dolfin team mailing list archive
-
dolfin team
-
Mailing list archive
-
Message #13035
Re: uBLASKrylovMatrix and Python callbacks: now the?fun begins
On Wed, Apr 15, 2009 at 08:30:17AM +0200, Johan Hake wrote:
> On Tuesday 14 April 2009 17:28:07 Anders Logg wrote:
> > On Tue, Apr 14, 2009 at 10:17:06AM -0500, Robert Kirby wrote:
> > > In most Trilinos packages, you also have the opportunity to set these
> > > parameters in code (though I think it's a case-by-case procedure rather
> > > than having a common Parametrized class). So there could be three
> > > methods for setting parameters:
> > > 1.) Directly in code
> > > 2.) a parameter list via XML
> > > 3.) command line
> > >
> > > I'm not sure if there are any advantages to having both 2 & 3 in place,
> > > but am happy to be persuaded.
> >
> > 1. Obvious
> >
> > 2. Good for storing default values system-wide and on a per-user basis in
> > say ~/.dolfin (but less obvious that it will be useful). But maybe most
> > of all to store the set of parameters used for a particular simulation
> > and remember those parameters.
> >
> > 3. Good for quickly changing an option without needing to edit any
> > file (either code or XML).
> >
> > The first option is already in place. The second would require little
> > work: XML input/output already working, just need to decide on a
> > location for the XML files. The third would require some work but not
> > that much.
>
> Looks good. I think it is worth while to have a breif look at what diffpack
> do, as Kent suggests. The command line parser would be nice if it could read
> out what parameters are defined and then define valid options based on these.
> I do not know if this is possible though :P
>
> With nested parameters I ment hierarchical parametrization. I did not know it
> was in place already! Well, then I use my own system in python instead ;)
If foo and bar are parametrized, then foo can make bar parametrized by
foo by doing
bar.set("parent", foo);
This is typically used for parametrizing member variables. For
example, VariationalProblem is a parent to its linear solver by doing
solver.set("parent", *this);
So when the linear solver requires a parameter and it's not found in
the solver itself, the parent is asked, then its parent etc until
finally the global parameter value is used.
--
Anders
> > Anyone is welcome to have a go at these, but I have limited time right
> > now myself.
>
> Unfortunatly I neither have the time...
>
> Johan
Attachment:
signature.asc
Description: Digital signature
References