[SciPy-dev] newcore and gcc4/gfortran (almost there)

Pearu Peterson pearu at scipy.org
Sun Nov 20 23:38:12 CST 2005



On Sun, 20 Nov 2005, Ravikiran Rajagopal wrote:

> Now, one more test passes - check_dot in scipy.linalg.test_blas. This 
> took some figuring out; I had my ATLAS compiled with -ff2c but not 
> SciPy. This caused a mismatch between the fortran code in 
> linalg/src/fblaswrap.f and the ATLAS library. After the addition of 
> -ff2c to the gfortran flag list, it succeeds.

Scipy assumes that ATLAS as well as LAPACK libraries are compiled without 
-ff2c.

> However, the code in lib/fblas/fblaswrap.f.src is a lot hairier and I 
> got lost trying to figure that one out.

Read

   pydoc scipy.distutils.from_template

that describes the syntax of Fortran .src files.

> Why are two wrappers (identical in intent as far as I can tell) built 
> for ATLAS libraries (one in lib and one in linalg)? Could they not be 
> unified? If so, I vote for the one in linalg since it works for me (and 
> presumably for everyone else too). If the two can be merged, how should 
> I go about doing it?

scipy.lib.lapack/blas will replace lapack/blas wrappers in scipy.linalg.
There is only few wrappers missing in scipy.lib.lapack to complete the
move of lapack wrappers. When they are implemented, lapack/blas wrappers 
will be removed from scipy.linalg.

The fact that linalg works better for you than lib.lapack may be due to 
more complete tests in lib.lapack. Wrappers in lib.lapack are in much 
better conditions than the ones in linalg in terms of unittests and 
documentation.

Pearu




More information about the Scipy-dev mailing list