Bug#831541: theano: single GradientError and WrongValues in tests on s390x, ppc64 and sparc

Daniel Stender debian at danielstender.com
Tue Feb 14 10:26:25 UTC 2017


On 14.02.2017 08:43, Rebecca N. Palmer wrote:
> Found the problem: the affected functions ( https://sources.debian.net/src/theano/0.8.2-4/theano/sparse/opt.py/#L862 , https://sources.debian.net/src/theano/0.8.2-4/theano/sparse/opt.py/#L1902 ) cast a pointer-to-intptr_t (64 bit) to a pointer-to-int (32-bit).
> 
> Which isn't just broken on big-endian systems, it's a strict aliasing violation *everywhere* (i.e. technically undefined behaviour with optimization on, which it is by default, though it appears to work in practice).
> 
> (I expect to post a patch tonight: the obvious has a potential overflow issue, and it also needs a c_code_cache_version change to make the fix be used in existing installs).

That would be sensational. #835531 doesn't keep up being RC, that would mean that we're through
to Stretch if this bug could be closed. Much thx in advance!

Daniel

-- 
4096R/DF5182C8
Debian Developer (stender at debian.org)
http://www.danielstender.com/



More information about the debian-science-maintainers mailing list