( _Replica, _un_m_rho );
const auto v3 = _mm256_add_ps ( v1, v2 );
const auto vect_proj = _mm256_sub_ps ( v3, _ambda );
//
// ON REALISE LA PROJECTION !!!
//
allVerified += ( syndrom & 0x01 );
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
const auto START = timer();
#endif
const auto latest = _mm256_loadu_ps(&latestProjVector[CumSumCheckDegree]);
const auto different = _mm256_sub_ps ( vect_proj, latest );
const auto maskAbsol = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF));
const auto absolute = _mm256_and_ps ( different, maskAbsol );
const auto despass = _mm256_cmp_ps( absolute, seuilProj, _CMP_GT_OS );
int skip = (_mm256_movemask_ps( despass ) & cDeg6) == 0x00; // degree 6
if( skip == false )
{
const auto _ztemp = mp.projection_deg6( vect_proj );
const auto _ztemp1 = _mm256_sub_ps(_ztemp, xpred );
const auto _ztemp2 = _mm256_sub_ps(_ztemp, _Replica );
const auto _ztemp3 = _mm256_mul_ps(_ztemp1, _rho);
const auto _ztemp4 = _mm256_mul_ps(_ztemp2, _un_m_rho);
const auto nLambda = _mm256_add_ps( _ambda, _ztemp3 );
const auto mLambda = _mm256_add_ps( nLambda, _ztemp4 );
_mm256_maskstore_ps(& Lambda[CumSumCheckDegree], mask_6, mLambda);
_mm256_maskstore_ps(&zReplica[CumSumCheckDegree], mask_6, _ztemp);
}
_mm256_maskstore_ps(&latestProjVector[CumSumCheckDegree], mask_6, vect_proj);
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
t_pj += (timer() - START);
#endif
CumSumCheckDegree += 6;
}else if( CheckDegree[j] == 7 )
{
const int cDeg7 = 0x7F;
const auto offsets = _mm256_loadu_si256 ((const __m256i*)&t_col1 [CumSumCheckDegree]);
const auto xpred = _mm256_mask_i32gather_ps (zero, OutputFromDecoder, offsets, _mm256_castsi256_ps(mask_7), 4);
const auto synd = _mm256_cmp_ps( xpred, dot5, _CMP_GT_OS );
const int test = (_mm256_movemask_ps( synd ) & cDeg7); // deg 7
const auto syndrom = _mm_popcnt_u32( test );
const auto _Replica = _mm256_loadu_ps( &zReplica[CumSumCheckDegree]);
const auto _ambda = _mm256_loadu_ps( &Lambda [CumSumCheckDegree]);
const auto v1 = _mm256_mul_ps ( xpred, _rho );
const auto v2 = _mm256_mul_ps ( _Replica, _un_m_rho );
const auto v3 = _mm256_add_ps ( v1, v2 );
const auto vect_proj = _mm256_sub_ps ( v3, _ambda );
//
// ON REALISE LA PROJECTION !!!
//
allVerified += ( syndrom & 0x01 );
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
const auto START = timer();
#endif
const auto latest = _mm256_loadu_ps(&latestProjVector[CumSumCheckDegree]);
const auto different = _mm256_sub_ps ( vect_proj, latest );
const auto maskAbsol = _mm256_castsi256_ps(_mm256_set1_epi32(0x7FFFFFFF));
const auto absolute = _mm256_and_ps ( different, maskAbsol );
const auto despass = _mm256_cmp_ps( absolute, seuilProj, _CMP_GT_OS );
int skip = (_mm256_movemask_ps( despass ) & cDeg7) == 0x00; // degree 7
if( skip == false )
{
const auto _ztemp = mp.projection_deg7( vect_proj );
const auto _ztemp1 = _mm256_sub_ps(_ztemp, xpred );
const auto _ztemp2 = _mm256_sub_ps(_ztemp, _Replica );
const auto _ztemp3 = _mm256_mul_ps(_ztemp1, _rho);
