Roadmap for CUDA / SciPy compatibility?

[sgbaird]

What needs to occur to integrate this into numba-scipy?
Are the versions of Numba supported by numba-scipy lagging behind stable versions of Numba?

[esc]

What needs to occur to integrate this into numba-scipy?

Hard to say. Perhaps we should start with a better definition of "this". Do you want numba-scipy to support CUDA?

Are the versions of Numba supported by numba-scipy lagging behind stable versions of Numba?

They shouldn't be, I guess. Is there something that creates this impression?

[sgbaird]

To your first question, yes.

At minimum, I imagine this would involve updating the CI and deciding on organization and design rules.

That's good to know. I think at first when I came to this repository about a month ago, the sparse documentation and long times since previous commits may have given that impression (at first, I don't think I realized this was an official Numba repo - though it's obvious to me now that it is), and at least compared with the main Numba repo. It has a bit of an impression of not being very actively maintained (obviously the core devs have a lot going on with many repos, Numba/Numba being top priority I imagine).

I didn't spend too much time on here originally because it didn't have the functionality I needed (distance metrics).

[sgbaird]

numba/numba#7368 (comment)

[JaRoSchm]

Hi,

I'm interested in using numba-scipy on the GPU too. Is there some workaround to use a scipy.special function (in my case only jv is needed) inside of a @cuda.jit function? Even a solution which is not generally applicable to all scipy functions would help me.

Thanks in advance!

[JaRoSchm]

Or could I somehow use the implementation from CUDA Math in numba? https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH__DOUBLE.html#group__CUDA__MATH__DOUBLE_1g643b020a4c66860acc8c0f0a76f7b67e
Or the implementation in mpmath? https://mpmath.org/doc/current/functions/bessel.html