Rusticl OpenCL Still Striving For Better Performance, SYCL & HIP Features

[phoronix]

Phoronix: Rusticl OpenCL Still Striving For Better Performance, SYCL & HIP Features

Karol Herbst at Red Hat who leads development on Mesa's Rust-written OpenCL "Rusticl" driver presented to share the progress made over the course of the year on this modern alternative to Gallium3D's Clover as well as some of the work still being pursued by this open-source OpenCL implementation for Gallium3D drivers...

https://www.phoronix.com/news/Rusticl-XDC-2023-Features

[Mathias]

What did they say about SyCL & HIP? Do they want to enable OpenCL for all SyCL/HIP targets or do they want to enable running SyCL/HIP kernels via Rusticl?

*edit* To answer my question:

chipStar is a tool for compiling and running HIP/CUDA on SPIR-V via OpenCL or Level Zero APIs.

So this actually lets you run HIP Kernels on any conformant OpenCL device. Since HIP is recompiled CUDA, this (chipStar) is quite interesting. They even have a section "Compilation of CUDA sources without changing the sources" in their documentation. Sweet. If RustiCL enables good OpenCL on many target architectures, it is theoretically possible to have universal CUDA compute on all devices.

Awesome. Don't wake me up from my dream and tell me this won't work well...

[dc_coder_84]

Hmmm... I wonder if this work will help with running folding@home on modern radeon GPUs? Anyboday knows?

[coder]

Keep up the good work!

[Jabberwocky]

Hmmm... I wonder if this work will help with running folding@home on modern radeon GPUs? Anyboday knows?

I wish I had answers. I only have questions myself.

I tried running a few of these projects 2 months ago on Windows using 6000 and 7000 series GPUs but didn't get any working. I went so far as to set "GPU only" under the scheduler but all the projects still used my CPU. I tried many things e.g. canceling existing jobs, reinstalling, setting profile settings on my account, reinstalling again... but couldn't get anything working on my GPU. I haven't tried the ROCm beta driver for Windows but that's another can of worms.

I hope this is easier under Linux with Rusticl but I have yet to try it. IIRC OpenCL is all you needed to run folding@home maybe things have changed?

[dc_coder_84]

I tried running a few of these projects 2 months ago on Windows using 6000 and 7000 series GPUs but didn't get any working.

Sorry to disappoint you but getting AMD GPUs to work on Linux with Folding@Home is super difficult and only few people managed to do so. Here you can find more info on the subject:

HOWTO: AMD Radeon RX 550 on Debian 10 with rocm driver
HOWTO: How I got my AMD card folding on Linux

I wasn't even able to download the necessary software because I don't use an LTS Ubuntu flavor. ^^

[Jabberwocky]

Sorry to disappoint you but getting AMD GPUs to work on Linux with Folding@Home is super difficult and only few people managed to do so. Here you can find more info on the subject:

HOWTO: AMD Radeon RX 550 on Debian 10 with rocm driver
HOWTO: How I got my AMD card folding on Linux

I wasn't even able to download the necessary software because I don't use an LTS Ubuntu flavor. ^^

Thanks for the info.

I would enjoy debugging this project if I had more time. For now I searched quickly. It looks like Folding@Home uses https://github.com/openmm/openmm . There's a lot of activity on the project. Some people even investigating HIP clang bugs: https://github.com/openmm/openmm/issues/4194

I'm not sure how the downstream process works. There's some info about the licensing here: https://foldingathome.org/support/fa...source/?lng=en quite abnormal but I understand given the burden around proof of work.

[marlock]

Folding@Home is sort of a golden standard but it's also far more restrictive than other similar initiatives, like BOINC, because it only runs projects that have more serialized steps, so each processing batch needs to be finished within a very narrow time constraint to enable the next round of batches to start, which leads to enabling participation by only a narrow selection of acceptable GPUs

Meanwhile BOINC has all sorts of loosely parallelizable projects adequate for running on all sorts of hardware, including cpu-only setups (when OpenCL is not available)

ps: there are a bunch of threads discussing why F@H isn't hosted inside BOINC yet, and this has been the target of a few attempts over the years

here is an old thread on Steam that discusses setting up F@H with OpenCL for AMD GPUs... (ah, covid... you're the worst...)

https://steamcommunity.com/app/221410/discussions/0/1862742067789070076/#c1862742137245661533

and be aware F@H does limit what GPU models the projects will run at, to a small high tier selection that can deliver fast results

[coder]

ps: there are a bunch of threads discussing why F@H isn't hosted inside BOINC yet, and this has been the target of a few attempts over the years

Didn't Google/DeepMind basically solve protein folding using a huge AI model?

https://deepmind.google/technologies/alphafold/

[marlock]

I didn't know that, thanks for the link!

From a cursory reading I imagine an AI model will need validation, etc, so the more deterministic approach is likely to continue for at least a while

I mean, we can at least assume no new folding problems would be submitted to the Folding@Home program if it wasn't still needed for some reason

I didn't read this anywhere yet, but my guess is they might reprioritize proteins with weirder setups instead of ones linked to prioritaty diseases... stuff that an AI trained on the current batch of data would be more likely to fold incorrectly because of it's more unusual builds