Fix highlights reconstruction issues #20183
Conversation
jenshannoschwalm
added
bugfix
|
Should this change the artifacts because it still looks the same as before for me? |
Some pre-demosaic code still requires snapping to sensor pattern.
jenshannoschwalm
force-pushed
the
fix_varios_highlights
branch
from
44c7c39 to
6c7f027
Compare
|
|
@TurboGit i would suggest to merge this now for master and 5.4.1 About the OpenCL issues, it used to work on my nvidia system but does not on AMD. After more reports i would know more - possibly we might want to disable on AMD for now. |
1. In some OpenCL kernels we want to make sure we read valid data making sure we don't take negatives/NaN from image. This has always been an issue with AMD hardware when using some optimizer flags. 2. Observed artefacts resulted from opacity outside the 0-1 range when rewriting the cfa data. This was happening on CPU and OpenCL codepath. 3. False color representation didn't work correctly in OpenCL code path.
jenshannoschwalm
changed the title
|
Friends, got it. For laplacian hightlights there were two issues:
@Jiyone you might also be interested :-) Release note: |
- also do mem checks
|
@TurboGit added 3rd commit, that's just simple maintenance so not necessarily for 5.4.1 |
Huu ? 🤔 |
|
Confirmed fixed for me |
TurboGit
left a comment
TurboGit
left a comment
There was a problem hiding this comment.
Was all good on my side, merged in master & 5.4.x branch. Thanks!
|
sorry wasn't able to check earlier on mac - I'm not sure if i exactly got the issue: |
|
@jenshannoschwalm : The commit Reintroduce demosaic snapping for stability has introduced 9 regressions:
I need your feedback to understand this. Why only those? I mean the commit changes roi x, y for all RAW, so why only those tests are affected? They are all sharing the same 2 RAWs (mire1.cr2 & xtransIV.raf) with many other tests. Probably because all of them are using flip/orientation or lens. Is that the proper explanation? TIA |
Absolutely :-) BTW i will do a rework on this making sure all pre-demosaic modules don't require this "snapping" |
5 participants
Some highlights code didn't work correctly, so an urgent fix. See #20167
Removing the snapper from demosaicing
modify_roi_inlead to various problems in highlights reconstruction module, so let's keep it. It got unnoticed likely because almost-always peoply just stay with opposed ... Arrgh. Anyway, reintroducing the snapper doesn't hurt at all as we over-calculate the canvas so smooth moving-around works still perfectly.To test here is
Busker in Shadow.zip
There is still one annoying issue here on AMD/rustiCL, if i zoom in or just drag the main canvas around there are very strong artefacts, Very bad colors, sometimes the whole screen is blu or red ...
So i would like you all to test on your systems and report. There are some parts of the OpenCL code that are definitely out-of-1.2 specs, for sure that worked correctly some time ago on my nvidia system.
@TurboGit @piratenpanda (yes the video was like what i have but much-much more annoying here) @MStraeten as on mac please.