x265 HEVC Encoder

[LigH]

x265 2.8+1-478bfe2b7673 (stable release, GCC 7.3.0, Win32+Win64)

last few fixes in analysis mode

[user1085]

Originally Posted by LigH

x265 2.8+1-478bfe2b7673 (stable release, GCC 7.3.0, Win32+Win64)

last few fixes in analysis mode

Do you also have a Linux build?

[Sagittaire]

Originally Posted by Ashok Kumar Mishra

Version 2.8
5. VMAF (Video Multi-Method Assessment Fusion)
Added VMAF support for objective quality measurement of a video sequence.
Enable cmake option ENABLE_LIBVMAF to report per frame and aggregate VMAF score. The frame level VMAF score does not include temporal scores.
This is supported only on linux for now.

VMAF is objective metric from Netflix:
https://medium.com/netflix-techblog/...c-653f208b9652

Really good idea to impliment this metric. VMAF seem have really good correlation with HVS. Why not include PSNR-HVS-M (seem have really good correlation too)?

Possible to introduce tune profil for VMAF or PSNR-HVS-M like for PSNR and SSIM?

[jd17]

Originally Posted by Sagittaire

Possible to introduce tune profil for VMAF or PSNR-HVS-M like for PSNR and SSIM?

I would like that too!

[Asmodian]

Any tune profile for VMAF should probably be the default, given that it correlates very well with the human visual system.

[user1085]

Anyone know how to build x265 2.8 with AVX512 enabled on Ubuntu?

I built x265 from 2.8 release tarball but x265 -V doesn't show me it's using AVX512 capability on a Intel Xeon

x265 [info]: HEVC encoder version 2.8
x265 [info]: build info [Linux][GCC 7.3.0][64 bit] 10bit
x265 [info]: using cpu capabilities: MMX2 SSE2Fast LZCNT SSSE3 SSE4.2 AVX FMA3 BMI2 AVX2

Here's the CPU info

uso@xvmub:~/x265_2.8/build/linux$ lscpu
Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              2
On-line CPU(s) list: 0,1
Thread(s) per core:  2
Core(s) per socket:  1
Socket(s):           1
NUMA node(s):        1
Vendor ID:           GenuineIntel
CPU family:          6
Model:               85
Model name:          Intel(R) Xeon(R) Platinum 8168 CPU @ 2.70GHz
Stepping:            4
CPU MHz:             2693.677
BogoMIPS:            5387.35
Virtualization:      VT-x
Hypervisor vendor:   Microsoft
Virtualization type: full
L1d cache:           32K
L1i cache:           32K
L2 cache:            1024K
L3 cache:            33792K
NUMA node0 CPU(s):   0,1
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology cpuid pni pclmulqdq vmx ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch pti tpr_shadow vnmi ept vpid fsgsbase bmi1 hle avx2 smep bmi2 erms rtm mpx avx512f avx512dq rdseed adx smap clflushopt avx512cd avx512bw avx512vl xsaveopt xsavec xsaves

[LigH]

AVX-512 is only enabled if you explicitly add --cpu avx512 to the command line; without it is not even checked for in CPU capabilities, apparently... so try:

x265 --cpu avx512 -V

Is it reported then?
__

Originally Posted by user1085

Do you also have a Linux build?

No, I was a) quite confident that Linux users know how to install and use GNU compilers (it's good enough documented on Multicoreware's Bitbucket), and b) not sure how portable one build under one specific distro is. Windows builds are quite portable, and knowledge about compiling from sources is probably not so common among Windows users.

[user1085]

Originally Posted by LigH

AVX-512 is only enabled if you explicitly add --cpu avx512 to the command line; without it is not even checked for in CPU capabilities, apparently... so try:

x265 --cpu avx512 -V

Is it reported then?
__



No, I was a) quite confident that Linux users know how to install and use GNU compilers (it's good enough documented on Multicoreware's Bitbucket), and b) not sure how portable one build under one specific distro is. Windows builds are quite portable, and knowledge about compiling from sources is probably not so common among Windows users.

No worries I figured out how to compile, didn't see the instructions earlier

Re:avx512, this worked for me

./x265 --asm avx512 -V

[LigH]

Oops, should have read the manual, "asm" instead of "cpu".

Thank you for confirming anyway.

[benwaggoner]

Originally Posted by Sagittaire

VMAF is objective metric from Netflix:
https://medium.com/netflix-techblog/...c-653f208b9652

Really good idea to impliment this metric. VMAF seem have really good correlation with HVS. Why not include PSNR-HVS-M (seem have really good correlation too)?

Possible to introduce tune profil for VMAF or PSNR-HVS-M like for PSNR and SSIM?

VMAF is our least-bad metrics, but it definitely has limitations and blind spots. It doesn't catch banding well, particularly in lower luma. It is only trained on x264 8-bit SDR, and without a lot of different psychovisual tuning, so it doesn't do a great job of discriminating between some kinds of psychovisual tradeoffs. And it's never seen the newer kind of artifacts in HEVC, so it doesn't know about customer preference between them. It doesn't know about <300 Kbps or >1080p. It was only tested against a particular display at 3x screen heights. The temporal comparison base metric is pretty weak.

