[petsc-users] Fieldsplit PC null pointer error on getksp
Mark Adams
mfadams at lbl.gov
Thu Jul 16 11:10:48 CDT 2015
On Thu, Jul 16, 2015 at 6:33 AM, Sander Land <sander.land at gmail.com> wrote:
> Yes, I figured I'd need to rewrite my dirichlet bc implementation to zero
> out the row rather than remove the dof completely. For now I'm just
> avoiding partial bcs.
> Using MatSetBlockSize is a problem because the matrix has stuff after the
> xyzxyz (hydrostatic pressure dofs and a few coupled things).
> MatGetSubMatrix inherits the block size of 1, and I can't change it
> afterwards.
>
Humm, this needs to work. Are you saying you get the (sub) matrix with an
IS of the non-pressure part? You should be able to set that block size to
3. If this get submatrix happens inside of FieldSplit then I think there
is a solution for this.
>
>
> Sander
>
> On Thu, Jul 16, 2015 at 12:26 AM, Mark Adams <mfadams at lbl.gov> wrote:
>
>> Note, you want to tell the matrix that it has these blocks with
>> MatSetBlockSize, and do not remove a single DOF (like for BCs). That is,
>> xyzxyxyz is not allowed. The block size is critical for GAMG and would
>> think HYPRE uses it also because it does OK on elasticity.
>>
>> On Wed, Jul 15, 2015 at 2:52 PM, Barry Smith <bsmith at mcs.anl.gov> wrote:
>>
>>>
>>> > On Jul 15, 2015, at 5:43 AM, Sander Land <sander.land at gmail.com>
>>> wrote:
>>> >
>>> > Hi Barry,
>>> >
>>> > I have some structures set up that map dofs to their respective matrix
>>> indices, and it wasn't difficult to change them.
>>> >
>>> > However, now this call:
>>> > MatMultAdd(ctxp->B,xp,bx,xxtmp);
>>> > is directed to "MatMultAdd_SeqAIJ_Inode" and gives a vector lock error
>>> at line 601 of mat/impls/aij/seq/inode.c
>>> > ierr = VecGetArray(zz,&z);
>>> > I think this is a bug and should be VecGetArrayRead as it doesn't need
>>> to be written to.
>>>
>>> Oh yes, this is a bug. I have fixed it in maint, master and next
>>> >
>>> > In the old xxxyyyzzz ordering this call went to MatMultAdd_SeqAIJ,
>>> which (via VecGetArrayPair) does use VecGetArrayRead. I'm not sure what is
>>> causing the different matrix formats.
>>>
>>> When the values are interlaced the matrix has "identical nodes"
>>> (i-nodes) which means certain rows have identical nonzero structure with
>>> there neighbors and the AIJ matrix code automatically switches to the more
>>> efficient Inode operations. This is why you only saw the bug with
>>> interlacing.
>>> >
>>> > I'm also still not sure at what point I can change the matrix block
>>> size to get the correct BoomerAMGSetNumFunctions.
>>> >
>>> > Thanks,
>>> > Sander
>>> >
>>> >
>>> >
>>> >
>>> >
>>> > On Tue, Jul 14, 2015 at 6:34 PM, Barry Smith <bsmith at mcs.anl.gov>
>>> wrote:
>>> >
>>> > > On Jul 14, 2015, at 11:37 AM, Sander Land <sander.land at gmail.com>
>>> wrote:
>>> > >
>>> > > Hi Matt,
>>> > >
>>> > > Doing things to the PC like this has worked for me for a long time.
>>> It just happened to be formerly KSPPREONLY/PCLU. Setting things like
>>> GMRES/PCNONE works as well with no errors, though it converges far too
>>> slowly.
>>> > >
>>> > > The actual preconditioner I am trying to form is one from recent
>>> literature and uses a slightly modified Schur complement split embedded in
>>> another Schur complement split with multiple solves of the inner system and
>>> very specific choices for sub-preconditioners. I don't think command line
>>> options will work very well here.
>>> > >
>>> > > I've since gone the way of just using PCSHELL and setting things up
>>> more explicitly. This seems to work, but i'm not quite sure how to set up
>>> boomeramg for one of the sub-problems.
>>> > >
>>> > > boomeramg is used on the matrix A derived from:
>>> > > MatGetSubMatrix(J,is_x ,is_x ,MAT_INITIAL_MATRIX,&ctxp->A);
>>> > > The degrees of freedom for J are xxxyyyzzzpw, with the size of
>>> xxxyyyzzz around 20k, p around 1k, and w 8 for my 'small' test problem.
>>> > >
>>> > > I can see that HYPRE_BoomerAMGSetNumFunctions takes its value from
>>> the matrix block size. This is 1 for J as it doesn't have a consistent
>>> block size. However, the number of functions for boomeramg should be 3.
>>> > > Changing the block size is not allowed by petsc, and there doesn't
>>> seem to be another way to reach HYPRE_BoomerAMGSetNumFunctions.
>>> > > What is the best way to handle this, and does it matter that the
>>> dofs are ordererd xxxyyyzzz rather than xyzxyzxyz ?
>>> >
>>> > I don't think BoomerAMG supports xxxyyyzzz, I think it will just
>>> get very confused trying to do algebraic multigrid with this form. Read up
>>> on it but my guess is you need to rewrite the code that prepares the system
>>> to do the interlacing.
>>> >
>>> > Barry
>>> >
>>> > >
>>> > >
>>> > > Thanks,
>>> > > Sander
>>> > >
>>> > >
>>> > >
>>> > > On Fri, Jul 10, 2015 at 7:44 PM, Matthew Knepley <knepley at gmail.com>
>>> wrote:
>>> > > On Tue, Jun 30, 2015 at 3:10 PM, Sander Land <sander.land at gmail.com>
>>> wrote:
>>> > > This gives
>>> > > [0]PETSC ERROR: Object is in wrong state
>>> > > [0]PETSC ERROR: Matrix must be set first
>>> > >
>>> > > The PC is from a SNESGetKSP, and I have called SNESSetJacobian
>>> before this.
