RUSTICL radeonsi produces artifacts in blown out highlights in darktable-view with demosaicing set to 3xMarkesteijn+VNG

[kasav81]

Is there an existing issue for this?

• I checked and did not find my issue in the already reported ones

Describe the bug

When setting the demosaicing algorithm to 3x Markesteijn + VNG on an otherwise unedited image blown out highlight show artifacts when processed on GPU using RUSTICL with the radeonsi driver, no such artifacts appear when selecting the iris driver and the artifacts aren't there on the exported image and vanish once the center view is zoomed >50%.
My system has an i7-13000 with an iGPU and an AMD Radeon RX6600. Because of problems with the recommended rocm-opencl and intel-compute-runtime which led to crashes I switched entirely to RUSTICL for OpenCL in darktable.
The AMD card is set to process the center view and the iGPU to process the little preview / histogram. Exporting is mainly done via the AMD card. Exported images appear normal.
When other modules get involved the artifacts get worse, creating almost artistic abstract images, but still only in the center view and while zoomed less than 50%.
Admittedly not a big problem as all other demosaicing options appear to work flawlessly.

Steps to reproduce

1. Set RUSTICL_ENABLE=radeonsi and enable RUSTICL in dt preferences + fast gpu to make sure everything gets processed on gpu
2. Open Fujifilm RAW file with blown highlights
3. Select demosaicing 3xMarkesteijn+VNG
4. Artifact appear in center view, the little preview image in the left upper corner appears normal, artifacts disappear when zooming in over 50% and aren't visible in exported image
5. With RUSTICL_ENABLE=iris everything takes considerably longer but there are no artifacts.

Expected behavior

Center image should be displayed without artifacts on lower zoom levels

Logfile | Screenshot | Screencast

Screenshot immediately after opening and selecting 3xMarkesteijn+VNG

After enabling diffuse/sharpen

Zoomed in

exported image

output of RUSTICL_ENABLE=radeonsi darktable -d common -d opencl
dt5.0.1.txt

Commit

No response

Where did you obtain darktable from?

darktable.org / GitHub release

darktable version

darktable 5.0.1

What OS are you using?

Linux

What is the version of your OS?

Gentoo / Kernel 6.6.74-gentoo-dist (64-bit) / KDE Plasma 6.2.5, Frameworks 6.10.0 / X11

Describe your system

Darktable was compiled using a slightly modified ebuild to enable gmic-support and with all gentoo patches disabled

Are you using OpenCL GPU in darktable?

Yes

If yes, what is the GPU card and driver?

AMD RX6600 + intel iGPU, amdgpu with mesa 24.3.3

Please provide additional context if applicable. You can attach files too, but might need to rename to .txt or .zip

No response

[jenshannoschwalm]

Can you see anything "unexpected" when using any module with a blended mask using the "details threshold" ?

[kasav81]

Can you see anything "unexpected" when using any module with a blended mask using the "details threshold" ?

I don't use masks a lot.
I tried an instance of contrast equalizer on the image I posted above with a parametric mask and changed the details threshold, there are small changes within the artifacts, but they occur with almost all modules I enable and disable no matter if masking is used or not. The most obvious change occurs with diffuse / sharpen and filmic rgb. Using the standard 1x Markesteijn algorithm I haven't noticed any artifacts, neither with my bayer sensor raw files irrespective of the selected demosaicing options.

[jenshannoschwalm]

Sorry i wasn't more precise. Do you see anything wrong while a) using a demosaicer that looks good to you for example with ayer images and and b) some details mask?

[kasav81]

I'm sorry, that's what I tried to say in the last sentence - I have played around with several images (bayer and xtrans) trying all demosaicing algorithms with contrast equalizer enabled and the details threshold slider at different positions but haven't noticed anything strange, especially no obvious artifacts in the images.

[da-phil]

@jenshannoschwalm don't you usually use the AMD OpenCL tag for those kind of issues 😄

[jenshannoschwalm]

How could i forget :-) I didn't do yet as still suspecting a rusticl problem.

Before bothering any of the rusticl developers we need a better and for-sure minimal reproducer what algorithm exactly is the problem. a) VNG - only on xtrans images? b) Markesteijn 1 or 3 c) the dual blending d) the details mask.

[jenshannoschwalm]

Remember #18042

[kasav81]

I am not sure if this helps, but I found another way to produce artifacts in a XTrans-Raw with a minimal history stack when enabling the "liquify" module - in this case I have artifacts in blown out highlights with every demosaicing option except for "frequency domain chroma" and those artifacts are also visible in the exported image.
I also tried this with a Bayer-Raw and was able to provoke artifacts with a minimal history stack with "liquify" enabled with all demosaicing options except for "RCD", when using "passthrough (monochrome)" those artifacts seemed randomly distributed (but also appearing in the blown out highlights). Interestingly most of those artifacts in the Bayer-Raws where only visible in the exported image or when turning on high quality mode in darkroom view (except for "passthrough" which also showed artifacts in normal processing mode).
Also with the XTrans-Raws the artifacts look quite different when using high quality vs. normal processing.
I have also upgraded to mesa 25.0.0.

Bayer-Raw with "passthrough (monochrome)" and liquify enabled as seen in the preview

Same part of the exported image

XTrans-Raw with artifact in blown out highlights in darktable with only "liquify" enabled and Markesteijn 1-pass

Artifacts in the exported image

[github-actions]

This issue has been marked as stale due to inactivity for the last 60 days. It will be automatically closed in 300 days if no update occurs. Please check if the master branch has fixed it and report again or close the issue.