[petsc-dev] MatZeroRows(), MatZeroRowsIS() have new arguments

[Jed Brown]

On Wed, Oct 27, 2010 at 12:27, Barry Smith <bsmith at mcs.anl.gov> wrote:

> Right, but how do you know what values you should put on the diagonal?


You have to look at the mesh and/or compute coefficients on the neighboring
elements.  This is one reason I never use MatZeroRows, but it's possible for
someone to have the information available.


> Should there be an option to just leave the value that is currently there?


Maybe, I was just pointing out the case that I assumed Lisandro was
concerned with when he suggested it.


> Should we change it so we always just leave the value that is already
> there?


That would seem quite surprising to me.  They need the matrix scaling to be
consistent with however they evaluate residuals, and it seems like an
awkward bit of work to compute the exact matrix diagonal during residual
evaluation.

Jed
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20101027/2f9aad8b/attachment.html>