Rewrite the ArbitraryDomain class.

Currently the ArbitraryDomain class only works when it is partitioned in z-direction. Rewrite it such that the global linear indexing works also with PARFFTX=TRUE and/or PARFFTY=TRUE.

@frey_m maybe related to the SAAMG parallelisation issue?

@adelmann Yes, according to the description. It is also mentioned in the list of #227.

I don't think that it's a valid argument to close an issue if it's in the list of #227. It's easy to forget if this issue is just on some list somewhere.

removed milestone

mentioned in merge request !371 (merged)

As discussed with @winklehner_d in !371 (comment 24025) the ArbitraryDomain is generally broken since OPAL-2.0. Hence, it would be good the rewrite also includes the fix of that bug.

mentioned in merge request !396 (merged)

mentioned in commit f48154af

added Bug label

changed milestone to %OPAL 2021.1

assigned to @winklehner_d and @frey_m

created merge request !406 to address this issue

mentioned in merge request !406

@gsell Were you able to reproduce the failing inside and intersection tests in the latest version (after merge)?

@winklehner_d I found some more issue. This time with the line-segment voxel intersection tests. There are special cases where it fails.

@gsell Quick question: If P lies directly on a boundary (i.e. |I - P| <= 1e-20 or whatever the number ends up being), is it counted as inside or outside? Asking because if it is counted as inside, I need to catch this special case during linear interpolation for the stencil creation, otherwise we have division by 0 there.

@winklehner_d if P is directly on a boundary, we have an intersection of the line segment defined by P and the reference point with a boundary. With the current implementation this means:

• P is considered to be outside if the reference point is inside
• P is considered to be inside if the reference point is outside

Another special case we have to deal with.

@gsell Ok, that is "good" behaviour for the SAAMG/ArbitraryDomain where P0 is inside. Then the distance |I - P| will never be 0 in the stencil calculation.

The reference point should always be inside or always be outside, we should make either one the standard, otherwise we set ourselves up for more errors. I vote for P0 always inside (for the above-mentioned reason).

@winklehner_d yes, the reference should always be inside. I'm currently adding more comments to the source code to make things more clear. I have to remember to add the case you mentioned (point exactly on the boundary).

@gsell where are with that? @winklehner_d any comments?

removed milestone

created branch 146-rewrite-the-arbitrarydomain-class-2 to address this issue