Evaluation of HEVC decoders (SW, Hybrid and HW)

[NikosD]

With Intel HW I always remember having 100% use of QuickSync MFX Decode engine, using GPA monitoring tool.

Also, my Nvidia monitoring tools always show me a slow progress in reaching 100% or 99% use of VPU.

Hybrid decoding on Intel HW, like HEVC decoding using Haswell, also uses GPU 100% when decoding is done by PowerDVD 15 decoder or MS DS/MFT H.265.

LAV can't keep up to 100%, as I have already posted in LAV filters thread.

[edigee]

I'm getting only black screen on those VP9 videos. Latest LAV ,GTX 960 ,DXVA (native or copyback), MPC-HC ,madVR or EVR CP ,W 7 64bit.
On software mode they work but with high CPU load.

[huhn]

you need the nightly lavfilter version.

edit: you wouldn't get an blackscreen with the older ones so you should have it...

[edigee]

you need the nightly lavfilter version.

edit: you wouldn't get an blackscreen with the older ones so you should have it...

I have 0.67.0.30
VP9 box is checked
Someone said here earlier it might be a driver issue as long as GTX 960 is credited with VP9 decoding capabilities.
I.m using 359.00 .The latest one gives me higher temps on video playback so I rolled back.

[nevcairiel]

VP9 only works in 361.43

[edigee]

VP9 only works in 361.43

Thanks ,good to know!
I'll wait for the next one though, with 361.43 my card runs at 60 degrees in some videos where "the old" 359.00 gives much lower temps. Only 40-45 degrees. Same settings ,same version of LAV ,MPC-HC ,madVR. I found on the specific driver thread on guru 3d forum that I'm not the only one reporting that behaviour.

[wanezhiling]

http://www.fudzilla.com/images/stories/2015/November/rtg-polarisgpu-2.jpg

Polaris’ display block will support features like HDMI 2.0a and DisplayPort 1.3, and its multimedia features will include H.265 Main 10 decode at resolutions up to 4K, and 4K H.265 encoding at up to 60 FPS.

[jones1913]

Polaris’ display block will support features like HDMI 2.0a and DisplayPort 1.3, and its multimedia features will include H.265 Main 10 decode at resolutions up to 4K, and 4K H.265 encoding at up to 60 FPS.
Finally a big step forward after years of stagnation.

Here are some more slides including the "H.265 Main 10" statement:
AMD announces Polaris architecture (http://videocardz.com/58031/amd-announces-polaris-architecture-gcn-4-0)

OT:
Some cool things may come at AMD side the next quarters:
Zen based APU with HBM to be AMD Carrizo successor (http://hexus.net/tech/news/cpu/89246-zen-based-apu-hbm-amd-carrizo-successor)
AMD Further Embraces Open Source with GPUOpen (http://wccftech.com/amd-further-embraces-open-source-with-gpuopen-for-games-and-compute)
AMD/Radeon Has Continued Making Much Linux Graphics Progress (http://www.phoronix.com/scan.php?page=news_item&px=Top-AMD-RTG-2015-News)
AMD @ SC15: Boltzmann Initiative Announced - C++ and CUDA Compilers for AMD GPUs (http://www.anandtech.com/show/9792/amd-sc15-boltzmann-initiative-announced-c-and-cuda-compilers-for-amd-gpus)

[aufkrawall]

Well, this was to be expected. It'd be more interesting if there is native VP9 support or any at all. I can't watch 4k 60fps VP9 in Chrome or Firefox since they are too slow. LAV works fine with my 2500k@4.8Ghz.

[sexus]

sup , anyone able to help me out with this issue , ive got a question about HEVC cuvid support

so ive tried to play a HEVC movie yesterday and my videoplayer application crashed , once ive unchecked HEVC support in cuvid it worked , im using 350.12 nvidia drivers and lav filters 0.65, please help

[edigee]

If your card can decode HEVC (GTX950/960, Intel Skylake) you only have to check the HEVC box on the LAV Video settings and you can use DXVA (copyback or native) for playing the videos. AFAIK Cuvid is not suitable for that ,didn't use it for ages anyway.
If you're on other platform without or with limited support for HEVC decoding check the HEVC box and "none" under "Hardware Decoder To Use". The video will play in software mode with higher CPU load ,of course.

