[vlc-devel] VideoToolbox green corruption bug

[Francois Cartegnie]

Le 07/07/2017 à 22:00, Oliver Collyer a écrit :

> I found that:
> 
> - helper_search_pps detects that the PPS is different. i_nal is normally 82 but every few seconds a PPS arrives of size 53 and this one triggers the change

So it's different. And in H264 spec, any pps with same id replaces the
previous one in transmission order.

> - despite the above, all entries of the h264_picture_parameter_set_t structure are the same, although I noticed that in h264_parse_picture_parameter_set_rbsp in h264_nal.c it only parses and reads in a few of them

Not the same. Both have different scaling lists.

> - if I disable the line that sets *p_config_changed = true it fixes the bug with no visible side-effects
> 
> So a couple of Qs to anyone familiar with this area:
> 
> 1) is it correct behaviour to restart the decoder whenever the PPS changes?

You're supposed to restart decoder because it is dumb and can't handle
SPS/PPS updates by itself.

> 2) if yes, what are the circumstances where it should restart the decoder/set config_changed to true, because it would seem that this shouldn't be one of them.

When the SPS/PPS parameters change. Which does here.

There's not much description about how to handle those cases...
That's "Apple standards" in high quality API.

You have to figure out how and if it will works, gather the hw
differences and add targeted hacks by yourself (You can't query
capabilities ! Which seems lots of unpublished internal conditional
switches in Foundation)

Since you're restarting both on Iframe and Iframe+1, there can only be
corrupted output.

> 3) if no, then it's a simple patch not to set *p_config_changed to true, correct?

No.

> Note that this issue seems specific to interlaced content. I've not seen it with progressive content and as you will see from the bug report it has been independently reported with other interlaced h.264 files too.

That's a corner case with what seems a broken H264 stream (if they need
to feed different scaling lists, then a different PPS should probably be
used). As they do transmit the same pps with each frame that's a strong
signal they don't really understand how it works.

Francois