Third Party Hardware > GPU Compute / Accelerators
AMD OpenCL / ROCm
Corvidae:
To potentially save someone from sinking tons of time into this when it might be doomed to fail: if you are planning on using a Polaris GPU, ROCm will not work on POWER currently due to the lack of PCIE atomics. According to this issue, Vega+ might be fine without them. I double checked if the lack of PCIE atomics was still an issue in 2023, and I would sadly say so:
--- Code: ---[ 14.145631] amdgpu 0000:01:00.0: amdgpu: PCIE atomic ops is not supported
[ 14.811332] kfd kfd: amdgpu: skipped device 1002:67c4, PCI rejects atomics 730<0
--- End code ---
chatcannon:
--- Quote from: MPC7500 on May 28, 2022, 05:01:06 pm ---By accident, I found this:
https://quickbuild.io/~raptor-engineering-public/+archive/ubuntu/rocm-power
--- End quote ---
I had a look at the packages included in the PPA. It seems that they have build the rocminfo tool but not any of the functionality to get ROCm OpenCL actually working. (Maybe I'm wrong there - I didn't look inside the deb packages, only at the package names.) So I tried it out on Gentoo and rocminfo now builds and runs with no modifications using the existing ebuild file.
rocm-opencl-runtime and rocclr still have arch-specific code for assembly and SSE intrinsics but maybe it is small enough to fix the remainder...
adaptl:
I tried ROCm years ago on this platform and it didn't work at all. Then I tried it on an x86 platform and it didn't work at all. Then I tried OpenCL and it was slow as hell.
My takeaway is that if you want to do reasonably portable GPGPU stuff learn WebGPU/Vulkan compute shaders. You can do also GPGPU stuff with opengles/webgl with "transform feedbacks" but its extremely limited and a total nightmare, speaking from experience.
Hasturtium:
--- Quote from: adaptl on February 15, 2025, 08:58:45 pm ---I tried ROCm years ago on this platform and it didn't work at all. Then I tried it on an x86 platform and it didn't work at all. Then I tried OpenCL and it was slow as hell.
My takeaway is that if you want to do reasonably portable GPGPU stuff learn WebGPU/Vulkan compute shaders. You can do also GPGPU stuff with opengles/webgl with "transform feedbacks" but its extremely limited and a total nightmare, speaking from experience.
--- End quote ---
OpenCL runs for contemporary Radeons in Fedora via RustiCL - managed to crack that nut a few months ago, but it took some doing. Clover is busted, as I think the firmware it expects to find hasn't been released for post-Navi hardware, and you have to blacklist it or it's likely to throw an error or hang your system outright.
lepidotos:
--- Quote from: adaptl on February 15, 2025, 08:58:45 pm ---My takeaway is that if you want to do reasonably portable GPGPU stuff learn WebGPU/Vulkan compute shaders. You can do also GPGPU stuff with opengles/webgl with "transform feedbacks" but its extremely limited and a total nightmare, speaking from experience.
--- End quote ---
Now if only the Blender team would do that. I do hear they are for viewport and Eevee, but Cycles depends on HIP if you're not using CUDA or CPU. Though I suppose render times shouldn't be too bad on the CPU in this case...
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version