inline
void store_nt(float *data) const
{
SHORTVEC_ASSERT_ALIGNED(data, 64);
_mm512_stream_ps(data + 0, val1);
_mm512_stream_ps(data + 16, val2);
}
inline
void load_aligned(const float *data)
{
SHORTVEC_ASSERT_ALIGNED(data, 64);
val1 = _mm512_load_ps(data + 0);
val2 = _mm512_load_ps(data + 16);
}
inline
void scatter(float *ptr, const int *offsets) const
{
__m512i indices;
SHORTVEC_ASSERT_ALIGNED(offsets, 64);
indices = _mm512_load_epi32(offsets);
_mm512_i32scatter_ps(ptr, indices, val1, 4);
indices = _mm512_load_epi32(offsets + 16);
_mm512_i32scatter_ps(ptr, indices, val2, 4);
}
inline
void gather(const float *ptr, const int *offsets)
{
__m512i indices;
SHORTVEC_ASSERT_ALIGNED(offsets, 64);
indices = _mm512_load_epi32(offsets);
val1 = _mm512_i32gather_ps(indices, ptr, 4);
indices = _mm512_load_epi32(offsets + 16);
val2 = _mm512_i32gather_ps(indices, ptr, 4);
}
inline
void store_nt(double *data) const
{
SHORTVEC_ASSERT_ALIGNED(data, 64);
_mm512_stream_pd(data, val);
}
inline
void load_aligned(const double *data)
{
SHORTVEC_ASSERT_ALIGNED(data, 64);
val = _mm512_load_pd(data);
}
