AV1 hardware decoder implementations

[Spyros]

Beyond the core, the Ryzen 7000 Series Processors feature an all-new 6nm I/O die, which enables hardware-accelerated video encode/decode*, light-duty graphics work and multi-display support.

* Video codec acceleration (including at least the HEVC (H.265), H.264, VP9, and AV1 codecs) is subject to and not operable without inclusion/installation of compatible media players.

Press Release

[ChaosKing]

Originally Posted by quietvoid

7 months have passed and the FireTV Stick 4K Max still doesn't do film grain synthesis correctly on playback.
At this point I've yet to watch anything in AV1 because of this.

I also don't think it supports AV1 through Netflix yet, but I haven't been subscribed for a couple months now.

I just updated my FireTV Stick Max and now I can see the generated grain by the av1 decoder in Kodi. It finally works!

When I hit pause in Kodi the grain still "moves" lol
I use Kodi 19 dev btw.

[quietvoid]

Originally Posted by ChaosKing

I just updated my FireTV Stick Max and now I can see the generated grain by the av1 decoder in Kodi. It finally works!

When I hit pause in Kodi the grain still "moves" lol
I use Kodi 19 dev btw.

Very cool, I just tested and it also works on my side.

[Beelzebubu]

Originally Posted by ChaosKing

When I hit pause in Kodi the grain still "moves" lol

That sounds rather weird. Certainly not quite what you'd expect given how grain works in the spec...

[benwaggoner]

Originally Posted by Beelzebubu

That sounds rather weird. Certainly not quite what you'd expect given how grain works in the spec...

Sounds like the FGS filter keeps running even when the video stops. Since it's an out-of-loop postprocessor, that sort of thing isn't unprecedented. In the implementations I'm aware of, the FGS filter isn't even part of the codec itself, and can be triggered by the same SEI metadata in other codecs as well.

[Beelzebubu]

Originally Posted by benwaggoner

Sounds like the FGS filter keeps running even when the video stops. Since it's an out-of-loop postprocessor, that sort of thing isn't unprecedented. In the implementations I'm aware of, the FGS filter isn't even part of the codec itself, and can be triggered by the same SEI metadata in other codecs as well.

True, except that the process is defined and the seed coded in the bitstream header. So if the seed is static (e.g. when paused, or because so coded), the grain looks the same every frame. So the above behaviour suggests the seed is changing even if the video is paused. Where does it get the new seeds from? And if from someplace else than the bitstream header, does that mean it looks anything like intended by the encoder?

(It's true that encoders like aomenc don't choose a specific seed based on reconstruction characteristics, they just randomly generate a number. But in theory, an encoder could select a seed expecting a particular reconstruction, and that's what I'd expect to see in the decoder as well then.)

[ChaosKing]

Originally Posted by Beelzebubu

True, except that the process is defined and the seed coded in the bitstream header. So if the seed is static (e.g. when paused, or because so coded), the grain looks the same every frame. So the above behaviour suggests the seed is changing even if the video is paused. Where does it get the new seeds from? And if from someplace else than the bitstream header, does that mean it looks anything like intended by the encoder?

I see what you mean. Basically the video stream and the grain stream are "out of sync" now.
Not sure who to blame here. The fire stick decoder or Kodi.

[benwaggoner]

Originally Posted by Beelzebubu

True, except that the process is defined and the seed coded in the bitstream header. So if the seed is static (e.g. when paused, or because so coded), the grain looks the same every frame. So the above behaviour suggests the seed is changing even if the video is paused. Where does it get the new seeds from? And if from someplace else than the bitstream header, does that mean it looks anything like intended by the encoder?

(It's true that encoders like aomenc don't choose a specific seed based on reconstruction characteristics, they just randomly generate a number. But in theory, an encoder could select a seed expecting a particular reconstruction, and that's what I'd expect to see in the decoder as well then.)

Good points. Perhaps it is looping, or falls back to randomizing a seed if the seed remains the same for too many frames?

[quietvoid]

FWIW sometimes pausing has dancing grain, sometimes it doesn't.

[benwaggoner]

Originally Posted by quietvoid

FWIW sometimes pausing has dancing grain, sometimes it doesn't.

Weird. Latest system update installed and everything?

What model number Fire Stick?

[quietvoid]

Originally Posted by benwaggoner

Weird. Latest system update installed and everything?

