[SciPy-user] lapack unresolved symbol question

David Cournapeau david@ar.media.kyoto-u.ac...
Tue Oct 16 21:28:14 CDT 2007


Karl Young wrote:
> Thanks again Robert; I "solved" the problem inspired by your suggestion 
> but maybe you or someone else has a suggestion for something more 
> sensible or elegant. The problem seems to be that there's a 
> liblapack.so.3 in /usr/lib (presumably not built against an 
> ATLAS-accelerated LAPACK) that gets picked up if I have /usr/lib in 
> LD_LIBRARY_PATH . I had forgotten that I'd had this problem before and 
> as a temporary solution had just taken /usr/lib out of LD_LIBRARY_PATH; 
> /usr/lib got put back in as a result of installing of something else 
> (hence the recurrence). Obviously removing /usr/lib from LD_LIBRARY_PATH 
> isn't a great "solution" so pardon my ignorance but is there a python 
> environment variable can I set to get python to look in /usr/lib/atlas 
> before /usr/lib (PYTHONPATH doesn't seem to work for this) ? Thanks,
PYTHONPATH won't work for this, indeed. This is outside python's hand, 
so no python solution here. I don't know why LD_LIBRARY_PATH contains 
/usr/lib: this is rather strange, and kind of defeat its purpose; 
LD_LIBRARY_PATH itself is a hack [1], though, so modifying it is not 
worse than using it in most cases. In your case, if it gets rewritten, 
it can be a pain, though.

One solution would be to launch a python shell from a shell command 
which set LD_LIBRARY_PATH correctly (that's what I would do in your 
case, but this just a workaround). Actually, there is something we could 
do on scipy's side to make all this easier, but it is not trivial to 
implement.

Alternatively, you could just install the rpm, which are available for FC5:

http://download.opensuse.org/repositories/home:/ashigabou/Fedora_Extras_5/

This includes numpy 1.0.4, scipy 0.6, blas/lapack, and a source rpm to 
ATLAS (you can get instructions here: 
http://www.scipy.org/Installing_SciPy/Linux

cheers,

David

[1]: http://xahlee.org/UnixResource_dir/_/ldpath.html


More information about the SciPy-user mailing list