It's the best thing we've got, but it definitely isn't perfect.

Also, the question of how to go from scores for individual frame to a per-clip score remains unanswered. A file that varies from 30-80 VMAF is going to be perceived as a lot worse than one that varies from 50-60, but the mean of the frames won't show any difference.

That's not a VMAF specific problem; it applies to all metrics where we are trying to figure out quality over time from individual frame metrics.

[jd17]

Wait a minute.... The Amazon video specialist does NOT think Netflix's metric is the holy grail?
I'm shocked! Shocked!
Kidding of course... I just think someone had to say it.

I do think Netflix abuses VMAF findings though.
Bitrates between 2000kbit/s and 3000kbit/s for 1080p x264 video are just not good. No matter what VMAF tells you...

[LigH]

The main limit is probably a bandwidth cap; restricted VBR. Some scenes will have a demand that can't be satisfied in a streaming usage. But quality tests will often be done for a local playback usage, I guess...

[jd17]

Originally Posted by LigH

The main limit is probably a bandwidth cap; restricted VBR. Some scenes will have a demand that can't be satisfied in a streaming usage.

Is this a comment on my complaint about the 2000-3000kbit/s?
If that is the case - I do not think bandwidth cap is a factor.
I was referring to the "highest" quality Netflix streams at 1080p for some content. You cannot get more, no matter how high your bandwidth is.
This not only used to be much higher in the past, it still is for some content with more grain.
Additionally, Amazon streams significantly higher 1080p AVC bitrates across the board, so bandwidth can hardly be a limitation...

I think it's just Netflix saving bandwidth, because most people either watch this stuff on smaller screens with a viewing distance too large, or they just do not care / do not see the flaws...
Either way - Netflix abuses it.

[nevcairiel]

The bitrate you target is really not related to the metric you use. You can use a metric to optimize encoding settings for any target bitrate to get the best out of the bits available to you.

[jd17]

You might be misunderstanding what I tried to say...
I know how metrics work.

I merely tried to point out that I think Netflix is abusing their internal VMAF findings to justify streaming at much lower bitrates.
I would argue they lowered their bitrate target (at least partly) as a result of their use of VMAF.

No matter how clean your video is, 1080p x264 will not be good at 2000-3000kbit/s, especially in the backgrounds.

Maybe you understand where I am coming from when you look at this blog:
https://medium.com/netflix-techblog/...n-7e99442b62a2

In theory, the arguments in that blog are well reasoned.
Use low bitrates for cartoons and clean video, use high bitrates for lots of movement, grain, and so on...

However, especially within the last year or so (look at recent releases of Netflix series for instance), they have moved to the aforementioned low bitrates as highest 1080p target bitrate. On live action video!

While 2000-3000kbit/s is indeed completely fine for Cartoon, it is not for live action video.

Is that VMAF's fault? Of course not.
But I think Netflix saw that the VMAF scores are not too bad for 1080p video at those bitrates, so they decided to lower their targets accordingly.


Sorry for the long off-topic discussion...
Originally, I just wanted to make a little joke and complain about Netflix's quality in the process...
I did not mean for it to drift so far away from x265!

[foxyshadis]

Sorry guys, this was getting way too offtrack. AVX512 vs Threadripper has been moved to the PC Hardware forum.

[IgorC]

Originally Posted by benwaggoner

Also, the question of how to go from scores for individual frame to a per-clip score remains unanswered. A file that varies from 30-80 VMAF is going to be perceived as a lot worse than one that varies from 50-60, but the mean of the frames won't show any difference.

It was solved in HVMAF

https://medium.com/netflix-techblog/...k-e19f1e3a277f

About Harmonic mean https://en.wikipedia.org/wiki/Harmonic_mean

Since the harmonic mean of a list of numbers tends strongly toward the least elements of the list, it tends (compared to the arithmetic mean) to mitigate the impact of large outliers and aggravate the impact of small ones.

P.S. It's not news at all. Geometric and harmonic means were in use since forever.

[K.i.N.G]

Hi, I have a possible request/idea.

Wouldn't it be possible to integrate a bitrate multiplier by mask (a greyscale image or video) as source? Preferably with an option to set strength/influence (so pure black doesn't always give 0% )
I think this would be quite useful. Not sure how difficult it would be to add.

It's just an idea, I don't want to come over as someone who 'demands' something from open source devs.

[LigH]

x265 2.8+10-a7bd0622ece5 (merge with stable)

several bug fixes (build errors, dyn.ref., scaled enc.)

[benwaggoner]

Originally Posted by IgorC

It was solved in HVMAF

https://medium.com/netflix-techblog/...k-e19f1e3a277f

About Harmonic mean https://en.wikipedia.org/wiki/Harmonic_mean

Looks like a promising improvement, but this stuff never gets SOLVED.

Averaged over a two hour movie, a harmonic mean can still not catch periods of terrible quality.