[Numpy-discussion] building numpy with atlas on ubuntu edgy
Charles R Harris
Wed Apr 18 13:35:38 CDT 2007
On 4/18/07, Keith Goodman <email@example.com> wrote:
> On 4/18/07, Charles R Harris <firstname.lastname@example.org> wrote:
> > On 4/18/07, Keith Goodman <email@example.com> wrote:
> > > I'd like to compile atlas so that I can take full advantage of my core
> > > 2 duo. Numpy dynamically links to the debian binary of atlas-sse that
> > > I installed. But the atlas website says that they recommend static
> > > linking.
> > >
> > > Which do you recommend, static or dynamic? Are there good directions
> > either?
> > I don't know which is best, although I suspect the statically linked
> > will be larger. It might seem that just pulling in the gemm routines
> > wouldn't add much, but they pull in lots of supporting routines. To get
> > numpy to link statically you will also probably need to have a directory
> > that contains only the *.a versions because the linker will default to
> > *.so if they are present; i don't think there is a way to specify the
> > -static flag to the gcc compiler. Maybe someone else knows how to do
> > For ATLAS, I believe the latest versions are also recommended because
> > stable version is so old.
> At the moment best is equal to easiest since I have never compiled
> atlas. Does anyone know of a howto on compiling atlas (dynamically
The instructions that come with ATLAS are useable, and with the
3.7.30version the process is pretty easy. Just make sure you read the
concerning relocatable libraries first, as you will need to add some flags
to the ./configure command line, -fPIC and a couple of others. After
compiling and install, IIRC, the libraries will be in /usr/local/lib/ATLAS,
which should be OK for debian. Then you need to add a file in /etc/ld.conf.d
with the name of the directory and run ldconfig in order to let the system
know where the libraries are and what they contain. Or you could save the
old /usr/lib/atlas directory and put the new libraries there and then run
ldconfig. I also found it necessary to delete the old numpy site-package
before reinstalling numpy.
I am not the expert here, so perhaps someone else will weigh in.
Besides speed I'm also interested in seeing if I can get rid of the
> repeatability problems I have with the debian atlas-sse2 binary.
> (Repeated calulations, as discuss on this list, give give difference
> results in numpy but not octave.)
I suspect this has something to do with resetting floating point flags, so
this might not get fixed.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Numpy-discussion