[SciPy-user] SciPy Ubuntu Dapper installation using Andrew Straw's packages

David Cournapeau david at ar.media.kyoto-u.ac.jp
Wed Oct 11 23:03:43 CDT 2006


Ryan Krauss wrote:
> So, Andrew's packages worked beautifully.  The information I was
> looking for is actually on
> http://scipy.org/Installing_SciPy/Linux .  I kept searching the list
> for the instructions but it is on the website (in case anyone else is
> ever looking here and finds this message).
>
> Are the packages using optimized Atlas and Lapack if I have them
> installed specific to my processor (sse2)?
An 'easy' way to check if you are using sse2 optimized atlas is to check 
which libraries the loader loads when loading numpy and scipy libraries. 
For example, for numpy:

ldd /usr/lib/python2.4/site-packages/numpy/core/_dotblas.so  | grep blas 
returns the atlas enabled libblas (in the /usr/lib/sse2 directory, but 
this is distribution dependent, I think)

>   Would I see any
> performance increases if I built from source?
>   
Only if your atlas is faster than the shipped one (stating the obvious:) 
). Concretely, this means that you are able to compile a faster atlas, 
with the correct extended lapack, and this is not trivial, specially if 
your cpu is not in the default. Recent (after 3.7.14, I think) atlas 
have a much easier to use configuration.

If you have a new intel core 2 duo, last atlas give outstanding results 
according to the main developer of Atlas:

http://sourceforge.net/mailarchive/forum.php?thread_id=30384800&forum_id=426

David


More information about the SciPy-user mailing list