Alliance for Open Media codecs

[nevcairiel]

0.1.1 was never released indeed, the changelog entries being created for it have been recycled into the 0.2.0 changelog.

[sneaker_ger]

Originally Posted by Wolfberry

The version has already bumped to 0.2.0

• libaom 1.0.0-1402-g442f429de
• libdav1d 0.2.0-493155af

Now dav1d is like 55% faster on i5-2500K (AVX, no AVX2) than libaom on Blackmagic Pocket Cinema "Nature" sample 1080p24.

[v0lt]

Originally Posted by Wolfberry

ffmpeg-4.2-N-93276-g3b23eb283a-win64-static

Test on this build. i5-3570K.

-t 10 Stream2_AV1_4K_22.7mbps.webm -benchmark -f null -
libaom-av1 - max 15 fps
libdav1d - max 13 fps
libdav1d -threads 4 -tilethreads 4 - max 18 fps
libdav1d -threads 8 -tilethreads 1 - max 18 fps

I see progress compared to last time.

Please give me a link to a good 10-bit AV1 sample.

[sneaker_ger]

Originally Posted by v0lt

Please give me a link to a good 10-bit AV1 sample.

http://download.opencontent.netflix....x=AV1/Chimera/

[nevcairiel]

Note that 10-bit is not an optimization target at all yet, since its not in use in the real world at all yet either.

[v0lt]

@nevcairiel
Why does libdav1d give a speed less than libaom-av1 with default settings? 8 and 10-bit. Poorly selected default settings?

i5-3570K, AV1 10-бит.

-ss 5 -t 20 Chimera-AV1-10bit-1920x1080-6191kbps.mp4 -benchmark -f null -
libaom-av1 - max 25 fps
libdav1d - max 21 fps
libdav1d -threads 4 -tilethreads 4 - max 25 fps
libdav1d -threads 8 -tilethreads 1 - max 26 fps

[soresu]

Dunno if anyone noticed the CNN restoration experiment commited to the "experimental" branch at the AOMedia git repo.
Link here.

[EwoutH]

@v0lt Which version of dav1d are you using? You can use `dav1d --version` to check.

A big factor is because your CPU doesn't support AVX2 instructions, where dav1d is most optimized for (Haswell (4000-series) and newer support it). Your CPU i5-3570K falls back on SSSE3, which recently got a lot faster with dav1d.

Also can you try -tilethreads 4 without a threads option?

Anyway there is at least one patch incoming that's not yet in FFmpeg that will improve performance with 10 to 30% percent, depending on content.

[dapperdan]

I interpreted volts question as being "if I can get faster results from dav1d by passing some settings, why can't it figure those out by default".

Presumably, if those patches land the default might beat libaom, but there would still be extra speed a ilavle as long as you knew the right options to pass.

There may well be a good reason for these not being the defaults though.

[Gravitator]

Through the synthetic test mode ffmpeg, I get higher rates of DAV1D than AOM. But in real conditions I get worse playing than AOM (via ffplay). We are waiting for optimization for SSSE3.

[sneaker_ger]

Originally Posted by EwoutH

Also can you try -tilethreads 4 without a threads option?

On i5-2500K with ffmpeg (binary from one post above) -tiletreads 4 (alone) is some 15% slower than -threads 6 -tilethreads 2. ("Blackmagic Pocket Cinema Camera 4K ‘Nature’-oAbB4dQOz4I" 1080p24)

[v0lt]

Originally Posted by EwoutH

@v0lt Which version of dav1d are you using? You can use `dav1d --version` to check.

ffmpeg-4.2-N-93276-g3b23eb283a-win64-static provided by Wolfberry. But the links are gone.

Originally Posted by Wolfberry

ffmpeg-4.2-N-93290-g88d0be1c0e-win64-static

Thank. This build works better.

i5-3570K

ffmpeg -hide_banner -t 10 -c:v <codec&parameters> -i Stream2_AV1_4K_22.7mbps.webm -benchmark -f null -
libaom-av1 - max 15 fps
libdav1d - max 16 fps
libdav1d -threads 4 -tilethreads 4 - max 24 fps
libdav1d -threads 8 -tilethreads 1 - max 21 fps

ffmpeg -hide_banner -ss 5 -t 20 -c:v <codec&parameters> -i Chimera-AV1-8bit-1920x1080-6736kbps.mp4 -benchmark -f null -
libaom-av1 - max 60 fps
libdav1d - max 64 fps
libdav1d -threads 4 -tilethreads 4 - max 105 fps
libdav1d -threads 8 -tilethreads 1 - max 87 fps

ffmpeg -hide_banner -ss 5 -t 20 -c:v <codec&parameters> -i Chimera-AV1-10bit-1920x1080-6191kbps.mp4 -benchmark -f null -
libaom-av1 - max 25 fps
libdav1d - max 21 fps
libdav1d -threads 4 -tilethreads 4 - max 26 fps
libdav1d -threads 8 -tilethreads 1 - max 27 fps

[benwaggoner]

Originally Posted by mzso

There's a solution.



Yay! Two more formats to sink into obscurity...

The time **IS** ripe for a new replacement for JPEG/GIF/PNG. We're looking at >>2x compression efficiency improvements. And HEIF is getting good traction in the Apple ecosystem.

I would prefer an AV1 in HEIF over a new file format, though. HEIF has a lot of great features as a container format.

[nevcairiel]

HEIFs tiling is pure cancer. It's a typical format designed by committee, overly complex for its own good.

[sneaker_ger]

dav1d 0.2.0 is final.

https://medium.com/@ewoutterhoeven/d...e-eac3e43868c2

[hajj_3]

DAV1D 0.2.0 final is out: https://code.videolan.org/videolan/dav1d/tags

Hopefully a new vlc player will be released soon with this integrated.

[utack]

Originally Posted by hajj_3

DAV1D 0.2.0 final is out: https://code.videolan.org/videolan/dav1d/tags

Hopefully a new vlc player will be released soon with this integrated.

Not yet
You can check here
https://git.videolan.org/?p=vlc.git;...es.mak;hb=HEAD
http://downloads.videolan.org/pub/videolan/dav1d/

[hin12]

Is 480p the highest quality option for all AV1 videos on YouTube? I'm seeing that a lot of popular videos are being encoded to AV1, but only at the 144p, 240p, 360p and 480p resolutions.

[sneaker_ger]

Yeah. But even Gangnam Style (3.3 Billion views) is only available in 720p with AV1 (1080p is H.264+VP9 only.), Despacito (6 Billion views) is limited to 480p AV1. On the AV1 beta playlist it seems 1080p is max for AV1, 1440p and 2160p are reserved for VP9.

[dapperdan]

When VP9 rolled out they talked about how delivering video to people with decent spec machines but terrible connectivity was a sweet spot where VP9 increased the amount of video watched. I'd guess they've run the numbers and they get more benefit from encoding the lower sizes in AV1 so that's what they prioritize.