Alliance for Open Media codecs

[Yups]

Originally Posted by Beelzebubu

It says profile 0 only, which is 4:2:0 and 8+10bit only. 4:4:4 is profile 1, 12-bit is profile 2 (along with 4:2:2 for 8+10bit).

Even in this leak there is a discrepancy. In the table there is 10 bit for video but the text says it's 8 bit for video. I'm keen to believe the table in this case but not sure. In the RKL-S slide there is 12 bit AV1, it could be just a mistake (confused with 12 Bit HEVC), it could refer to still image or it could be that the Gen12LP in RKL-S uses a newer decoding unit than TGL-U.

[Beelzebubu]

Originally Posted by Yups

Even in this leak there is a discrepancy. In the table there is 10 bit for video but the text says it's 8 bit for video.

I think you're misreading a table converted to text. The table says "profile 0 (8-bit and 10-bit 4:2:0) 4k60 video, 16k still (HW decode)". The text says "AV1 codec support<newline>Profile 0 (10-bit 4:2:0)<tab>16k (still picture)<newline>Profile 0 (8-bit 4:2:0)<tab>4k x 2k (video)". I inserted the <tab> because it matches what the table says: 4k60 (4k x 2k) video, 16k still (picture) and each supporting profile 0 (8-bit & 10-bit, 4:2:0).

[Sagittaire]

Originally Posted by Beelzebubu

aomenc --denoise-noise-level= does the same as a FFT denoiser + examples/noise_model + aomenc --film-grain-table=.

Well seem not work. I have denoise but seem not have FGM in output.

You have command line exemple?

[Beelzebubu]

Originally Posted by Sagittaire

Well seem not work. I have denoise but seem not have FGM in output.

Please show your commandlines.

[Sagittaire]

Originally Posted by Beelzebubu

Please show your commandlines.

for exemple.

aomenc.exe -o ToS-1000.ivf C:\ToS_1920x800_xdither.y4m --ivf --i420 --width=1920 --height=800 --fps=25000/1000 --passes=2 --pass=1 --fpf=stats.log --target-bitrate=960 --maxsection-pct=4000 --buf-sz=12000 --buf-initial-sz=8000 --buf-optimal-sz=10000 --end-usage=vbr --cpu-used=3 --verbose --tune=ssim --psnr --q-hist=30 --bit-depth=10 --bias-pct=75 --kf-max-dist=120 --aq-mode=2 --test-decode=fatal --limit=1332 --skip=214 --denoise-noise-level=25

aomenc.exe -o ToS-1000.ivf C:\ToS_1920x800_xdither.y4m --ivf --i420 --width=1920 --height=800 --fps=25000/1000 --passes=2 --pass=2 --fpf=stats.log --target-bitrate=960 --maxsection-pct=4000 --buf-sz=12000 --buf-initial-sz=8000 --buf-optimal-sz=10000 --end-usage=vbr --cpu-used=3 --verbose --tune=ssim --psnr --q-hist=30 --bit-depth=10 --bias-pct=75 --kf-max-dist=120 --aq-mode=2 --test-decode=fatal --limit=1332 --skip=214 --denoise-noise-level=25

the noise is dithering. Perhaps not adapted denoising value for that? I try with --denoise-noise-level=5,10,50 and no result.

[Beelzebubu]

Originally Posted by Sagittaire

for exemple.



the noise is dithering. Perhaps not adapted denoising value for that? I try with --denoise-noise-level=5,10,50 and no result.

Dithering should be fine. Does --help include a description for --denoise-noise-level? It might be built without CONFIG_DENOISE (in which case the --help output will be missing also). Otherwise I'm not entirely sure, it has worked fine for me. Does it work on 8-bit material?

[Sagittaire]

Originally Posted by Beelzebubu

Dithering should be fine. Does --help include a description for --denoise-noise-level? It might be built without CONFIG_DENOISE (in which case the --help output will be missing also). Otherwise I'm not entirely sure, it has worked fine for me. Does it work on 8-bit material?

well seem work only if you have noise_model.exe in same directory than aomenc.exe ... ?

[benwaggoner]

Originally Posted by soresu

Isn't all this pointless when hardly any TVs actually support 12 bit though?

Certainly for the moment, yes. Perhaps 12-bit could become a default with VVC or AV2.

[soresu]

Oh man, the new Cortex X1 CPU core has twice the NEON units of the A78 and previous Axx cores.

At 4x 128 bit units that's some real grunt for an off the shelf design, should be noice for encoding.

[mzso]

What kind of beast of a computer would you need to play 8k AV1?
I wanted to check out a few 8k videos on Youtube, but I accidentally downloaded AV1 streams, and the videos ate up my R5 1600. Even VP9 did...
I wonder if any mere mortal will have such a machine in the next decade.

[ChaosKing]

Beast machine? You used the wrong player!
In my browser it is almost smooth, with some dopped frames
In mpv it is perfect with ~ 55% cpu

My pc: ryzen 2600 (no oc) + gtx 1070
8K Video: https://www.youtube.com/watch?v=ChOhcHD8fBA

[huhn]

your GPU can decode this video...

[ChaosKing]

gtx 1070 does not have a HW AV1 decoder...
And mpv uses the dav1d decoder which works on cpu only.

