#![cfg(target_arch = "x86_64")]
use std::arch::x86_64::{
__m256i, _mm256_cmpgt_epi64, _mm256_loadu_si256, _mm256_movemask_pd, _mm256_set1_epi64x,
_mm512_cmpgt_epi64_mask, _mm512_loadu_si512, _mm512_set1_epi64,
};
const AVX512_LANES: usize = 8;
const AVX2_LANES: usize = 4;
#[target_feature(enable = "avx512f")]
pub unsafe fn scan_gt_i64_avx512(data: &[i64], target: i64) -> usize {
let mut matches = 0_usize;
let target_vec = _mm512_set1_epi64(target);
let chunks = data.chunks_exact(AVX512_LANES);
let remainder = chunks.remainder();
for chunk in chunks {
let data_vec = unsafe { _mm512_loadu_si512(chunk.as_ptr().cast()) };
let mask: u8 = _mm512_cmpgt_epi64_mask(data_vec, target_vec);
matches += mask.count_ones() as usize;
}
matches + super::portable::scan_gt_i64(remainder, target)
}
#[target_feature(enable = "avx2")]
pub unsafe fn scan_gt_i64_avx2(data: &[i64], target: i64) -> usize {
let mut matches = 0_usize;
let target_vec = _mm256_set1_epi64x(target);
let chunks = data.chunks_exact(AVX2_LANES);
let remainder = chunks.remainder();
for chunk in chunks {
let data_vec: __m256i = unsafe { _mm256_loadu_si256(chunk.as_ptr().cast()) };
let cmp = _mm256_cmpgt_epi64(data_vec, target_vec);
let as_pd = unsafe { std::mem::transmute::<__m256i, std::arch::x86_64::__m256d>(cmp) };
let bits = _mm256_movemask_pd(as_pd);
matches += bits.unsigned_abs().count_ones() as usize;
}
matches + super::portable::scan_gt_i64(remainder, target)
}