[Numpy-discussion] doc bug in numarray 1.1

[Stephen Walton]

On Mon, 2004-09-13 at 16:32, Chris Barker wrote:

> I just installed numarray 1.1... All went well with setup.py. Then I 
> decided to try to build it against an atlas lapack. I found, under the 
> Linear Algebra section:
> 
> """
> the setup procedure needs to be modified to force the lapack_lite module 
> to be linked against those rather than the builtin replacement functions.

This does seem like a documentation bug.

> addons.py
>
> so I tried:
> 
> export USE_LAPACK=true
> python setup.py build --gencode
> 
> Now it appears to be compiling linear_algebra differently, as I now get 
> a linking error, can't find:
> 
> f90math, fio, or f77math

Amazingly enough, I just found this problem today myself, and I also
confess it is My Fault (tm), as I was unclear in a previous post to thie
forum.  These libraries are specific to the commercial Absoft Fortran
compiler.  If you change the lapack_libs assignment in addons.py to

    lapack_libs = ['lapack', 'f77blas', 'cblas', 'atlas', 'g2c', 'm']
 
then the command you used above will build against ATLAS with g77.

Building against ATLAS is definitely worthwhile.  On my little laptop
(mobile AMD Athlon XP2800+) the time to solve a 1000x1000 random array
went from 10.7 to 0.9 seconds, and that was just using the prebuilt
Linux_ATHLON ATLAS tarball from the scipy.com site, not one I compiled
myself optimized for my computer.

Steve Walton