4 ms·
Knowing adgjlsfhk1's work, yes, this would be for correctness — specifically measuring error in ULPs. Most frequently, adgjlsfhk1 pushes Julia's numeric routin
by mbauman 4y ago
Knowing adgjlsfhk1's work, yes, this would be for correctness — specifically measuring error in ULPs. Most frequently, adgjlsfhk1 pushes Julia's numeric routines to errors below 0.5 ULPs — that is, perfect correctly rounded behavior.
- adgjlsfhk1 4y agoI actually am not a believer in perfect rounding. It tends to have a high performance cost, and IMO isn't that useful.