[SciPy-user] optimizing numpy on Linux?
Mon May 19 06:52:16 CDT 2008
> I was able to install atlas on 64 bit feisty/gutsy AMD64. Trick is
> building everything yourself. I used gfortran and followed Robin's
> install help page:
If you use Hardy, you do not need to build everything by yourself, even
if you use gfortran instead of g77 (but again, unless you really need
gfortran, say for OpenMP, avoid it on ubuntu, since the ABI is still g77
and not gfortran). The trick is to install libatlas* (which uses
gfortran compatible ABI) instead of atlas* (g77 ABI).
> I had find and delete all of the old libs first. That turned out to be
> quite a task and most of troubles as I had put libs
> in /usr/lib /usr/local/lib /usr/home/osman/lib , 32 and 64 ... etc
> At the end I did have a working scipy with umfpack and sfepy.
One useful tool to avoid this kind of trouble is stow:
http://www.gnu.org/software/stow/. Stow can be used to install many
concurrent versions at the same time, by using soft links (it works on
any UNIX). You cannot use several versions at the same time, though. It
has the extremely important advantage of being able to uninstall
anything installed under it. I use it to test many different
environments to build numpy/scipy (ATLAS, MKL, NETLIB, different
compilers, versions, etc...). It can also be used to track numpy and
scipy svn, while having a known, working version which can be re-enabled
in a second.
You should never install things in /usr (except /usr/local), that's one
of the best way to break your system, BTW.
More information about the SciPy-user