New scheduling directive to disallow partitioning.

[mcourteaux]

Solves #7881.
The changes are numerous but always the same. In summary:

• Just pass in the flag that sets allowing partitioning from the scheduling Dim to the For IR.
• For::make() takes extra argument allow_partitioning.
• All Lowing passes are updated to just pass this flag along or set it appropriately.
• Make an additional .disallow_partitioning(VarOrRVar) function in Stage.

I will start working with this now, but this can be reviewed. Leaving it at WIP for a while until I have tested some more with it.

[abadams]

After thinking about this, I agree that this should be a scheduling primitive, and I agree with the way you've implemented it. We should think pretty hard about the name and interface though, because we're stuck with scheduling primitives forever. I'll flag this to discuss at the dev meeting so we can brainstorm the name. Let me know if you'd like an invite (assuming the time zone works for you). It's every other Friday at 1pm pacific time, and the next one is next week on the inauspicious 13th. If you can't make it we'll report back here with our thoughts and we can continue the discussion in text form.

[abadams]

Any additions here also require updates to serialization and deserialization code. I suggest just copying how some other field is handled (i.e. dim_type).

[mcourteaux]

In general, one of my thoughts about this was that I'd probably prefer .allow_partitioning(), rather than to disallow it, but I implemented it like this for now to be backwards compatible with how scheduling works today. Even stronger, rather than "allowing", you might "instruct it", with a warning when it can't find a good partitioning split.

Regarding the dev meeting, sure, I'd be happy to attend that. For the timezone, 1pm pacific time is 10pm for me, so I could join every now and then when it's relevant. 🙂

[mcourteaux]

Yeah, I should compile locally with serialization enabled. Haven't tried that yet...

[abadams]

Maybe we want an enum rather than a flag in case new strategies for partitioning come up in future. Initially it would be Never, Always (which would error or warn as you suggest), and Auto (the default). Then the call might be something like .partition(y, Partition::Never)... I don't like the "Partition::" part of that very much though.

I guess we could also have partition_always(y), partition_auto(y). partition_never(y) as sugar.

[mcourteaux]

Sounds good. Perhaps PartitionPolicy with indeed Never, Always, Auto. The Policy mimics better the naming scheme of TailStrategy.