[aufkrawall]

Test GTX 960 on crowd_run_2160p50_VP9_250Mbps.mkv

Decode:
68-77-80 /3% CPU/94% VPU
Are the first values GPU load? If yes, it splits load between GPU & VPU part instead of GPU & CPU like with older Maxwell HEVC hybrid decoder?

[huhn]

MIN, AVG and MAX FPS. there is no GPU load it is a real hardware decoder.

[aufkrawall]

Ok. I could swear that at first it has been said it would be hybrid.

[nevcairiel]

We assumed so, but it turns out the GTX960 has the full decoder chip from the Tegra line, which has had VP9 support for a while - it was just never exposed on PCs yet. So yes, its full hardware acceleration.

[aufkrawall]

Thanks for clarifying!

[CruNcher]

If your card can decode HEVC (GTX950/960, Intel Skylake) you only have to check the HEVC box on the LAV Video settings and you can use DXVA (copyback or native) for playing the videos. AFAIK Cuvid is not suitable for that ,didn't use it for ages anyway.
If you're on other platform without or with limited support for HEVC decoding check the HEVC box and "none" under "Hardware Decoder To Use". The video will play in software mode with higher CPU load ,of course.

In certain Edge Cases cases CUVID can actually be usefull depending on your hardware config but most probably if that ist he case the CPU will be even more efficient and overall most probably then even consume less power alone ;)

Because a huge Difference in CUVID decoding mode is that the Frequency isn't going to change dynamically like it does in DXVA this can mean depending on the Bitstream complexity on a Hybrid Decoder that it might be more efficient if you use CUVID with a higher Frequency then DXVA with a changing Frequency, because for the Bitstream complexity the DXVA Frequency might be to low to maintain a Stable Playback result without dropouts ;)

It's a pretty funny thing on the other side the Frequency for a lower Complexity Bitstream (or if that Bitstream is not Hybrid Decoded) might be way overshoot and you get very inefficient compared to DXVA fot that same Bitstream ;)

How this looks in Practice :)


DXVA Native Changing Frequency (Hybrid Decoder)
http://i2.sendpic.org/t/wU/wUfliF0cQPUXzKg8yVjEjFgURhj.jpg (http://sendpic.org/view/2/i/zfWRonzJtl5Sz0yD5Io5BE9PaEg.png)


CUVID Fixed Frequency (Hybrid Decoder)
http://i2.sendpic.org/t/q8/q8udcrT0AIOfMLdf3rPgBbqwb8z.jpg (http://sendpic.org/view/2/i/ffz97YluX0ijnPqsCHEkzPKcukQ.png)

For DXVA the Frequency of the GPU it can be 200 mhz constant wouldn't matter then the Frequency of the DSP but for a Hybrid Decoder that uses the Shader Units that obviously doesn't applies anymore and when you then come with DXVA ideas for such a Bitstream complexity (heavily fluctuating frequency or to low frequency) it goes heavily wrong ;)

So for the Hybrid Decoder it would need very very good Dynamic Frequency Tuneing for the most efficient results, the same Strategy as with DXVA can't be used here ;)

[NikosD]

It would be the first time ever that they extended a codec beyond its "first" profile, and after releasing the spec at that. Which is why I think its unlikely - but like I said, maybe they'll surprise us!
It would probably require someone other than YouTube starting to pick up VP9 first, so it gets more wide-spread usage.

The new DXVA Checker added that change:

Version 3.12.0 (2016/01/20)
Added "VP9_VLD_10bit_Profile2" as an alternative name of Decoder Device
Minor fixes

Is it Microsoft's or Vendor oriented ?

[nevcairiel]

Microsoft doesn't define such a device, no.

[huhn]

7. Restricted-Mode Profiles
The following restricted-mode profiles for DXVA operations for VP8 and VP9 video decoding are
defined. The GUIDs that identify these profiles will be defined in the header file dxva.h. Additional
restricted-mode profiles for DXVA operations may be defined in the future.
7.2 DXVA_ModeVP9_VLD_10bit_Profile2 Profile
This profile supports the features necessary for a decoder that conforms to VP9 profile 2 for 10 bit video.
In this profile, the accelerator performs bitstream parsing, inverse quantization scaling, inverse transform
processing, motion compensation, and deblocking with the support of 4:2:0 chroma subsampling.
All data buffers shall contain only data that is consistent with the constraints specified for VP9 profile 2
for 10 bit video (including the constraint that BitDepthMinus8Luma and BitDepthMinus8Chroma shall
both be equal to 2).
The associated GUID definition for the corresponding entry in the dxva.h header file is as follows:
// {A4C749EF-6ECF-48AA-8448-50A7A1165FF7}
DEFINE_GUID(DXVA_ModeVP9_VLD_10bit_Profile2,
0xa4c749ef, 0x6ecf, 0x48aa, 0x84, 0x48, 0x50, 0xa7, 0xa1, 0x16, 0x5f,
0xf7);


source: https://www.microsoft.com/en-us/download/details.aspx?id=49188

the pdf it self has a date from 2015 12 14
the site said 2015 01 13

[nevcairiel]