const auto _ztemp4 = _mm256_mul_ps(_ztemp2, _un_m_rho);
const auto nLambda = _mm256_add_ps( _ambda, _ztemp3 );
const auto mLambda = _mm256_add_ps( nLambda, _ztemp4 );
_mm256_maskstore_ps(& Lambda [CumSumCheckDegree], mask_7, mLambda);
_mm256_maskstore_ps(&zReplica [CumSumCheckDegree], mask_7, _ztemp);
}
_mm256_maskstore_ps(&latestProjVector[CumSumCheckDegree], mask_7, vect_proj);
//
// MEASURE OF THE PROJECTION EXECUTION TIME
//
#ifdef PROFILE_ON
t_pj += (timer() - START);
#endif
CumSumCheckDegree += 7;
}else{
exit( 0 );
}
}
//
// MEASURE OF THE CN LOOP EXECUTION TIME
//
#ifdef PROFILE_ON
t_cn += (timer() - starT);
#endif
#ifdef PROFILE_ON
t_ex += 1;
//FILE *ft=fopen("time.txt","a");
//fprintf(ft,"%d \n", t_cn/t_ex);
//fprintf(ft,"%d %d %d \n", t_cn, t_vn, t_pj);
//fclose(ft);
#endif
if(allVerified == 0)
{
mAlgorithmConverge = true;
mValidCodeword = true;
break;
}
}
//
// MEASURE OF THE NUMBER OF EXECUTION
//
// #ifdef PROFILE_ON
// t_ex += 1;
// #endif
}
void molec_quadrant_neighbor_interaction_fma(molec_Quadrant_t q, molec_Quadrant_t q_n, float* Epot_)
{
#ifdef __AVX2__
const __m256 sigLJ = _mm256_set1_ps(molec_parameter->sigLJ);
const __m256 epsLJ = _mm256_set1_ps(molec_parameter->epsLJ);
const __m256 Rcut2 = _mm256_set1_ps(molec_parameter->Rcut2);
const int N = q.N;
const int N_n = q_n.N_pad;
__m256 Epot8 = _mm256_setzero_ps();
__m256 _1 = _mm256_set1_ps(1.f);
__m256 _2 = _mm256_set1_ps(2.f);
__m256 _24epsLJ = _mm256_mul_ps(_mm256_set1_ps(24.f), epsLJ);
for(int i = 0; i < N; ++i)
{
const __m256 xi = _mm256_set1_ps(q.x[i]);
const __m256 yi = _mm256_set1_ps(q.y[i]);
const __m256 zi = _mm256_set1_ps(q.z[i]);
__m256 f_xi = _mm256_setzero_ps();
__m256 f_yi = _mm256_setzero_ps();
__m256 f_zi = _mm256_setzero_ps();
for(int j = 0; j < N_n; j += 8)
{
// count number of interactions
if(MOLEC_CELLLIST_COUNT_INTERACTION)
++num_potential_interactions;
// load coordinates and fores into AVX vectors
const __m256 xj = _mm256_load_ps(&q_n.x[j]);
const __m256 yj = _mm256_load_ps(&q_n.y[j]);
const __m256 zj = _mm256_load_ps(&q_n.z[j]);
__m256 f_xj = _mm256_load_ps(&q_n.f_x[j]);
__m256 f_yj = _mm256_load_ps(&q_n.f_y[j]);
__m256 f_zj = _mm256_load_ps(&q_n.f_z[j]);
// distance computation
const __m256 xij = _mm256_sub_ps(xi, xj);
const __m256 yij = _mm256_sub_ps(yi, yj);
const __m256 zij = _mm256_sub_ps(zi, zj);
const __m256 zij2 = _mm256_mul_ps(zij, zij);
const __m256 r2 = _mm256_fmadd_ps(xij, xij, _mm256_fmadd_ps(yij, yij, zij2));
// r2 < Rcut2
const __m256 mask = _mm256_cmp_ps(r2, Rcut2, _CMP_LT_OQ);
// if( any(r2 < R2) )
if(_mm256_movemask_ps(mask))
{
const __m256 r2inv = _mm256_div_ps(_1, r2);
const __m256 s2 = _mm256_mul_ps(_mm256_mul_ps(sigLJ, sigLJ), r2inv);
const __m256 s6 = _mm256_mul_ps(_mm256_mul_ps(s2, s2), s2);
const __m256 s12 = _mm256_mul_ps(s6, s6);
const __m256 s12_minus_s6 = _mm256_sub_ps(s12, s6);
const __m256 two_s12_minus_s6 = _mm256_sub_ps(_mm256_mul_ps(_2, s12), s6);
Epot8 = _mm256_add_ps(Epot8, _mm256_and_ps(s12_minus_s6, mask));
