[Numpy-discussion] Da Blas

Paul F. Dubois pauldubois at home.com
Thu Sep 14 07:10:53 CDT 2000


Well, I certainly didn't mean to flame anyone in particular, so I'm sorry
that Frank was offended. I suppose I got irritated when I tried to solve one
problem and simply created one that was harder to solve. I don't think any
of us quite appreciated how bad the situation is. No good deed goes
unpunished.

I tried yesterday to help someone internally with an SGI system. They had
two new wrinkles on this:

a. They had their own blas but not their own lapack, in /usr/lib. Well, I
had allowed for that in my second version, but of course the imprecision of
the -L stuff makes it possible to get the wrong one that way.

b. I did get the /usr/lib version of the blas, and boy was it the wrong one,
since apparently there are two binary modes on an SGI and it was not -n32
like Python expected. Neither was the lapack_lite one. I had to edit the
Makefile. That problem could be avoided using distutils to do the compile.

The solution was to edit setup.py to remove /usr/lib from the search list.

The problem with simply using the desired library as a binary instead of a
library is that you would load all of it. I do not know how much control you
can get with the -l -L paradigm, especially in a portable way.





More information about the Numpy-discussion mailing list