It both support SSE2
and AVX2
and much faster than regular implementations.
[ a, b, c, d .... x, y, z]
compare a and b,
compare b and c,
compare c and d
[ a, b, c, d .... x, y, z]
load [a, b, c, d] into chunk0,
load [b, c, d, d] into chunk1,
compare chunk0 and chunk1,
load [d, e, f, g] into chunk0,
load [e, f, g, g] into chunk1,
compare chunk0 and chunk1 again
size | generic | sse | avx2 |
---|---|---|---|
4 | 7.4974 ns | 8.3876 ns | 1.1984 ns |
32 | 65.302 ns | 40.632 ns | 16.469 ns |
64 | 121.92 ns | 71.861 ns | 30.551 ns |
128 | 237.09 ns | 144.08 ns | 66.191 ns |
256 | 466.77 ns | 273.48 ns | 99.640 ns |
1024 | 1.8486 µs | 1.0634 µs | 410.66 ns |
4096 | 7.3648 µs | 4.2080 µs | 6.1038 µs |
16384 | 29.458 µs | 16.702 µs | 6.1038 µs |
65536 | 117.82 µs | 66.982 µs | 24.355 µs |