Running on Intel GPUs

[kris-rowe]

Is there interest in having support for running on Intel GPUs?
This could be done using the Julia OpenCL or oneAPI packages.

I am willing to help with development and benchmarking on this front.

[glwagner]

Does KernelAbstractions support Intel GPUs? If its supported by KernelAbstractions then support can be developed for Oceananigans. There's some work in progress for AMD GPUs now on #2540.

One challenge is finding resources for continuous integration testing.

[kris-rowe]

It seems like KernelAbstractions is in the process of adding oneAPI support. (See this PR)

On the CI front, it should be easy to verify builds work since the oneAPI toolkits can be checkout using apt or similar.
For a complete CI pipeline, I could chat with some contacts at Intel to see if it is possible to launch jobs to their DevCloud.

[francispoulin]

Sounds like fun!

[glwagner]

On the CI front, it should be easy to verify builds work since the oneAPI toolkits can be checkout using apt or similar.

But we need the hardware, right?

I could chat with some contacts at Intel to see if it is possible to launch jobs to their DevCloud.

That could be interesting!

[kris-rowe]

Linking this PR comment since it is a relevant data point:
#3475 (comment)

[kris-rowe]

There is a group at Argonne investigating the performance of Oceananigans on the Aurora supercomputer in case there are others interested in running on Intel GPUs as well.

[glwagner]

Curious to hear how that goes and if it makes sense to try to get an allocation on Aurora!

[francispoulin]

I am very interested to hear how this goes and happy to help set things up and even run things if you like.