Fix strides for 0-sized arrays

[alexfikl]

Found one last issue with 0-sized arrays in inducer/arraycontext#91.

This was messing up the code generation on the pytato side and it would fail in the invoker when checking strides.

[kaushikcfd]

Thanks, could you please also add a tiny regression that fails on current main.

[alexfikl]

could you please also add a tiny regression that fails on current main.

Definitely! Added one in bd5ed02.

That fails on main because cl.array.to_device reuses the strides that numpy gives (here (8, 8)) and loopy computed the strides differently (here (0, 8)), resulting in a mismatch.

[kaushikcfd]

LGTM, thanks!

[inducer]

I'm not loving this, on multiple levels.

• If we merge this, the cost of the conditional is incurred on potentially every single array access.
• I'm not sure why this logic needs to live here (and why code in the invokers don't suffice). Could you explain?
• Negative strides are totally fine: np.arange(10)[::-1].strides -> (-8,) (Execution targets don't support negative strides #504)
• Zero strides are also, in principle, well-defined.

[alexfikl]

I'm not sure why this logic needs to live here (and why code in the invokers don't suffice). Could you explain?

Mostly because this was the direction I went down the rabbit hole. I'll work on putting it in the invokers, since there's no reason that wouldn't work, as far as I can tell.

Negative strides are totally fine: np.arange(10)[::-1].strides -> (-8,) (#504)

Do any of the other targets support them? Haven't really tried that, but agree that in principle there's no reason why they can't work.

Zero strides are also, in principle, well-defined.

When would those come up? Haven't managed to convince numpy to make some, although for 0-sized arrays, just setting ary.strides to something shouldn't break the world, right?

[inducer]

Do any of the other targets support them?

🤷 No reason they shouldn't, though. :)

When would those come up?

stride_tricks? As intermediates in broadcasting? Not sure.

[alexfikl]

@inducer Moved the stride stuff to the invokers in 23f8e7b.

It still feels quite hacky and/or incomplete, but it works™. Let me know what you think.

[alexfikl]

Superseded by #584.