Today I learned that f64::tan (or maybe f64::to_radians?) produces different results (given the same inputs) on my ARM MacBook and on x86_64 Linux.
The math consistently returns -3.40476137196133 on one, and -3.4047613719613308 on the other. At least it's deterministic. Happening on both debug and release builds.
=> More informations about this toot | More toots from adam_chal@hachyderm.io
OK it's definitely the tan call. https://github.com/adamchalmers/fp-repro if anyone is interested in a repro.
=> More informations about this toot | More toots from adam_chal@hachyderm.io
@adam_chal Yeah there's no particular guarantee that floating point operations are exactly the same across machines
=> More informations about this toot | More toots from rain@hachyderm.io This content has been proxied by September (3851b).Proxy Information
text/gemini