[petsc-users] MatRARt

Hong Zhang hzhang at mcs.anl.gov
Thu Jul 19 09:30:29 CDT 2012


> > Can you test MatRARt on
> > petsc-3.3/src/mat/examples/tests/ex94 with attached matrix data file by
> >
> > ./ex94 -f0 medium -f1 medium
> >
> > Do it crash?
> Yes, it does crash in the same exact fashion as the code I sent.  This
> is 3.3-p2.  petsc-dev also crashes, but I haven't pulled changes in
> about a month (and the server is down right now).

Hmm, ex94 runs well on my tests.
May I have your ~petsc/configure.log?

> > From flops analysis and experiments, MatRARt() is not as efficient as
> > MatPtAP,
> > likely costs twice as MatPtAP.
> I can just as easily construct P, so this isn't a big deal.  How would
> MatPtAP compare to a hypothetical MatRAP?  Is there a reason this
> routine doesn't exist?

Using petsc aij matrix format, sparse inner-product used in
A*Rt (MatMatTransposeMult)
costs twice as A*P (MatMatMult).
Out-product used in Pt*A (MatTransposeMatMult) requires extensive
data movement. I'm planning to implement
MatRAP (MatMatMatMult), but have not been able to work on it.
Suggestions are welcome.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120719/36db808c/attachment.html>

More information about the petsc-users mailing list