[nevcairiel]

My 7900X (OC) can also play that video in 8K without drops, at 40% usage at most - straight in Chrome 83. Make sure its not your GPU that struggles downscaling the 8K video.

With Ryzen making 8 cores available to "mere mortals" for relatively low pricing, I don't think you need any particular "beast" right now, nevermind the next decade.

[benwaggoner]

Originally Posted by nevcairiel

My 7900X (OC) can also play that video in 8K without drops, at 40% usage at most - straight in Chrome 83. Make sure its not your GPU that struggles downscaling the 8K video.

With Ryzen making 8 cores available to "mere mortals" for relatively low pricing, I don't think you need any particular "beast" right now, nevermind the next decade.

How well are the fast SW decoders scaling with multiple cores? VP9 struggled to get much parallelism due to some unfortunate serialization in the loop filters and the lack of a clear reference structure that allowed for parallel decoding of non-reference frames.

AV1 is certainly going to be more parallizable than VP9. But how far have the best decoders gone so far?

[huhn]

Originally Posted by ChaosKing

gtx 1070 does not have a HW AV1 decoder...
And mpv uses the dav1d decoder which works on cpu only.

the video is as far as i can see only VP9 8 bit.
if there is a AV1 version i could try it with an R7 3700X which could yield up to ~2.5x more performance then a r5 2600 in this special case

Originally Posted by benwaggoner

How well are the fast SW decoders scaling with multiple cores? VP9 struggled to get much parallelism due to some unfortunate serialization in the loop filters and the lack of a clear reference structure that allowed for parallel decoding of non-reference frames.

AV1 is certainly going to be more parallizable than VP9. But how far have the best decoders gone so far?

this doesn't seam to be the case anymore or it's not as bad as it was in the past. while the microsoft VP9 decoder has issues using more then 8 threads ffmpeg in lavfilter doesn't have this issues easily using 70-90% CPU usage on a 8 core 16 thread CPU. it sometimes falls to 50% or lower CPU usage but with usually over 100 FPS.

[ChaosKing]

I think I activated AV1 on YT, it was under /testtube or so. Just google it.
I used youtube-dl.exe to download the video, it was the av1 version.

@nevcairiel I use Vivaldi and have currently over 100 tabs open xD It can have an effect on performance

Originally Posted by benwaggoner

AV1 is certainly going to be more parallizable than VP9. But how far have the best decoders gone so far?

https://www.phoronix.com/scan.php?pa....7-Performance

Over 300fps for 4k is not bad
I hope 10bit decoding will be faster on ARM soon too. My 4k fire tv stick struggles with 1080p 10bit video

[huhn]

i'm still confused by YTDL doesn't showing it before. but this make far more sense now to me.
the stream is very easy to decode i get about 40~ FPS with 50% CPU usage using lavfilter from mpc-hc 1.9.4.

[soresu]

Originally Posted by ChaosKing

I hope 10bit decoding will be faster on ARM soon too. My 4k fire tv stick struggles with 1080p 10bit video

10 bit isn't likely to get much better on ARM64 as most of the NEON asm has been written at this point - the problem is that all Fire TV products use 32 bit ARM Android as a base due to laziness on their part IMHO, therefore you won't get as much performance out of it as a phone or tablet with the same HW spec.

They probably will fill out the ARM32 NEON asm over time, though I'm not sure if the performance will match the ARM64 equivalent code path.

10 bit on x86 is another matter entirely - as far as I am aware there are zero SIMD assembly optimisations currently for AVX2, SSSE3 or SSE2 where 10+ bpc video is concerned.

[Beelzebubu]

Originally Posted by benwaggoner

How well are the fast SW decoders scaling with multiple cores? VP9 struggled to get much parallelism due to some unfortunate serialization in the loop filters and the lack of a clear reference structure that allowed for parallel decoding of non-reference frames.

AV1 is certainly going to be more parallizable than VP9. But how far have the best decoders gone so far?

I don't think this is true. From a decoder's point-of-view, the reference structure (most notably the cross-frame entropy dependency) and postfilter dependency (tile-crossing) between VP9 and AV1 are the same, and equally parallelizable. dav1d uses the same techniques for frame threading as FFmpeg's native VP9 decoder (ffvp9) and achieves siimilar concurrency multipliers as dav1d. Tile threading works the same. See e.g. graphs at https://youtu.be/WgfklAi50nM?t=1378 (VP9 vs. AV1) and https://youtu.be/WgfklAi50nM?t=345 (AV1 tile vs. frame vs. both multithreading), and note that these results are over a year old, dav1d has surpassed ffhevc (SW) decoding speed since then. See also the documentation for the threading model in dav1d.

The practical problem in ffvp9 is that it decided (to fit in FFmpeg's more static design) to only allow one threading type (frame or tile) instead of multiple concurrently (frame and tile) like dav1d does. That's the only reason dav1d scales better with multi-threading. We could have resolved that, but it was decided that ffvp9 was fast enough and it wasn't worth it.

(I can explain libaom's and libvpx' threading models if you want to learn more, but since they are a subset of dav1d/ffvp9, I was assuming this would be enough. I'm not familiar with gav1's threading model.)