const __m256 fr = _mm256_mul_ps(_mm256_mul_ps(_24epsLJ, r2inv), two_s12_minus_s6);
const __m256 fr_mask = _mm256_and_ps(fr, mask);
// update forces
f_xi = _mm256_fmadd_ps(fr_mask, xij,f_xi);
f_yi = _mm256_fmadd_ps(fr_mask, yij,f_yi);
f_zi = _mm256_fmadd_ps(fr_mask, zij,f_zi);
f_xj = _mm256_fnmadd_ps(fr_mask,xij,f_xj);
f_yj = _mm256_fnmadd_ps(fr_mask,yij,f_yj);
f_zj = _mm256_fnmadd_ps(fr_mask,zij,f_zj);
// store back j-forces
_mm256_store_ps(&q_n.f_x[j], f_xj);
_mm256_store_ps(&q_n.f_y[j], f_yj);
_mm256_store_ps(&q_n.f_z[j], f_zj);
}
}
// update i-forces
float MOLEC_ALIGNAS(32) f_array[8];
_mm256_store_ps(f_array, f_xi);
q.f_x[i] += f_array[0] + f_array[1] + f_array[2] + f_array[3] + f_array[4] + f_array[5]
+ f_array[6] + f_array[7];
_mm256_store_ps(f_array, f_yi);
q.f_y[i] += f_array[0] + f_array[1] + f_array[2] + f_array[3] + f_array[4] + f_array[5]
+ f_array[6] + f_array[7];
_mm256_store_ps(f_array, f_zi);
q.f_z[i] += f_array[0] + f_array[1] + f_array[2] + f_array[3] + f_array[4] + f_array[5]
+ f_array[6] + f_array[7];
}
float MOLEC_ALIGNAS(32) E_pot_array[8];
_mm256_store_ps(E_pot_array, Epot8);
// perform reduction of potential energy
*Epot_ += 4
* molec_parameter->epsLJ*(E_pot_array[0] + E_pot_array[1] + E_pot_array[2]
+ E_pot_array[3] + E_pot_array[4] + E_pot_array[5]
+ E_pot_array[6] + E_pot_array[7]);
#endif
}
/* since sin256_ps and cos256_ps are almost identical, sincos256_ps could replace both of them..
it is almost as fast, and gives you a free cosine with your sine */
void sincos256_ps(v8sf x, v8sf *s, v8sf *c) {
v8sf xmm1, xmm2, xmm3 = _mm256_setzero_ps(), sign_bit_sin, y;
v8si imm0, imm2, imm4;
#ifndef __AVX2__
v4si imm0_1, imm0_2;
v4si imm2_1, imm2_2;
v4si imm4_1, imm4_2;
#endif
sign_bit_sin = x;
/* take the absolute value */
x = _mm256_and_ps(x, *(v8sf*)_ps256_inv_sign_mask);
/* extract the sign bit (upper one) */
sign_bit_sin = _mm256_and_ps(sign_bit_sin, *(v8sf*)_ps256_sign_mask);
/* scale by 4/Pi */
y = _mm256_mul_ps(x, *(v8sf*)_ps256_cephes_FOPI);
#ifdef __AVX2__
/* store the integer part of y in imm2 */
imm2 = _mm256_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
imm2 = _mm256_add_epi32(imm2, *(v8si*)_pi32_256_1);
imm2 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_inv1);
y = _mm256_cvtepi32_ps(imm2);
imm4 = imm2;
/* get the swap sign flag for the sine */
imm0 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_4);
imm0 = _mm256_slli_epi32(imm0, 29);
//v8sf swap_sign_bit_sin = _mm256_castsi256_ps(imm0);
/* get the polynom selection mask for the sine*/
imm2 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_2);
imm2 = _mm256_cmpeq_epi32(imm2, *(v8si*)_pi32_256_0);
//v8sf poly_mask = _mm256_castsi256_ps(imm2);
#else
/* we use SSE2 routines to perform the integer ops */
COPY_IMM_TO_XMM(_mm256_cvttps_epi32(y),imm2_1,imm2_2);
imm2_1 = _mm_add_epi32(imm2_1, *(v4si*)_pi32avx_1);
imm2_2 = _mm_add_epi32(imm2_2, *(v4si*)_pi32avx_1);
imm2_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_inv1);
