Vega 64 + Linux: Is OpenCL 2.0 supported?

[blueberry]

I am currently using the old Catalyst drivers on Arch Linux with R9 290X. My programs require some OpenCL 2.0 features, mainly on-line in-memory kernel compilation, non-uniform work group sizes, and device enqueue_kernel.

Recently I have acquired RX Vega 64, but I still did not try it because I'm confused about OpenCL 2.0 support. It does not help that AMD drivers are in state of flux, with all possible combinations of OS+AMDGPU(PRO)+ROCm+PAL and whatnot. From various online sources I understood that some parts of OpenCL 2.0 are supported in post-Catalyst drivers, but not on all hardware and not on all operating systems.

Is there any software combination on Linux, open or closed source, that supports OpenCL 2.0 (at least the features that I mentioned) on AMD Vega 64?

[bridgman]

I believe full OpenCL 2.0 is coming fairly soon - I'll see if I can publish more details.

At the moment I believe we have 1.2+ runtime and 2.0 kernel language on Linux.

[blueberry]

I have seen that 1.2+ runtime and 2.0 kernel language in ROCm documentation, but, unfortunately, there are important parts that are missing in that 2.0 kernel language; notably enqueue_kernel. Did you refer to ROCm in your message, or something else? I was hoping that (some of) the still closed-source drivers have more complete support for OpenCL 2.0.

Anyway, it'd be great if you can publish at least some rough estimation for OpenCL 2.0...

[bridgman]

I was referring to something else - the PAL implementation shipped with the packaged drivers.

I'm surprised the ROCm version doesn't have kernel enqueue - maybe that's because kernel enqueue is hard on all the other platforms even though it should be easy on ROCm (shader code can write command packets directly to a user queue without host involvement). I'll see what I can find out.