OpenCL

[neworderofjamie]

I've been working on adding OpenCL support to our code generator (GitHub - genn-team/genn: GeNN is a GPU-enhanced Neuronal Network simulation environment based on cod... ) and the generated code is now working on NVIDIA and Intel devices but, on AMD devices, it is failing rather mysteriously. 

I have reduced the attached on-GPU initialization code to a minimal reproduction case. If the printf in the kernel is commented out, then initializeKernel correctly initializes d_xPost and it is displayed correctly. If not nans are printed.

We have reproduced this issue on an iMac with the following clinfo:

Platform Name Apple
Platform Vendor Apple
Platform Version OpenCL 1.2 (Apr 18 2019 20:03:31)

Device Name AMD Radeon Pro 570 Compute Engine
Device Vendor AMD
Device Vendor ID 0x1021c00
Device Version OpenCL 1.2
Driver Version 1.2 (Jan 23 2020 07:52:24)

And a Windows machine with 3004.8 drivers and a RX 580.

Am I doing something stupid or is this a compiler bug? If so any workarounds would be much appreciated.

Thanks in advance for your help

[dipak]

Thank you for reporting it and providing the reproducible test-case. I've whitelisted you and moved this post to OpenCL forum.

It looks like a new driver (Adrenalin 20.5.1) is available for Windows10. Could you please try it once and share your observation? 

By the way, I ran the above code on my laptop with Adrenalin 20.5.1, but could not reproduce the issue.

For any OpenCL issue on Mac, please contact Apple support team.

Thanks,

[neworderofjamie]

Thank you for the prompt response - newer drivers do indeed seem to solve these issues on Windows! I have posted in the Apple developer forums, but not yet received any response - you wouldn't happen to have any ideas of a better means of contacting Apple OpenCL support?

[dipak]

Thank you for this update.

Regarding your query related to Apple OpenCL support, I think, you have already posted in the right place. If you don't get any response, please check if there is any other option to directly submit a bug report against this issue.

Thanks.