fast_inv_sqrt
This was made purely for fun and testing crates.io publishing, but may actually be usable.
InvSqrt32 trait provide inv_sqrt32() function for primitive numeric types. InvSqrt64 provides inv_sqrt64().
How do I use it?
Cargo.toml:
[]
= "1.0.0"
Code:
extern crate fast_inv_sqrt;
use InvSqrt32;
use InvSqrt64;
Benchmarks
Benchmarks require nightly compiler.
"ref" benchmarks use f1 / f2.sqrt() "impl" benchmarks use f1 * f2.inv_sqrtXX()
$ cargo bench --features 'nightly'
test test32::bench_plain_impl ... bench: 8 ns/iter (+/- 0)
test test32::bench_plain_ref ... bench: 13 ns/iter (+/- 0)
test test32::bench_real_impl ... bench: 8 ns/iter (+/- 0)
test test32::bench_real_ref ... bench: 13 ns/iter (+/- 0)
test test64::bench_plain_impl ... bench: 10 ns/iter (+/- 0)
test test64::bench_plain_ref ... bench: 20 ns/iter (+/- 0)
test test64::bench_real_impl ... bench: 10 ns/iter (+/- 0)
test test64::bench_real_ref ... bench: 20 ns/iter (+/- 0)
Feature 'omit-checking' disables checks of if value passed is_sign_positive() and is_normal(), and will produce invalid results for denormalized and negative values, but this is fast:
test test32::bench_plain_impl ... bench: 2 ns/iter (+/- 0)
test test32::bench_plain_ref ... bench: 13 ns/iter (+/- 0)
test test32::bench_real_impl ... bench: 3 ns/iter (+/- 0)
test test32::bench_real_ref ... bench: 13 ns/iter (+/- 0)
test test64::bench_plain_impl ... bench: 2 ns/iter (+/- 0)
test test64::bench_plain_ref ... bench: 20 ns/iter (+/- 0)
test test64::bench_real_impl ... bench: 3 ns/iter (+/- 0)
test test64::bench_real_ref ... bench: 20 ns/iter (+/- 0)