Correct-downscaling causes ewa_lanczossharp to introduce a half-pixel shift when downsampling sometimes

[Artoriuz]

Important Information

• mpv version: mpv v0.37.0-234-g57d68262
• Platform and Version: Windows 11 Pro 22H2
• Source of the mpv binary: Shinchiro
• If known which version of mpv introduced the problem: No idea
• GPU model, driver and version: 6600XT Adrenalin 24.1.1
• Possible screenshot or video of visual glitches: All below

Reproduction steps

magick convert rei.png -filter lanczossharp -distort resize 75% im_ewa_lanczossharp.png

mpv --no-config --vo=gpu-next --gpu-api=vulkan --no-hidpi-window-scale --pause=yes --screenshot-format=png --deband=no --dither-depth=no --screenshot-high-bit-depth=no --dscale=ewa_lanczossharp --linear-downscaling=no --window-scale=0.75 rei.png

mpv --no-config --vo=gpu-next --gpu-api=vulkan --no-hidpi-window-scale --pause=yes --screenshot-format=png --deband=no --dither-depth=no --screenshot-high-bit-depth=no --dscale=ewa_lanczossharp --linear-downscaling=no --window-scale=0.75 --correct-downscaling=no rei.png

Expected behavior

mpv's output should probably roughly match whatever ImageMagick is doing.

Actual behavior

mpv's output with correct-downscaling turned on appears to have a half-pixel shift towards the top-left corner (rounding issue somewhere?). Screenshot with it turned off also provided to show that there's no shift in this case, so the problem seems related to turning it on. Filters with integer supports seem fine as well.

Log file

output.txt

Sample files

[Artoriuz]

Update: After talking to @llyyr about this he couldn't reproduce it on his system and I couldn't reproduce it with gpu-api=opengl either (tested on both AMD and Intel), so it's likely a driver windows-only problem.

Edit: Can confirm both Vulkan and OpenGL are working as intended on Linux.

[llyyr]

I couldn't reproduce this on X11/Wayland or with a Windows build running through wine. Seems to be driver/Windows related since Artoriuz can only reproduce it with d3d11/vulkan as well.

[kasper93]

I haven't tested myself, but I've seen recently similar issue. That could be related and explain on px offset.

See:
Blinue/Magpie#803
bjin/mpv-prescalers@bfcf48a

Maybe @bjin remembers more detail about it. But since it has 0.49 workaround, I think he didn't get to the root cause of it?

If we really have rounding issues around half values, well it should be fixed somehow, because it may produce weird issues here and there.

~EDIT:

Interesting. The gather version of polar sampling is indeed broken. Works fine with compute and direct sampling. Here is good read about gather issues with 16.8 fixed point conversion https://www.reedbeta.com/blog/texture-gathers-and-coordinate-precision/ but I must say, I couldn't make it work correctly even with this knowledge, still missing something. (edit: still interesting, but not really important in our case)

Also probably on some systems the gather version is not used, that's why it works. I don't see why it would behave differently on different platforms, with similar hardware.

[kasper93]

Will be fixed by https://code.videolan.org/videolan/libplacebo/-/merge_requests/647