imm2_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_inv1);
COPY_XMM_TO_IMM(imm2_1,imm2_2,imm2);
y = _mm256_cvtepi32_ps(imm2);
imm4_1 = imm2_1;
imm4_2 = imm2_2;
imm0_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_4);
imm0_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_4);
imm0_1 = _mm_slli_epi32(imm0_1, 29);
imm0_2 = _mm_slli_epi32(imm0_2, 29);
COPY_XMM_TO_IMM(imm0_1, imm0_2, imm0);
imm2_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_2);
imm2_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_2);
imm2_1 = _mm_cmpeq_epi32(imm2_1, _mm_setzero_si128());
imm2_2 = _mm_cmpeq_epi32(imm2_2, _mm_setzero_si128());
COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
#endif
v8sf swap_sign_bit_sin = _mm256_castsi256_ps(imm0);
v8sf poly_mask = _mm256_castsi256_ps(imm2);
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(v8sf*)_ps256_minus_cephes_DP1;
xmm2 = *(v8sf*)_ps256_minus_cephes_DP2;
xmm3 = *(v8sf*)_ps256_minus_cephes_DP3;
xmm1 = _mm256_mul_ps(y, xmm1);
xmm2 = _mm256_mul_ps(y, xmm2);
xmm3 = _mm256_mul_ps(y, xmm3);
x = _mm256_add_ps(x, xmm1);
x = _mm256_add_ps(x, xmm2);
x = _mm256_add_ps(x, xmm3);
#ifdef __AVX2__
imm4 = _mm256_sub_epi32(imm4, *(v8si*)_pi32_256_2);
imm4 = _mm256_andnot_si128(imm4, *(v8si*)_pi32_256_4);
imm4 = _mm256_slli_epi32(imm4, 29);
#else
imm4_1 = _mm_sub_epi32(imm4_1, *(v4si*)_pi32avx_2);
imm4_2 = _mm_sub_epi32(imm4_2, *(v4si*)_pi32avx_2);
imm4_1 = _mm_andnot_si128(imm4_1, *(v4si*)_pi32avx_4);
imm4_2 = _mm_andnot_si128(imm4_2, *(v4si*)_pi32avx_4);
imm4_1 = _mm_slli_epi32(imm4_1, 29);
imm4_2 = _mm_slli_epi32(imm4_2, 29);
COPY_XMM_TO_IMM(imm4_1, imm4_2, imm4);
#endif
v8sf sign_bit_cos = _mm256_castsi256_ps(imm4);
sign_bit_sin = _mm256_xor_ps(sign_bit_sin, swap_sign_bit_sin);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
v8sf z = _mm256_mul_ps(x,x);
y = *(v8sf*)_ps256_coscof_p0;
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(v8sf*)_ps256_coscof_p1);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(v8sf*)_ps256_coscof_p2);
y = _mm256_mul_ps(y, z);
y = _mm256_mul_ps(y, z);
v8sf tmp = _mm256_mul_ps(z, *(v8sf*)_ps256_0p5);
y = _mm256_sub_ps(y, tmp);
y = _mm256_add_ps(y, *(v8sf*)_ps256_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
v8sf y2 = *(v8sf*)_ps256_sincof_p0;
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(v8sf*)_ps256_sincof_p1);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(v8sf*)_ps256_sincof_p2);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_mul_ps(y2, x);
y2 = _mm256_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
v8sf ysin2 = _mm256_and_ps(xmm3, y2);
v8sf ysin1 = _mm256_andnot_ps(xmm3, y);
y2 = _mm256_sub_ps(y2,ysin2);
y = _mm256_sub_ps(y, ysin1);
xmm1 = _mm256_add_ps(ysin1,ysin2);
xmm2 = _mm256_add_ps(y,y2);
/* update the sign */
*s = _mm256_xor_ps(xmm1, sign_bit_sin);
*c = _mm256_xor_ps(xmm2, sign_bit_cos);
}
/* evaluation of 8 sines at onces using AVX intrisics
The code is the exact rewriting of the cephes sinf function.
