Closed KristofferC closed 5 years ago
Against --math-mode=fast
. This is terribly unsound, considering things like the sinpi
issue.
@fastmath
is fair and idiomatic. I don't think a lot of real-world projects should use --math-mode=fast
. Other languages use fastmath, but this is fair in C: You set the flags per source-file, not per process.
Also against --check-bounds=no
. We want to showcase how boundscheck elision works. If Base
has an issue where some critical boundscheck cannot be elided from user code, then we need an additional @propagate_inbounds
in Base
, and should fairly report the slower speed until the PR has landed, either via backport point release or via release of 1.1.
Ok, default arguments make sense.
Only need to make sure JULIA_NUM_THREADS
is set.
Maybe O3? But it has actually been a long time since I've seen something actually becoming faster because of it ;)
Yeah, after SLP being in O2 Im not sure what O3 does really.
Clearly,
--check-bounds=no
. How about--math-mode=fast
? If tests pass, arguably we should run with fast math mode. What do other languages do?