On Wed, Jan 2, 2013 at 3:21 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> Matt, you know how we always tell people not to use KSP (or even Lapack) > for 2x2 problems? Surely the same advice applies to SNES, but your point > location code is doing it. > Yep. The wrapping overhead is far less with some Newton iterations, and we have no good alternative. > Also, please remember to test that complex actually works before declaring > it fixed. I just cleaned up a bunch more stuff, including pain invoked by > SNES operating in PetscScalar for those 2x2 point location problems. > > > https://bitbucket.org/petsc/petsc-dev/commits/0213b4504a412b7304d6032d6047079603a8c2c8 > We need to get mixed-mode working. This kind of programming is far to inconvenient and ugly, especially for small side problems like these that are never ever ever intended to support complex. > To be vectorization-friendly, I would write point location to operate > directly on a batch of points, wrapped in the little Newton loop (not > SNES). Maybe this implementation of yours is a reference that you're > planning to replace? > Yes, this is for a very small number of point locations necessary for seismic sources, not intended for inter-mesh interpolation or anything that fills up the mesh. Matt -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20130102/52afcc43/attachment.html>
