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

Jeff Clagg snacky at ikaruga.co.uk
Sun Sep 26 13:36:50 CEST 2004


I would also like to discuss the frameref listing in the manpage:

> .B frameref=<1\-15>
> Number of previous frames used as predictors in a P frame (default: 1).
> This is effective in Anime, but seems to make little difference in
> live-action source material.
> Some decoders are unable to deal with large frameref values.

First of all, the range is incorrect. Right now, the upper limit is
defined at 100. 15 just happens to correspond with the point at which
libavcodec will no longer decode it.

Secondly, I'd like to encourage others to discuss their own findings
about frameref values, as well as retract some of the comments I've made.

Earlier I stated I often saw no difference or even small filesize
increases when changing frameref to a higher value. I'm now changing my
story. My earlier claims came from two causes:

1) older x264 was biased against using older reference frames
2) using CABAC seems to neutralize the extra coding overhead caused by
using an unnecessarily large frameref value.

Indeed it seems that a lot of my lukewarm experiences with high frameref
values can be fixed by always using CABAC. The problem is that I was
trying to do some very simple tests in which I only specified the
frameref option and a fixed quant. My belief is that without CABAC, an
unnecessarily high frameref only ends up wasting bits on syntax elements,
adding more than enough extra bits to cancel out the tiny amounts of
bits saved. But when CABAC is turned on, this penalty is reduced to
almost nothing, or even eliminated completely.

Next I would like to discuss fullinter. Quite simply, I have yet to pin
down any benefit from it. Can anyone show two otherwise identical encodes
in which fullinter resuits in same/decreased bitrate with same/increased
PSNR? In every single instance, I get significantly increased bitrate
(which is expected) with no improvement in PSNR.




More information about the MPlayer-dev-eng mailing list