dolfin team mailing list archive

[Robert Kirby <kirby@xxxxxxxxxxxxxxx>]

On Oct 26, 2004, at 5:07 PM, Anders Logg wrote:

> On Tue, Oct 26, 2004 at 02:00:30PM -0500, Matthew Knepley wrote:
>
> >      I, of course, would advocate using an IDL (since that is what I 
> > do).
> > However, I really want to take issue with the idea that we NEED 
> > pieces in
> > C++. If SWIG has shown anything, its that the implementation language 
> > should
> > not be important and wrappers are fairly easy. Much more important is 
> > the
> > interface and the packaging.
> >
> >               Matt
>
> That's a good point. We don't need to have any code in C++.
> But there are good reasons for providing a C++ *interface*:
>
> If we cannot provide a complete toolset that covers every problem,
> people will have to implement some parts of the code themselves, and
> people prefer to use C++. With people I mean people like ourselves in

*some* people prefer C++.  Other people prefer some variant of fortran. 
 Others will want a bash interface...Hence, providing C is nice since 
most languages have figured out how to play well with it.  Then, if 
people want to add stuff on their own, they can do whatever they want.  
We have to be careful when we start accepting contributions back from 
the general public.

> the business of developing new algorithms and methods for PDEs, but not
> necessarily in the business of developing tools. I don't think you can
> say too people who want to implement a new algorithm that they should
> go ahead and create a low-level library for that feature and then
> create a wrapper for their new library.

What do we mean by "algorithm"?  do we mean sparse data access, linear 
solver, or variational form?  The granularity of algorithm affects 
where people have to do their writing.

> It all depends on what you want to do. If you want to create the best
> possible tools, then optimized low-level libraries with wrappers might
> be the best way. But if you want to make your system
> (user-)extensible, I would like to see a C++ interface, since it
> allows you to implement fairly efficient new algorithms in the same
> language as the interface of the toolset you are using.