Google VP9 "Next Generation Open Video" information posted

[Motenai Yoda]

on all samples in recent comparison of mine, 0.053bpp (rl) and 0.030bpp (anime), x265 10bit (no-sao) got the best in both, vp9 10bit is sometimes better than x265 10bit (no-sao) only on anime, but suffer for inconsistent (very low) quality, especially on high complexity + high correlation scenes, although shows very "high quality" on flashy/no-correlation ones and almost no banding at all (unlike x265), it still smooth a lot on movie sample.

Increase qcomp (bias-pct) to 1.0 (100) help a bit, but I'd like some sort of psy/aq/deblock tuning. (aq1/aq2 looks a little worse)

ps I'm talking about 2pass+altref only as 1 pass crf quality is beaten by x264 too.

[benwaggoner]

Originally Posted by dapperdan

Have VP9 (and AV1) been working on doing better on VMAF? With Netflix adopting them it wouldn't be that surprising if tweaks or bug-fixes that particularly affect VMAF got prioritized.

The question then is if it is "teaching to the test" and therefore not a valid quality increase, or if VMAF is good enough to direct them towards genuine improvements that agree with subjective viewers.

VMAF looks pretty promising. But it was trained just on Netflix's standard bitrates at 1080p SDR. So it's applicability to <300 Kbps, >8 Mbps, UHD, and HDR is undetermined. Also, it was only trained on (IIRC) a few dozen ~10 sec clips. Netflix has shared some of their test clips, but not all of them, so it's hard to know if some classes of content are under- or un-represented.

I also have some concern that the base metrics it uses are largely spatial only. While VMAF does include a temporal comparison of adjoining frames, it is pretty limited, and I worry there may be kinds of temporal distortions that it might not pick up on.

That said, these are relatively theoretic concerns, and VMAF certainly appears to outperform better than standbys like the mean SSIM, and way better than PSNR.

[mzso]

How do I get libvpx to utilize more of my cpu?
The CPU usage with my new 6 core CPU is pretty pathetic. <20% I increased the -threads -tile-columns options in ffmpeg, but it didn't do anything notable.

(lib264 gets to about 50% CPU usage. Even though apparently it doesn't have any multi-threaded encoding settings.)

[wiak]

Originally Posted by mzso

How do I get libvpx to utilize more of my cpu?
The CPU usage with my new 6 core CPU is pretty pathetic. <20% I increased the -threads -tile-columns options in ffmpeg, but it didn't do anything notable.

(lib264 gets to about 50% CPU usage. Even though apparently it doesn't have any multi-threaded encoding settings.)

you need libvpx git with row-mt
https://nwgat.ninja/compiling-libvpx...ng-on-windows/

[mzso]

Originally Posted by wiak

you need libvpx git with row-mt
https://nwgat.ninja/compiling-libvpx...ng-on-windows/

Is there an ffmpeg build with that? (I'm familiar with ffmpeg.)

[LigH]

You could try running jb-alvarado's media-autobuild_suite for yourself and check if the result contains all you need. If not, suggest it in its issue tracker.

Remember, if you build a "non-free" ffmpeg regardless of all licenses, keep it for yourself, don't distribute it. The "zeranoe compatible" option already contains more libraries than most people need ...

P.S.: Just checked one I built on August 23, it reports:

libvpx-vp9 encoder AVOptions:
...
  -row-mt            <boolean>    E..V.... Row based multi-threading (default auto)

So I assume a current MABS build and current Zeranoe builds should contain current VPX libs which support it.

[mzso]

@LigH

Thanks. I'll check it out.

[LigH]

New file in my VPx builds archive: v1.6.1-1258-gc8f6e7b99 (MABS: MSYS2/MinGW, GCC 7.2.0)

VP10 as separate encoder branch seems to be abandoned.

[LigH]

New build: v1.6.1-1451-gc58f01724

[hajj_3]

new Libvpx 1.7.0 released: https://www.phoronix.com/forums/foru...mizations-more

[LigH]

It may have been "released"; unfortunately, branch "mandarinduck" was not yet merged with branch "master", thus tag "v1.7.0" does not yet apply to the master branch. So it's not yet available without manually switching a branch.

[mzso]

Does anyone know typically how long it takes to get into ffmpeg? By the way does ffmpeg add the libvpx library centrally or is it usually decided by whoever makes the build? (eg: zeranoe builds)

[JEEB]

Originally Posted by mzso

Does anyone know typically how long it takes to get into ffmpeg? By the way does ffmpeg add the libvpx library centrally or is it usually decided by whoever makes the build? (eg: zeranoe builds)

There are no official binaries so whomever is building picks whatever version that gets used in a binary. The only case where some work is required from FFmpeg's side is when the API is changed enough to break build, and thus API usage changes are required.

[LigH]

When it gets available in the master branch, I will try to build ffmpeg as well as vpxenc using the media-autobuild_suite. I will report if it worked for me. If successful, the vpx binaries will be published on MediaFire; but my ffmpeg build is "non-free", I have to keep it for myself.

[mzso]

Originally Posted by LigH

but my ffmpeg build is "non-free", I have to keep it for myself.

Only zealots care.

[LigH]

vpx v1.7.0-69-g77108f500 is available.

[Selur]

it even adds two new options,..

 --corpus-complexity=<arg>   corpus vbr complexity midpoint

and

--tune-content=<arg>        Tune content type
                                                      default, screen, film

sadly that is all the information about those options.
Does anyone know:
a. what 'corpus-complexity' is for and what are valid values for it? seems to be a boolean, so 1&0 should be the valid values
b. what content did they have in mind for the types 'default, screen, film'? (wild guessing: screen = screen captures of simple windows; film = movies; default = anything else)

Cu Selur

[LigH]

VPx v1.7.0-110-gedc9a4687

[mzso]

I just checked that the zeranoe ffmpeg builds now have 1.7.0. CPU utilization is still crap. 30% or less in total with my R5 1600.

[Selur]

@mzso: have you tried a higher '--cpu-used' value? (using higher values fives me more cpu usage, ~70%+ on R7 1800X with 1080p content, but still freaking slow encoding speeds)