#[cfg(target_arch = "x86")]
use std::arch::x86::*;
#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;
#[inline]
#[target_feature(enable = "sse4.1")]
fn grad_pack_vector(
left: &[i32; 4],
top: &[i32; 4],
topleft: &[i32; 4],
px: &[i32; 4],
out: &mut [u32; 4],
) {
let zero = _mm_setzero_si128();
let left = unsafe { _mm_loadu_si128(left.as_ptr() as *const __m128i) };
let top = unsafe { _mm_loadu_si128(top.as_ptr() as *const __m128i) };
let topleft = unsafe { _mm_loadu_si128(topleft.as_ptr() as *const __m128i) };
let px = unsafe { _mm_loadu_si128(px.as_ptr() as *const __m128i) };
let ac = _mm_sub_epi32(left, topleft);
let ab = _mm_sub_epi32(left, top);
let bc = _mm_sub_epi32(top, topleft);
let grad = _mm_add_epi32(ac, top);
let d = _mm_xor_si128(ab, bc);
let clamp = _mm_blendv_epi8(left, top, _mm_cmpgt_epi32(zero, d)); let s = _mm_xor_si128(ac, bc);
let pred = _mm_blendv_epi8(clamp, grad, _mm_cmpgt_epi32(zero, s));
let res = _mm_sub_epi32(px, pred);
let packed = _mm_xor_si128(_mm_slli_epi32::<1>(res), _mm_srai_epi32::<31>(res));
unsafe { _mm_storeu_si128(out.as_mut_ptr() as *mut __m128i, packed) };
}
#[target_feature(enable = "sse4.1")]
pub(crate) fn grad_pack_interior(cur: &[i32], prev: &[i32], out: &mut [u32], gw: usize) {
assert!(cur.len() >= gw && prev.len() >= gw && out.len() >= gw);
if gw <= 1 {
return;
}
let (left, left_tail) = cur[..gw - 1].as_chunks::<4>();
let (top, top_tail) = prev[1..gw].as_chunks::<4>();
let (topleft, topleft_tail) = prev[..gw - 1].as_chunks::<4>();
let (px, px_tail) = cur[1..gw].as_chunks::<4>();
let (out, out_tail) = out[1..gw].as_chunks_mut::<4>();
let (left_pairs, left_single) = left.as_chunks::<2>();
let (top_pairs, top_single) = top.as_chunks::<2>();
let (topleft_pairs, topleft_single) = topleft.as_chunks::<2>();
let (px_pairs, px_single) = px.as_chunks::<2>();
let (out_pairs, out_single) = out.as_chunks_mut::<2>();
for ((((left, top), topleft), px), out) in left_pairs
.iter()
.zip(top_pairs)
.zip(topleft_pairs)
.zip(px_pairs)
.zip(out_pairs)
{
grad_pack_vector(&left[0], &top[0], &topleft[0], &px[0], &mut out[0]);
grad_pack_vector(&left[1], &top[1], &topleft[1], &px[1], &mut out[1]);
}
for ((((left, top), topleft), px), out) in left_single
.iter()
.zip(top_single)
.zip(topleft_single)
.zip(px_single)
.zip(out_single)
{
grad_pack_vector(left, top, topleft, px, out);
}
for ((((&w, &n), &nw), &value), out) in left_tail
.iter()
.zip(top_tail)
.zip(topleft_tail)
.zip(px_tail)
.zip(out_tail)
{
let ac = w.wrapping_sub(nw);
let bc = n.wrapping_sub(nw);
let grad = ac.wrapping_add(n);
let clamp = if (w.wrapping_sub(n) ^ bc) < 0 { n } else { w };
let pred = if (ac ^ bc) < 0 { grad } else { clamp };
let r = value.wrapping_sub(pred) as i64;
*out = ((r << 1) ^ (r >> 63)) as u32;
}
}