use std::arch::aarch64::*;
#[target_feature(enable = "neon")]
pub(crate) fn downsample_rows_neon(top: &[f32], bottom: &[f32], dst: &mut [f32]) -> usize {
assert!(top.len() / 2 >= dst.len());
assert!(bottom.len() / 2 >= dst.len());
let done = downsample::<2>(top, bottom, dst);
done + downsample::<1>(&top[2 * done..], &bottom[2 * done..], &mut dst[done..])
}
#[inline]
#[target_feature(enable = "neon")]
fn downsample<const N: usize>(top: &[f32], bottom: &[f32], dst: &mut [f32]) -> usize {
const LANES: usize = 4;
let quarter = vdupq_n_f32(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 = vld1q_f32(top.as_ptr().add(offset));
let t1 = vld1q_f32(top.as_ptr().add(offset + LANES));
let b0 = vld1q_f32(bottom.as_ptr().add(offset));
let b1 = vld1q_f32(bottom.as_ptr().add(offset + LANES));
let top_pairs = vpaddq_f32(t0, t1);
let c = vuzp1q_f32(b0, b1);
let d = vuzp2q_f32(b0, b1);
let sum = vaddq_f32(vaddq_f32(top_pairs, c), d);
vst1q_f32(dst.as_mut_ptr().add(lane * LANES), vmulq_f32(sum, quarter));
}
}
done += N * LANES;
}
done
}