Ondrej Certik wrote:
On Mon, Jul 21, 2008 at 7:19 PM, rpmu...@gmail.com <rpmu...@gmail.com> wrote:
On Jul 21, 3:24 am, Robert Cimrman <cimr...@ntc.zcu.cz> wrote:
It is an umfpack installation issue, as Ondrej has pointed out.
I have read your messages at scipy-users and it seems that you have some path problems. Do you have any updates on this? What if you manually add the umfpack directory to your sys.path?
r. Update: I now get the same errors on the Mac and on RHEL4.
I don't know whether they're path features or not. The umfpack libraries are static (.a) rather than dynamic (.so) libraries, which normally means you don't have to worry about path, right?
On Debian, the umfpack is an .so package: /usr/lib/libumfpack.so.3.1.0
The python extension modules must be dynamic .so libraries. This may be the problem... See http://www.python.org/doc/ext/compilation.html
The umfpack library can be static though, you just have to have something like scipy/sparse/linalg/dsolve/umfpack/__umfpack.so in your scipy installation, i.e. the umfpack wrappers I wrote for scipy, and the .a lib must be linked to it. The exact path depends on your scipy version, the newest is to have umfpack module as a scikit.
Have a look at http://projects.scipy.org/pipermail/scipy-user/2008-February/015500.html See also http://www.scipy.org/Installing_SciPy/Linux, maybe it helps.
Thus far I haven't gotten any replies on the scipy lists, so I'm still waiting on both platforms.
I am sorry for these problems. The umfpack is needed though. Doesn't RHEL have a scipy package already build with umpack?
All these things makes me happy I chose Debian for all my computers, it just works.
Same on gentoo.
The problem is that umfpack is the default solver in sfepy, as it can solve any system (symmetric positive definite or completely general, as far as it is not singular) that fits in memory and is very fast.
You can play with the code without it, but the other interfaced solvers (e.g. iterative solvers shipped with SciPy) are not so robust and one usually must do some preconditioning.
For example, a Poisson equation can be solved very fast with
solver_0 = { 'name' : 'ls', 'kind' : 'ls.scipy_iterative',
'method' : 'cg',
'i_max' : 1000,
'eps_a' : 1e-12,}
instead of the default
solver_0 = { 'name' : 'ls', 'kind' : 'ls.umfpack', }
r.