TST: Skip test_rolling_var_numerical_issues on power platforms #50349
Conversation
mroeschke added Testing 
| | ||
| | ||
| @pytest.mark.xfail(is_platform_arm() and not is_platform_mac(), reason="GH 38921") | ||
| @pytest.mark.xfail( |
There was a problem hiding this comment.
I guess we have to figure out what's wrong in here at some point
There was a problem hiding this comment.
Yeah definitely. Maybe we should reconsider if it's also worth using a sliding window algorithm if it's susceptible to these numerical precision issues
There was a problem hiding this comment.
Hm I think this is a compiler issue. Something is optimised that triggers the problem.
There was a problem hiding this comment.
Right a compiler flag seems to match the consensus in issue ( #38921 ).
Namely it looks like some optimization causes a form of fma to get called, which evaluates incorrectly. This implementation in Cython seems to demonstrate the issue ( #38921 (comment) ).
Am guessing this issue would also show up on macOS ARM (even though that hasn't come up yet).
There was a problem hiding this comment.
Agreed. At least for this PR I am just hoping to skip this test for now until a fix is investigated
There was a problem hiding this comment.
Yep my comment was meant more generally.
yes this also happens on Mac, already validated this
| Thx @mroeschke |
| Thanks Matthew! 🙏 |
3 participants
Discovered while working on conda-forge/pandas-feedstock#147. cc @jakirkham