Putting a peculiar multilevel scheme into DMMG?

[Jed Brown]

On Sun, 13 Dec 2009 18:06:04 -0600, Matthew Knepley <knepley at gmail.com> wrote:
> I will explain the history since I am to blame for this one. I needed
> coarsenHierarchy because repeated coarsening can cause degradation of
> the meshes. Thus all coarse meshes must be made at once. I added the
> hierarchy option, but only the coarsen branch was used.

I suspected something like that.  So one real difference with hierarchy
is that (in principle) DMRefine can move to a different communicator.
Is this flexibility used (it seems like this could be quite powerful for
additive multigrid, but I'm not aware of such a thing working now)?  If
so, DMRefineHierarchy and DMCoarsenHierarchy should also take an array
of communicators (passing NULL would mean to use the same communicator).

> > DMCoarsenHierarchy is implemented for Mesh, but the array is currently
> > leaking (DMMGSetDM forgets a PetscFree).  Is it indeed preferred that
> > the caller does not allocate the array (despite knowing how much is
> > needed) but is responsible for freeing it (I ask because this is clumsy
> > for a single level of refinement).  Either way, I'll document the choice
> > and fix the leak.
> >
> 
> Cool. I guess we could have caller allocation, but that is harder to
> check for correctness.

The alternative (unless we want two ways to do the same thing) seems
much worse:

  DM *tmp;
  DMRefineHierarchy(dmc,1,&tmp);
  dmf = *tmp;
  PetscFree(tmp);

Jed