[Ffmpeg-cvslog] [ffmpeg]: r5417 - in /trunk: libavcodec/snow.c libavcodec/snow.h tests/ffmpeg.regression.ref tests/rotozoom.regression.ref

[Corey Hickey]

lorenm at natsuki.mplayerhq.hu wrote:
> Author: lorenm
> Date: Sun May 28 23:44:47 2006
> New Revision: 5417
> 
> Log:
> Snow multiple reference frames
> (bitstream is not backwards compatible, even if refs aren't used)

Wow, two big snow patches in one day. It must be Christmas already. :)

Loren mentioned that this is very content dependent, but here are the
PSNR and user times for the second pass of my usual test clip. There
wasn't any significant variation of video stream size, so I'm omitting
that line to keep it clean.

refs=1
PSNR: Y:41.66, Cb:44.04, Cr:44.63, All:42.38
user    1m5.722s

refs=2
PSNR: Y:41.74, Cb:44.04, Cr:44.64, All:42.44
user    1m10.520s

refs=3
PSNR: Y:41.77, Cb:44.06, Cr:44.64, All:42.47
user    1m15.733s

refs=4
PSNR: Y:41.79, Cb:44.07, Cr:44.66, All:42.49
user    1m20.727s

refs=5
PSNR: Y:41.80, Cb:44.06, Cr:44.65, All:42.49
user    1m26.234s

refs=6
PSNR: Y:41.81, Cb:44.08, Cr:44.66, All:42.50
user    1m32.960s

refs=7
PSNR: Y:41.80, Cb:44.08, Cr:44.65, All:42.50
user    1m37.030s

refs=8
PSNR: Y:41.81, Cb:44.08, Cr:44.66, All:42.51
user    1m42.654s


Looks like the encoding time increases linearly and the the PSNR
increases logarithmically. The speed reduction isn't all that bad,
except when using vme=8, which compounds the speed reduction:

refs=1:vme=8
PSNR: Y:42.35, Cb:44.55, Cr:45.11, All:43.03
user    7m2.596s

refs=8:vme=8
PSNR: Y:42.46, Cb:44.56, Cr:45.11, All:43.11
user    23m20.679s


-Corey