Safe and branchless error-free transformation algorithms for floating point numbers.
Now supports types which impl num_traits::Float and size of bits of its significand is odd (Because Float does not offer its size).
With nightly compiler, execute
$ cargo +nightly bench
If your CPU has fma target-feature,
$ RUSTFLAGS='-C target-cpu=native' cargo +nightly bench --features use-fma
If your CPU does not have fma and compile and run with this command, ***_fma will work very slowly due to software emulation of fma.
twosum| algorithm | time (ns) / operation |
|-----------------------|-----------------------|
| twosum(not safe) | 1.7530 |
| safetwosum_branch | 4.4827 |
| safetwosum_straight | 4.6438 |
| safetwosum_fma | 2.4749 |
split| algorithm | time (ns) / operation |
|------------------------------|------------------------|
| split(not safe) | 0.8915 |
| safesplit_branch(not safe) | 1.1918 |
| safesplit_straight | 3.3454 |
twoproduct| algorithm | time (ns) / operation |
|---------------------------|-----------------------|
| twoproduct(not safe) | 2.9322 |
| safetwoproduct_branch | 3.7826 |
| safetwoproduct_straight | 12.9883 |
| safetwoproduct_fma | 0.7125 |