On 04/16/2014 09:44 AM, Jed Brown wrote:
Eric Chamberland <[email protected]> writes:
I was writing (new) code which do the firsts MatSetValues after the
MatXAIJSetPreallocation. We have to do this because the "real"
non-zeros will be added later by a mix of ADD_VALUES and
INSERT_VALUES... which would prevent us to "lock" the matrix
(MatSetOption(aMatricePETSc, MAT_NEW_NONZERO_LOCATION_ERR, PETSC_TRUE)
because this option must be passed after the first MatAssemblyEnd()...
but all the non-zeros are "triggered" only after this "mixed" assembly...
In other words, we have to do a "fake" assembly with all "0" to cover
all non-zeros that will be in fact added later... So we have to create
many "fake" elementary matrices to pass for assembly... that is why
Patrick was asking if the feature was supported... It would save us from
creating the fake matrices... Or maybe there is a better solution?...
Just create one large-enough buffer containing all zeros. For example,
PetscScalar *values;
PetscCalloc1(1000,&values);
for (...) {
MatSetValues(...,values,INSERT_VALUES);
}
PetscFree(values);
Is this okay?
Yes this can be ok.
I have a new question now : what would be the fastest way to do the
first assembly: with INSERT_VALUES or ADD_VALUES? Right now we are
doing ADD_VALUES but I imagine it have to hold all the "0" which will be
"added" later... maybe "INSERT_VALUES" will forget about overlapping
values??... Anyway, the code I am rewriting now will hold all the
non-zeros for each "block" of similar lines... so I don't have
overlapping values... in sequential... but in parallel I think I will
have some... Anyway, any pros/cons between INSERT_VALUES or ADD_VALUES
for this first assembly?
Thanks,
Eric
