[MPlayer-dev-eng] [PATCH] synch to new x264 revision

[Loren Merritt]

On Fri, 24 Sep 2004, Guillaume POIRIER wrote:

>>> qcomp helps when set to 0.6 50% of times
>>
>> qcomp=0.6 is the default. What other values is this compared to?
>
> (0.4, 0.5, 0.6, 0.7, 0.8). Isn't it great that default is _that_ good?
> ;-)

The default was picked by running that same test on 2 random movies I had 
on my HD.
I'm currently trying to improve 2 pass ratecontrol. The optimal qcomp may 
change soon.

>>> subq helps when set to 5 or 3 50% of times
>>> and helps when set to 0 33% of times
>>
>> really? subq=0 should _always_ be worse than subq=1 by a large margin,
>> (0 searches with only halfpel precision, but still takes the bits to
>> store qpel) and subq>1 should _always_ be better, though the amount of
>> difference depends on the source video.
>
> I also though it was weird. But I thought it was maybe like qpel in
> "regular" MPEG-4 that doesn't help PSNR but improves the
> precision/details of the picture.
> Is that the same?

qpel improves the precision of motion compensation, which means that it 
takes fewer bits to code texture afterwards. This would always help in 
MPEG-4 ASP too, except that it takes extra bits (2 bits per motion 
vector, which is a significant fraction of the total bitrate) to store 
that extra motion precision.
On the other hand, H.264 has no option to turn off qpel, so you always pay 
those bits, even if the encoder doesn't use the extra precision. 
That's with VLC. CABAC may be able to compress the qpel bits if they're 
always 0.

>>> iframes help when set to 2 47% of times, when set to 3 33% of times
>>
>> what option is "iframes"? do you mean "idrframe"?
>
> Yeah, That's what I meant. Sorry.
> That confirms that anime seems to like idrframes.

Note that IDR-frames are identical to I-frames if bframe=0 and 
frameref=1.

--Loren Merritt