#[cfg(target_arch = "x86")]
use std::arch::x86::*;
#[cfg(target_arch = "x86_64")]
use std::arch::x86_64::*;
use crate::filter::{UvLumaTapHbd, WienerTapHbd};
#[inline(always)]
fn load8_u16_i32(p: &[u16]) -> (__m128i, __m128i) {
unsafe {
let v = _mm_loadu_si128(p.as_ptr().cast());
let lo = _mm_cvtepu16_epi32(v);
let hi = _mm_cvtepu16_epi32(_mm_srli_si128(v, 8));
(lo, hi)
}
}
#[inline(always)]
fn finish8_u16(dst: &mut [u16], slo: __m128i, shi: __m128i, bitdepth_max: i32) {
unsafe {
let rnd = _mm_set1_epi32(64);
let zero = _mm_setzero_si128();
let max = _mm_set1_epi32(bitdepth_max);
let vlo = _mm_min_epi32(
_mm_max_epi32(_mm_srai_epi32(_mm_add_epi32(slo, rnd), 7), zero),
max,
);
let vhi = _mm_min_epi32(
_mm_max_epi32(_mm_srai_epi32(_mm_add_epi32(shi, rnd), 7), zero),
max,
);
let packed = _mm_packus_epi32(vlo, vhi);
_mm_storeu_si128(dst.as_mut_ptr().cast(), packed);
}
}
#[inline(always)]
fn load4_u16_i32(row: &[u16], idx: usize) -> __m128i {
unsafe { _mm_cvtepu16_epi32(_mm_loadl_epi64(row[idx..].as_ptr() as *const __m128i)) }
}
#[inline(always)]
fn gather4_u16_i32(row: &[u16], idx: usize, step: usize) -> __m128i {
if step == 1 {
load4_u16_i32(row, idx)
} else {
let arr = [
row[idx],
row[idx + step],
row[idx + 2 * step],
row[idx + 3 * step],
];
unsafe { _mm_cvtepu16_epi32(_mm_loadl_epi64(arr.as_ptr() as *const __m128i)) }
}
}
#[inline(always)]
fn finish4_u16(dst: &mut [u16], x: usize, s: __m128i, bitdepth_max: i32) {
unsafe {
let v = _mm_min_epi32(
_mm_max_epi32(
_mm_srai_epi32(_mm_add_epi32(s, _mm_set1_epi32(64)), 7),
_mm_setzero_si128(),
),
_mm_set1_epi32(bitdepth_max),
);
let packed = _mm_packus_epi32(v, v);
_mm_storel_epi64(dst[x..].as_mut_ptr() as *mut __m128i, packed);
}
}
#[target_feature(enable = "sse4.1")]
pub(crate) fn ns_wiener_fir_run_hbd_sse41(
dst: &mut [u16],
center: &[u16],
col0: usize,
taps: &[WienerTapHbd],
n: usize,
bitdepth_max: i32,
) {
let mut x = 0usize;
while x + 8 <= n {
let c = col0 + x;
debug_assert!(c + 8 <= center.len());
let (mlo, mhi) = load8_u16_i32(¢er[c..]);
let mut slo = _mm_slli_epi32(mlo, 7);
let mut shi = _mm_slli_epi32(mhi, 7);
let two_mlo = _mm_add_epi32(mlo, mlo);
let two_mhi = _mm_add_epi32(mhi, mhi);
for t in taps {
let cp = (c as i32 + t.dx) as usize;
let cm = (c as i32 - t.dx) as usize;
debug_assert!(cp + 8 <= t.row_p.len() && cm + 8 <= t.row_m.len());
let (alo, ahi) = load8_u16_i32(&t.row_p[cp..]);
let (blo, bhi) = load8_u16_i32(&t.row_m[cm..]);
let coef = _mm_set1_epi32(t.coef);
slo = _mm_add_epi32(
slo,
_mm_mullo_epi32(_mm_sub_epi32(_mm_add_epi32(alo, blo), two_mlo), coef),
);
shi = _mm_add_epi32(
shi,
_mm_mullo_epi32(_mm_sub_epi32(_mm_add_epi32(ahi, bhi), two_mhi), coef),
);
}
finish8_u16(&mut dst[x..], slo, shi, bitdepth_max);
x += 8;
}
while x < n {
let c = col0 + x;
let m = center[c] as i32;
let mut s = m << 7;
for t in taps {
let a = t.row_p[(c as i32 + t.dx) as usize] as i32;
