Evaluation of HEVC decoders (SW, Hybrid and HW)

[huhn]

VP9 is checked by default in lavfilter for a while now and yes i tried bothd3d11 and dxva2 for that video it falls back to software.
according to GPU-Z the clock was at 6 mhz. even when i "boost" the clock through other means (madVR) the speed didn't change.

when hardware decoding works the video decoder was at 99-100 % load in taskmanager.

i'm on 19.9.2 and they changed something:
System instability may be experienced on some Radeon RX 5700 series graphics system configurations when watching video content in a web browser.

i would usually get a driver crash when playing a video with hardware acceleration. i didn't got that with that new version.
one of the reasons i'm not in this thread earlier. there was simply no point. i take it back just crashed...

on old VP9 video i used to trigger a limitation in lavfilter using huge CPU queues it always uses software decoding and that file clearly worked on my 1060. mpc-hc with youtubedl works and VP9 decoding is used for 1080p.

[NikosD]

VP9 is checked by default in lavfilter for a while now and yes i tried bothd3d11 and dxva2 for that video it falls back to software.
according to GPU-Z the clock was at 6 mhz. even when i "boost" the clock through other means (madVR) the speed didn't change.

when hardware decoding works the video decoder was at 99-100 % load in taskmanager. I think VP9 was not checked by default on AMD cards, but anyway.

In order to find out the real clocks (and many other parameters) for video decoding on any chip (Intel,nVidia, AMD), I would try HWiNFO v6.10 or newer.

It seems like a broken driver for HW acceleration of VP9, but if suddenly VP9 acceleration is working on a browser (Edge, Chrome etc) playing YouTube and DXVA Checker is still falling back to SW decoding, then you should mention it to DXVA Checker developer to update his app.

[huhn]

some files work some don't nothing todo with dxva checker.

video clock in hwinfo is greyed out.

[DMU]

8K_Ex_sample.mp4
39 45 48

Please explain to me what these numbers mean or where they come from?

[huhn]

it's the same format as the rest of the tests.
min avg max fps.

the test show the card can only do 8k 30.

[DMU]

Thank you, huhn.
What kind of test do you mean?

[huhn]

we test the speed of hardware decoder here.
the test is done using dxva checker.

[NikosD]

Does anyone have access to an Intel Gen11 iGPU laptop or other product ?
I would like to see its 8K HEVC/VP9 decoding speed.
I don't know if there is anything already posted on Net.

[littleD]

Some Gen11 info from whitepapers:
The Media FF baseline configuration increases in Gen11to include 2 MFX (up from 1 in Gen9) along with 1 VQE and 1 SFC. The 2 MFX units can be used for better concurrency such as video playback and video encoding simultaneously. Alternatively, workloads can utilize both engines concurrently to increase performance or reduce clock frequencies to increase battery life. With Gen11, MFX VP9 decode bitdepth support is increased up to10bits which is required for HDR video scenarios. Additionally, both HEVC and VP9 have been improved to support higher quality chroma subsampling of 4:2:2 and 4:4:4 for both HEVC and VP9. 4:2:2 benefits users in high-end video production while 4:4:4 is optimal for screen content such as text\document screen recording and sharing. MFX encoding on Gen11 introduces VP9 support and significantly improves HEVC encoding compression efficiency over Gen9. Gen11 VQE expands hardware denoise bitdepth to 10bits. High dynamic range (HDR) workloads will benefit from new software programmable 3DLUT tables such has HDR2HDR and HDR2SDR tone mapping.
"New" VP9 10 bit decode info is repeated every new generation :)

[NikosD]

peru is software only
What do you think of this conversation ?
https://github.com/GPUOpen-LibrariesAndSDKs/AMF/issues/186

[huhn]

my navi GPU fallbacks to software decoding on some 720p files...
it looks like it doesn't likes old VP9 encodes and 8k simply doesn't work. BTW peru 4k VP9 profile 2 is software only too.
so he want me to contact AMD just like this:
https://community.amd.com/thread/218278
not saying he is wrong but they clearly don't care in that example.
AMF doesn't support VP9 yet. Work in progress.
http://i.imgur.com/AaVqVKv.png

only been 3 year. but why should you add something that newer was suppose to work anyway.

[aufkrawall]

my navi GPU fallbacks to software decoding on some 720p files...