Precision is excellent as long as x < 8192 (I did not bother to
take into account the special handling they have for greater values
-- it does not return garbage for arguments over 8192, though, but
the extra precision is missing).
Note that it is such that sinf((float)M_PI) = 8.74e-8, which is the
surprising but correct result.
*/
v8sf sin256_ps(v8sf x) { // any x
v8sf xmm1, xmm2 = _mm256_setzero_ps(), xmm3, sign_bit, y;
v8si imm0, imm2;
#ifndef __AVX2__
v4si imm0_1, imm0_2;
v4si imm2_1, imm2_2;
#endif
sign_bit = x;
/* take the absolute value */
x = _mm256_and_ps(x, *(v8sf*)_ps256_inv_sign_mask);
/* extract the sign bit (upper one) */
sign_bit = _mm256_and_ps(sign_bit, *(v8sf*)_ps256_sign_mask);
/* scale by 4/Pi */
y = _mm256_mul_ps(x, *(v8sf*)_ps256_cephes_FOPI);
/*
Here we start a series of integer operations, which are in the
realm of AVX2.
If we don't have AVX, let's perform them using SSE2 directives
*/
#ifdef __AVX2__
/* store the integer part of y in mm0 */
imm2 = _mm256_cvttps_epi32(y);
/* j=(j+1) & (~1) (see the cephes sources) */
// another two AVX2 instruction
imm2 = _mm256_add_epi32(imm2, *(v8si*)_pi32_256_1);
imm2 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_inv1);
y = _mm256_cvtepi32_ps(imm2);
/* get the swap sign flag */
imm0 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_4);
imm0 = _mm256_slli_epi32(imm0, 29);
/* get the polynom selection mask
there is one polynom for 0 <= x <= Pi/4
and another one for Pi/4<x<=Pi/2
Both branches will be computed.
*/
imm2 = _mm256_and_si128(imm2, *(v8si*)_pi32_256_2);
imm2 = _mm256_cmpeq_epi32(imm2,*(v8si*)_pi32_256_0);
#else
/* we use SSE2 routines to perform the integer ops */
COPY_IMM_TO_XMM(_mm256_cvttps_epi32(y),imm2_1,imm2_2);
imm2_1 = _mm_add_epi32(imm2_1, *(v4si*)_pi32avx_1);
imm2_2 = _mm_add_epi32(imm2_2, *(v4si*)_pi32avx_1);
imm2_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_inv1);
imm2_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_inv1);
COPY_XMM_TO_IMM(imm2_1,imm2_2,imm2);
y = _mm256_cvtepi32_ps(imm2);
imm0_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_4);
imm0_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_4);
imm0_1 = _mm_slli_epi32(imm0_1, 29);
imm0_2 = _mm_slli_epi32(imm0_2, 29);
COPY_XMM_TO_IMM(imm0_1, imm0_2, imm0);
imm2_1 = _mm_and_si128(imm2_1, *(v4si*)_pi32avx_2);
imm2_2 = _mm_and_si128(imm2_2, *(v4si*)_pi32avx_2);
imm2_1 = _mm_cmpeq_epi32(imm2_1, _mm_setzero_si128());
imm2_2 = _mm_cmpeq_epi32(imm2_2, _mm_setzero_si128());
COPY_XMM_TO_IMM(imm2_1, imm2_2, imm2);
#endif
v8sf swap_sign_bit = _mm256_castsi256_ps(imm0);
v8sf poly_mask = _mm256_castsi256_ps(imm2);
sign_bit = _mm256_xor_ps(sign_bit, swap_sign_bit);
/* The magic pass: "******"
x = ((x - y * DP1) - y * DP2) - y * DP3; */
xmm1 = *(v8sf*)_ps256_minus_cephes_DP1;
xmm2 = *(v8sf*)_ps256_minus_cephes_DP2;
xmm3 = *(v8sf*)_ps256_minus_cephes_DP3;
xmm1 = _mm256_mul_ps(y, xmm1);
xmm2 = _mm256_mul_ps(y, xmm2);
xmm3 = _mm256_mul_ps(y, xmm3);
x = _mm256_add_ps(x, xmm1);
x = _mm256_add_ps(x, xmm2);
x = _mm256_add_ps(x, xmm3);
