Google VP9 "Next Generation Open Video" information posted

[mzso]

Originally Posted by 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)

Doesn't help. Though all cores are utulised. Per core utilization is very poor. Some cores go up to 30+% average. Some are around 15%

[Selur]

Here's the command line I used:

ffmpeg -y -loglevel fatal -threads 8 -r 24000/1001 -analyzeduration 200M -probesize 200M -i "H:\00005.m2ts" -map 0:0 -an -sn  -vsync 0 -strict -1 -pix_fmt yuv422p10le  -f yuv4mpegpipe - | vpxenc --codec=vp9 --row-mt=1 --passes=1 --pass=1 --end-usage=cq --cq-level=18 --target-bitrate=15000 --profile=3 --good --cpu-used=2 --min-q=0 --max-q=63 --undershoot-pct=0 --buf-sz=6 --buf-initial-sz=4 --buf-optimal-sz=5 --drop-frame=0 --resize-allowed=0 --kf-min-dist=0 --kf-max-dist=250 --auto-alt-ref=0 --noise-sensitivity=0 --sharpness=0 --static-thresh=0 --tile-columns=2 --tile-rows=1 --min-gf-interval=0 --max-gf-interval=0 --threads=32 --width=1920 --height=1080 --i422 --color-space=unknown --input-bit-depth=10 --bit-depth=10 -o "H:\Temp\15_10_58_7610_01.vp9" -

Cu Selur

[mzso]

Originally Posted by Selur

Here's the command line I used:
Cu Selur

Unless you point out something relevant this is just a lot of noise to me.

I use something like this.

ffmpeg -i <input> -acodec libopus -b:a 260k -g 30 -vcodec libvpx-vp9 -b:v 0 -threads 12 -tile-columns 4 -cpu-used 1 -crf 30 out.webm

Increasing cpu-used did nothing.

[Selur]

Seeing the settings you use, enabling row-mt might help.

[Jamaika]

I see the wrong approach. The test should be performed on the itself VPX to eliminate accidental ffmpeg errors. Input y4m.
Support speedup in sse4 has been added for the VP8 codec.

[mzso]

Originally Posted by Selur

Seeing the settings you use, enabling row-mt might help.

It helps a bit when cpu-used is 1 (not when it's 8). So I guess it makes higher quality encodings a little faster.

Let's hope that AOM will be more readily multi-processing optimized.

[Selur]

I would go for 1 or 2 (https://www.webmproject.org/docs/encoder-parameters/) not more,..

[LigH]

For AOM, a range of 5..8 was recommended to me to speed up remarkably. But VPx may either not support such levels, or it may decrease the quality too much...

[Selur]

vpxenc nowadays supports:

--cpu-used=<arg>           	CPU Used (-16..16)

but from my experience with older vpxenc versions I would normally stick to 1-3 not more.

[mzso]

Does -tile-columns and -tile-rows serve any purpose, when using -row-mt?
I don't see any difference in encoding speed and cpu utilization.

[Selur]

iirc they mainly determine how well threading is possible during decoding and row-mt was responsible for threading during encoding,..

[Mawazi]

Originally Posted by mzso

Does -tile-columns and -tile-rows serve any purpose, when using -row-mt?
I don't see any difference in encoding speed and cpu utilization.

I experienced a decent boost in encoding speed with --row-mt=1, --tile-columns=6, and --threads=16. This is with a 4-core, 8-hyperthreads Intel. As far as I know, --tile-rows doesn't do much, as VP9's multithreaded encoding is column based.

[mzso]

Originally Posted by Mawazi

I experienced a decent boost in encoding speed with --row-mt=1, --tile-columns=6, and --threads=16. This is with a 4-core, 8-hyperthreads Intel. As far as I know, --tile-rows doesn't do much, as VP9's multithreaded encoding is column based.

How about if you drop tile-columns? It made no difference to me. Row-mt gave the cpu utilization boost.

[mzso]

Is it normal for stuff that fades in to look this crappy, or is something going wrong on my end?
https://drive.google.com/open?id=1x3...goGBj1CRbFJ_JH

(It sucks with AVC too BTW: https://drive.google.com/open?id=1n-..._dj3P98PPCJ7R4)

[LigH]

New upload: VPx v1.7.0-120-g167594414

[Mawazi]

Originally Posted by mzso

How about if you drop tile-columns? It made no difference to me. Row-mt gave the cpu utilization boost.

My initial testing with a number of permutations lead me to the numbers that I use, though I don't know (and can't find any documentation for) the default number for tile-columns once row-mt is activated. I assume it's at least two or you wouldn't see any difference at all.

[Mawazi]

Is there a way to add color-range (limited/tv vs full/pc) to vp9 encodes using vpxenc? I require pc range for an encode I'm doing. I know this is possible using ffmpeg, but I'm unsure if color-range is a part of the vp9 bitstream or just part of the container, or both I suppose, and without compiling my own ffmpeg there is no row-mt or tune film options in the binaries of ffmpeg I can track down, so I'd like to use vpxenc.

[mzso]

Originally Posted by Mawazi

Is there a way to add color-range (limited/tv vs full/pc) to vp9 encodes using vpxenc? I require pc range for an encode I'm doing. I know this is possible using ffmpeg, but I'm unsure if color-range is a part of the vp9 bitstream or just part of the container, or both I suppose,

You mean color range metadata? You can't "add" color range itself.

Originally Posted by Mawazi

and without compiling my own ffmpeg there is no row-mt or tune film options in the binaries of ffmpeg I can track down, so I'd like to use vpxenc.

This isn't true. Plain old zeronae builds have row-mt, just used it a few days ago.
Also there's a "-tune-content film" option.

[mzso]

By the way. I tested lossless encoding with vp9, but it had some 36% higher filesize (570MB vs 420MB) than a lossless encode with libx264.

Is VP9 this poorly suited to lossless encoding, or have something likely gone awry?

[LigH]

VPx v1.7.0-177-g2640f2507