Periodic index spaces

[streeve]

Add local grid index space generation for periodic boundaries that matches the current boundary index space concept. This completes all current boundary types:

• System boundary (non-periodic): boundaryIndexSpace
• Periodic "system boundary": periodicIndexSpace
• MPI boundary (not a system boundary): sharedIndexSpace

TODO: add tests to avoid issues similar to #615 due to assumed "equivalent" boundary spaces

Closes #762

[streeve]

@JStewart28 I still need to add tests, but wanted to push this and make sure this was in the right direction

[JStewart28]

sharedIndexSpace and boundaryIndexSpace look good. Question about periodicIndexSpace: If the offset is in a shared index space along a periodic boundary, boundaryIndexSpace returns, for example, for i min/max and j min/max: (0, 2), (9, 12), and periodicIndexSpace returns (0, 2), (9, 11) for the same offset. Why is the upper bound for periodicIndexSpace one lower?

[streeve]

sharedIndexSpace and boundaryIndexSpace look good. Question about periodicIndexSpace: If the offset is in a shared index space along a periodic boundary, boundaryIndexSpace returns, for example, for i min/max and j min/max: (0, 2), (9, 12), and periodicIndexSpace returns (0, 2), (9, 11) for the same offset. Why is the upper bound for periodicIndexSpace one lower?

I found a mistake here - initially I wrote returned the boundaryIndexSpace, but by definition the periodicIndexSpace is just a sharedIndexSpace which happens to be on a boundary (it's shared with a neighbor rank). But I'm confused on your comparison (which may or may not be because of my previous mistake): isn't any given boundary either a periodic or non-periodic, where one of the two index spaces is empty?

[JStewart28]

sharedIndexSpace and boundaryIndexSpace look good. Question about periodicIndexSpace: If the offset is in a shared index space along a periodic boundary, boundaryIndexSpace returns, for example, for i min/max and j min/max: (0, 2), (9, 12), and periodicIndexSpace returns (0, 2), (9, 11) for the same offset. Why is the upper bound for periodicIndexSpace one lower?

I found a mistake here - initially I wrote returned the boundaryIndexSpace, but by definition the periodicIndexSpace is just a sharedIndexSpace which happens to be on a boundary (it's shared with a neighbor rank). But I'm confused on your comparison (which may or may not be because of my previous mistake): isn't any given boundary either a periodic or non-periodic, where one of the two index spaces is empty?

What I mean is for the same halo distance - in this case, 2 cells - and on a periodic boundary, in a periodicIndexSpace the j indices are 9 and 10, but in a sharedIndexSpace the j indices are 9, 10, and 11. sharedIndexSpace has the behavior we intend. Shouldn't the ghost index space be the same size for shared, periodic spaces whether or not they are on only an MPI boundary or an MPI+periodic boundary? Please correct me if I am missing something.