[FFmpeg-devel] [PATCH 07/12] ratecontrol: fix stuffing bits for 30000/1001 fps one frame vbv

[Michael Niedermayer]

On Thu, Jul 05, 2018 at 11:49:19PM +0200, Michael Niedermayer wrote:
> On Thu, Jul 05, 2018 at 08:23:24AM -0700, Baptiste Coudurier wrote:
> > Hi Michael
> > 
> > On Thu, Jul 5, 2018 at 3:43 AM, Michael Niedermayer <michael at niedermayer.cc>
> > wrote:
> > 
> > > On Wed, Jul 04, 2018 at 11:35:09AM -0700, Baptiste Coudurier wrote:
> > > > ---
> > > >  libavcodec/ratecontrol.c | 12 ++++++------
> > > >  1 file changed, 6 insertions(+), 6 deletions(-)
> > > >
> > > > diff --git a/libavcodec/ratecontrol.c b/libavcodec/ratecontrol.c
> > > > index 49d169ba25..28bdddbad1 100644
> > > > --- a/libavcodec/ratecontrol.c
> > > > +++ b/libavcodec/ratecontrol.c
> > > > @@ -705,16 +705,16 @@ int ff_vbv_update(MpegEncContext *s, int
> > > frame_size)
> > > >          rcc->buffer_index += av_clip(left, min_rate, max_rate);
> > > >
> > > >          if (rcc->buffer_index > buffer_size) {
> > > > -            int stuffing = ceil((rcc->buffer_index - buffer_size) / 8);
> > > > +            int stuffing = rcc->buffer_index - buffer_size;
> > > >
> > > > -            if (stuffing < 4 && s->codec_id == AV_CODEC_ID_MPEG4)
> > > > -                stuffing = 4;
> > > > -            rcc->buffer_index -= 8 * stuffing;
> > > > +            if (stuffing < 32 && s->codec_id == AV_CODEC_ID_MPEG4)
> > > > +                stuffing = 32;
> > > > +            rcc->buffer_index -= stuffing;
> > > >
> > > >              if (s->avctx->debug & FF_DEBUG_RC)
> > > > -                av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n",
> > > stuffing);
> > > > +                av_log(s->avctx, AV_LOG_DEBUG, "stuffing %d bytes\n",
> > > stuffing>>3);
> > > >
> > > > -            return stuffing;
> > > > +            return stuffing>>3;
> > > >          }
> > >
> > > This would break VBV for most codecs
> > > buffer_index is a double, the if() condition is triggered if the current
> > > frame
> > > is too small. All the rounding has to be up as in the ceil() that was
> > > there.
> > > The stuffing that is added, is minimal, so a decrease would be expected to
> > > break cases occasionally
> > >
> > > Can you explain how this code fails, in what exact case
> > > (max/min rate/ buffer_size/fps) ?
> > >
> > 
> > Basically It doesn't work for IMX50, requiring exact frame size of 208541
> > bytes at 30000/1001 fps
> > and vbv delay of 3003
> 
> Is exactly 208541 required or is it "up to 208541"? Tim said in
> https://ffmpeg.org/pipermail/ffmpeg-user/2015-June/026872.html
> that its "up to"
> 
> up to should be easy achievable by adjusting the bitrates and vbv buffer
> size to values consistent and below 208541. This would then result in a
> mix of 208541 and 208540 bytes. (always 208541 does not result in a integer
> bitrate so its tricky to specify and also due to float use could then still
> be somewhat unreliable
> (208541 bytes and 30000/1001 fps result in a bitrate of 49999840.159840159840159840159840...)
> 
> iam not sure if theres a combination of parameters to achieve always 208541
> currently.
> If not it may be an option to add a parameter to force a specific frame size
> directly to the ratecontrol code.

After looking a bit, you can achive 208541 with
-vcodec mpeg2video   -minrate 49999841 -maxrate 49999841 -b:v 49999841 -bufsize 1668329  -rc_init_occupancy 1668329    -f image2 -an  delthisimg%d.mpeg2

not sure if this has issues but it seems to consistently produce the intended
frame size

[...]

-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Complexity theory is the science of finding the exact solution to an
approximation. Benchmarking OTOH is finding an approximation of the exact
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: not available
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20180706/fae80f92/attachment.sig>