Alliance for Open Media codecs

[Selur]

I wish aomenc/vpxenc had GOP-level parallelism.

Which would require 2pass encoding and a fixed gop structue (in regard to the gop sizes), iirc 2nd pass normally should be able to overwrite GOP to archive vbv limits (not totally sure).

[Gravitator]

ffmpeg -hide_banner -t 10 -c:v libaom-av1 -i 1.mp4 -benchmark -f null - (43 fps)
ffmpeg -hide_banner -t 10 -c:v libdav1d -i 1.mp4 -benchmark -f null - (52 fps)
ffmpeg -hide_banner -t 10 -c:v libdav1d -threads 1 -tilethreads 2 -i 1.mp4 -benchmark -f null - (61 fps)
ffmpeg -hide_banner -t 10 -c:v libdav1d -threads 2 -tilethreads 2 -i 1.mp4 -benchmark -f null - (65 fps)

[lvqcl]

Originally Posted by v0lt

sse2, sse4.1?

It seems that one of dav1d developers said: "we don't care about mmx/sse2 support anyway" (link). Have no idea about sse4.1.

[SmilingWolf]

Originally Posted by lvqcl

It seems that one of dav1d developers said: "we don't care about mmx/sse2 support anyway" (link). Have no idea about sse4.1.

BBB is part of TwoOrioles, so it might have been referred to the company based on its userbase.
Still, MMX is hardly relevant nowadays. SSE4.1 as the lowest bar doesn't sound too unreasonable

Also relevant: https://code.videolan.org/videolan/d.../15#note_22262

[NikosD]

Originally Posted by lvqcl

It seems that one of dav1d developers said: "we don't care about mmx/sse2 support anyway"
Have no idea about sse4.1.

Originally Posted by SmilingWolf

Still, MMX is hardly relevant nowadays. SSE4.1 as the lowest bar doesn't sound too unreasonable.

MMX is too old and not that beneficial as it can reach only 64bits (maybe 80bits max)
SSEx should be the base as it is 128bit with very fast implementation on all CPUs of the last 10 years.
Especially SSE2 is mandatory for x64 architecture.
From the last link it's obvious that dav1d developers targeted AVX2 for 256bit acceleration using ASM, but not exclusively.
They are going to optimise for SSEx later.
So no worries, I think.

[marcomsousa]

if they want to go with 4k and 8k videos they have to use AVX2.

[Nintendo Maniac 64]

Originally Posted by NikosD

Especially SSE2 is mandatory for x64 architecture.

You can also usually safely target SSE3 (no, not SSSE3) as well since it's supported on all DDR2-capable 64bit x86 CPUs and newer.

(the only 64bit x86 CPUs that don't support SSE3 are some socket 754 and 939 Athlon 64s which used DDR1)

[Mystery Keeper]

Originally Posted by Selur

Which would require 2pass encoding and a fixed gop structue (in regard to the gop sizes), iirc 2nd pass normally should be able to overwrite GOP to archive vbv limits (not totally sure).

I'm totally fine with that. I usually use 2pass anyway. And, of course, I meant I wish they had it as an option.

[LigH]

@ Nintendo Maniac 64:

Even AMD Athlon64/Phenom (K8-K10 arch.) support some SSE3; but x264/x265 does not use it, considers their implementation as "too slow", I believe.

[marcomsousa]

SSE3-optimised av1_nn_predict


https://aomedia.googlesource.com/aom...6f313f27b1c501

I have developed a SIMD-optimised neural network implementation using
SSE3. I have also added functional equivalence tests between this and
the original implementation. I added aom_clear_system_state() to a few
places where FPU operations are used after av1_nn_predict.

Speed-ups over the original C implementation for various network shapes:
10x64x16: 1.72x
12x12x1: 2.72x
12x24x1: 2.35x
12x32x1: 3.34x
18x24x4: 0.94x
18x32x4: 0.93x
4x16x1: 2.01x
8x16x1: 1.89x
8x16x4: 2.02x
8x24x1: 2.77x
8x32x1: 2.98x
8x64x1: 3.76x
9x32x3: 1.08x
4x8x4: 1.66x

A few awkwardly-shaped networks are slightly slower: these could be
padded to more convenient sizes to use the SIMD kernels.

