Rusticl Support For AMD RadeonSI Driver Being Worked On

[phoronix]

Phoronix: Rusticl Support For AMD RadeonSI Driver Being Worked On

Rusticl as the Rust-written OpenCL implementation for Mesa's Gallium3D as a newer and modern-focused CL alternative to the existing "Clover" code may soon see mainline support for working with the RadeonSI Gallium3D driver for supporting modern AMD graphics processors...

https://www.phoronix.com/news/Rusticl-RadeonSI-WIP

[zboszor]

In the same MR, there's already a question about using ACO for Rusticl in the future. Probably when the radeonsi driver itself was switched to ACO.

[tildearrow]

Typo:

This weekend Karol opened a Mesa merge request with "critical fixes" for the RadeonSI driver. This incliuded the ac/llvm code to support non-32-bit sized workgroup IDs, assigning locations before passing the NIR to drivers from Rusticl kernels, and Rusticl limiting the global memory to 2GB.

[Alliancemd]

This could make it more beginner friendly, knowing that they don't have to deal with memory management bugs.
Let's see where it goes but it sounds exciting

[Michael]

Typo:

Thanks

[Black_Fox]

Would Rusticl be able to run Stable Diffusion on AMD GPUs instead of ROCm?

[_r00t-]

I'd love to see Davinci Resolve running on Linux with Rusticl.

[dragorth]

Would Rusticl be able to run Stable Diffusion on AMD GPUs instead of ROCm?

There are implementations of Stable Diffusion for both pyTorch, the most common one, and a new tensorflow based one. There are OpenCL backends for both of these frameworks, but I don't know if any of the available models are trained on them. My guess would be not, since most of the cloud hosting where these are tested are Nvidia GPUs.

You could create your own version that is for OpenCL specifically, however.

Here is that new tensorflow version, in case you didn't see it.

https://github.com/divamgupta/stable-diffusion-tensorflow

[karolherbst]

I'd love to see Davinci Resolve running on Linux with Rusticl.

I looked into it, but it needs some CL extensions, which also require some more work: cl_khr_gl_sharing and cl_khr_image2d_from_buffer. Especially the former is quite a bit of work.

[Quackdoc]

Can't wait to test this