Zooming out produces black artefacts

[kofa73]

Describe the bug

When using ctrl+scroll to zoom out, if rotate and perspective is used with cropping, black areas show up.

Steps to reproduce

1. Get the image and sidecar from https://tech.kovacs-telekes.org/files/2023-10-14-zoom-out/
2. Zoom in/out using Ctrl+wheel. For maximum effect, zoom in strongly, move the preview area to the edge of the image, then zoom in.
3. Observe the artefacts.

Expected behavior

Show no artefacts.

Logfile | Screenshot | Screencast

https://tech.kovacs-telekes.org/files/2023-10-14-zoom-out/capture.mp4

Commit

No response

Where did you install darktable from?

self compiled

darktable version

4.5.0+908~g84906440ed

What OS are you using?

Linux

What is the version of your OS?

Ubuntu 23.04

Describe your system?

X11, Nvidia

Are you using OpenCL GPU in darktable?

Yes

If yes, what is the GPU card and driver?

NVidia 1060/6GB, 525.125.06

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

Same behaviour with OpenCL on and off.
With OpenCL off, I get a huge log on the console, starting with:

157.1727 [dt_interpolation_compute_pixel4c] problem at (607,252) in 25fx348  1157.1728 [dt_interpolation_compute_pixel4c] problem at (608,252) in 25fx348  1157.1728 [dt_interpolation_compute_pixel4c] problem at (609,251) in 25fx348  1157.1728 [dt_interpolation_compute_pixel4c] problem at (607,253) in 25fx348
1157.1728 [dt_interpolation_compute_pixel4c] problem at (608,253) in 25fx348  1157.1728 [dt_interpolation_compute_pixel4c] problem at (609,253) in 25fx348  1157.1728 [dt_interpolation_compute_pixel4c] problem at (607,254) in 25fx348  1157.1729 [dt_interpolation_compute_pixel4c] problem at (608,254) in 25fx348 

[jenshannoschwalm]

1. Clearly a log formatting error here (missing newline and printing hex)
2. I recently added that debug log info, until then we silently wrote a zero. For some reason we are out of valid input data to do the pixelwise interpolation.

I tried here but can't reproduce so far. Could you check:

1. issue depending on any of the interpolators? (we had issues with lanc3 until recently)
2. depending on active modules (ashift or crop)?
3. depending on darkroom main window size or the border size?

[kofa73]

This is how the log looks like on the screen:

Interpolators were the default:

I get the same with both set to bilateral:

Same issue if I resize the window (from maximised on a 4K screen to small):

Simply turning off ashift (that's rotate and perspective, right?) produces:

Then, zooming in/out:

There are no new log entries in the terminal, though.

Turning rotate and perspective back on, with no cropping in the module produces interpolation error messages on the console. Zooming out fully only shows part of the image (compare the navigation preview with the editor preview -- this is with zoomed out as far as Ctrl+scroll goes):

[kofa73]

I'll try to bisect later today.

[pehar1]

It took a while until I could reproduce... , about one minute of zooming and panning.
I could reproduce occasionally (very rare) if

• in "Rotate and perspective" "Automatic cropping" is set to "Largest area" or "Original format" (independent of interpolator used, checked Lanczos3, Bilinear and Bicubic)
• the artefacts are dependent on the position of the mouse cursor on the canvas during CTRL-scrolling. Centering the mouse corsor on the canvas I never could reproduce, with positions near the edges it is much more likely. The black areas get larger the more one move the mouse cursor in the direction of the image corners during scrolling.

If I do not commit the "Largest area" cropping in "Rotate and perspective" and CTRL-scroll out, I get a very stange effect (possibly related) :

Set image size to "Fit" (ALT-3)

Move the mouse cursor clearly in the direction of the upper left corner, then CTRL-scroll out

Set image size to "Fit" (ALT-3) again, then move the mouse cursor clearly in the direction of the bottom right corner, then CTRL-scroll out

I could not reproduce if

• in "Rotate and Perspective" "Automatic cropping" is set to "Off" and cropping is done with module "Crop"

[kofa73]

Thanks for confirming! Strangely though, I can reproduce this any time, with small or big window, with or without OpenCL. Running a bisect now, starting from 4.4.2.

