Evaluation of HEVC decoders (SW, Hybrid and HW)

[NikosD]

Nvidia just announced the new Nvidia Turing GPU microarchitecture at Siggraph, wonder what Feature Set hardware decoder does it have, I like Volta?

do you really care about a 10k cards hardware decoder?

...so i don't except anything new from nvidia this round.

The reason I asked about Titan V is because I had 1150 (Ti) or 2050 (Ti) in mind - I don't know what name will be Nvidia's final decision.

I thought that Titan V most likely has the same video decoder with Volta/Turing/Ampere - whatever Nvidia architecture decides to put in gaming platform.

The gaming cards are very close to release, so I guess we are just going to wait and see.

Video decoding is not very efficient for CUDA cores. NVIDIA and other GPUs have dedicated encoding and decoding modules.

You seem that you weren't been following this thread for a long time.

In the context of this thread, whenever someone is talking about a GPU regarding video performance or features, he is actually referring to VPU - that dedicated ASIC that you mentioned too.

But thanks for the post.

[huhn]

volta is quite old and nvidia has different decoder for the 1080 and the 1050 so more than enough time to add something. so no i don't think it will be the same.

as you already said the decoder from volta is different at least in name from the pascal decoder so the decoder really has to be different for these upcoming cards. if it will be something meaningful i doubt it. maybe is just named different because of the shrink.

my command about 10k cards is about the upcoming quadros. volta/titan V is at least some kind of consumer grade card at least to some people out there.

[NikosD]

From Anandtech:
Finally, while NVIDIA only briefly touched upon the subject, we do know that their video encoder block, NVENC, has been updated for Turing.
The latest iteration of NVENC specifically adds support for 8K HEVC encoding.
Meanwhile NVIDIA has also been able to further tune the quality of their encoder, allowing them to achieve similar quality as before with a 25% lower video bitrate

[alex1399]

No signs for a new hardware decoder block

[sneaker_ger]

The latest pascal decoders leave little to be desired, anyways (except AV1). Encoder improvements would be more important.

[alex1399]

Maybe they could add some hardware enhancer (https://forum.videohelp.com/threads/389555) and make the HEVC decode stuffs looks better.

[huhn]

you can destroy your image quality by using the sharpener in the nvidia control panel with a video renderer that supports it by your choice...

[Cary Knoop]

Maybe they could add some hardware enhancer (https://forum.videohelp.com/threads/389555) and make the HEVC decode stuffs looks better.
Did you actually read the full topic?

Sharper is not better!

[Forteen88]

Maybe they could add some hardware enhancer (https://forum.videohelp.com/threads/389555) and make the HEVC decode stuffs looks better."Sharpening increases high frequency detail (well, in normal circumstances - you could also sharpen/emphasise low frequenzy detail), which then requires more bitrate for the same quantiser (whereas increasing low frequency detail doesn't increase the needed bitrate significantly)."
So it's better to sharpen during playback than in the encode, if you want to sharpen.

[NikosD]

Guys, if someone sells his kidney in the black market in order to buy a new RTX card, please run DXVA Checker in order to check features and performance and share this info with us.

Also, I'm interested in the release month of 2050 (Ti) and if someone knows if 2050/2060 will have Tensor and Ray Tracing cores inside like 2070 and above.

[Zetti]

Are you sure 2050/2060 is planned??

[wanezhiling]

https://www.phoronix.com/scan.php?page=news_item&px=More-Vega-20-Enablement-Linux
- Support for the 2nd UVD instance on Vega 20 (previously mentioned on Phoronix, there will be two UVDs with Vega 20).

[NikosD]

Crazy.

Probably a 2xGPU implementation.

[foobar2000]

dxva with hdr tone mapping at the same time, will this cause frame drop?

[huhn]

i don't see a reason why DXVA should do that.

[NikosD]

Don't worry, AMD is not sentenced to death yet, I will buy a Ryzen 5 2400G soon.It's been just seven months (!) but did you buy that Ryzen 5 2400G ?

I was looking at your results of your Ryzen 5 2500U with slow single channel DDR4-2133 MHz RAM and they were disappointing.

I wonder if desktop version of VCN 1.0 inside an APU (Raven Ridge) has changed anything regarding features and especially performance compared to laptops.

[wanezhiling]

No i didn't.
Others tested 2400G with dual channel DDR4-2400, no luck.

[huhn]

could be the same type of 4.4:4 as AVC lossless only.

[NikosD]

nVidia, by adding VP9 12bit, H.264 8K and HEVC 444, even if it's only for lossless, has the most fully featured hardware decoder ever.

But it's still missing AV1 which YouTube has already started leveraging in order to convert everything.

