Re: [Pysparse-developers] [Pysparse-users] reading the source code of pysparse

[Dominique O. <dom...@gm...>]

On Sat, Feb 11, 2012 at 04:30, Oz Nahum Tiram <na...@gm...> wrote:
> Hi,
>
> First, thanks for the great work on this peace of software. I have
> been using it for more than 3 years now and I very happy.
> I have a few question and I hope this is not
> I am still very new to C programming and I was wondering what does the
> following statement for expamle do:
>
> n2b = F77(dnrm2)(&n, b, &ONE);/* Norm of rhs vector, b */
>
> Is this some kind of type casting ?

This is a macro that adjusts the call to various types of Fortran
compilers (this particular call is a call to the BLAS library, which
is written in Fortran). Some Fortran compilers add a trailing
underscore to symbols, some add two, some add none, etc.

> The second question is, now that I have a working application with
> pysparse that I'd like to accelerate,
> I am thinking of rewriting my code in C. I know there are C solvers
> for bicgstab, but I was wondering how
> much work is it just to convert the bicgstab.c from pysparse to plain
> C without the python dependencies.
> Are there any hints here?

I wouldn't recommend coding bicgstab all over again. The only costly
operations in Bi-CGSTAB (and other Krylov-type methods) are vector
operations (most often, addition of vectors) and operator-vector
products (e.g., A*x or A.T*x). I would say that to speed things up,
you'll want to speed up your operator-vector operations; they are the
dominant cost.

You can take a look at PyKrylov (https://github.com/dpo/pykrylov)
which contains a pure Python implementation of Bi-CGSTAB and allows
you to input your operator in different ways (a Pysparse matrix being
one of them). For instance, you could implement your operator in C or
in Cython and that should speed things up. I believe that is the way
to go.

Of course, vector operations in PyKrylov could also be speeded up with
Cython. That's been on my list for a while.

I hope this helps.

-- 
Dominique