[FFmpeg-devel] [PATCH] yuv4mpeg: add rough duration estimate and seeking.

[wm4]

On Mon, 22 Jun 2015 11:23:55 +0200
Hendrik Leppkes <h.leppkes at gmail.com> wrote:

> On Mon, Jun 22, 2015 at 11:02 AM, wm4 <nfxjfg at googlemail.com> wrote:
> > On Sun, 21 Jun 2015 18:34:33 -0400
> > "Ronald S. Bultje" <rsbultje at gmail.com> wrote:
> >
> >> Hi,
> >>
> >> On Sun, Jun 21, 2015 at 5:17 PM, Hendrik Leppkes <h.leppkes at gmail.com>
> >> wrote:
> >>
> >> > On Sun, Jun 21, 2015 at 10:41 PM, Ronald S. Bultje <rsbultje at gmail.com>
> >> > wrote:
> >> > > ---
> >> > >  libavformat/yuv4mpeg.h    |  1 +
> >> > >  libavformat/yuv4mpegdec.c | 26 +++++++++++++++++++-------
> >> > >  2 files changed, 20 insertions(+), 7 deletions(-)
> >> > >
> >> >
> >> > What happens if a seek does not end up on a perfect frame boundary?
> >> > Wouldn't that entirely screw over any further reading?
> >> >
> >> > So is it really that "rough", or does it actually work precise enough
> >> > to do that?
> >>
> >>
> >> It works for all y4m files I see in the wild, but the problem is that the
> >> y4m spec is a little confusing. The frame header magic is "FRAME", followed
> >> by optional per-frame options, and then a "\n". This is always empty (i.e.
> >> "FRAME\n"), but in theory it could be non-empty and then this doesn't work.
> >>
> >> I don't know how much we care about such theoretical files to make a more
> >> difficult dur/seek implementation.
> >
> > Why not use generic seeking mode? Add AVFMT_GENERIC_INDEX to the format
> > flags. The utils.c will do seeking by doing byte seeks and using the
> > byte position of previous packets. If you seek forward, it will read
> > and skip packets until the target is reached (for parts of the file
> > that have not been indexed yet). This should be very reliable, but of
> > course seek speed will depend on I/O bandwidth.
> 
> Generic seeking is rather expensive, and if a format is already strict
> CBR, doing a seek based on that is probably better.
> I wonder how hard it would be to write a re-sync function to find the
> next FRAME header when a seek does actually end up on the wrong spot.

OK, but I'd rather have slow seeking than possibly randomly inaccurate
seeking. Maybe the seek mode could be switched on AVFMT_FLAG_FAST_SEEK.