Matt,
How difficult would it be to rework DMPLEX to allow the use of VecGhost? We have performance problems with GPUs with simple DMNETWORK models because the code spends more time uselessly copying the local part of the vector to another vector in global to local and local to global; more than 1/2 the time of the total simulation.
Firedrake already does this because they "vec ghost" their vectors by default. Here is what you need:
When you create the PetscSection, by default it orders the unknowns according to the default point numbering. This
is what causes the ghost unknowns to be mixed in with the local unknowns. However, PetscSection allows you to set
a point permutation
This determines the order of dogs by iterating through points in this permutation, and you can put all shared points at the end.
How do I know what are shared points to put at the end? Couldn't DMPLEX do this automatically with an option? Where is the Firedrake code that does this with DMPLEX so I can see it?
Yes, it could. I was thinking this as I wrote you, and also thinking how tired I am :)
Here is how to do it. Suppose you only care about ghost points (Firedrake wants three layers: totalling internal, roots, and ghosts), then there is a simple test
DMGetLocalSection(dm, &s)
PetscSectionGetChart(s, &pStart, &pEnd);
DMGetPointSF(dm, &sf);
PetscSFGetGraph(sf, NULL, &nleaves, &leaves, NULL)
for (p = pStart; p < pEnd; ++p) {
PetscFindInt(p, nleaves, leaves, &loc);
if (loc >= 0) {I am a ghost point}
}
If you also care about roots (shared points that you own), you also need
PetscSFComputeDegreeBegin(sf, &rootdegree);
PetscSFComputeDegreeEnd(sf, &rootdegree);
and then also check
rootdegree[p] > 0
If you get it working, I will put it inside Plex and make an option to turn it on.
Thanks,
Matt
Does this make sense?
Thanks,
Matt
Barry
--
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