AMD ROCm 5.2.1 Released For The Open-Source Radeon GPU Compute Stack

[JacekJagosz]

Beautiful 4 years of official support
On the other hand, Intel's compute support still works on Skylake (2015, 7 years of support and still going)

Actually the oldest supported architecture is Broadwell, so even a generation earlier! So Intel has definitely the best GPU support, but on the other hand, their compute stack was recently broken for a few months with the newest LLVM, and over all it could be better maintained. Also their support for Cycles is mostly on yet unreleased ARC GPUs, not even any of their iGPUs. But Intel is definitely going the right way, I wish AMD can catch up.

[Hibbelharry]

For intel and support for older hardware see this:

https://www.phoronix.com/news/Intel-CR-22.28.23726

[Mystro256]

For anyone unaware, ROCm OpenCL is in default Fedora repos now! No need to add the Copr or deal with a manual install. I imagine it won't have 5.2.1 immediately.

FYI:

https://bodhi.fedoraproject.org/updates/FEDORA-2022-922dcc59cd

I could patch the stack to allow for Polaris and prior, but the support is a bit limited and you can enable it by doing:

export ROC_ENABLE_PRE_VEGA=true

(or maybe it's 1 instead of true?)

I don't mind patching it in the Fedora package if there's demand, it's a one line change:

https://github.com/ROCm-Developer-Tools/ROCclr/blob/develop/utils/flags.hpp#L248

[JacekJagosz]

FYI:

https://bodhi.fedoraproject.org/updates/FEDORA-2022-922dcc59cd

I could patch the stack to allow for Polaris and prior, but the support is a bit limited and you can enable it by doing:

export ROC_ENABLE_PRE_VEGA=true

(or maybe it's 1 instead of true?)

I don't mind patching it in the Fedora package if there's demand, it's a one line change:
https://github.com/ROCm-Developer-To...flags.hpp#L248

AUR package, Debian and even Solus has been doing it for a while, with no problems so far, so I think it is worth to enable it, it shouldn't affect newer cards anyways.

[Mystro256]

AUR package, Debian and even Solus has been doing it for a while, with no problems so far, so I think it is worth to enable it, it shouldn't affect newer cards anyways.

The only issue I'm aware of is that some older HW, such as "Lexa" (RX 540/550/630/640) is really broken. So if you have two openCL implementations, ROCm and non-ROCm (say Mesa's OCL), where non-ROCm works but has a lower openCL capability, the OCL ICD will automatically select ROCm's OpenCL based on the higher capability (I think) and crash. As far as I know, this was the motivation to switching the default to false, but I might be making some incorrect assumptions.

Granted, I guess you could argue that someone can just export ROC_ENABLE_PRE_VEGA to disable this on a case by case basis.

Anyway, I'll follow suit with the rest of the distros on my next update.