I also wrote an AVX/AVX2 implementation but on these relatively small
networks it was barely faster than the SSE3 code.

[Nintendo Maniac 64]

Originally Posted by LigH

Even AMD Athlon64/Phenom (K8-K10 arch.) support some SSE3

...but this is exactly what I alluded to?

Athlon 64 CPUs are available on socket 754, 939, and AM2; 754 and 939 used DDR1 memory while AM2 used DDR2, and all AM2 CPUs support SSE3.

(there are some socket 754 and 939 CPUs that support SSE3, though it's kind of hit and miss).

Phenom for reference requires at least DDR2.

[LigH]

I'm sorry, I don't know socket numbers... - so we looked at the same topic from different angles.

[v0lt]

Originally Posted by Wolfberry

I ran the same test as above and get 16/38/46 fps.
What is the CPU you use for testing?
It might be related to the AVX2 code used in dav1d.

Intel i5-3570k (SSE4.1, SSE4.2, AVX), Windows 7 Sp1 x64.

[SmilingWolf]

Status report!
Previous edition: //www.zs-x.com/showthread.ph...49#post1852449
Whatever paragraph I don't repeat here can be assumed to be the same as in the aforementioned post

First of all: graphs! Click to enlarge
Y axis: chosen metric
X axis: bits per pixel

720p:


1080p:


BD rates for 720p:

x264 -> rav1e (yeah you read that right!)
        RATE (%)  DSNR (dB)
 MSSSIM -0.736889 0.0375593
PSNRHVS -5.5274   0.375081

rav1e -> x265
        RATE (%) DSNR (dB)
 MSSSIM -26.5291 1.29942
PSNRHVS -27.1134 1.70509

x265 -> libaom
        RATE (%) DSNR (dB)
 MSSSIM -18.9088 0.7852
PSNRHVS -15.3123 0.761791

BD rates for 1080p:

x264 -> rav1e (yeah you read that right again!)
        RATE (%) DSNR (dB)
 MSSSIM -4.92009 0.235151
PSNRHVS -7.23088 0.473125

rav1e -> x265
        RATE (%) DSNR (dB)
 MSSSIM -26.7063 1.16103
PSNRHVS -28.0007 1.53902

x265 -> libaom
        RATE (%) DSNR (dB)
 MSSSIM -26.486  0.938124
PSNRHVS -21.7431 0.905916

Encoders:
x264 157-2935-545de2f
x265 2.9-4-471726d3a046
rav1e 0.1.0-702-ab4d23e2
libaom 1.0.0-908-g3a607f7b0

Cmdlines:
x264 --preset veryslow --tune ssim --crf 16 -o test.x264.crf16.264 orig.i420.y4m
x265 --preset veryslow --tune ssim --crf 16 -o test.x265.crf16.hevc orig.i420.y4m
rav1e --low_latency false -o test.rav1e.cq80.ivf --quantizer 80 -s 2 --tune psnr orig.i420.y4m
aomenc --frame-parallel=0 --tile-columns=3 --auto-alt-ref=1 --cpu-used=4 --tune=psnr --passes=2 --threads=2 --end-usage=q --cq-level=20 --test-decode=fatal -o test.av1.cq20.webm orig.i420.y4m

Notes:
So as you can see, the rav1e and aomenc cmdlines have been slightly adjusted to take advantage of the bugfixes and updates from the last months.
In particular, rav1e has been gifted by Frank Bossen the ability to create a B-pyramid, which almost single handedly decreed rav1e's advantage over x264.
A word of warning on this last point: it's still kind of a mixed bag. In very flat, static scenes like PresageFlowerWalk x264 still rules by quite a margin, while rav1e takes the crown in clips like F.Y.C and PresageFlowerFight