let b = t.row_m[(c as i32 - t.dx) as usize] as i32;
s += (a + b - 2 * m) * t.coef;
}
dst[x] = ((s + 64) >> 7).clamp(0, bitdepth_max) as u16;
x += 1;
}
}
#[target_feature(enable = "sse4.1")]
pub(crate) fn pc_wiener_fir_run_hbd_sse41(
dst: &mut [u16],
center: &[u16],
center_coef: i32,
col0: usize,
taps: &[WienerTapHbd],
n: usize,
bitdepth_max: i32,
) {
let mut x = 0usize;
while x + 8 <= n {
let c = col0 + x;
debug_assert!(c + 8 <= center.len());
let (mlo, mhi) = load8_u16_i32(¢er[c..]);
let cc = _mm_set1_epi32(center_coef);
let mut slo = _mm_mullo_epi32(mlo, cc);
let mut shi = _mm_mullo_epi32(mhi, cc);
for t in taps {
let cp = (c as i32 + t.dx) as usize;
let cm = (c as i32 - t.dx) as usize;
debug_assert!(cp + 8 <= t.row_p.len() && cm + 8 <= t.row_m.len());
let (alo, ahi) = load8_u16_i32(&t.row_p[cp..]);
let (blo, bhi) = load8_u16_i32(&t.row_m[cm..]);
let coef = _mm_set1_epi32(t.coef);
slo = _mm_add_epi32(slo, _mm_mullo_epi32(_mm_add_epi32(alo, blo), coef));
shi = _mm_add_epi32(shi, _mm_mullo_epi32(_mm_add_epi32(ahi, bhi), coef));
}
finish8_u16(&mut dst[x..], slo, shi, bitdepth_max);
x += 8;
}
while x < n {
let c = col0 + x;
let m = center[c] as i32;
let mut s = m * center_coef;
for t in taps {
let a = t.row_p[(c as i32 + t.dx) as usize] as i32;
let b = t.row_m[(c as i32 - t.dx) as usize] as i32;
s += (a + b) * t.coef;
}
dst[x] = ((s + 64) >> 7).clamp(0, bitdepth_max) as u16;
x += 1;
}
}
#[target_feature(enable = "sse4.1")]
#[allow(clippy::too_many_arguments)]
pub(crate) fn ns_wiener_uv_fir_run_hbd_sse41(
dst: &mut [u16],
c_center: &[u16],
co: usize,
ctaps: &[WienerTapHbd],
l_center: &[u16],
lo: usize,
ltaps: &[UvLumaTapHbd],
lstep: usize,
n: usize,
bitdepth_max: i32,
) {
let mut x = 0usize;
while x + 4 <= n {
let cb = co + x;
let m = load4_u16_i32(c_center, cb);
let two_m = _mm_add_epi32(m, m);
let mut s = _mm_slli_epi32(m, 7);
for t in ctaps {
let a = load4_u16_i32(t.row_p, (cb as i32 + t.dx) as usize);
let b = load4_u16_i32(t.row_m, (cb as i32 - t.dx) as usize);
let coef = _mm_set1_epi32(t.coef);
s = _mm_add_epi32(
s,
_mm_mullo_epi32(_mm_sub_epi32(_mm_add_epi32(a, b), two_m), coef),
);
}
let lb = lo + x * lstep;
let lc = gather4_u16_i32(l_center, lb, lstep);
for t in ltaps {
let lv = gather4_u16_i32(t.row, (lb as i32 + t.ldx) as usize, lstep);
let coef = _mm_set1_epi32(t.coef);
s = _mm_add_epi32(s, _mm_mullo_epi32(_mm_sub_epi32(lv, lc), coef));
}
finish4_u16(dst, x, s, bitdepth_max);
x += 4;
}
while x < n {
let cc = co + x;
let m = c_center[cc] as i32;
let mut s = m << 7;
for t in ctaps {
let a = t.row_p[(cc as i32 + t.dx) as usize] as i32;
let b = t.row_m[(cc as i32 - t.dx) as usize] as i32;
s += (a + b - 2 * m) * t.coef;
}
let lcx = lo + x * lstep;
let lc = l_center[lcx] as i32;
for t in ltaps {
let lv = t.row[(lcx as i32 + t.ldx) as usize] as i32;
s += (lv - lc) * t.coef;
}
dst[x] = ((s + 64) >> 7).clamp(0, bitdepth_max) as u16;
x += 1;
}
}