use std::arch::x86_64::*;
#[target_feature(enable = "avx2")]
pub(crate) fn downsample_rows_avx2(top: &[f32], bottom: &[f32], dst: &mut [f32]) -> usize {
assert!(top.len() / 2 >= dst.len());
assert!(bottom.len() / 2 >= dst.len());
let done = downsample::<4>(top, bottom, dst);
done + downsample::<1>(&top[2 * done..], &bottom[2 * done..], &mut dst[done..])
}
#[inline]
#[target_feature(enable = "avx2")]
fn downsample<const N: usize>(top: &[f32], bottom: &[f32], dst: &mut [f32]) -> usize {
const LANES: usize = 8;
let quarter = _mm256_set1_ps(0.25);
let mut done = 0;
let (vectors, _) = dst.as_chunks_mut::<LANES>();
let (blocks, _) = vectors.as_chunks_mut::<N>();
for (block, dst) in blocks.iter_mut().enumerate() {
let dst = dst.as_flattened_mut();
let x = block * N * LANES;
for lane in 0..N {
unsafe {
let offset = 2 * (x + lane * LANES);
let t0 = _mm256_loadu_ps(top.as_ptr().add(offset));
let t1 = _mm256_loadu_ps(top.as_ptr().add(offset + LANES));
let b0 = _mm256_loadu_ps(bottom.as_ptr().add(offset));
let b1 = _mm256_loadu_ps(bottom.as_ptr().add(offset + LANES));
let top_pairs = _mm256_hadd_ps(t0, t1);
let c = _mm256_shuffle_ps::<0x88>(b0, b1);
let d = _mm256_shuffle_ps::<0xdd>(b0, b1);
let sum = _mm256_add_ps(_mm256_add_ps(top_pairs, c), d);
let value = _mm256_castsi256_ps(_mm256_permute4x64_epi64::<0xd8>(
_mm256_castps_si256(_mm256_mul_ps(sum, quarter)),
));
_mm256_storeu_ps(dst.as_mut_ptr().add(lane * LANES), value);
}
}
done += N * LANES;
}
done
}