F.Y.C, x264 -> rav1e:
        RATE (%) DSNR (dB)
 MSSSIM -18.451  1.01281
PSNRHVS -25.7463 2.03419

PresageFlowerFight, x264 -> rav1e:
        RATE (%) DSNR (dB)
 MSSSIM -31.4953 1.80761
PSNRHVS -31.0827 2.27546

PresageFlowerWalk, x264 -> rav1e:
        RATE (%) DSNR (dB)
 MSSSIM 66.2264 -1.70084
PSNRHVS 70.8208 -2.28853

(as always, a negative BD rate means improvement, positive means regression)

Considerations about times with libaom:
I'm using my desktop PC to run all the encodes. It is also my main study/work PC, so the times can come quite off. Plus, I run multiple encodes in parallel, which further messes up timings.
HOWEVER, between annoying bugs and a lot of stuff, the first report did cost me nearly a week of time (this includes having to re-run some encodes because sh*t happened) ONLY to encode with libaom.
Taking advantage of the recent bugfixes and improvements I have been able to rework my workflow and bring down that time to a couple days only, WITHOUT having to touch the --cpu-used parameter and no night time encoding.
All in all, I am pretty satisfied.

This concludes my (bi-monthly?) report.
As always, I'm open to any kind of feedback to improve my comparisons and my encodes.

[benwaggoner]

So, what's everyone's favorite AV1 decoder app on Windows? Chrome looks to be not converting from video to PC range correctly (blacks are washed out, contrast is low, etcetera). Is there a nightly of something that does AV! correctly for apples-apples?

[SmilingWolf]

Originally Posted by benwaggoner

So, what's everyone's favorite AV1 decoder app on Windows? Chrome looks to be not converting from video to PC range correctly (blacks are washed out, contrast is low, etcetera). Is there a nightly of something that does AV! correctly for apples-apples?

VLC 3.0.5 (Nightly). I fixed my nVidia settings just today because I had that same problem while playing back the ToS fragment I use for the tests. Plays out correctly now.
In alternative, ffplay for quick stuff when I already have a bunch of command prompts open in the right path.

[LigH]

I use almost only MPC-HC. Which uses LAV Filters with a direct API. It was able to play AV1 clips from the YouTube beta playlist and some tiny own encodes (I don't have powerful CPU's available). So, only a limited experience, yet, but it appears to work.

[SmilingWolf]

32/64bits binaries (GCC 9.0):
av1-1.0.0-941-gd2a592e1c: https://mega.nz/#!F5Am2KyK!9aQ6_7mM2...6_OaZahvKCHPWQ

[mandarinka]

Originally Posted by Mystery Keeper

I wish aomenc/vpxenc had GOP-level parallelism. When each thread is encoding one GOP, and then they are stitched together. That would make use of all CPU power without compromising quality/compression.

You could get the same results by splitting manually into X parts end encode them separately at once. I'm not sure how much does libvpx/libaom count with that. It works great with x264 and x265 (using raw output at least).

[mandarinka]

Originally Posted by LigH

@ Nintendo Maniac 64:

Even AMD Athlon64/Phenom (K8-K10 arch.) support some SSE3; but x264/x265 does not use it, considers their implementation as "too slow", I believe.

SSE3 is not particularly useful for multimedia and it's just a few instructions introduced in Presscot P4 and Venice 90nm K8.

You probably mean SSSE3 (SSS instead of SS) aka "Suplemental SSE3" which is a confusing and dumb name. Probably should have been SSE4 but got renamed for marketing reasons. Or SSE3 was not supposed to be SSE3 originally.

SSSE3 is very useful for encoding and decoding, but only comes on Core 2 chips, and Bobcat/Bulldozer and later cores from AMD. K10 and K8 end at the not-so-important SSE3.
(Note that x265 actually needs SSSE3 + SSE4 to be useful, you are barred from most of assembly optimization if you only have SSSE3, like with 65nm Core 2s or pre-Sandy Bridge Pentium/Celeron).