non-local values being dropped in MatSetValues

[Barry Smith]

    Stephan,

    I'm sorry for your wasted time finding this bug.  I have fixed it  
in the Mecurial version of PETSc 3.0.0 and in petsc-dev. It will be  
fixed in the next 3.0.0 patch that we release.

    Barry

On Jul 17, 2009, at 7:44 AM, Stephan Kramer wrote:

> Hello
>
> We've spend sometime debugging a problem were in the assembly of a  
> parallel MPIAIJ matrix, some values that were created on a process  
> other than the owner of the row seemed to disappear. I think I  
> narrowed it down to what I think is a bug in MatSetValues_MPIAIJ,  
> but please tell me if I'm wrong.
>
> The situation is the following: I'm calling MatSetValues with the  
> flag ADD_VALUES and with matrix option MAT_IGNORE_ZERO_ENTRIES. I'm  
> inserting  multiple values at once, multiple columns and rows, so I  
> provide a rank-2 matrix of values. As I'm calling this from fortran  
> I'm also using MAT_COLUMN_ORIENTED. Now for provided rows that are  
> not owned by the process, it jumps to mpiaij.c:394 (line numbers as  
> in petsc-dev). On line 399, it checks for zero entries, but only  
> checks the very first entry of the (non-owned) row. If however other  
> entries of that same row are nonzero, the entire row is still  
> dropped. Note that this is independent of row_oriented/ 
> column_oriented as line 396 does exactly the same.
>
> If I don't set the option MAT_IGNORE_ZERO_ENTRIES the problem  
> disappears. In that case however we would either have to preallocate  
> substantially more nonzeros, or complicate the matrix assembly in  
> our code by taking out the zero entries ourselves and call  
> MatSetValues for each entry seperately.
>
> Your help would be much appreciated,
> Cheers
> Stephan
>
>
> -- 
> Stephan Kramer  <s.kramer at imperial.ac.uk>
> Applied Modelling and Computation Group,
> Department of Earth Science and Engineering,
> Imperial College London