Evil silent updates!
I didn't bother to re-download it because the page still had the same date.

Wonder which hardware will support it however.

[huhn]

unprofessional. the page still shows it is version 1.0

[NikosD]

I was right then ;)

Microsoft surprised us and for the first time made a second profile for VP9 (10bit) after the support of the first one.

All of this year's new architecture GPUs will support it, i think, and Intel and Nvidia will support it in hybrid mode even in the today's architecture, probably.

Wait for the new drivers for Skylake and Nvidia 950/960 GPUs!

[Nullack]

Anyone got a VP9 10 bit sample to trial for this stuff? Thanks

[P.J]

This player is back again (H.265 decoder coming soon)
https://mirillis.com/en/products/splash-free-hd-video-player-all-features.html

:rolleyes:

[wanezhiling]

http://i.imgur.com/Nus4VGT.png

Off topic, is the 'Video Acceleration Settings' option valid on ATI only?

[NikosD]

Probably yes.

I'm not in front of my PC right now, but ATI has some other exclusive features in DXVA checker, like GPU load.

[ionutm80]

Hi to all,

Just a dumb question: is lavfilters fully optimised for Haswell hevc 8bit hybrid decoding?
I'm asking this because I have tried playing a hevc file: http://demo-uhd3d.com/fiche.php?cat=uhd&id=126 , 23.8 Mb/s, 30 fps on my Intel NUC Core I3 4010U, HD 4400 and I have noticed something interesting: GPU was hoovering around 60% with random spikes to 99% and CPU usage varied between 40-60%. Indeed the video was choppy but my impression was that with a little more usage of both CPU and GPU it would have been playable. Tried the same with PowerDVD 15 but the spikes to 99% in GPU usage were more often than using MPC-HC + LAV DXVA native.

Any thoughts on this?

[NikosD]

I have myself reported a few times that LAV seems more optimised for Nvidia's hybrid HEVC decoding than Intel's

The behavior you describe is common to all Intel iGPUs.

Using PowerDVD 15, the opposite is true.
It seems that it is more optimised for Intel than Nvidia.

Generally speaking, those 4K HEVC 8 bit files are a little bit out of hybrid's decoding capabilities using common iGPUs from Intel.

Maybe those rare Iris iGPUs of Haswell and Broadwell could handle them better.

Of course for pure HW decoders from all brands (AMD, Intel, Nvidia) those 8 bit HEVC clips are piece of cake.

[ionutm80]

Of course for pure HW decoders from all brands (AMD, Intel, Nvidia) those 8 bit HEVC clips are piece of cake.

I know :). I have another computer with GTX 960 and I was able to play smoothly all the samples from http://demo-uhd3d.com/.

What I still hope is for an add-on mini pci-e card specifically dedicated for HEVC decoding to emerge (as was the Chrystal Broadcom HD for H.264). A lot of people with Haswell and Broadwell laptops, NUCs will probably not want to upgrade their PCs just to be able to play H.265 files.

[NikosD]

Very interesting new version of DXVA Checker v3.13.0, including a few things requested by many users:

Version 3.13.0 (2016/02/11)
Added support to select GPU
Added support to confirm GPU Usage and Video Engine Usage for NVIDIA GeForce series
Removed some D3D11 TraceLog events
Updated DSF/MFT Viewer
Minor changes

[P.J]

Very interesting new version of DXVA Checker v3.13.0, including a few things requested by many users:

Version 3.13.0 (2016/02/11)
Added support to select GPU
Added support to confirm GPU Usage and Video Engine Usage for NVIDIA GeForce series
Removed some D3D11 TraceLog events
Updated DSF/MFT Viewer
Minor changes

It doesn't work, can't use iGPU :rolleyes:

[nevcairiel]

Switching seems to work fine between my two NVIDIA cards, neat.

