[Pysparse-users] Re: Creating a CSR matrix

[Roman G. <rom...@gm...>]

Hi

Implementing a C API for a Python extension is one of the more obscure
programming tasks IMHO.

To expose additional pysparse functions using the C API the
spmatrix_api.h needs to be extended. This is however an tedious and
error prone task.

I wrote a simple Python script to generate spmatrix_api.h. You can
play around with this script if you want. (see attached file)

Cheers
-- Roman


On 6/22/05, Dominique Orban <dom...@gm...> wrote:
> On 6/22/05, Roman Geus <rom...@gm...> wrote:
> > Hi Dominique
> >
> > Sorry for not replying. I actually forgot to register myself for this l=
ist
> > :-/
> >
> > > Hello,
> > >
> > > I am linking my C code to PySparse and would like to create a CSR
> > > matrix using newCSRMatObject. I don"t see why the prototype for that
> > > function does not appear in the headers unless SPMATRIX_MODULE is
> > > #define"d.
> >
> > > If I #define SPMATRIX_MODULE before #include-ing csr_mat.h, I obtain
> > > tons of warning and error messages.
> >
> > newCSRMatObject() is not supposed to be visible outside of pysparse.
> >
> > > Why not have newCSRMatObject"s prototype in spmatrix_api.h like, e.g.=
,
> > > newLLMatObject? Same for newSSSMatObject.
> >
> > Since there is no way to create CSR and SSS matrices, other than
> > building an LL matrix first and then converting it, I don't really see
> > any benefit in exposing these functions. Am I missing something?
>=20
> Hello Roman,
>=20
> There may not be any benefits from Python. However, from C, you could
> obviously create instances of CSR and SSS matrices without the
> overhead of first creating an ll_mat and then converting it (which
> takes some time when the matrices get large). If you know in advance
> that you're going to create a matrix that will not be modified (in my
> case: the factors of a system) but that you will use (eg for
> backsolves), it would be nice to be able to skip the overhead of
> creating an ll_mat.
>=20
> I see it as a limitation of PySparse right now because my C kernel
> isn't able to call directly  other low-level language subroutines that
> expect CSR or SSS matrices. Instead, I must take the long route
> through ll_mat land and that has time and computational consequences.
>=20
> I am not sure how to expose those functions though as tweaking the
> source code produces lots of errors and warnings. I must not be
> tweaking right.
>=20
> Dominique
>