in lattices/lattice_utils.cpp [243:257]
static inline __m128 masked_read (int d, const float *x)
{
assert (0 <= d && d < 4);
__attribute__((__aligned__(16))) float buf[4] = {0, 0, 0, 0};
switch (d) {
case 3:
buf[2] = x[2];
case 2:
buf[1] = x[1];
case 1:
buf[0] = x[0];
}
return _mm_load_ps (buf);
// cannot use AVX2 _mm_mask_set1_epi32
}