[P.J]

For iGPU HD4600, Playback doesn't work at all and Decode works only in CPU mode.
GTX960 (connected to monitor) + HD4600

[NikosD]

As of new Nvidia drivers family 361 (December 2015), using a Windows 10 November Update OS you can get Fermi DX12 (WDDM 2.0) support.

Nvidia, silently, updated their drivers to support Fermi cards in single GPU and non-SLI modes, but not SLI-mode yet.
Maybe that is the reason they didn't advertise their achievement so much, they wrote it in release notes though.

I tried latest 361.91 released yesterday on my GT610 and it works!

I got a WDDM 2.0 driver model with Feature Level support 11.0 and D3D API 11.2/11.3

[nevcairiel]

Note that those 361 drivers only provide WDDM 2.0 support, not DX12 support yet. Its an important step on the way, but its not the final goal yet.
Full DX12 support would show the appropriate D3D API in dxdiag, but right now its still 11.2/3

[NikosD]

Are you sure that Fermi will get full DX12 support ?

[nevcairiel]

Are you sure that Fermi will get full DX12 support ?

That depends what you call "full", but NVIDIA has said it will get DX12 support, that means at least API support to benefit from the lower CPU overhead.

[NikosD]

I thought that WDDM 2.0 was enough for low CPU overhead.

[NikosD]

Speaking of DX12, the other low level API based on AMD's Mantle API - Vulkan - released yesterday by Khronos, as the Next Generation OpenGL successor.

In this case the table of GPUs supporting Vulkan 1.0 is this:

AMD GCN 1.0 & Later

Intel Broadwell (Gen 8.5) & Later

NVIDIA Kepler & Later

[chros]

As of new Nvidia drivers family 361 (December 2015), using a Windows 10 November Update OS you can get Fermi DX12 (WDDM 2.0) support.
What does it mean exactly? only 8bit @ fullHD resolution?
Is it only on Win10?
(I'm asking because of my GT650m on Win8.1)

Speaking of DX12, the other low level API based on AMD's Mantle API - Vulkan - released yesterday by Khronos, as the Next Generation OpenGL successor.
Sorry, what's this for?
Thanks

[NikosD]

Never mind.
It's off - topic and I wouldn't like to be discussed any more.

[Yups]

Speaking of DX12, the other low level API based on AMD's Mantle API - Vulkan - released yesterday by Khronos, as the Next Generation OpenGL successor.

In this case the table of GPUs supporting Vulkan 1.0 is this:

AMD GCN 1.0 & Later

Intel Broadwell (Gen 8.5) & Later

NVIDIA Kepler & Later


Broadwell is Gen8

[vinnytx]

Please, P.J or someone else can tell me how to show this control panel?

CLCvd.ax? No luck with 750Ti too :(

http://i.imgur.com/h9FyKY0.png





I have only this

http://i.imgur.com/Uu9HyYO.jpg

[P.J]

Press Ctrl+Shift before opening it.

[NikosD]

Latest DXVA Checker v3.13 doesn't look compatible with Nvidia Optimus technology.

To be more exact, my Nvidia 740M card in a laptop using Optimus technology, falls back to SW decoding using both LAV and MS DS decoders in DXVA2 native mode.

I even forced DXVA Checker to run on Nvidia card through Nvidia's Control Panel, but still with no success.

It falls back to SW mode.

Could anyone else test if GPU switching technology from Nvidia or AMD is compatible with latest DXVA Checker v3.13 ?

[wanezhiling]

http://bluesky23.yukishigure.com/en/DXVAChecker.html
v3.14 should have fixed that.

[NikosD]

I've been talking with DXVA Checker's developer about the problem and when I saw the new version, I thought the same thing.

Now the problem is that I don't have the laptop available to check it out :)

[P.J]

https://mirillis.com/en/products/splash-free-hd-video-player-history.html

25/03/2016 Splash 2.0.2

NEW FEATURE: Hardware accelerated HEVC/H.265 video decoding support added for NVIDIA and Intel graphics cards

Old right now :rolleyes:

[P.J]

Can someone play a 4K 10bit HEVC 59.940fps video like Samsung_UHD_Ride_on_Board.ts in EVR/EVR-CP mode with GTX 950/960 2GB on a native 4K monitor/TV?
Someone told me that it requires at least GTX 970 or maybe GTX 960 4GB :rolleyes: