On Aug 9, 2013, at 4:42 PM, Jed Brown <[email protected]> wrote:
> Barry Smith <[email protected]> writes:
>> Generally there should be some non zeros in each row, we could
>> probably just use 2.0*a->nz - m
>
> When using cprow, this should always be exact, right?
I believe so.
>
> If we are not using cprow because there weren't enough empty rows, but
> there are still some empty rows, it won't be exact, but it's likely not
> far off for typical matrices. It would be a shame, however, that we
> could now compute a negative number of flops. Should the check for
> cprow store the number of zero rows?
We don't actually check for compressed rows by default, presumably because
it takes a little time to scan through the rows (they are check for the
off-diagonal part of Mat_MPIXXAIJ()). But note that MatAssemblyEnd_SeqXXAIJ()
always has the loop
/* reset ilen and imax for each row */
for (i=0; i<mbs; i++) {
ailen[i] = imax[i] = ai[i+1] - ai[i];
}
so we could capture the count at that point. We could then use that value to
decide when to compress rows always and throw away the optional checking for
compressed rows. And use that value to get the accurate flop count for
MatMult_SeqXXAIJ()
I can make the change to master if this is the right approach?
Barry
