Here's a simplistic solution, but one which might provide the most users with what they're expecting, which I think is important for a component like DMStag. It simply changes the behavior of the functions to set uniform coordinates to set all local coordinates, including ghosts, using linear extrapolation. In the periodic case, that means that the redundant boundary coordinates are what you would expect (in this case, 2pi insteaf of 0). https://gitlab.com/petsc/petsc/-/merge_requests/3804 <https://gitlab.com/petsc/petsc/-/merge_requests/3804>

Am 01.04.2021 um 16:37 schrieb Jed Brown <[email protected]>:

Matthew Knepley <[email protected]> writes:

...

We use it to identify that the mesh is periodic and in what directions, and use the length if we have to figure out the coordinates.

...

Jed may argue that he wants you to retain the far point and use L2G to eliminate it, but that sounds like a lot more work.

Computational work or implementation work?

implementation.

I think this is more implicit DMPlex assumptions than a fundamental issue.

...

...

I already have logic in DMStag to support something related, which was required for L2G with INSERT_VALUES in the periodic, 1-rank case, where multiple local points can respond to the same global point. So building the local coordinates with the far point and then doing L2G doesn't sound bad to implement, at least.

https://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/DMSTAG/DMStagPo...

That may work. It is a problem in Plex because that strategy destroys the topological queries. However, it sounds like you do not have that problem.

Patrick, if you're still in early stages, I would encourage you to support nodally affine maps in the L2G, which allows things like rotated periodicity for vector-valued fields.

https://gitlab.com/petsc/petsc/-/issues/566