[FFmpeg-devel] [Ffmpeg-devel] [RFC] ratecontrol buffer size magic

[Michael Niedermayer]

Hi

On Tue, Oct 16, 2007 at 12:17:26PM +0200, Baptiste Coudurier wrote:
> Hi Michael,
> 
> Michael Niedermayer wrote:
> > Hi
> > 
> > On Tue, Sep 25, 2007 at 12:00:02PM +0200, Baptiste Coudurier wrote:
> >> Hi
> >>
> >> Michael Niedermayer wrote:
> >>> Hi
> >>>
> >>> On Wed, Apr 25, 2007 at 12:21:52PM +0200, Baptiste Coudurier wrote:
> >>>> Michael Niedermayer wrote:
> >>>>> Hi
> >>>>>
> >>>>> On Thu, Apr 12, 2007 at 11:07:10AM +0200, Baptiste Coudurier wrote:
> >>>>>> Hi
> >>>>>>
> >>>>>> Michael Niedermayer wrote:
> >>>>>>> Hi
> >>>>>>>
> >>>>>>> On Wed, Apr 11, 2007 at 01:47:55PM +0200, Baptiste Coudurier wrote:
> >>>>>>>> Hi
> >>>>>>>>
> >>>>>>>> Michael Niedermayer wrote:
> >>>>>>>>> Hi
> >>>>>>>>>
> >>>>>>>>> On Thu, Apr 05, 2007 at 06:51:00PM +0200, Baptiste Coudurier wrote:
> >>>>>>>>>> Hi,
> >>>>>>>>>>
> >>>>>>>>>> I stumbed accross a weird behaviour of the CBR ratecontrol lately,
> >>>>>>>>>> basically doing:
> >>>>>>>>>>
> >>>>>>>>>>  ffmpeg -i file.mpg -vcodec mpeg2video -minrate 30000k -maxrate 30000k
> >>>>>>>>>> -b 30000k -intra -flags +ildct+low_delay -dc 10 -flags2
> >>>>>>>>>> +ivlc+non_linear_q -ps 1 -qmin 1 -qmax 12 -top 1 -vglobal 4 -acodec
> >>>>>>>>>> pcm_s16be -f mov -bufsize 1200000 -rc_init_occupancy 1200000 cbr.mov
> >>>>>>>>>>
> >>>>>>>>>> Constraints are so in order to achieve vbv delay as 3600.
> >>>>>>>>>> Now ffmpeg will only give me qp as 24. While tweaking I changed some
> >>>>>>>>>> magic in ratecontrol.c (see patch attached) and now qp have decent/good
> >>>>>>>>>> values.
> >>>>>>>>>>
> >>>>>>>>>> I dig into svn history and found out that r2601 bumped */2 to */3 to
> >>>>>>>>>> improve CBR ratecontrol. Can anyone explain in detail why that magic
> >>>>>>>>>> does improve behaviour ?
> >>>>>>>>> yes, your vbv buf size is VERY significantly too small it can barely
> >>>>>>>>> hold a single frame, increase it by a factor of 100 ...
> >>>>>>>>>
> >>>>>>>>> i hope theres no spec which mandates such nonsense parameters ...
> >>>>>>>>>
> >>>>>>>>>
> >>>>>>>>>> Can that parameter be user setable ? 
> >>>>>>>>> if you make it user settable ...
> >>>>>>>> Is something like that ok ?
> >>>>>>> yes except that name of the variable and the comment
> >>>>>> +
> >>>>>> +    /**
> >>>>>> +     * Ratecontrol available buffer space coefficient applied when
> >>>>>> choosing qp.
> >>>>> does anyone understand what this variable does based on this description?
> >>>>> no?
> >>>>> if no then patch rejected
> >>>> Humm ok another try:
> >>>>
> >>>>     /**
> >>>>      * Coefficient applied to buffer space, by ratecontrol,
> >>>>      * for limiting QP (bits2qp) when encoding with a buffer size.
> >>>>      * - encoding: Set by user.
> >>>>      * - decoding: unused
> >>>>      */
> >>>>     float rc_buffer_space_to_qp_coeff;
> >>> hmm its better
> >>> but still
> >>> the description says therese that parameter which is a coefficient
> >>> and that is applied to the buffer space, whats still unclear
> >>> * buffer? -> vbv buffer
> >>> * buffer space? is that the size of the whole buffer? of the free space
> >>>   left?
> >>> * applied? if id say ill apply a coeffificint to a number 95% of the math
> >>>   people would lynch me, its scaled or multiplied or something like that
> >>>   or if its a more complex operation then that should also be said more
> >>>   clearly
> >>> * also what is the now changed "buffer size" used for?
> >>>
> >> Another try:
> >>
> >> /**
> >>   * Ratecontrol factor applied to vbv buffer free space when limiting QP
> >> (bits2qp).
> >>   * With a maximum rate: free space / factor;
> >>   * With a minimum rate: free space * factor;
> >>   * - encoding: Set by user.
> >>   * - decoding: unused
> >>   */
> >>   float rc_vbv_to_qp_factor;
> >>
> >> {"rc_vbv_to_qp_factor", "ratecontrol vbv buffer free space to qp
> >> factor", OFFSET(rc_vbv_buf_free_space_to_qp_factor), FF_OPT_TYPE_FLOAT,
> >> 3.0, FLT_MIN, FLT_MAX, V|E},
> >>
> >> Im trying to keep name somewhat short.
> > 
> > please split this variable in 2
> > 2 good names would be:
> > 
> > rc_max_available_vbv_use (default 1.0/3)
> > that is attempt use at maximum 1/3 of what we can use without a underflow
> > 
> > rc_min_vbv_overflow_use (default 3)
> > that attempt to use at least 3 times the amount we must use to prevent
> > a vbv overflow
> > 
> 
> Patch attached.

ok

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Democracy is the form of government in which you can choose your dictator
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20071016/162651be/attachment.pgp>