Add 2d split unit test

[jdinan]

Here's a sketch of the unit test I had in mind for 2d split. It seems to work fine on the SHMEMX_TEAM_WORLD cases and crash on the even_team cases. It also fails with npes == 1 for both cases. I had expected this case to work, should it not?

[jdinan]

There were, indeed, some bugs in my unit test. I think they are mostly sorted out now, and hopefully the new PE index calculation makes more sense.

[davidozog]

I'm still seeing some errors here...

@jdinan - It looks like the problem is that xrange needs to be greater than npes. I don't think this is clearly specified in the spec though... When xrange > npes, the 2D split doesn't make sense to me... For example, if npes is 1 and xrange 2, do you expect just 1 x team? Or do you expect 2 x teams with the 2nd one equal to SHMEM_TEAM_INVALID?

[jdinan]

Oh, interesting. What does the spec say about xrange? I think I was expecting that xrange > npes would just result in a single xteam that is not full, similar to what happens when npes % xrange > 0.

[davidozog]

Yeah, it does kinda make sense for the xteam to be "unfilled" in this case, just like some yteams would be unfilled when npes % xrange > 0. My reading/understanding of the spec is that xrange is allowed to be (any?) non-negative number, but it's not explicit that the xteam is unfilled when xrange < npes. Is that your interpretation?

For what it's worth, I think it'd be simpler to require that xrange be less than or equal to npes, but perhaps that's not users' preference. I'll make a ticket to discuss this on the spec repo.

[jdinan]

Given that this is currently allowed, would something simple like if (xrange > parent->size) xrange = parent->size at the beginning of 2d split support this case?

[davidozog]

I think a one-liner like that wouldn't work because it messes with the "size" of the xteam (xsize here). But I'm sure something relatively simple can be done. I'll try it out.

Ha, nevermind - yeah I think it'll work...