static void __stdcall
comb_mask_1_simd(uint8_t* dstp, const uint8_t* srcp, const int dpitch,
const int spitch, const int cthresh, const int width,
const int height) noexcept
{
const uint8_t* sc = srcp;
const uint8_t* sb = sc + spitch;
const uint8_t* sd = sc + spitch;
const V cth = set1_i16<V>(static_cast<int16_t>(cthresh));
const V all = cmpeq_i8(cth, cth);
constexpr int step = sizeof(V) / 2;
for (int y = 0; y < height; ++y) {
for (int x = 0; x < width; x += step) {
V xb = load_half<V>(sb + x);
V xc = load_half<V>(sc + x);
V xd = load_half<V>(sd + x);
xb = sub_i16(xb, xc);
xd = sub_i16(xd, xc);
xc = andnot(mulhi(xb, xd), mullo(xb, xd));
xc = cmpgt_u16(xc, cth, all);
store_half(dstp + x, xc);
}
sb = sc;
sc = sd;
sd += (y < height - 2) ? spitch : -spitch;
dstp += dpitch;
}
}
always_inline VecFloat ldexp_float(VecFloat const & x, typename VecFloat::int_vec const & n)
{
typedef typename VecFloat::int_vec int_vec;
const VecFloat exponent_mask = VecFloat::gen_exp_mask();
const VecFloat exponent = exponent_mask & x;
const VecFloat x_wo_x = andnot(exponent_mask, x); // clear exponent
int_vec new_exp = slli(n, 16+7) + int_vec(exponent); // new exponent
VecFloat new_exp_float(new_exp);
VecFloat ret = x_wo_x | new_exp_float;
return ret;
}
always_inline VecFloat frexp_float(VecFloat const & x, typename VecFloat::int_vec & exp)
{
typedef typename VecFloat::int_vec int_vec;
const VecFloat exponent_mask = VecFloat::gen_exp_mask();
const VecFloat exponent = exponent_mask & x;
const VecFloat x_wo_x = andnot(exponent_mask, x); // clear exponent
const int_vec exp_int(exponent);
exp = srli(exp_int, 16+7) - int_vec(126);
return x_wo_x | VecFloat::gen_exp_mask_1();
}
always_inline VecType vec_cos_float(VecType const & arg)
{
typedef typename VecType::int_vec int_vec;
const typename VecType::float_type four_over_pi = 1.27323954473516268615107010698011489627567716592367;
VecType abs_arg = arg & VecType::gen_abs_mask();
VecType y = abs_arg * four_over_pi;
int_vec j = y.truncate_to_int();
/* cephes: j=(j+1) & (~1) */
j = (j + int_vec(1)) & int_vec(~1);
y = j.convert_to_float();
/* sign based on quadrant */
int_vec jm2 = j - int_vec(2);
VecType sign = slli(andnot(jm2, int_vec(4)), 29);
/* polynomial mask */
VecType poly_mask = VecType (mask_eq(jm2 & int_vec(2), int_vec(0)));
/* black magic */
static float DP1 = 0.78515625;
static float DP2 = 2.4187564849853515625e-4;
static float DP3 = 3.77489497744594108e-8;
VecType base = ((abs_arg - y * DP1) - y * DP2) - y * DP3;
/* [0..pi/4] */
VecType z = base * base;
VecType p1 = (( 2.443315711809948E-005 * z
- 1.388731625493765E-003) * z
+ 4.166664568298827E-002) * z * z
-0.5f * z + 1
;
/* [pi/4..pi/2] */
VecType p2 = ((-1.9515295891E-4 * z
+ 8.3321608736E-3) * z
- 1.6666654611E-1) * z * base + base;
VecType approximation = select(p1, p2, poly_mask);
return approximation ^ sign;
}
always_inline VecType vec_select(VecType lhs, VecType rhs, VecType bitmask)
{
return andnot(bitmask, lhs) | (bitmask & rhs);
}
always_inline VecType vec_select(VecType lhs, VecType rhs, VecType bitmask)
{
const VecType result = andnot(bitmask, lhs) | (bitmask & rhs);
return result;
}
