Bug#1043048: openblas: gives wrong results on mips64el, ignores test failures

[Sébastien Villemot]

Dear Rebecca,

Thanks for your report.

Le samedi 05 août 2023 à 09:25 +0100, Rebecca N. Palmer a écrit :
> Package: libopenblas0-pthread
> Version: 0.3.23+ds-2
> Control: affects -1 src:statsmodels
> Severity: serious
> Justification: the default BLAS should NOT silently give wrong answers
> (i.e. if there's no easy way to actually fix this, please switch the 
> default on mips64el back to atlas, and consider removing this package 
> from mips64el)
> 
> statsmodels recently (between 0.14.0+dfsg-1 and -2) started to FTBFS on 
> mips64el with multiple wrong test results.  The most obviously relevant 
> change between those is that the installed BLAS changed from atlas to 
> openblas.
> 
> openblas' own tests on mips64el ( 
> https://buildd.debian.org/status/fetch.php?pkg=openblas&arch=mips64el&ver=0.3.23%2Bds-2&stamp=1686760279&raw=0 
> ) have 64 instances of "FATAL ERROR - COMPUTED RESULT IS LESS THAN HALF 
> ACCURATE".  (I don't know why this isn't failing the build, which is 
> possibly a bug in itself.)
> 
> openblas upstream are not _obviously_ aware of this.  Given the 
> existence of .github/workflows/mips64.yml, this suggests it _may_ be 
> nontrivial to reproduce in qemu.

It looks like version 0.3.21+ds-4 is not affected by this issue and
passes its testsuite. Can you possibly check whether statsmodels builds
against that version?

My guess is that this bug is caused by the switch to the MIPS64_GENERIC
kernel that I made in version 0.3.22+ds-1. If this is indeed the cause,
then an easy short term fix is to roll back this change and go back to
the SICORTEX kernel. For a longer term fix, I would need to work with
upstream to determine why the MIPS64_GENERIC kernel is broken.

Also note that in any case, using ATLAS is not a good solution because
we are considering its removal, see the thread on debian-science at .¹
BLIS is a better alternative to OpenBLAS.

I also agree that the fact that the OpenBLAS testsuite fails without
triggering an FTBFS is abnormal. I’m surprised by this, and this should
be investigated.

Cheers,


¹
https://lists.debian.org/msgid-search/4311acc16afb473599c79bd5b17a8b734c2f8d2b.camel@debian.org
-- 
⢀⣴⠾⠻⢶⣦⠀  Sébastien Villemot
⣾⠁⢠⠒⠀⣿⡁  Debian Developer
⢿⡄⠘⠷⠚⠋⠀  https://sebastien.villemot.name
⠈⠳⣄⠀⠀⠀⠀  https://www.debian.org


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: This is a digitally signed message part
URL: <http://alioth-lists.debian.net/pipermail/debian-science-maintainers/attachments/20230805/5865a776/attachment-0001.sig>