Decoding error on encoded Krøyer_PS.y4m from subset1 at high quantizers

[rzumer]

The command line options are --tiles 4 --quantizer [252|220], everything else is default. Lower quantizer values (among the default quantizers tested on AWCY) do not seem to generate corrupted bitstreams. I have not tested much besides that.

dav1d output:

dav1d 0.4.0-47-g556890b - by VideoLAN
Error decoding frame: Invalid argument
No data decoded

aomdec output:

Warning: Failed to decode frame 1: Corrupt frame detected
Warning: Additional information: Corrupted segment_ids

[ycho]

Oh, it was a corrupted bitstream, not a desync.

[rzumer]

It seems this is tiling related. Currently bisecting with 2x2 tiles and quantizer of 250...

[rzumer]

This is an old issue. At some point tiling was broken for frame sizes not divisible by some number if I recall correctly, and after this was fixed for most clips Krøyer_PS remained broken, though instead of the encoder crashing it produces a corrupt bitstream.

[ycho]

OMG, we can send this for use by the director of next Episode of Matrix!

[rzumer]

First broken in [5271249e90a8fd93554a706f358e2d035a77080e] Initial full-scale large LRU RDO with integrated CDEF.

[rzumer]

[2090440ef595d79ade5924755b42338f695e2414] Add LRF setup for non-power-of-two tile sizes (fixes #1642) fixed the encoding crash but this clip at high quantizers has produced an invalid bitstream ever since then.

It might have something to do with the clip's resolution somehow, it is 960x1164p.

[tdaede]

looping in @xiphmont

[rzumer]

Can reproduce the same issue with ffmpeg -i nyan.y4m -vf scale="960x1164", so it seems resolution dependent.

Changing the width to 961 fixes the issue but, but small changes to the height does not.

[rzumer]

Setting lrf_base_shift to 2 regardless of the quantizer fixes the issue for this clip.

That might be fine as an interim solution.