[FFmpeg-devel] [PATCH 3/5] ffmpeg: flush and drain video filters.

Nicolas George nicolas.george at normalesup.org
Tue Mar 13 11:56:54 CET 2012

Le quartidi 24 ventôse, an CCXX, Michael Niedermayer a écrit :
> > Yes, but the problem here is FIFO. It does not satisfy this:
> > > > Pushing a balanced flux unrequested frames on its inputs (what balanced
> > > > means being filter-dependant) should not cause frames to accumulate anywhere
> > > > before or inside the filter.
> If you push a balanced mix into a box and take nothing out it will
> accumulate inside
> if you push in and pull out then the above filter graph does not
> have an accumulation problem

You are being too laconic, and I can not decide whether you are describing
things as they are or as they should be.

Things, currently, are not like that. Most filters, in fact, do not have
pull methods at all, and behave exactly like I stated: when you push to
them, they push to the next.

My point is that if all filters were made to behave like that (instead of
most, as it is currently), it would make the code simpler (less duplication
between poll_frame and request_frame) and move all scheduling problems to
the sinks, where they are simpler to solve (no need to recurse).

> For this to work an application must run its demuxer and protocol in
> a seperate thread. Not all applications do this, some are single
> threaded.
> With a single theraded application at the filters input a request_frame
> can either call the demuxer and decoder or not call the demuxer to get
> more data

The same reasoning applied to poll_frame too exactly as well. And to prove
my point, imagine you can write a poll_frame that behaves as you want. Then
I can write:

int request_frame() {
    if what_you_did_in_poll_frame == 0
	return EAGAIN

and I have a request_frame that behaves as I want.

The difference is that with that design, the behaviour does not change
whether the surrounding code has called avfilter_poll_frame or not. For the
record, we currently have exactly that problem in the vf_thumbnail, that is
what patch 3/6 in the other thread fixes.

> With poll_frame() we can find out which output of a filter graph has
> data buffered inside the filter graph. With a request_frame() that
> returns EAGAIN on an empty filter graph you have a infinite loop

Why would we have an infinite loop? EAGAIN breaks out of the loop exactly as
poll_frame returning false would do. Confer the code of the patch: it does
exactly that. The code is either:

while(poll_frame()) {


while(request_frame() != AVERROR(EAGAIN)) {

The difference is that the filters themselves can be made simpler. I know
for sure for having implemented vf_tile for both versions:
The same applies to vf_thumbnail: with this patchset applied, its poll_frame
can simply be removed and it will work just as well. I believe the same
could go for yadif and tinterlace, but I am less familiar with their
workings. This is even worst vf_select: not only could poll_frame go away,
but also could all the buffering logic. 

I am sure so much code simplification can only be appealing to you.


  Nicolas George
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: Digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20120313/b4625ae9/attachment.asc>

More information about the ffmpeg-devel mailing list