Have you tried in browser? I can't get D3D11VA/DXVA2 for VP9 to work outside of browsers on Navi.
VAAPI on Linux works however.

[NikosD]

My navi GPU fallbacks to software decoding on some 720p files...

It looks like it doesn't likes old VP9 encodes and 8k simply doesn't work.

BTW peru 4k VP9 profile 2 is software only too.

So he want me to contact AMD just like this:
https://community.amd.com/thread/218278 If I had a Navi card I would open a thread like I did back then.
You reminded me the good old days of RX 470 :D

Anyway, I updated my post to AMF github, but remember that we are talking about AMF SDK which is mainly encoding (and decoding of course) API, or better say transcoding.

They could easily not update AMF for HW decoding of VP9 right now, but at the same time DXVA2/D3D11VA could work like a charm or better say should work.

I think it's a drivers problem for VP9 HW decoding but if I were you I would open a thread on drivers forum of AMD with specific VP9 samples (from 720p to 8K)

[huhn]

Have you tried in browser? I can't get D3D11VA/DXVA2 for VP9 to work outside of browsers on Navi.
VAAPI on Linux works however.

works in browser not with lavfilter same for VP9 profile 2.

[NikosD]

works in browser not with lavfilter same for VP9 profile 2. Then I suggest to post to LAV filters thread regarding VP9 HW decoding of Navi because it looks like the AMD guy is right.

But I'm not sure how AMD is using standard DXVA2/D3D11VA decoding and not proprietary and LAV filters fail to decode it in HW.

Only nevcairiel and AMD can answer to that.

[nevcairiel]

Feel free to send me any cards that don't work and I can tell you.

[NikosD]

Not me, I did my best to bring the two worlds closer (SW - LAV filters and HW - Navi cards) by posting to the AMD AMF GitHub.

It's up to you guys...

[aufkrawall]

mpv shows the same behavior as lav filters, so the problem apparently is in the AMD driver.

[NikosD]

AMD probably has changed something regarding VP9 HW decoding and the developers should contact AMD according to the AMD guy of the thread I posted above, IIUC.

[nevcairiel]

Thats why we have standards like DXVA, so that we don't need vendor-specific information for this to work. Thats the entire point of something universal like this over the whole Linux situation where you need to deal with 3 vendor APIs.
LAV doesn't really have any vendor-specific DXVA stuff at all, outside of some compatibility flags for very old GPUs (since they often supported only very limited bitstreams). Any more recent hardware has typically just worked fine out of the box. VP9 DXVA support is implemented based on Microsoft documents, not anything provided by Intel or NVIDIA.

Its also imaginable that an application needs to be whitelisted in the AMD driver to allow this. Maybe try renaming mpc-hc or so to Chrome.exe and see if it works.

In any case, I plan to work on improving logging for LAV, so that users can enable debug logs more easily, maybe that can help to shed some light in the future.

[huhn]

with DXVA2 copyback MPC-HC shows [H/W] indicator at playback start for about a sec only with files that should be able to use hardware decoding.

[aufkrawall]

Thats why we have standards like DXVA, so that we don't need vendor-specific information for this to work. Thats the entire point of something universal like this over the whole Linux situation where you need to deal with 3 vendor APIs.

VAAPI works with free drivers of both Intel & AMD, so it's actually sufficient to support this. And unlike AMD's weird intransparent Windows driver, it just works with any program utilizing it on Navi. ;)

[NikosD]

It's better not to play power games, we all want the things to just work.

I have done the "middle man" between SW devs and HW OEMs many times in order to fix things, but this time I just don't have a Navi card to play that role again.

It's not only AMD.

I remember Intel introducing a new DXVA2 mode of VC-1 HW decoding that LAV filters should adapt in order to work.

Or something regarding 10bit rendering of Intel's drivers, that I don't remember exactly.

Anyway, I don't know the case here, but let's all move one step forward to solve the issue instead of standing to our positions "we do everything perfectly, it's the others fault" even if it's true.

[NikosD]

Intel has essentially cancelled 10nm desktop CPUs with this statement: We continue to make great progress on 10nm, and our current roadmap of 10nm products includes desktop. after the rumors.
I think Charlie of Semiaccurate is HUGE and extremely accurate too.
Respect to Charlie.

