use core::arch::wasm32::*;
#[target_feature(enable = "simd128")]
pub(crate) fn downsample_rows_wasm(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 = "simd128")]
fn downsample<const N: usize>(top: &[f32], bottom: &[f32], dst: &mut [f32]) -> usize {
const LANES: usize = 4;
let quarter = f32x4_splat(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 = v128_load(top.as_ptr().add(offset).cast());
let t1 = v128_load(top.as_ptr().add(offset + LANES).cast());
let b0 = v128_load(bottom.as_ptr().add(offset).cast());
let b1 = v128_load(bottom.as_ptr().add(offset + LANES).cast());
let a = i32x4_shuffle::<0, 2, 4, 6>(t0, t1);
let b = i32x4_shuffle::<1, 3, 5, 7>(t0, t1);
let c = i32x4_shuffle::<0, 2, 4, 6>(b0, b1);
let d = i32x4_shuffle::<1, 3, 5, 7>(b0, b1);
let sum = f32x4_add(f32x4_add(f32x4_add(a, b), c), d);
v128_store(
dst.as_mut_ptr().add(lane * LANES).cast(),
f32x4_mul(sum, quarter),
);
}
}
done += N * LANES;
}
done
}