Mark, the problem was that the Inode structure was not learning about the diagonal change in MatScale(). Fixed now:
https://bitbucket.org/petsc/petsc-dev/commits/44a71c6d4b84406696d7c21549916099ada394b8 I also modified the example to get rid of all the preprocessor crud with the same functionality (now at run-time instead of compile time). https://bitbucket.org/petsc/petsc-dev/commits/a4ec749136a25a44f7afdd293fac7c273adfaed0 On Tue, Dec 18, 2012 at 4:02 PM, Mark F. Adams <mark.adams at columbia.edu>wrote: > I'm getting a floating point exception in the eigen estimator for cheby on > the 2nd solve of ex56 (you need to uncomment the define of TwoSolve) when I > use sor PC (runex56 uses jacobi so you need to change that). I see this on > Cray and my Mac. > > I also found that the eigen estimator was called before the first solve, > and worked fine, but I set the eigen values so I don't want it computing > eigen values. > > I can understand cheby being the eager beaver and recomputing eigen values > when I set a new matrix, but this not the semantics that I prefer. > > So there seem to be three issues: > > 1) seems the floating point exception. This only happens with a non-debug > build. I'm also using 64 bit indices. > > 2) computing eigenvalues even though I set them. > > 3) I would like to tell cheby not to recompute eigenvalues. > > Mark > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20130102/e679b3cb/attachment.html>
