KSPSolveTranspose and PCFieldSplit

[Barry Smith]

On Tue, 26 Sep 2006, Randall Mackie wrote:

> Barry,
> 
> Okay, I assume this means it will be done at some point in the near future?
> 
> I tried taking the transpose in my code, and then using KSPSolve, but then
> I was getting error messages about zero pivots. I then added
> -pc_shift_positive_definite and it ran, but it didn't seem to do any better
> than KSPSolveTranspose, although I haven't been able to spend a lot of time
> on that.
> 
> I'll just add another KSP in my code, one to use the fieldsplit PC for
> the first problem, and then another one to just use ILU for the transpose
> problem, at least until you get the patch for fieldsplit PC.

  To use it you'll have to use the development copy of PETSc
http://www-unix.mcs.anl.gov/petsc/petsc-as/developers/index.html
or wait for our 
next release which is a few months away.

  Sorry,

     Barry

In general there is nothing that says if B is a good preconditioner for 
A that B' is a good preconditioner for A', though usually you would think
it would be.


> 
> Thanks, Randy
> 
> 
> Barry Smith wrote:
> > 
> > On Tue, 26 Sep 2006, Randall Mackie wrote:
> > 
> > > When I try to use KSPSolveTranspose with PCFieldSplit, I get the following
> > > errors:
> > > 
> > > [0]PETSC ERROR: PCApplyTranspose() line 483 in
> > > src/ksp/pc/interface/precon.c
> > > [0]PETSC ERROR: No support for this operation for this object type!
> > > [0]PETSC ERROR:  !
> > > [0]PETSC ERROR: KSPInitialResidual() line 56 in
> > > src/ksp/ksp/interface/itres.c
> > > [0]PETSC ERROR: KSPSolve_BCGS() line 41 in src/ksp/ksp/impls/bcgs/bcgs.c
> > > [0]PETSC ERROR: KSPSolveTranspose() line 562 in
> > > src/ksp/ksp/interface/itfunc.c
> > > 
> > 
> >    This is just "lazyness", I hadn't implemented the apply transpose. I've
> > added it now to PETSc-dev; the problem is it cannot be made into an
> > easy patch since it involves several nontrivial changes.
> > 
> >     
> > > So then I thought in my code I would just take the transpose of A and call
> > > KSPSolve, but then I get the following errors:
> > > 
> > > [0]PETSC ERROR: MatGetSubMatrix_SeqAIJ() line 1686 in
> > > src/mat/impls/aij/seq/aij.c
> > > [0]PETSC ERROR: Nonconforming object sizes!
> > > [0]PETSC ERROR: Cannot reuse matrix. wrong no of nonzeros!
> > > [0]PETSC ERROR: MatGetSubMatrices_SeqAIJ() line 1799 in
> > > src/mat/impls/aij/seq/aij.c
> > > [0]PETSC ERROR: MatGetSubMatrices() line 4828 in
> > > src/mat/interface/matrix.c
> > > [0]PETSC ERROR: MatGetSubMatrix() line 5438 in src/mat/interface/matrix.c
> > > [0]PETSC ERROR: PCSetUp_FieldSplit() line 162 in
> > > src/ksp/pc/impls/fieldsplit/fieldsplit.c
> > > [0]PETSC ERROR: PCSetUp() line 798 in src/ksp/pc/interface/precon.c
> > > [0]PETSC ERROR: KSPSetUp() line 234 in src/ksp/ksp/interface/itfunc.c
> > > [0]PETSC ERROR: KSPSolve() line 334 in src/ksp/ksp/interface/itfunc.c
> > > 
> > > 
> >     I'm guessing you used MAT_SAME_NONZERO_PATTERN? But the transpose won't
> > have the same nonzero pattern?
> > 
> > BTW: Did you find explicitly forming the transposes and solving worked
> > better
> > than just using KSPSolveTranspose()?
> > 
> >   Barry
> > 
> > > Is there some reason why the PCFIeildSplit is not compatible with
> > > KSPSolveTranspose?
> > > 
> > > I'm solving with the following options:
> > > 
> > > 
> > >          -em_ksp_truemonitor \
> > >          -em_ksp_type bcgs \
> > >          -em_pc_type fieldsplit \
> > >          -em_pc_fieldsplit_default \
> > >          -em_pc_fieldsplit_type multiplicative \
> > >          -em_fieldsplit_pc_type bjacobi \
> > >          -em_fieldsplit_sub_pc_type ilu \
> > >          -em_fieldsplit_sub_pc_factor_levels 4 \
> > >          -em_fieldsplit_sub_pc_factor_fill 10 \
> > > 
> > > Randy
> > > 
> > > 
> > 
> 
>