[petsc-dev] DMMG without libpetscsnes?

[Jed Brown]

On Thu, 7 Jan 2010 14:52:41 -0600, Barry Smith <bsmith at mcs.anl.gov> wrote:

>     There would be no SNES in this context!  One would be able to  
> cleanly and simply (more so then today) put geometric multigrid inside  
> a Schure fieldsplit preconditioner. If one can't then we screwed up  
> our new design.

I think we're agreeing.  My comment was just a retort to the claim that
linear problems are silly.  They are silly if that's your end goal, and
I believe they are still silly when using a scheme where the quality of
your solution depends on the quality of the linear solve (as in a
semi-implicit scheme), but having a nice interface for linear MG is
important when it goes inside a preconditioner.  In contrast, I think
that all special treatment of linear problems could be stripped out of
TS with no ill consequences.

Jed