On Aug 25, 2013, at 11:32 PM, Jed Brown <[email protected]> wrote:
> Barry Smith <[email protected]> writes: >> Not that axpy is DIFFERENT_NONZERO_PATTERN which means a very slow >> MatAXPY(). If I understand things correctly after the first time >> through here *A will have a nonzero pattern at least as big as Arhs? > > Yes, but what happens when Arhs is reassembled with a different nonzero > pattern? PetscErrorCode (*TSRHSJacobian)(TS,PetscReal,Vec,Mat*,Mat*,MatStructure*,void*); Which is available as flg2 in TSComputeIJacobian() So the only thing missing is know that this is a repeat call to > >> If so how can the code be organized so that slow MatAXPY() is not used >> for an eternity and switches to using subset for much faster >> performance? > > We just need a reliable way to know when the nonzero pattern of Arhs > changes.
