[ci] Run Linux OpenCL tests against POCL instead of the AMD App SDK

[jgiannuzzi]

As discussed in #5252 (comment), I think that using PoCL for the Linux gpu tests would make a lot of sense, as it is open source and actively maintained, as opposed to the AMD App SDK.

The dask tests are currently disabled for the gpu task, but they in my local testing the do succeed with PoCL. Should we maybe (re-)enable them?

[StrikerRUS]

The dask tests are currently disabled for the gpu task, but they in my local testing the do succeed with PoCL. Should we maybe (re-)enable them?

Here is the reason why they are disabled: #3708 (comment). I think for this PR we shouldn't enable them to not delay #5252. In a follow-up PR we'll definitely try to re-enable them and run multiple times to ensure their stability.

[StrikerRUS]

How about Windows, is PoCL available there?

[StrikerRUS]

Could you please update GPU Targets Table and OpenCL SDK Installation with PoCL in this PR as well?

[StrikerRUS]

Thank you so much for this PR!

Just some minor comments below.

[jgiannuzzi]

How about Windows, is PoCL available there?

It's available, but unfortunately not as an ICD, which means that we wouldn't be able to test the integrated OpenCL builds, and that the other OpenCL builds would not test something representative.

[StrikerRUS]

It's available, but unfortunately not as an ICD,

OK, got it! Thank you for the explanation.

[jgiannuzzi]

Could you please update GPU Targets Table and OpenCL SDK Installation with PoCL in this PR as well?

I have updated the GPU Targets Table. I haven't updated the Windows docs, as I don't think we can recommend using PoCL on that platform.

[StrikerRUS]

I haven't updated the Windows docs, as I don't think we can recommend using PoCL on that platform.

Given this your previous comment #5282 (comment), this sounds reasonable. Thank you!

[StrikerRUS]

Awesome! Thanks so much!

[jameslamb]

Love this change, thanks very much for all the hard work @jgiannuzzi !

[StrikerRUS]

The dask tests are currently disabled for the gpu task, but they in my local testing the do succeed with PoCL. Should we maybe (re-)enable them?

Here is the reason why they are disabled: #3708 (comment). I think for this PR we shouldn't enable them to not delay #5252. In a follow-up PR we'll definitely try to re-enable them and run multiple times to ensure their stability.

Unfortunately, with PoCL Dask tests fail with OSError: dlopen: cannot load any more object with static TLS error in both 14.04 and 20.04 Ubuntu.
#5285 (comment)

[jgiannuzzi]

Unfortunately, with PoCL Dask tests fail with OSError: dlopen: cannot load any more object with static TLS error in both 14.04 and 20.04 Ubuntu. #5285 (comment)

My understanding is that the dask tests only fail with this error on Ubuntu 14.04 with Python 3.10. I ran them locally with Python 3.8 and 3.9 and they work fine. Do we maybe want to set PYTHON_VERSION again on the Linux gpu_source job?

On Ubuntu 20.04, the tests seem to fail because of a timeout. Enabling those tests on a fake GPU is expected to take more time. Maybe setting timeoutInMinutes like we do for the aarch64 job would help?

I'm happy to create a PR for this if you'd like.

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.