use crate::noise::{NOISE_AC_MASK, texture_score_from_simd_sads};
use std::arch::x86_64::*;
#[target_feature(enable = "avx2")]
pub(crate) fn noise_laplacian_stats_avx2(patch: &[f32; 64]) -> [f32; 2] {
let zero = _mm256_set1_ps(0.0);
let quarter = _mm256_set1_ps(-0.25);
let five = _mm256_set1_ps(5.0);
let sign = _mm256_set1_ps(-0.0);
let left = _mm256_setr_epi32(1, 0, 1, 2, 3, 4, 5, 6);
let right = _mm256_setr_epi32(1, 2, 3, 4, 5, 6, 7, 6);
let mut rows = [[zero; 3]; 8];
for (src, dst) in patch.as_chunks::<8>().0.iter().zip(&mut rows) {
let center = unsafe { _mm256_loadu_ps(src.as_ptr()) };
*dst = [
_mm256_permutevar8x32_ps(center, left),
center,
_mm256_permutevar8x32_ps(center, right),
];
}
let mut abs = [0.0; 64];
for y in 0..8 {
let top = &rows[if y == 0 { 1 } else { y - 1 }];
let center = &rows[y];
let bottom = &rows[if y == 7 { 6 } else { y + 1 }];
let mut filtered = _mm256_add_ps(zero, _mm256_mul_ps(quarter, top[0]));
filtered = _mm256_sub_ps(filtered, top[1]);
filtered = _mm256_add_ps(filtered, _mm256_mul_ps(quarter, top[2]));
filtered = _mm256_sub_ps(filtered, center[0]);
filtered = _mm256_add_ps(filtered, _mm256_mul_ps(five, center[1]));
filtered = _mm256_sub_ps(filtered, center[2]);
filtered = _mm256_add_ps(filtered, _mm256_mul_ps(quarter, bottom[0]));
filtered = _mm256_sub_ps(filtered, bottom[1]);
filtered = _mm256_add_ps(filtered, _mm256_mul_ps(quarter, bottom[2]));
unsafe {
_mm256_storeu_ps(
abs.as_mut_ptr().add(y * 8),
_mm256_andnot_ps(sign, filtered),
)
};
}
crate::noise::noise_laplacian_stats_from_abs(patch, &abs)
}
#[target_feature(enable = "avx2")]
pub(crate) fn noise_shrink_avx2(coeffs: &mut [f32; 64], t2: f32) {
if t2 <= 0.0 || t2.is_nan() {
return;
}
let threshold = _mm256_set1_ps(t2);
let zero = _mm256_set1_ps(0.0);
for (p, mask) in coeffs
.as_chunks_mut::<8>()
.0
.iter_mut()
.zip(NOISE_AC_MASK.as_chunks::<8>().0)
{
let c = unsafe { _mm256_loadu_ps(p.as_ptr()) };
let c2 = _mm256_mul_ps(c, c);
let gain = _mm256_div_ps(_mm256_sub_ps(c2, threshold), c2);
let shrunk = _mm256_mul_ps(c, _mm256_max_ps(gain, zero));
let mask = unsafe { _mm256_loadu_si256(mask.as_ptr().cast()) };
unsafe {
_mm256_storeu_ps(
p.as_mut_ptr(),
_mm256_blendv_ps(c, shrunk, _mm256_castsi256_ps(mask)),
)
};
}
}
#[target_feature(enable = "avx2")]
pub(crate) fn noise_texture_score_avx2(patch: &[f32; 64]) -> f32 {
let mut sads = [0.0; 20];
let sign = _mm256_set1_ps(-0.0);
for yb in (0..4).step_by(2) {
let mut sad = _mm256_setzero_ps();
for cy in 0..4 {
for cx in 0..3 {
let center = _mm256_set1_ps(patch[(2 + cy) * 8 + 2 + cx]);
let lo = unsafe { _mm_loadu_ps(patch.as_ptr().add((yb + cy) * 8 + cx)) };
let hi = unsafe { _mm_loadu_ps(patch.as_ptr().add((yb + 1 + cy) * 8 + cx)) };
let pixels = _mm256_set_m128(hi, lo);
sad = _mm256_add_ps(sad, _mm256_andnot_ps(sign, _mm256_sub_ps(center, pixels)));
}
}
unsafe {
_mm_storeu_ps(sads.as_mut_ptr().add(yb * 5), _mm256_castps256_ps128(sad));
_mm_storeu_ps(
sads.as_mut_ptr().add((yb + 1) * 5),
_mm256_extractf128_ps::<1>(sad),
);
}
}
texture_score_from_simd_sads(patch, sads)
}