I am a bit confused. The black patches are on the right.
Sorry I meant new implementation is on the right of course. I swear I know the difference between left and right :p
From the tests…
I can confirm the speedup, but results are now significantly different than before. In images with many rough surfaces I think the results are visually equal, if not better. But for images with…
Since you removed the outer loop for channels this todo is now done. Nicely done ;)
Thanks for looking into this patch.
I approached this as a bug after fixing #105590, not as a direct user request, so I don't have a strong opinion on how anamorphic input should be handled.…
Thank you @Sergey ! I prepared some files for regression testing already, I will commit them to svn after submitting this patch
@Sergey thanks for the tips. I updated the patch with the minor changes. Sure, feel free to push further changes! I am planning to look into the performance this weekend.