[RESOLVED] Darktable render speed with openCL


i got my Framework 13 with Ryzen 7640 2 Weeks ago and quite happy. I like the device and i think it was a good decission. Almost everythink worked out of the box even with Linux Mint 21.2. The known fingerprint issue was fixed soon - thanks to the community.

But one point i am dissapointed about is the rendering speed with darktable. Compared to my 7 years old DELL 7559 with Nvidia 960M i expected a noticable speedup but - Pustekuchen. CPU-rendering is slightly faster then the old NVidia combination, but when i enable openCL there is a massive slowdown. Exporting an image takes 3-4 times longer (20 seconds) than with CPU only, mouse pointer movement is laggy and the device gets quickly really hot. Might be an driver issue, but what is your experience?

Greetings, Marco

Check out this thread, maybe you need more recent ROCm (and I think it is only included with AMD proprietary drivers)?

Then again there is this:

“The integrated GPUs in AMD APUs are not officially supported targets for ROCm.”

Thanks for your reply. I found this page too but its almost one year old.

I did not expect such bad performance, since clinfo and rcominfo both show all devices correctly and Darktable reports it is using openCL. I have the latest ROCm 5.7.1 installed. When ROCm does not officially support AMD integrated GPUs, what is then the best way get them working with darktable?

You can try this but I haven’t tested it: export HSA_OVERRIDE_GFX_VERSION=11.0.0

To make ROCM think that your 780M is a 7900XTX which is supported. Other than that you will have to wait until official support. Or you could just use it on Windows which has a working implementation (not a good answer I know).

1 Like

Try rusticl: https://wiki.archlinux.org/title/GPGPU#Rusticl
Works for my AMD dedicated card. Not sure what the needed package name is in Mint. Needs mesa >=23.1, I think.

Thanks for your replies.

i’ve done some more testing and decided to buy an additional 16GB memory module. Never before i experienced more the need to dual channel memory since this is my first device with an iGPU. Available GPU memory is now 4GB instead 2GB and Geekbench6 results raised 20% to 30%. But more important, Darktable runs much much better now. Even installed Rocm 6.0 yesterday and everything seems to be fine.

Hope that findings helps other.

PS: Darktable received some updates in the meantime to extend openCL support. I think this had impact too.

1 Like

Thanks to the community for their help, I will mark this as resolved.