[MPlayer-dev-eng] [PATCH] changes in vf semantics

[Ivan Kalvachev]

On Tue, 23 Nov 2004 21:56:10 +0100
Reimar D_ffinger <Reimar.Doeffinger at stud.uni-karlsruhe.de> wrote:

> Hi,
> > > [...]
> > > > > the desired destination size or just write it somewhere to vf_instance_s 
> > > > > and let it be. get_image would be then called with the current and next 
> > > > > vf as argument, instead of next and x/y-size.
> > > > 
> > > > again you totally misunderstand the video layer. i don't have time to
> > > > test and debug this stuff right now, but imo someone like myself,
> > > > ivan, or michael who understands it should be the one to fix it,
> > > > rather than you adding crazy hacks based on misunderstandings.
> > > 
> > > Oh, I'm sure we would not mind that. If only it looked like this is
> > > ever going to happen. This bug is know at least since months, and
> > > nobody gave even any useful comment.
> > 
> > could you do some research then and find out why it doesn't happen
> > with vd_ffmpeg.c?
> 
> No, I can't because it does happen with vd_ffmpeg.c as well (maybe not
> in all cases though). See my second patch in the "segfault in vf_scale"
> thread.

It usually doesn't happen with normal ffmpeg codecs because most of them
support emu_edge, this they don't need alignment.

The problem is real and quite messy.
Actually there are few problems:

First. Alignment

The horizontal line alignment is not an issue as it is easily to workaround by flags.
The real problem is vertical block alignment, as it is not specified anywhere.

As really really messy workaround we could make 16 lines height alignment when 
horizontal flag alignment is set.

Second.
Video initialization.
By legacy reasons the vo is initialized before even first mpi structure is filled.
This makes vo_xv and probably other vo allocate memory, init devices
even before knowing what alignment they may need.

Because of this, we should pass the aligned dimensions to vo config().
But this may lead to problems as we also pass visible dimensions (aspect) at the same time.
(e.g. off by few lines creates horrible scaling artifacts on cheap old video cards,
like the one I used to have). For fixed size vo it may lead to black bars, but this is not really issue.


Finally.

Probably it would be good idea to add an alignment parameters to the mpcodecs_config_vo()
and remove them from get_image() (e.g. making flags informative, not settable).
This seams to be also the less intrusive way.
Indeed, it will need changing of all vo-s.

Probably an better way to fix it is to pass mpi structure to vo_config(),
as mpi already have both visible and "physical" dimensions and alignment info.
An better and more generic alignment parameters could be added.
Probably adding aspect info to mpi would be interesting change,too ;)



I insist to discuss this here in maillist as it give more time to think about issues,
more time to check the code and test things.


Wish You Best
   Ivan Kalvachev
  iive

p.s.
i wrote this mail right before going to sleep, if I overlooked
something, point it kindly;)