/* Evaluate the first polynom (0 <= x <= Pi/4) */
y = *(v8sf*)_ps256_coscof_p0;
v8sf z = _mm256_mul_ps(x,x);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(v8sf*)_ps256_coscof_p1);
y = _mm256_mul_ps(y, z);
y = _mm256_add_ps(y, *(v8sf*)_ps256_coscof_p2);
y = _mm256_mul_ps(y, z);
y = _mm256_mul_ps(y, z);
v8sf tmp = _mm256_mul_ps(z, *(v8sf*)_ps256_0p5);
y = _mm256_sub_ps(y, tmp);
y = _mm256_add_ps(y, *(v8sf*)_ps256_1);
/* Evaluate the second polynom (Pi/4 <= x <= 0) */
v8sf y2 = *(v8sf*)_ps256_sincof_p0;
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(v8sf*)_ps256_sincof_p1);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_add_ps(y2, *(v8sf*)_ps256_sincof_p2);
y2 = _mm256_mul_ps(y2, z);
y2 = _mm256_mul_ps(y2, x);
y2 = _mm256_add_ps(y2, x);
/* select the correct result from the two polynoms */
xmm3 = poly_mask;
y2 = _mm256_and_ps(xmm3, y2); //, xmm3);
y = _mm256_andnot_ps(xmm3, y);
y = _mm256_add_ps(y,y2);
/* update the sign */
y = _mm256_xor_ps(y, sign_bit);
return y;
}
/* natural logarithm computed for 8 simultaneous float
return NaN for x <= 0
*/
v8sf log256_ps(v8sf x) {
v8si imm0;
v8sf one = *(v8sf*)_ps256_1;
//v8sf invalid_mask = _mm256_cmple_ps(x, _mm256_setzero_ps());
v8sf invalid_mask = _mm256_cmp_ps(x, _mm256_setzero_ps(), _CMP_LE_OS);
x = _mm256_max_ps(x, *(v8sf*)_ps256_min_norm_pos); /* cut off denormalized stuff */
// can be done with AVX2
imm0 = _mm256_srli_epi32(_mm256_castps_si256(x), 23);
/* keep only the fractional part */
x = _mm256_and_ps(x, *(v8sf*)_ps256_inv_mant_mask);
x = _mm256_or_ps(x, *(v8sf*)_ps256_0p5);
// this is again another AVX2 instruction
imm0 = _mm256_sub_epi32(imm0, *(v8si*)_pi32_256_0x7f);
v8sf e = _mm256_cvtepi32_ps(imm0);
e = _mm256_add_ps(e, one);
/* part2:
if( x < SQRTHF ) {
e -= 1;
x = x + x - 1.0;
} else { x = x - 1.0; }
*/
//v8sf mask = _mm256_cmplt_ps(x, *(v8sf*)_ps256_cephes_SQRTHF);
v8sf mask = _mm256_cmp_ps(x, *(v8sf*)_ps256_cephes_SQRTHF, _CMP_LT_OS);
v8sf tmp = _mm256_and_ps(x, mask);
x = _mm256_sub_ps(x, one);
e = _mm256_sub_ps(e, _mm256_and_ps(one, mask));
x = _mm256_add_ps(x, tmp);
v8sf z = _mm256_mul_ps(x,x);
v8sf y = *(v8sf*)_ps256_cephes_log_p0;
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_log_p1);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_log_p2);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_log_p3);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_log_p4);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_log_p5);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_log_p6);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_log_p7);
y = _mm256_mul_ps(y, x);
y = _mm256_add_ps(y, *(v8sf*)_ps256_cephes_log_p8);
y = _mm256_mul_ps(y, x);
y = _mm256_mul_ps(y, z);
tmp = _mm256_mul_ps(e, *(v8sf*)_ps256_cephes_log_q1);
y = _mm256_add_ps(y, tmp);
tmp = _mm256_mul_ps(z, *(v8sf*)_ps256_0p5);
y = _mm256_sub_ps(y, tmp);
tmp = _mm256_mul_ps(e, *(v8sf*)_ps256_cephes_log_q2);
x = _mm256_add_ps(x, y);
x = _mm256_add_ps(x, tmp);
x = _mm256_or_ps(x, invalid_mask); // negative arg will be NAN
return x;
}
