Bcast object wrappers, attempt 2

[inducer]

Second attempt at #280, much simpler technically.

I'm mostly seeking early feedback on this. It still needs, at least:

• Documentation
• Tests
• Fixed types

This could help fill the hole that's left by the plan to make implicit broadcasting of data class containers across actx array types illegal. See inducer/grudge#377 for an example use, specifically shortcuts.py. (Note that I've updated the grudge PR to use this.)

[majosm]

Noticed one thing. Otherwise, I'm on board. 👍

[majosm]

One other thing that just occurred to me: do we need to add some unpacking for the case where the broadcastee is itself a container? (For example, adding a cv of scalars to a cv of elems x dofs data. Should there be some code that recurses into the broadcastee and produces a cv of Bcast*s or something like that?)

[inducer]

One other thing that just occurred to me: do we need to add some unpacking for the case where the broadcastee is itself a container? (For example, adding a cv of scalars to a cv of elems x dofs data. Should there be some code that recurses into the broadcastee and produces a cv of Bcast*s or something like that?)

That's a fair question. I think I would prefer to wait for a specific use case and keep it simple until that point. One happy fact about this approach is that we can introduce additional Bcast* objects with distinct behavior when needed.

[inducer]

FYI @majosm: Also added you as co-author. I used your initial code from #307 after all. 🙂 Forgot about that initially.

[alexfikl]

Looks good to me too! Worth adding the simple non-recursive Bcast version too?

[inducer]

Worth adding the simple non-recursive Bcast version too?

I'm not in favor. I brought over the tests from #280, replaced all the Bcast objects with the one introduced here, and that did the trick. So I'm not super sold we need anything else. The cost differential between the recursing and not ought to be a few isinstance checks---I think that's fairly minor in the grand scheme of things.

[inducer]

Thanks all for your help with this!