I think AV1 is going to replace VP9 for YouTube.

[huhn]

while AV1 will accomplish VP9 one day well at least it should do that.
the AV1 encoder has to get optimise i mean just remember the release of HEVC the image quality was not that great to say it friendly so this will take a couple of months-years.

VP9 will not leave us anytime soon we still don't have full support on all mainstream GPUs yet.

and we simply don't know about AV1 support for turing or let me say it this way i can't remember nvidia saying the 960 will support VP9. but i doubt it it's just to early.

[mariner]

Greetings.

Quick impression after short test of an entry level low power HD620 system running w7: handles 4k well, but struggles with 8k. The 8KHEVC10 (4in1) clip wouldn't play in DXVA Checker. As this is my first x64 system, appreciate some advice in setting the system properly for video playback evaluation:

1. Is it still necessary to run DirectX End-User Runtimes (June 2010)?
2. Where can I find x64 version of proppage.dll and evrprop.dll? And how to get them registered?
3. How to set up LAV video filter to take advantage of the HD620?
4. Any other free decoders available for testing?
5. Should video size be set to native or size of display monitor when running playback performance in DXVA Checker?
6. How to set up HDR on w7?
7. Other tips?

Many thanks and best regards,

[clsid]

3. Select DXVA2 native
6. That is not supported on W7 with Intel GPU

[NikosD]

1. Is it still necessary to run DirectX End-User Runtimes (June 2010)?
Good question.
I always do this although I don't know if it's actually necessary.
2. Where can I find x64 version of proppage.dll and evrprop.dll? And how to get them registered?
Never used them. Where are they useful to?
3. How to set up LAV video filter to take advantage of the HD620?
Like clsid said, use DXVA2 native for best performance.
4. Any other free decoders available for testing?
If you want DirectShow decoders like LAV filters, I think MPC-BE developers are still developing stand-alone DS filters you could try.
If you want complete players, I would suggest PotPlayer and maybe VLC.
5. Should video size be set to native or size of display monitor when running playback performance in DXVA Checker?
In my performance tests I always use a fixed resolution for playback performance regardless of the original resolution of the source.
This resolution is 1280x720.
6. How to set up HDR on w7?
I think clsid is right, Win 7 doesn't support HDR.
7. Other tips?
Well, I can't see the screenshot posted above, but can you take a look of the results that others users have already tested in my collection of test video clips regarding 4K and 8K and tell me what is your difference ?

I have all of them gathered together for easy comparison, just find the link in the initial post of this thread.

Can you give me full specs of your system ? (RAM, CPU, PC/LAPTOP model etc.)

[mariner]

Greetings clsid and Nikos, thanks for the kind reply.

1. If my memory is correct proppage allows you to see the pin properties in graphedit, while evrprop displays various info of the evr renderer in use. The copies I have are 32bit, what's the procedure for registering them? Are there 64bit version available?

2. LAV filters:
(a) Is VP9 supported? As the default has it un ticked.
(b) Isn't quicksync the native HD620 DXVA decoder? When selected, tt seems LAV falls back to CPU decoding. Why?
(c) Does one lose hardware deinterlacing capability when DXVA2 Native is selected?

3. 8k video test
(a) 8k HEVC10 is problematic. Crashed DXVA Checker , and wouldn't play in mpc if dxva enabled.
(b) 8k VP9profile2 not tested. Test samples welcome.
(c) 8k HEVC8 and VP9 seems well behaved in DXVA checker, but not mpc.

4. Test system is i7 7500u/HD620/16GB DDR3 RAM/w7x64. As the graphics drivers is a hacked version of 4680, what sort of artifacts/issues could one come across?

Hope these cut and paste jobs are intelligible. Suggestion and comments most welcome.

[clsid]

Yes, VP9 is supported.
QuickSync uses the GPU the same way as DXVA. QuickSync option in LAV is deprecated and does not support HEVC and VP9. Don't use it!
I doubt you have a 8k screen, so just forget about 8k.

[NikosD]

2. LAV filters:
(a) Is VP9 supported? As the default has it un ticked.
(b) Isn't quicksync the native HD620 DXVA decoder? When selected, tt seems LAV falls back to CPU decoding. Why?
(c) Does one lose hardware deinterlacing capability when DXVA2 Native is selected?

VP9 is supported by both hardware and software but not for QuickSync.

The developer of QS is egur who has left doom9 years ago and nobody ever developed further his work.
So QS supports only MPEG2, VC-1/WMV3, H.264.

Regarding hardware deinterlacing, I think the opposite is true.
When you leverage DXVA native, hardware deinterlacing is automatically used.