What model number Fire Stick?

Yea, however it's not consistently reproducible. I managed to do it after 10+ attempts just now.
Sometimes there's also a frame skip after pause. I use Kodi but it makes use of MediaCodec.

This is the 4K Max, OS is 7.2.9.2 (PS7292/2984).
I'm not sure what other model number there is, other than the serial number.

[ChaosKing]

Originally Posted by quietvoid

Yea, however it's not consistently reproducible. I managed to do it after 10+ attempts just now.
Sometimes there's also a frame skip after pause. I use Kodi but it makes use of MediaCodec.

This is the 4K Max, OS is 7.2.9.2 (PS7292/2984).
I'm not sure what other model number there is, other than the serial number.

Dancing grain is gone with the latest Kodi 20 beta.

[quietvoid]

Originally Posted by ChaosKing

Dancing grain is gone with the latest Kodi 20 beta.

The build I have installed is from 2022-09-10 git, which was after the last tagged alpha.
So unless it was fixed in git in the last 10 days, it still happens even if rarely.

[benwaggoner]

Originally Posted by ChaosKing

Dancing grain is gone with the latest Kodi 20 beta.

Good to know!

But odd. Maybe Kodi was doing its own SW AV1 decode or something?

[hajj_3]

a new 'chromecast with google tv' has been released which uses the S805X2 chipset which can hardware decode av1. The chipset can only output 1080p not 4k, for all codecs not just AV1. Google's cast codec support page (https://developers.google.com/cast/docs/media) doesn't show this new model so not sure if AV1 decoding is enabled or not but the hardware should be capable of it.

[hajj_3]

Nvidia 4000 RTX series supports AV1 hardware encoding up to 8k60p: https://www.neowin.net/news/nvidia-r...google-chrome/

[Yups]

Originally Posted by hajj_3

Nvidia 4000 RTX series supports AV1 hardware encoding up to 8k60p: https://www.neowin.net/news/nvidia-r...google-chrome/

8k60 is not perfect but they might improve it with newer drivers.

The 8K results have some surprises in store because the RTX 4090 which even has a dual AV1 engine drops as much as the Ampere card regarding P1. Something seems to be wrong with the driver. The AMD card even drops related to the average frame rates. The frame rate suffers from strong stuttering as the 1% percentile is barely more than one FPS.

https://www.capframex.com/tests/AV1%...l%20Arc%20A770

[benwaggoner]

Originally Posted by hajj_3

Nvidia 4000 RTX series supports AV1 hardware encoding up to 8k60p: https://www.neowin.net/news/nvidia-r...google-chrome/

Even if unfixed, that's unlikely to be a serious limitation in this time frame. The global media industry is pretty darn committed to 4Kp24 for scripted entertainment. 4K because costs go up rapidly with more pixels, and the visual benefit of >4K for scripted entertainment is far from demonstrated. Lots of big feature films still do most or all of their CGI effects at 2K. And the commitment to 24p for the "film look" is almost religious. And at 24p, the default shutter speed is 1/48th sec, so almost any motion of camera or content introduces enough motion blur to eliminate any theoretical extra detail 8K could offer.

Live sports is much more interested in high frame rates, so I expect 4Kp120 to be a much bigger deal than 8Kp60 over the next 5+ years.

[nevcairiel]

I would take "dual engines" more like "can work on two streams at the same time without performance loss", not necessarily "twice as fast on one stream", so I wouldn't automatically say its a clear bug, unless NVIDIA has specifically stated otherwise.

Also while of course relevant because its used on eg. YouTube, in my experience browsers are not the best choice to fully explore hardware capabilities.
Not sure if thats the test methodolgy (like not excluding the first second of startup glitches), but not even able to produce fully fluid 4K 60 fps is either a software or a testing failure, because a dedicated video player certainly can render 4k 60 fps without any frame drops.

[Yups]

Originally Posted by nevcairiel

I would take "dual engines" more like "can work on two streams at the same time without performance loss", not necessarily "twice as fast on one stream", so I wouldn't automatically say its a clear bug, unless NVIDIA has specifically stated otherwise.

Exactly, on my Arc A380 and also my Iris Xe only one of the two decoder is working in this video. The second decoder won't give extra performance on a single video, obviously same for RTX 4090.