On Mon, 2022-01-24 at 19:37 +0000, Michael Gutteridge wrote: > We have an interesting situation where a test suite that checks the > value of e^-inf is behaving differently depending on Numpy version. > The minimal test I'm using is: > > #!/usr/bin/env python3 > import numpy as np > np.seterr(under='raise') > print(np.exp(np.NINF)) > > With numpy version 1.19.5 and 1.20.3 we consistently get "0.0" as the > result. Later versions- 1.21 and 1.22- raise an underflow error: > > Traceback (most recent call last): > File "./under.py", line 4, in <module> > print(np.exp(np.NINF)) > FloatingPointError: underflow encountered in exp > > A further confounding factor is that on our slightly older processors > we don't get this underflow, instead returning 0. The Python and GCC > toolchains is the same thanks to environment modules and the OS and > patches are the same on both generations, but the older processors > (Intel Xeon E5-2667) return 0 and the newer (Intel Xeon Gold 6254) > return underflow with the newer numpy versions. > > I think my primary question is what is the expected result for this > function? I sort of expect this to be defined as zero but I've been > unable to find documentation of this fact. Underflow does sort of > make sense in this situation as well. However, the inconsistency is > concerning and I'm uncertain if it would affect results from other > calculations. >
Thanks for the report. Could you open an issue about this (And maybe ping @r-devulap on it)? Underflow warnings are ignored by default, so this would sneak under the radar easily. Just to be clear: If this is not an easy fix, I am not sure we should bother. Getting FPEs (floating point errors) right 100% is tricky, also because compilers do get them wrong often enough (it improved, but older compilers or math libraries can be problematic, e.g. older clang versions get things wrong a lot). That said, I do expect you are right and no warning should be given here and that there is a problem in the SIMD version (acceleration for more advanced CPU instruction sets). But the question is how deep that issue is :). Cheers, Sebastian > Thanks much > > - Michael > _______________________________________________ > NumPy-Discussion mailing list -- numpy-discussion@python.org > To unsubscribe send an email to numpy-discussion-le...@python.org > https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ > Member address: sebast...@sipsolutions.net >
signature.asc
Description: This is a digitally signed message part
_______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-le...@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: arch...@mail-archive.com
