DAV1D vs. LIBGAV1 Performance - Benchmarking Google's New AV1 Video Decoder

[coder]

Uh, are you okay?

I'm great! What's not is that duplication of effort you cite. That nonsense should be criminal, amiright? Such a waste of resources!

While, we're at it, we should probably outlaw all opensource OS' besides Linux. One opensource OS should be enough for everybody.

[profoundWHALE]

OMG. How dare they duplicate effort. Marx would be horrified!

You should buy some Alphabet stock, go to the next shareholder meeting, and get a vote going to put a stop to this utter nonsense!

Uh, are you okay?

[Toggleton]

there is a first single thread test of libgav1 vs libaom vs dav1d https://www.reddit.com/r/AV1/comments/dfga7z/av1_decoder_comparison_2019oct7_libgav1_libaom/

on the Raspberry Pi 4 https://docs.google.com/spreadsheets...gid=1359401180

[coder]

The problem is that as far as we can tell, they haven't really done anything different and dav1d is already VERY good.

OMG. How dare they duplicate effort. Marx would be horrified!

You should buy some Alphabet stock, go to the next shareholder meeting, and get a vote going to put a stop to this utter nonsense!

[profoundWHALE]

Especially with regards to AV1: the same philosophy that begot SVT-AV1/rav1e/libaom. A single piece of software shouldn't be everything to everyone. People should be free to pursue different approaches ("Write it in Rust!" "Assembly for every supported platform!" "Portable C only!" "Do it JavaScript!") without being told how to to spend their time, and this exploration lets different projects learn from each other: they can adopt what works & ignore what flounders.

The problem is that as far as we can tell, they haven't really done anything different and dav1d is already VERY good.

[archsway]

I don't really get the point, though. Any CPU that's fast enough to decode it with ARMv7 is going to be a 64-bit core. Right now, Pi is probably one of the main reasons for continued interest in ARMv7. However, as both the v3 and v4 hardware are 64-bit capable, I think it's only a matter of time before Raspbian goes 64-bit.

They would have to get a few multiples of performance improvements to enable even 1080p decode on any true 32-bit ARM cores, and even that is only going to work on the few, fastest ones out there.

Applying those two merge requests for dav1d (and compiling with LTO), I get 24 fps on the 1080p "Summer Nature" demo, on a 32-bit ARM CPU. NEON support in dav1d is nowhere near complete, and scalability work is definitely [hopefully] not finished, so one day it might be possible to watch 1080p AV1, but 720p probably already plays okay today.

And why do you say "even 1080p"? 720p and 480p are still things which exist, and they aren't going away any time soon.

[coder]

https://code.videolan.org/videolan/dav1d/issues/215 If you look here dav1d on arm32 is not done Yet and if you look on the Merge requests it is Work in Progress

I don't really get the point, though. Any CPU that's fast enough to decode it with ARMv7 is going to be a 64-bit core. Right now, Pi is probably one of the main reasons for continued interest in ARMv7. However, as both the v3 and v4 hardware are 64-bit capable, I think it's only a matter of time before Raspbian goes 64-bit.

They would have to get a few multiples of performance improvements to enable even 1080p decode on any true 32-bit ARM cores, and even that is only going to work on the few, fastest ones out there.

[Toggleton]

Neither of those are nearly good enough for targeting ARMv7. I'd hazard a guess that they don't plan to support AV1 on it.

Anyway, thanks for the data. Now, it would be nice to see it on ARMv8 (anybody?). Too bad we can't use Pi v4 (at least, not on Raspbian, which still runs in ARMv7 mode).

https://code.videolan.org/videolan/dav1d/issues/215 If you look here dav1d on arm32 is not done Yet and if you look on the Merge requests it is Work in Progress

https://code.videolan.org/videolan/dav1d/merge_requests/809

https://code.videolan.org/videolan/dav1d/merge_requests/811

Did try to compile libgav1 on my Odroid N2(arm64) on Manjaro but it does fail ...

[discordian]

Strange results when going from 1080p to 4k. For example, the Ryzen 5 3600X goes from 384 fps to 172 fps. But 4k is 4 times 1080p, not half.

Anyone have an idea why?

More independent blocks to process in parallel. You should measure effective cpu time (accumulated time spent on all cores), the more cores you throw at it, the more core time you are wasting.

Likely you will never decode 1080p video with more than 4 threads, if you care about efficiency. Otherwise you run into amdahls law.

[AndyChow]

Strange results when going from 1080p to 4k. For example, the Ryzen 5 3600X goes from 384 fps to 172 fps. But 4k is 4 times 1080p, not half.

Anyone have an idea why?