On Fri, Jan 23, 2009 at 3:19 PM, Derek Gaston <fried...@gmail.com> wrote:
> Ok - all good arguments... and I'm seeing some of your reasoning now.... but
> let me show you the use case I'm thinking of which might better explain why
> I think the LinearSolver link is useful...
>
> I'm needing to solve individual systems _just_ for preconditioning
> purposes... and they use a preconditioner _as a solver_. If a
> Preconditioner was _not_ a LinearSolver this would be what I would have to
> do (and is close to what I'm doing now... although there is even more manual
> manipulation).
>
> ApplyPreconditioner(Vector x, Vector & y)
> {
> LinearImplicitSystem sys;
> sys.add_variable("u");
>
> AssemblePrecMatrix(sys.matrix);
>
> copyCorrectPiecesOfX(x, sys.rhs);
>
> Preconditioner pre(sys.matrix);
> pre.apply(sys.rhs, sys.solution);
>
> copyCorrectPiecesOfSolution(sys.solution, y);
> }
>
> This really doesn't look too bad... the problem is that there's
> datastructures that aren't used... or are redundant. For instance there's a
> KSP object inside of that LinearImplicitSystem that never gets used... and
> it has it's own PC object... again that never gets used. Then there's the
> fact that you have to call apply() and pass in two things that are both held
> within the system.
I know that's a simple example but it really doesn't look too bad. It
seems that x, and y are only temporary, so if they became members of
Preconditioner and you called the ApplyPreconditioner function solve()
instead, I think it would be very close to what you want...
Also, I wouldn't worry too much about unused PC/KSP objects. They
can't possibly take up much memory or take much time to build
up/destroy, can they?. In any event, an ImplicitSystem may be more
applicable in your example (assuming it's not a virtual base class, I
didn't check!) if System::solve() is *never* to be called. The System
hierarchy doesn't get a LinearSolver until the LinearImplicitSystem
level.
--
John
------------------------------------------------------------------------------
This SF.net email is sponsored by:
SourcForge Community
SourceForge wants to tell your story.
http://p.sf.net/sfu/sf-spreadtheword
_______________________________________________
Libmesh-devel mailing list
Libmesh-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/libmesh-devel
