Attn @inazarenko

Any comments on this at all?

Can we not just support output.dim(2).set_stride(...) and co. on Output<Func> ?

Can we not just support output.dim(2).set_stride(...) and co. on Output ?

No, because the Output<Func> might not have any realization if the Generator has been inlined into another Generator.

Why can't output.dim(1).set_stride mean the same thing as output.dimensions_and_alignment().dim(1).set_stride? I.e. it gets ignored if there's no realization.

gets ignored if there's no realization

The thing I'm concerned about is that it's easy to use dim(0).extent(), etc mistakenly, and then get weird or incorrect results that are hard to debug. Let me see if I can reliably ensure that using these inappropriately will produce a hard+clear failure at compile time.

Maybe a bit pie in the sky, but I'd love to try to reduce/eliminate the inconsistency between internal and output stages in general, which might implicitly make this problem better.

For example, it seems like we should be able to use bound and reorder_storage to implement interleaved output buffers, the way you would for an internal stage. The fact that we have a different language for describing this for internal and external stages is unfortunate IMO and a source of confusion for myself at least.

Basically, I think it would be awesome to get rid of set_min/extent_stride completely, and use reorder_storage, align_bounds, and bound on output stages instead.

+1

I think it would be awesome to get rid of set_min/extent_stride completely, and use reorder_storage, align_bounds, and bound on output stages instead.

SGTM. (Do we have the ability to fully replicate everything we need already, or do we need to enrich the scheduling language a bit?)

That said: will all of these work on Funcs that may be inlined? (e.g. if I call align_storage on a Func with no storage?)

That said: will all of these work on Funcs that may be inlined?

In particular, Funcs wrapped around buffers in the implementation of Input<Func> are always inlined.