[Yups]

Intel has essentially cancelled 10nm desktop CPUs with this statement: after the rumors.
I think Charlie of Semiaccurate is HUGE and extremely accurate too.
Respect to Charlie.


In this statement Intel confirmed there will be desktop CPUs and not the other way around. About Charlie, just 1 year ago he claimed 10nm is completely dead. We already have Icelake and we get multiple 10nm products the next year from the very low end to server and discrete graphics, saying he is extremely accurate is a laughable joke in this context. Charlie is a known Intel hater, also he needs bait articles for his subscribers, it's part of his job, don't trust everything he says.

[NikosD]

10nm for desktop is DOA.

Intel will skip it for 7nm (hopefully)

Charlie is a living God.

[mariner]

VP92 support

Finally succumbed to the temptations of remembrance day offerings and picked up a couple of cheap HDMI2.0 cards for my decade old LGA775 antiques. Here are some screenshots while I dig up the 8k nasa clip for testing.

[Klaus1189]

"Attachments Pending Approval"

please upload to external filehoster

[mariner]

1030 or 550?

The gt1030 and rx550 are the two most affordable HDMI2.0 cards on the market aimed at those on a tight budget running a low-power setup. The 1030 claims 8k support for both hevc10 and vp92, but only 4Khevc10 for the 550 (see previous post). Test clips used are the 24fps/66mbps/hevc NASA samples in 8K and 4K. No 8K vp92 clips are available atm.

The hardware decoding capability of the rx550 clearly lags behind the gt1030 on the basis of these tests. The anemic performance is a surprise, perhaps other users had better experience with this highly rated card.

8K
CPU: Intel(R) Core(TM)2 Quad CPU Q8400 @ 2.66GHz
GPU: NVIDIA GeForce GT 1030
Decoder: HEVCVideoExtension
Decoder Device: HEVC_VLD_Main10
Frames: 4535
FPS: 53.820 [30-56]
CPU Usage: 4 [1-9] %
GPU 3D Engine Usage: 0 [0-8] %
GPU VideoDecode Engine Usage: 98 [78-100] %


4K
CPU: Intel(R) Core(TM)2 Quad CPU Q8400 @ 2.66GHz
GPU: NVIDIA GeForce GT 1030
Decoder: HEVCVideoExtension
Decoder Device: HEVC_VLD_Main10
Frames: 4535
FPS: 164.349 [118-189]
CPU Usage: 6 [3-12] %
GPU 3D Engine Usage: 0 [0-2] %
GPU VideoDecode Engine Usage: 97 [85-98] %

CPU: Intel(R) Core(TM)2 Duo CPU E7600 @ 3.06GHz
GPU: Radeon RX550/550 Series
Decoder: HEVCVideoExtension
Decoder Device: HEVC_VLD_Main10
Frames: 4535
FPS: 72.404 [60-80]
CPU Usage: 35 [17-53] %
GPU 3D Engine Usage: 0 [0-4] %
GPU VideoDecode Engine Usage: 99 [84-100] %

[huhn]

the whole 900 series has HDMI 2.0 ports and the 960 and 950 have HEVC 10 bit decoder. used they should be cheaper the 1030 is bad at price performance ratio.

the polaris cards are well known to be able to reach up 60 FPS for HEVC 10 bit and in your case they didn't even reach that.

here is a vp9 profile 2 8k HDR video:
https://www.youtube.com/watch?v=1La4QzGeaaQ
the new navi cards are supposed to do that but to no ones surprise they can't.

polaris cards where very briefly shown with VP9 support which worked in hybrid using chrome turning the GPU into a heater for just playing a video(just use a CPU and forget the decoder) no one really cares and test this pretty useless feature.
i still love this blunt lie of an image: http://i.imgur.com/AaVqVKv.png

[NikosD]

The polaris cards are well known to be able to reach up 60 FPS for HEVC 10 bit and in your case they didn't even reach that. His numbers are 60/72/80 for 4K HEVC 10 bit, so he is fine as he can obviously hit 4K60 goal.

Here is a vp9 profile 2 8k HDR video:

https://www.youtube.com/watch?v=1La4QzGeaaQ

the new navi cards are supposed to do that but to no ones surprise they can't. Still no browser can do the trick ?
Edge, Chrome, Firefox ?
And what about Win 10 built-in "Movies something" player if you download the clip somehow ?
Although I think there is no way to grab the 8K60 VP92 files from YouTube.

