On Fri, Aug 31, 2018 at 2:20 PM Smith, Barry F. <[email protected]> wrote:
PETSc developers,
There are a variety of "interpolation" modules in PETSc but the documentation is scattered (mostly missing). Could everyone who knows anything about the various modules provide a little information about which modes exist as interfaces and which have actual supporting code and expected usage? Anything duplicative?
1) nested DM to DM (mesh to mesh) of the same type of DM seems to be supported for DMDA and DMPLEX using DMCreateInterpolation(). But what is DMPlexComputeInterpolatorNested(DM, DM, Mat, void *), how is it different? (Note usr context is not used)
DMCreateInterpolation() says nothing about nesting, and in fact lets people create arbitrary algebraic interpolation. For DMPlex, we support both nested and non-nested MG (Patrick says there is a subtle bug in non-nested), but my tests pass for this. The Nested case is obviously much easier and has a fixed element matrix, whereas non-nested uses point location.
2) non-nested DM to DM. DMPlexComputeInterpolatorGeneral(DM, DM, Mat, void *); (and what does "local portion" mean?) (also the usr context? is not used).
We can take out "local portion". That was me talking to myself.
3) DM to a set of points (mesh to points) with DMInterpolationInfo and the routines DMInterpolationEvaluate() etc. Is this fully implemented for DMPLEX, DMDA?
This relies on the DMLocatePoints(), which I think was implemented for both, but if it isn't for DMDA, its trivial and we should do it.
Parallel, does the user need to know which process the points are on or is that all figured out?
That is figured out. Actually, its most of the reason that this exists. However, I wrote this a long time ago. It should go away in favor of DMSwarm.
4) Points to a DM. Is this supported (should be?) by DMSWARM? In fact should 3) work with DMSWARM as the set of points and not have its own construct (DMInterpolationInfo)?
This is supported. However, what is currently in there is custom code Dave wrote which only works for P1. We now have code that does this for any element and in all dimensions. Its in a branch and will get merged shortly (next week) since it passes all tests. However, we need some more time to fully integrate it into the interface currently in DMSwarm. Another problem is that Dave has a different definition of points than the plasma people. We need a nice way to switch between these perspectives when doing interpolation, which is what will take us a little time when we are integrating.
5) Points to points? (Done indirectly by interpolating to a DM then back to the other points)?
Yes, that is how I would do it.
There is a routine DMPlexInterpolate() is this mis-named/confusing thing? Interpolate seems to mean something slight different here.
Yes. Unfortunately, the same word is used by topology people. I am willing to change this since very few people use it, when they do its only used once, and its completely different. It refers to figuring out the edges and faces automatically in a mesh when you get only cells and vertices. Better name? Thanks, Matt
All of this explanation could go into the users manual (or FAQ for now).
Thanks for any explanations,
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 https://www.cse.buffalo.edu/~knepley/ <http://www.cse.buffalo.edu/~knepley/>