>>> > >
>>> > > If this is embedded in a SNES, it is very messy to yank it out and
>>> do thing manually. However, if you
>>> > > set the matrix in the PC manually, it might work.
>>> > >
>>> > > However, I would encourage you to try an alternate scheme and see if
>>> you think it is easier. We have
>>> > > started to use the DM object to hold this kind of information. For
>>> FieldSplit, I think it is easier to tell
>>> > > a DMShell about your field division, and then have it pass this on
>>> to PCFIELDSPLIT. So you would
>>> > >
>>> > > a) Create a DMSHELL
>>> > >
>>> > > b) Create a PetscSection and call DMSetDefaultSection()
>>> > >
>>> > > c) Call SNESSetDM()
>>> > >
>>> > > d) use command line options to configure the split
>>> -pc_fieldsplit_0_fields 0 -pc_fieldsplit_1_fields 1
>>> > >
>>> > > Then everything should work, and can work in any embedded context.
>>> Does this make sense?
>>> > >
>>> > > Thanks,
>>> > >
>>> > > Matt
>>> > >
>>> > > Thanks,
>>> > > Sander
>>> > >
>>> > > On Tue, Jun 30, 2015 at 2:09 PM, Matthew Knepley <knepley at gmail.com>
>>> wrote:
>>> > > On Tue, Jun 30, 2015 at 11:43 AM, Sander Land <sander.land at gmail.com>
>>> wrote:
>>> > > I am trying to use the Schur complement preconditioner in petsc, but
>>> am encountering a null argument error calling PCFieldSplitGetSubKSP.
>>> > > This only happens on PC_COMPOSITE_SCHUR, the multiplicative/additive
>>> options do return a KSP array.
>>> > >
>>> > > Error message:
>>> > >
>>> > > [0]PETSC ERROR: --------------------- Error Message
>>> --------------------------------------------------------------
>>> > > [0]PETSC ERROR: Null argument, when expecting valid pointer
>>> > > [0]PETSC ERROR: Null Object: Parameter # 1
>>> > > [0]PETSC ERROR: See
>>> http://www.mcs.anl.gov/petsc/documentation/faq.html for trouble
>>> shooting.
>>> > > [0]PETSC ERROR: Petsc Release Version 3.6.0, Jun, 09, 2015
>>> > > [0]PETSC ERROR: ./elecmech on a arch-linux2-cxx-debug
>>> > > [0]PETSC ERROR: Configure options --with-shared-libraries=1
>>> --with-debugging=1 --download-openmpi=1 --with-clanguage=c++
>>> --download-fblaslapack=1 --download-scalapack=1 --download-blacs=1
>>> --download-suitesparse=1 --download-pastix=1 --download-superlu=1
>>> --dowload-essl=1 --download-ptscotch=1 --download-mumps=1 --download-lusol=1
>>> > > [0]PETSC ERROR: #1 MatSchurComplementGetKSP() line 317 in
>>> petsc-3.6/src/ksp/ksp/utils/schurm.c
>>> > > [0]PETSC ERROR: #2 PCFieldSplitGetSubKSP_FieldSplit_Schur() line
>>> 1264 in petsc-3.6/src/ksp/pc/impls/fieldsplit/fieldsplit.c
>>> > > [0]PETSC ERROR: #3 PCFieldSplitGetSubKSP() line 1668 in
>>> petsc-3.6/src/ksp/pc/impls/fieldsplit/fieldsplit.c
>>> > >
>>> > > Code snippet:
>>> > >
>>> > >
>>> ISCreateGeneral(PETSC_COMM_SELF,xp_dofs,&ii[0],PETSC_COPY_VALUES,&is_xp);
>>> > > ISCreateGeneral(PETSC_COMM_SELF,all_dofs -
>>> xp_dofs,&ii[xp_dofs],PETSC_COPY_VALUES,&is_wk);
>>> > >
>>> > > PCSetType(pc,PCFIELDSPLIT);
>>> > > PCFieldSplitSetType(pc,PC_COMPOSITE_SCHUR);
>>> > > PCFieldSplitSetIS(pc,"xp",is_xp);
>>> > > PCFieldSplitSetIS(pc,"wk",is_wk);
>>> > > int n;
>>> > > KSP* ksps;
>>> > > PC subpc;
>>> > > PCFieldSplitGetSubKSP(pc,&n,&ksps);
>>> > >
>>> > >
>>> > > ii here is simply an array with ii[i] = i. There is probably a
>>> better way to simply indicate two blocks of different size, but I couldn't
>>> find it.
>>> > >
>>> > > It looks like the PC is not setup. Can you try calling PCSetUp(pc)
>>> before GetSubKSP()?
>>> > >
>>> > > Thanks,
>>> > >
>>> > > Matt
>>> > >
>>> > > Thanks,
>>> > > Sander
>>> > >
>>> > >
>>> > >
>>> > > --
>>> > > What most experimenters take for granted before they begin their
>>> experiments is infinitely more interesting than any results to which their
>>> experiments lead.
>>> > > -- Norbert Wiener
>>> > >
>>> > >
>>> > >
>>> > >
>>> > > --
>>> > > What most experimenters take for granted before they begin their
>>> experiments is infinitely more interesting than any results to which their
>>> experiments lead.
>>> > > -- Norbert Wiener
>>> > >
>>> >
>>> >
>>>
>>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20150716/20dfc09c/attachment.html>
More information about the petsc-users
mailing list