[FFmpeg-devel] [PATCH v2] avformat/hls Implement support for using AVSEEK_FLAG_BACKWARD when seeking

[Gustav Grusell]

On Sat, Jun 12, 2021 at 6:51 PM Lingjiang Fang <vacingfang at foxmail.com>
wrote:

> On Sat, 12 Jun 2021 14:16:22 +0200
> Gustav Grusell <gustav.grusell at gmail.com> wrote:
>
> > On Fri, Jun 11, 2021 at 6:29 PM Lingjiang Fang
> > <vacingfang at foxmail.com> wrote:
> >
> > > On Sat,  5 Jun 2021 22:42:26 +0200
> > > Gustav Grusell <gustav.grusell at gmail.com> wrote:
> > >
> > > > Before, seeking in hls streams would always seek to the next
> > > > keyframe after the given timestamp. With this fix, if
> > > > AVSEEK_FLAG_BACKWARD is set, seeking will be to the first
> > > > keyframe of the segment containing the given timestamp. This
> > > > fixes #7485.
> > > >
> > > > Signed-off-by: Gustav Grusell <gustav.grusell at gmail.com>
> > > > ---
> > > >  libavformat/hls.c | 9 +++++++++
> > > >  1 file changed, 9 insertions(+)
> > > >
> > > > diff --git a/libavformat/hls.c b/libavformat/hls.c
> > > > index 8fc6924c90..fb8c9b071a 100644
> > > > --- a/libavformat/hls.c
> > > > +++ b/libavformat/hls.c
> > > > @@ -2197,6 +2197,15 @@ static int hls_read_packet(AVFormatContext
> > > > *s, AVPacket *pkt) break;
> > > >                      }
> > > >
> > > > +                    /* If AVSEEK_FLAG_BACKWARD set and not
> > > > AVSEEK_FLAG_ANY,
> > > > +                     * we return the first keyframe encountered
> > > > */
> > > > +                    if (pls->seek_flags & AVSEEK_FLAG_BACKWARD &&
> > > > +                        !(pls->seek_flags & AVSEEK_FLAG_ANY) &&
> > > > +                        pls->pkt->flags & AV_PKT_FLAG_KEY) {
> > > taking account of the case of using flag backward and flag any
> > > together, a logic like this is better:
> > >
> > > if ( pls->seek_flags & AVSEEK_FLAG_BACKWARD &&
> > >     (pls->seek_flags & AVSEEK_FLAG_ANY ||
> > >      pls->pkt->flags & AV_PKT_FLAG_KEY)) {
> > >    ....
> > > }
> > >
> > >
> > Thanks for your feedback!
> > I'm not sure about this, I think it makes more sense if seeking with
> > AVSEEK_FLAG_BACKWARD and AVSEEK_FLAG_ANY behaves the same as when
> > using only AVSEEK_FLAG_ANY, that is to seek to as close as possible
> > as the exact timestamp, regardless if that is a keyframe or not. If
> what do you mean "AVSEEK_FLAG_BACKWARD and AVSEEK_FLAGAV_ANY behaves the
> same as when using only AVSEEK_FLAG_ANY"?
> I think the default behavior of seek is seek to key frame, where
> AVSEEK_FLAGAV_ANY means we can accept seeking to non-key frame.
>
>
Sorry I think you can disregard what I wrote. I was thinking of the case
where the user would set both BACKWARDS and ANY flags, and
forgot that if the user sets the BACKWARDS flag, the ANY flag will be added
for other playlists than the one containing the stream
searched in. I think this made me miss your point.
So what I meant was that if seeking with both AVSEEK_FLAG_BACKWARD and
AVSEEK_FLAG_ANY, optimally, I would expect the result to
be to search to the nearest frame before or exactly at the seek timestamp.
Note that this is not how i behaves with my patch, it will
seek to the first frame with a timestamp equal to or greater to the seek
timestamp. In my opinion this is probably good enough, and is
in any case the same behaviour as before the patch.
So if the user has etAVSEEK_FLAG_ANY we would not then want to search to
the first frame of the segment, even if we have the AVSEEK_FLAG_BACKWARD
set. But of course this is different if AVSEEK_FLAG_ANY has been set
because this is not the seek playlist.


> > I'm not mistaken, with your proposed change, using both
> > AVSEEK_FLAG_BACKWARD and AVSEEK_FLAG_ANY would always seek to the
> > first frame of the segment which is always a keyframe, so behaviour
> > would be the same as using only AVSEEK_FLAG_BACKWARD.
> when the first packet of a segment is not key-frame(rarely seen, but
> is possible), the behavior will be different.
> - AVSEEK_FLAG_BACKWARD will go on seeking until find a key frame
> - AVSEEK_FLAG_BACKWARD + AVSEEK_FLAG_BACKWARD should return immediatly
>
>
Ah yes you are correct. I was thinking that the HLS specification requires
an IDR frame at the start of every (video) segment, but now that I looked
it up that is not strictly the case.
Apples HLS authorign specification does require it though.


there is a similar logic at the end of this if{} block.
>
> >
> >
> > > > +                        pls->seek_timestamp = AV_NOPTS_VALUE;
> > > > +                        break;
> > > > +                    }
> > > > +
> > > >                      tb = get_timebase(pls);
> > > >                      ts_diff = av_rescale_rnd(pls->pkt->dts,
> > > > AV_TIME_BASE, tb.den, AV_ROUND_DOWN) -
> > >
> > > BTW, when I used seek backward before, what I want is to get exactly
> > > packets of the specified seek time, not only the specified stream
> > > but also all streams I need. To achieve this, I will put this patch
> > > lines before this if block, like this:
> > >
> > > if (pls->seek_timestamp == AV_NOPTS_VALUE)
> > >     break;
> > >
> > > if ( pls->seek_flags & AVSEEK_FLAG_BACKWARD &&
> > >     (pls->seek_flags & AVSEEK_FLAG_ANY ||
> > >      pls->pkt->flags & AV_PKT_FLAG_KEY)) {
> > >    ....
> > > }
> > >
> > > if (pls->seek_stream_index < 0 ||
> > > ...
> > >
> > >
> > Not sure I'm following. If I'm not mistaken, the hls spec does not
> > require audio and video segments to be aligned. Given this, it seems
> > to me that the above proposed change could lead to incorrect seek for
> > some streams when using AVSEEK_FLAG_BACKWARDS, since we would get the
> > first frame of the segments for all streams, and for audio streams
> > this might not match the timestamp of the first segment of the video
> > stream.
> it's hard to seek precisely when we want to seek backward, we just
> guarantee we seek to a place before the given time, like what this
> patch does.
> the way I proposed is to follow the flag as much as possible, and behave
> similarly no matter audio and video are in different playlists or in
> one, because it only affects the stream in the same playlist.
>

Ah, it took me a while, but now I just realized you have a very valid
point. If audio is in different playlist than video and we seek in video
stream, the audio playlist will seek
with AVSEEK_FLAG_ANY and with my patch may seek to a timestamp that comes
quite some time (a few seconds or so) after the timestamp we seeked to in
the video stream. This seems undesirable, I will rethink and rewrite the
patch.


>
> anyway, it's just a suggestion, hoping more discussion :)
>
>
Many thanks for your input, very helpful :) .

Cheers,
Gustav