[petsc-users] multigrid implementation problem

[Jed Brown]

On Sun, Jun 10, 2012 at 7:54 PM, Nakib Haider Protik <nprot048 at uottawa.ca>wrote:

> Okay, so here is the code that works. This gives the correct result, but I
> want to make sure that it is not giving the correct result accidentally.
>

1. I don't know how to emphasize this enough: you need to learn C to write
C code.

2. Please run in Valgrind, it will show you several memory leaks here.

3. Please run with -ksp_view so you can see what gets used. If you did
this, it would be obvious that multigrid was not being used.


> Thank you very much for your time.
>
>    KSPCreate(PETSC_COMM_WORLD, &ksp);
>    KSPCreate(PETSC_COMM_WORLD, &cksp);
>    KSPCreate(PETSC_COMM_WORLD, &uksp);
>    KSPCreate(PETSC_COMM_WORLD, &dksp);
>

^^ You only need to create ksp here, delete the other lines


>    KSPSetType(ksp, KSPGMRES);
>    KSPSetType(cksp, KSPGMRES);
>    KSPSetType(uksp, KSPGMRES);
>    KSPSetType(dksp, KSPGMRES);
>

^^ You don't need any of these


>
>    KSPGetPC(ksp, &pc);
>    KSPGetPC(cksp, &pc);
>    KSPGetPC(uksp, &pc);
>    KSPGetPC(dksp, &pc);
>

^^ delete all but the first of these


>
>    KSPSetFromOptions(ksp);
>
>    PCSetType(pc, PCMG);
>    PCMGSetLevels(pc, 2, PETSC_NULL);
>    PCMGSetType(pc, PC_MG_MULTIPLICATIVE);
>    PCMGSetCycleType(pc, PC_MG_CYCLE_V);
>    MatDuplicate(A, MAT_COPY_VALUES, &P);
>    PCMGSetCyclesOnLevel(pc, 0, 1);
>    PCMGSetCyclesOnLevel(pc, 1, 1);
>
>     PCMGGetCoarseSolve(pc, &cksp);
>    PCMGGetSmootherDown(pc, 0,  &dksp);
>    PCMGGetSmootherUp(pc, 1, &uksp);
>

^^^ delete the last three lines here, they aren't doing anything


>
>    PCMGSetInterpolation(pc, 1, P);
>    PCMGSetRestriction(pc, 1, P);
>    PCMGSetResidual(pc, 0, PCMGDefaultResidual, P);
>    PCMGSetResidual(pc, 1, PCMGDefaultResidual, P);
>

There is no way that you want to be using P for interpolation/restriction
and as the residual on levels and as the preconditioning matrix on the line
below. The matrices are not even the right size. The reason the above runs
is because nothing you have done with "pc" is being used: it refers to a
"dksp" which you are leaking without using.


>
>     KSPSetOperators(ksp, A, P, SAME_NONZERO_PATTERN);
>     KSPSolve(ksp, breal, xreal_harm);
>    KSPSolve(ksp, bimag, ximag_harm);
>    //////////////////////////////////////////////////////////
>
> > On Sun, Jun 10, 2012 at 7:32 PM, Nakib Haider Protik
> > <nprot048 at uottawa.ca>wrote:
> >
> >> Okay. But if I understood correctly, doing the following would also be
> >> wrong since I will have made three extra outer KSP objects. I am not
> >> sure
> >> I understand what these functions want.
> >>
> >>  KSP cksp, uksp, dksp;
> >>  PCMGGetCoarseSolve(pc, &cksp);/**/
> >>  PCMGGetSmootherDown(pc, 0, &dksp);/**/
> >>  PCMGGetSmootherUp(pc, 1, &uksp);/**/
> >>
> >
> > This is fine, all of these are getting references to inner objects
> without
> > overwriting the outer objects. Of course you don't need to get these
> > references if you aren't going to modify them.
> >
> > I recommend that you experiment with algorithms using the command line
> and
> > check that it does what you expect using -ksp_view.
> >
>
>
> --
> Nakib :)
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120610/c43add10/attachment.html>