[jenshannoschwalm]

Ok, i can reproduce the log data (only reported if non-opencl).

I still can't confirm the artefacts though.

Please note while bisecting, don't care about the logs, that has been added recently.

[kofa73]

kofa@eagle:~/darktable$ git bisect good
09fa2c909b8a00a9648ed17d0550a2867a2b10e7 is the first bad commit
commit 09fa2c909b8a00a9648ed17d0550a2867a2b10e7
Author: Diederik ter Rahe <dterrahe@yahoo.com>
Date:   Mon Oct 2 17:02:12 2023 -0400

    concentrate zoom handling in develop.c

[jenshannoschwalm]

@dterrahe any quick idea on this ?

[dterrahe]

Steps to reproduce

1. Get the image and sidecar from https://tech.kovacs-telekes.org/files/2023-10-14-zoom-out/

Would you be able to provide a more minimal xmp (that doesn't give me a missing lut file error) to help speed up reproduction? Unless of course this only happens with slow xmps (using multiple d&s instances).

@dterrahe any quick idea on this ?

My first guess, if this only happens when the cursor is outside the image area while the zoom is adjusted, that the code that tries to keep the same point under the cursor is at fault. Or that the image doesn't get correctly repositioned afterwards. I'm currently working on deep further adjustments in this area, so rather than trying to fix this in code I'm about to replace anyway, I'd like to have a quick test case that I can use to verify that the new code doesn't suffer the same problem.

[jenshannoschwalm]

Another idea here would be bad roi_in or roi_out resulting from slightly different scales. That might lead to out of bounds of output data ... i also couldn't reproduce yet, the logs are just misleading (as long as roi_in is correct but we can't test that in the interpolation function)

[pehar1]

Removed all the "heavy" modules, removed "LUT 3D", removed "Color look up table", compressed history stack on kofa's xmp. Can still reproduce. Moving the mouse cursor outside the center (near edges) I can reproduce easier.
The attached xmp is not zipped, extension zip only added to be able to upload.
2023-10-14_08-19-06_P1020106_01.RW2.xmp.zip

edit : if it is of interest : Ubuntu 22.04, dt 4.5.0+911-g4e44bca430
Memory : MemTotal: 65756236 kB
Graphics Card : Product Name : NVIDIA GeForce RTX 2060
Graphics Card : Driver Version : 470.199.02
OpenCL installed : Device OpenCL C Version OpenCL C 1.2
OpenCL activated : yes
Xorg : Version: 1:7.7+23ubuntu2
Desktop : LXQt
GTK+ : 3.24.33
gcc (Ubuntu 12.3.0-1ubuntu1~22.04) 12.3.0
CMAKE_BUILD_TYPE : release

[kofa73]

@dterrahe, @pehar1 : Sorry about the heavy modules. I've removed the old sidecar and uploaded a minimal one. As @pehar1 , the problem can be reproduced with the minimal set, too.

[dterrahe]

@pehar1 thanks a lot!

Submitting clear bug reports like this, with easy ways to reproduce, whenever they encounter strange behavior in a development version, is one of the most helpful assistance users can provide to the project. You've directly helped to make the next release better.

I found the issue and implemented a fix that solves it with my testing. It will be included in a large intrusive PR that I'll hopefully submit today, but if it looks like that will take some time to merge I will also submit a separate PR just addressing this. Your feedback on the big PR would be appreciated though... ;-)

[dterrahe]

Sorry @kofa73 ! I didn't mean to not mention you and hadn't seen your update before pressing send. What I said above of course was meant for both of you (and all other proactive testers). Many thanks again!