[huhn]

AMD has some special needs anyway:
https://abload.de/img/amdatitsbestwtjpl.png
how could they even mess this up?

even in a browser the 8k clip is software. and even if it would work with one application that doesn't change anything it has to work in any that's what spec are for.

[NikosD]

AMD has some special needs anyway: This screenshot is one of the weirdest ever posted here.
There is a 55% video encoding figure, instead of decoding.
The Ethernet figure is huge ~57 Mbps, but stats of YouTube are really extreme.
I can see there an almost 168 Mbps bandwidth, if I'm not mistaken.
How fast is your internet line and what kind ?
Do you have fiber ?

[huhn]

200 mbit don't really care about more i'm pretty sure i could get 400 if not 1000 mbit for the same money but not risking my ipv4.
it's a coax cable connection and at some point fiber in the "middle of nature".

the point is AMD reports encoding... i simply can't blame microsoft for this it works fine with nvidia and intel.

[Klaus1189]

Where do you live, that there is 1000 Mbit/s available? What happens if you are risking ipv4? Then ipv6? What downsides comes with it?

[huhn]

last post about internet.
1000 mbit:
https://www.unitymedia.de/privatkunden/kombipakete/2play/


you always have an ipv4 address but old internet connection can have a not shared one so you can for example host a game server with ipv4 which is not really possible with a shared ipv4 address.
if an shared ipv4 address get blocked all user that share this ipv4 address are blocked for that service and it is shared with easily 1000+ user if not 10000+.
it will be utterly irrelevant in the future and ipv6 will work for everything.

[NikosD]

My two cents for my Internet and please don't make this thread an Internet connection speed survey.

I have a VDSL2- FTTC (Fiber To The Cabinet) 50Mbps connection since late 2015 and two months ago 100Mbps and 200Mbps became available to me, as the cabinet is about 10 meters away from the building, but they are both really expensive.

FTTH (Fiber To The Home) at 1000 Mbps is very expensive and very rare here, so I have to wait a little more for 200 Mbps with a reasonable price and probably a lot more for 1 Gbps.

[littleD]

Amd drivers are usually the mess. They can enable some features, because they announced so, but later disable, because not working as intended or not fully compliant, or they just give up.

Can anybody share ultimate 8K60fps HDR video Clip?

1Gbit is not premium here anymore. Its available for home already. We have tests 2 and 5 Gbit.

[DMU]

VP9.2 (https://drive.google.com/open?id=12-KIpCd5unep03Y0FoIgLyaABN3v1CAZ)

[huhn]

that's 4k and you have a ton of GPU load that shouldn't be there.

[NikosD]

Intel Gen12/Xe GPU in Tiger Lake supports HEVC 12-bit, 4:2:2 & 4:4:4 fixed function hardware decoding. ...and VP9 12bit 4:2:2 & 4:4:4.
Interesting, it reached Turing's decoding features.
We have to wait for decoding speed and I'm interested in HW encoding quality too of Intel Gen12.
But of course Ampere is going to be announced in less than a month with probably a new decoder/ encoder.
Could Ampere's decoder support AV1 decoding in pure fixed-function HW ?
Will see..

[NikosD]

Huge upgrade for Intel architecture, implementing the 10nm Tiger Lake to 14nm, as Rocket Lake.

But we haven't seen Comet Lake yet, so what is happening inside Intel nowadays ?

According to Wiki, Tiger Lake supports:
HEVC 12-bit, 4:2:2 & 4:4:4 fixed-function hardware decoding &
VP9 12-bit & 4:4:4 fixed-function hardware decoding.

Now, that Videocardz leak, adds AV1 too.

And all of these in late 2020 without even Comet Lake out yet?

Don't think so...

[sol-compiler]

Having problems with 50/60fps 4k HEVC decoding using Intel Graphics HD 630 in LAV video decoder using D3D11 copyback in conjunction with an Nvidia GPU for MadVR. 4k30 works fine. Is there an inherent bandwidth limit to the Intel+Nvidia combination or is it a software issue that might be fixed? Thanks

[littleD]

Do not use dx11 copyback. It is slow on intel. Use instead dx11 "native" or rather Automatic option. If You would still have problem, then reply with more information.

[aufkrawall]

He can only use copyback with the IGP as it seems the display is connected to the dGPU.
Copyback puts additional load on the GPU, it's simply too much for the IGP with that high resolution and frame rate.

