use core::arch::x86_64::*;
#[target_feature(enable = "avx2")]
pub(crate) fn polish_count_nonzero_avx2(values: &[i32]) -> usize {
let mut count = 0;
let mut i = 0;
while i + 8 <= values.len() {
let v = unsafe { _mm256_loadu_si256(values.as_ptr().add(i).cast()) };
let eq = _mm256_cmpeq_epi32(v, _mm256_setzero_si256());
count += _mm256_movemask_ps(_mm256_castsi256_ps(eq)).count_ones() as usize;
i += 8;
}
values.len() - count - values[i..].iter().filter(|&&v| v == 0).count()
}
#[target_feature(enable = "avx2")]
pub(crate) fn polish_add_scaled_avx2(row: &mut [f32], delta: &[f32], sign: f32) {
assert_eq!(row.len(), delta.len());
let factor = _mm256_set1_ps(sign);
let mut i = 0;
while i + 8 <= row.len() {
let v = unsafe { _mm256_loadu_ps(row.as_ptr().add(i)) };
let d = unsafe { _mm256_loadu_ps(delta.as_ptr().add(i)) };
let scaled = _mm256_mul_ps(d, factor);
let sum = _mm256_add_ps(v, scaled);
unsafe { _mm256_storeu_ps(row.as_mut_ptr().add(i), sum) };
i += 8;
}
for (v, &d) in row[i..].iter_mut().zip(&delta[i..]) {
*v += sign * d;
}
}