[Ffmpeg-cvslog] CVS: ffmpeg/libavcodec mpegvideo.c, 1.503, 1.504 mpegvideo.h, 1.234, 1.235 h263.c, 1.298, 1.299

[Corey Hickey]

Guillaume POIRIER wrote:
> Hi,
> On 2/3/06, Corey Hickey <bugfood-ml at fatooh.org> wrote:
> 
> [..]
> 
> 
> 
>>vmax_b_frames=3:vb_strategy=2
>>  pass 1:
>>PSNR: Y:41.45, Cb:44.92, Cr:45.40, All:42.37
>>user    729m40.550s
>>  pass 2:
>>PSNR: Y:42.25, Cb:45.24, Cr:45.99, All:43.11
>>user    188m33.290s
>>frame counts: I:   3424  1%  P:  65680 33%  B: 126962 64%
>>average qps:  I: 2.607185    P: 3.043849    B: 5.030922
>>
>>This one takes the crown PSNR-wise, with a +0.07 dB increase over using
>>turbo. Compared to vmax_b_frames=2:vb_strategy=2:brd_scale=0 (without
>>turbo), though, it doesn't really look better. Some parts looked a bit
>>better and some looked worse, but, on average they both looked about the
>>same.
> 
> 
> Ok, this looks like turbo would benefit from a slight lifting.
> Currently, turbo pretty much disables all encoding options but keeps
> the given max number of bframes unchanged.
> 
> I'd like to have 2 modes for turbo, just like x264:
> turbo=1 for a speed-up without any noticeable change in quality
> turbo=2 for the greatest quality with some decrease in quality, which
> would be current turbo mode.
> 
> However, I don't know what options are supposed to have the most
> influence on getting the best first pass stats.
> I guess leaving in trellis and mbd=2 should make the encode be closer
> to how the encode would look like without turbo mode... therefore it
> should produce better stats. But are there any other particular option
> that really shouldn't be left out?

mbd=2 and trell are probably the two biggest ones, yes; I would also
suggest allowing subcmp<=2 and maybe cmp<=2.

I'm not really sure about diamond size. Values of diamond size such that
-3 < x < 3 don't seem to make much of a difference to encoding time and
sometimes give slight PSNR improvements. I doubt it matters either way.

qpel can give a substantial improvement on any movie with scrolling
credits, but quite a large performance drop. You might compromise and
allow qpel if the user specifies it, but force subq to 3 or lower. Be
sure to test it...

mv0, v4mv, and cbp can give slight improvements at a slight performace
drop. Probably not worth it for either turbo mode.

I haven't tested qprd in ages. When I first did, I found it gave
slightly higher PSNR but had some transient problems wherein a single
frame, once in a while, would look really bad. I can't seem to dig up
the original mail I sent.... Anyway, checking it on a test clip now, it
seems to give +0.03 dB PSNR and increases encoding time 65%.

qns can stay out, of course.

-Corey