On Sun, Feb 26, 2012 at 21:10, Hong Zhang <[email protected]> wrote:
Vaclav:
I feel that it is quite incorrect that MatMerge calls MatDestroy(&inmat) since this corrupts "garbage collection" - inmat is destroyed but only nullified in MatMerge. It remains unchanged outside so subsequent MatDestroy call fails. I think that there is no reason for this destroy, if am right, MatMerge does not corrupt the inmat so it can be reused and then destroyed manually.
MatMerge(): Creates a single large PETSc matrix by concatenating local sequential matrices from each processor. It was written for petsc internal use (private function), and only works for mpiaij format. Thanks for the suggestion. I'll remove MatDestroy(&inmat) from MatMerge().
I think it should also be renamed to something like MatCreateMPIAIJConcatenateSeqAIJ(). It doesn't make semantic sense for general matrices because those need not be stored using a row partition.
The second question is - what is the principal difference between MatMerge and MatMerge_SeqsToMPI - I am confused a bit.
MatMerge_SeqsToMPI - Creates a MPIAIJ matrix by adding sequential matrices from each processor .
And I have finally one note about MatMerge_SeqsToMPI manual page: there is instead a synopsis of MatMerge_SeqsToMPINumeric because this comment is located above the MatMerge_SeqsToMPINumeric function in source.
I fixed this in petsc-3.2, and will merge to petsc-dev.
We may need rename these two functions for better understanding.
MatCreateMPIAIJSumSeqAIJ()?