Implement CupyArrayContext

[matthiasdiener]

Please squash

[matthiasdiener]

@inducer: We had briefly discussed adding support for object arrays via a wrapper (?), but I'm not sure how that should look like.

[matthiasdiener]

Turns out this was caused by the incomplete array container support, fixed in e9c42dd. Thanks to @majosm for the help!

[matthiasdiener]

This is ready for a first look @inducer

[alexfikl]

Was curious a bit about this and left some comments. Hope that's ok!

What's the use of adding a cupy backend? I thought eager ones don't really give good performance and we already have both the pyopencl and the jax ones that can use CUDA (behind some abstraction layers, at least).

[matthiasdiener]

Was curious a bit about this and left some comments. Hope that's ok!

What's the use of adding a cupy backend? I thought eager ones don't really give good performance and we already have both the pyopencl and the jax ones that can use CUDA (behind some abstraction layers, at least).

Thank you for the review!

I don't have experience with jax, and we generally don't run mirgecom with it. Regarding the pyopencl-eager backend(s), it does not support all operations that we need (e.g., broadcasting), so we can't run all our drivers with it. This only leaves us with numpy or pytato backends. pytato has generally high compile times, so it's sometimes not great for debugging etc. Numpy time step times are often much slower than with cupy, as indicated above.

[alexfikl]

I don't have experience with jax, and we generally don't run mirgecom with it. Regarding the pyopencl-eager backend(s), it does not support all operations that we need (e.g., broadcasting), so we can't run all our drivers with it. This only leaves us with numpy or pytato backends. pytato has generally high compile times, so it's sometimes not great for debugging etc. Numpy time step times are often much slower than with cupy, as indicated above.

Ah.. broadcasting.. the bane of everyone's existence! Fair enough :(

[matthiasdiener]

This is ready for another review @inducer @alexfikl

[matthiasdiener]

Out of curiosity, what are the implementations for debugging/profiling? (also mentioned in doc/)

[alexfikl]

I don't think there are any at the moment (?). Unless passing a CommandQueue with profiling enabled counts there..

It would be nice to have some fancier profiling for things though 😁

[alexfikl]

Left a little nitpick, but otherwise this looks good to me! Thanks!