[SciPy-dev] Progress with linalg2

eric eric at scipy.org
Sun Mar 3 16:47:36 CST 2002


>
> Hi Eric,
>
> On Sun, 3 Mar 2002, eric wrote:
>
> > 1.  My atlas is missing some routines that you have in generic_clapack.pyf.
> >     Specifically:
> >
> >         clapack_xgetri
> >         clapack_xpotri
> >         clapack_xlauum
> >         clapack_xtrtri
> >
> >     where x stands for the various types.  Are these in your ATLAS?  I have
a
> >     fairly recent version, but perhaps I need to upgrade?
>
> Yes, my ATLAS have them. I am using ATLAS-3.3.13 and I didn't realize that
> it may contain functions not included in the stable ATLAS. Sorry about
> that.

I downloaded the latest (3.3.14) and it built fine under cygwin.  It sounds like
this is the last one before a stable release for ATLAS, so the problem with
missing functions is soon to go away.

>
> > 2.  I get segfaults when trying to run the tests.  It looks like they happen
> >     for both C and Fortran (I commented out contiguous cases to test
Fortran).
> >     Do you think this is a windows specific issue?  I am using the latest
> >     f2py CVS.
>
> Try using only flapack. Because the segfaults may be caused by the
> 1. issue above. To be sure that solve() will not use clapack functions,
> you need to remove clapack.so file, commenting some test cases out may not
> be enough.

Good news!  After building with the new ATLAS, everything works fine.

Here is the output on my machine:

C:\home\ej\wrk\scipy\linalg2>python tests\test_basic.py
..
      Solving system of linear equations
      ==================================
      |    continuous     |   non-continuous
----------------------------------------------
 size |  scipy  | Numeric |  scipy  | Numeric
   20 |   0.49  |   0.54  |   0.49  |   0.65     (secs for 2000 calls)
  100 |   0.98  |   2.11  |   0.86  |   2.46     (secs for 300 calls)
  500 |   1.08  |   3.13  |   1.06  |   3.28     (secs for 4 calls)
 1000 |   3.56  |  28.32  |   3.59  |  28.39     (secs for 2 calls)
.
----------------------------------------------------------------------
Ran 3 tests in 81.427s

OK

So, we get similar results for the scipy/ATLAS stuff.  It looks like the reason
I remembered such a difference in scipy/Numeric is that I'm getting such lousy
performance from Numeric's linpack_lite routines.  I wonder why our Numeric
performance is so different??  I'm running the latest 21.0b1 .exe on the Numeric
website.  Are you linking your Numeric against an optimized set of linpack
routines?

Anyway, it doesn't matter to much.  The important thing is that linalg2 is
working on W2K just fine.

see ya,
eric






More information about the Scipy-dev mailing list