`numba-cuda` `DeviceNDArray` deprecation

[brandon-b-miller]

Description of new feature

Hi AwkwardArray maintainers! numba-cuda dev here. There's been some efforts recently to split off the pieces of numba-cuda that aren't pure compiler functionality, and as part of that we're moving away from supporting numba-cuda's array library features. In particular, after NVIDIA/numba-cuda#546 is released, you'll observe deprecation warnings using numba-cuda APIs that create DeviceNDArray objects, such as those used in this set of awkwardarray tests.

We're recommending users migrate to alternative libraries that are purpose built for working with GPU arrays, such as CuPy, when allocating device data, etc. There looks to be some cupy usage in awkward already, and I'm wondering if it'd be feasible to simply switch to using CuPy arrays in the relevant tests. I am happy to work on a PR for this if that seems feasible.

[ianna]

@brandon-b-miller - thanks for heads-up! We use numba-cuda to lower ak.Array type (e.g. its C++ generated hashed view -- every array has a different internal structure and a different type) to GPU. It's a no-copy -- a few pointers numba-cuda extension -- that we pass to the jit decorated function. It allows us to pass the ak.Array to jited functions. I'm not sure how we can do that with CuPy. I think, it's a different thing. We should discuss it. You are welcome to contribute a PR.

[brandon-b-miller]

Hi @ianna, that makes sense, thanks for the context. I think ak.Array should be safe and unaffected by the deprecation. The change should only affect the tests that use to_device and copy_to_host to pass a "result" array that's later inspected. I am hoping a cupy array can serve this purpose instead, generally, numba-cuda kernels should work the same way for any passed data structure that supports __cuda_array_interface__.