[SciPy-dev] Re: Linalg2

[Travis Oliphant]

> Hi Travis,
>
> On Sat, 23 Mar 2002, Travis Oliphant wrote:
> > I'm trying to help you with linalg2 in my spare time, but I'm having a
> > time
>
> That's great. I have been quite short with my spare time lately..
>
> > getting what's there to work.  I think linalg is important and it's not
> > worth my time to try and fix the problems in linalg caused by the f2py
> > changes.
> >
> > Currently I get an error on import.
> >
> > exceptions.ImportError:
> > /usr/lib/python2.1/site-packages/scipy/linalg/clapack.so: undefined
> > symbol: clapack_sgetri
> >
> > What version of ATLAS do you need?
> > I think I have the latest stable release.  Do I need the unstable
> > release?
>
> Yes, I am using ATLAS-3.3.14. It seems from scanning atlas mailing
> lists that the ATLAS team is close to releasing a new stable release.
> On the other hand, it is possible to separate wrappers to new routines in
> ATLAS so that one can use linalg2 also with the current stable release.
> So, the question is whether to require latest "unstable" (that looks
> quite stable to me) ATLAS or to introduce hooks to linalg2 that later
> become obsolete. I would prefer the first approch provided that new
> stable ATLAS will be released soon. What do you think?
>

If we phase in linalg2, then this idea will probably work.  

> While you start testing with linalg2 and find unimplemented features,
> could we make a hotlist of what are the most important functions that need
> to be implemented ASAP, like the ones that are used in other
> scipy modules. I could then releatively quickly implement the required
> wrappers. For example, I'll try to wrap the svd algorithm today. Anything
> else?

There aren't that many functions implemented in linear algebra that require 
wrappings.   Schur decomposition, QR decomposition, Cholesky, and SVD 
decomposition would pretty much do it.