David, It might be easier to just rebuild the whole matrix from scratch: you would in effect be doing all that with disassembling and resetting the preallocation. MatSetType(mat,MATMPIAIJ) or PetscObjectGetType((PetscObject)mat,&type); MatSetType(mat,type); followed by MatXAIJSetPreallocation(...); should do. Dmitry.
On Sun Feb 22 2015 at 4:45:46 PM Barry Smith <[email protected]> wrote: > > Do not call for SeqAIJ matrix. Do not call before the first time you have > preallocated and put entries in the matrix and done the > MatAssemblyBegin/End() > > If it still crashes you'll need to try the debugger > > Barry > > > On Feb 22, 2015, at 4:09 PM, David Knezevic <[email protected]> > wrote: > > > > Hi Barry, > > > > Thanks for your help, much appreciated. > > > > I added a prototype for MatDisAssemble_MPIAIJ: > > PETSC_INTERN PetscErrorCode MatDisAssemble_MPIAIJ(Mat); > > > > and I added a call to MatDisAssemble_MPIAIJ before > MatMPIAIJSetPreallocation. However, I get a segfault on the call to > MatDisAssemble_MPIAIJ. The segfault occurs in both serial and parallel. > > > > FYI, I'm using Petsc 3.5.2, and I'm not using a non-debug build (though > I could rebuild PETSc in debug mode if you think that would help figure out > what's happening here). > > > > Thanks, > > David > > > > > > > > On Sun, Feb 22, 2015 at 1:13 PM, Barry Smith <[email protected]> wrote: > > > > David, > > > > This is an obscure little feature of MatMPIAIJ, each time you > change the sparsity pattern before you call the MatMPIAIJSetPreallocation > you need to call MatDisAssemble_MPIAIJ(Mat mat). This is a private > PETSc function so you need to provide your own prototype for it above the > function you use it in. > > > > Let us know if this resolves the problem. > > > > Barry > > > > We never really intended that people would call > MatMPIAIJSetPreallocation() AFTER they had already used the matrix. > > > > > > > On Feb 22, 2015, at 6:50 AM, David Knezevic < > [email protected]> wrote: > > > > > > Hi all, > > > > > > I've implemented a solver for a contact problem using SNES. The > sparsity pattern of the jacobian matrix needs to change at each nonlinear > iteration (because the elements which are in contact can change), so I > tried to deal with this by calling MatSeqAIJSetPreallocation and > MatMPIAIJSetPreallocation during each iteration in order to update the > preallocation. > > > > > > This seems to work fine in serial, but with two or more MPI processes > I run into the error "nnz cannot be greater than row length", e.g.: > > > nnz cannot be greater than row length: local row 528 value 12 > rowlength 0 > > > > > > This error is from the call to > > > MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz); in > MatMPIAIJSetPreallocation_MPIAIJ. > > > > > > Any guidance on what the problem might be would be most appreciated. > For example, I was wondering if there is a problem with calling > SetPreallocation on a matrix that has already been preallocated? > > > > > > Some notes: > > > - I'm using PETSc via libMesh > > > - The code that triggers this issue is available as a PR on the > libMesh github repo, in case anyone is interested: > https://github.com/libMesh/libmesh/pull/460/ > > > - I can try to make a minimal pure-PETSc example that reproduces this > error, if that would be helpful. > > > > > > Many thanks, > > > David > > > > > > > > >