[NikosD]

Having problems with 50/60fps 4k HEVC decoding using Intel Graphics HD 630 in LAV video decoder using D3D11 copyback in conjunction with an Nvidia GPU for MadVR. What is your nVidia GPU ?
If it supports 4K HEVC decoding in HW, I would try to do everything on dGPU, leaving Intel's iGPU outside.
Is it a laptop with Optimus ?
If not, try to disable iGPU from BIOS to save system RAM and try again with nVidia only.

[littleD]

Having problems with 50/60fps 4k HEVC decoding using Intel Graphics HD 630 in LAV video decoder using D3D11 copyback in conjunction with an Nvidia GPU for MadVR. 4k30 works fine. Is there an inherent bandwidth limit to the Intel+Nvidia combination or is it a software issue that might be fixed? Thanks
He can only use copyback with the IGP as it seems the display is connected to the dGPU.
Copyback puts additional load on the GPU, it's simply too much for the IGP with that high resolution and frame rate.Right. I misinterpeted the quiestion. I have the same. To make nvidia work with madvr at conunction with intel dxva decoding i need to use dx11 copyback too. The bad thing is that copyback itself is taxing memory bus. I tested yesterday, cannot pass 4k30fps with fluid playback.

What is your nVidia GPU ?
If it supports 4K HEVC decoding in HW, I would try to do everything on dGPU, leaving Intel's iGPU outside.
Is it a laptop with Optimus ?
If not, try to disable iGPU from BIOS to save system RAM and try again with nVidia only.Some optimus systems like mine for egzample do not have available hardware decoding on dedicated nvidi gpu unfortunately.

[NikosD]

This is the end of the first era of "Ryzen Effect".

On January 2017 Intel said to the world that the best desktop processor was Core i7 - 7700K a 4C/8T CPU at 350$
On March 2017 AMD with Zen architecture and Ryzen implementation destroyed 6 years of Core iX pattern with its multi-core, multi-threaded approach.

A lot of people thought that Intel could react with a "secret weapon" as we were all thinking that a lot of processing power was left on the table from Intel during all those years 2011 - 2017.

Comet Lake-S even using 14nm (++...+) is essentially a Skylake architecture of 2015 and it's the end of this particular road for Intel.

A 10C/20T Intel CPU of ~500$ in 2020 versus a 10C/20T Intel CPU of 1700$ in 2017 is the maximum gain for the user after AMD's Ryzen arrival.

Now the gap closes in all desktop sections Core i3 vs Ryzen 3, Core i5 vs Ryzen 5, Core i7/i9 vs Ryzen 7/9 besides the top 16C/32T

But from now on Intel has no other moves, regarding 14nm and Skylake architecture.
Comet Lake-S is the "all in", last card of Intel.

Next step is either delivering a real new 10nm architecture for the desktop or having the fate of AMD's rough years 2010 - 2017.

We' ll see...

[Lumiel]

Hey guys, I almost forgot what I wanted to say after 5 days of waiting :( Basically I was just wondering what is the status of VP9 Hardware Decoding for AMD 5700 XT (Navi)? I did some research last week and I noticed that it's not working on most software or it's slow. It looks like that only Microsoft has a fast implementation (Edge and Movies & TV) and it is either not supported or dropping a lot of frames on other browsers and media players, both on Windows and Linux (MPV - MPC-HC - VLC). I also noticed that Microsoft was recommending to use Media Foundation (via IMFTransform) for implementing VP9 HW decoding but I did search in LAV filters Github and it seems there are no plans for supporting it?
Is there any way I could help with testing and debugging? I've also downloaded DXVA checker but I see the checkbox for Hardware VP9 is not selected for some reason.

In theory the HW decoder should support just 8K@24 but Microsoft implementation is pushing it near 60FPS. Sorry if it's the wrong topic but I already saw some posts about it here and I was curious what's keeping most players behind.

[NikosD]

Basically I was just wondering what is the status of VP9 Hardware Decoding for AMD 5700 XT (Navi)? I don't have a Navi card but I tried to help asking AMD here:
https://github.com/GPUOpen-LibrariesAndSDKs/AMF/issues/186
Read the conversation there and take a look at that GitHub forum to see if anything new has appeared.

Maybe @huhn knows more.