3. 8k video test
(a) 8k HEVC10 is problematic. Crashed DXVA Checker , and wouldn't play in mpc if dxva enabled.
(b) 8k VP9profile2 not tested. Test samples welcome.
(c) 8k HEVC8 and VP9 seems well behaved in DXVA checker, but not mpc.

I don't have HD 6xx iGPU but maybe Win 7 and hacked driver could cause all these problems.
I was trying to find out VP9 10bit samples with no luck.
Try YouTube, if you manage to get the stream of VP9 profile 2.

Did you try any other player ? (MPC-BE, PotPlayer, VLC)

I would definitely upgrade my system to Windows 10 with proper drivers, maybe Win 7 is too old for these formats.

Hope these cut and paste jobs are intelligible. Suggestion and comments most welcome.

None of your screenshots are visible.

Doom9, like all forums actually, is very slow approving attachments like these.

We always use external image servers like my favorite postimages.org

[mariner]

8k mpc playback.

While DXVA checker suggested these 8k clips are well behaved, mpc tells a different story. What's going on?

[NikosD]

By releasing the new Turing Tesla T4, we learned today that fixed function decoder and encoder of Turing cards - even gaming cards obviously - besides having many new features could be very fast too.

Tesla T4 delivers breakthrough performance for AI video applications, with dedicated hardware transcoding engines that bring twice the decoding performance of prior-generation GPUs.

T4 can decode up to 38 full-HD video streams, making it easy to integrate scalable deep learning into video pipelines to deliver innovative, smart video services.

[Asmodian]

It seems like that hardware decoder is for AI to be able to use almost any video for training data?!? That is pretty impressive, I hope (and assume) consumer cards have the same decoder. :)

[nevcairiel]

It would probably not be too hard for them to just double up on decoder hardware specifically for such a specialized card. But another week and we'll possibly know either way.

[Klaus1189]

8k mpc playback.

While DXVA checker suggested these 8k clips are well behaved, mpc tells a different story. What's going on?


None of your screenshots are visible.

Doom9, like all forums actually, is very slow approving attachments like these.


Like NikosD already pointed out, we cannot open your files.
Do it like that:
I created a screenshot (https://drive.google.com/file/d/1B-R1h4cfhP8HFpmqqc-wjq8UzggMKCyy/view?usp=sharing), here you can see I can not open your files

[littleD]

8k mpc playback.

While DXVA checker suggested these 8k clips are well behaved, mpc tells a different story. What's going on?Win10 hd620 latest drivers

latest mpc hc dxva2 native in lav
4k 10bit hdr 60fps plays fluently
8k 8bit 24 fps plays fluently but its on the edge
8k 8bit 60 fps is slow (not so much, but visible desync with audio)

Dxva cb or d3d11 is slow with intel and lav filters so do not use those options

but using win10 Movies and Tv app i get
8k 8bit 60fp fluently wit rare/occasional glitch. This app uses d3d11 mft filter



I cannot download HDR files from Youtube, only 8k 8bit. Is this hdr real at all? Does chrome cheat or changing metadata or what.

[nevcairiel]

A fully optimized D3D11 video playback chain is the most efficient you can get on a PC. Unfortunately DirectShow is too old to really support that out of the box. You might be able to achieve some improvements using D3D11 native in LAV Video with madVR, but madVR internally still uses D3D9 a lot, so that loses a bit of the efficiency again there, because the image needs to be transfered from D3D11 to D3D9 again.

[NikosD]

I cannot download HDR files from Youtube, only 8k 8bit. Is this hdr real at all? Does chrome cheat or changing metadata or what.

It seems that nobody can.

But they are real for sure at least for 4K, because my 4K HDR10 AndroidTV plays them fine as HDR clips using YouTube with real and obvious difference from 8bit non HDR clips.

@nevcairiel
Good post.

[wanezhiling]

https://communities.intel.com/thread/128950
I hate intel..:devil:

[nevcairiel]

You can definitely download 4K HDR streams with youtube-dl. Just 8K seems to be either hiding or in fact only be available in 8-bit SDR.

[littleD]

Thanks nevcairiel for this clarifications. Since almost all praise d3d11 then it must be intel drivers or hardware shortcoming. I will try madvr with some fast downscaling algorithm and see results ;) Also Madvr gives proper colors with hdr on laptop screen. Evr custom not :/

my last test with mpc
8k 10bit hdr 24 fps file is fluid in playback on intel hd620
Sample from vimeo


EDIT:

How about to make mpc hc shader: rec. 2020 >> 709 (HD) or whatever renderer could expect???
Just to correctly display colours in EVR Custom?

[nevcairiel]

D3D11 is definitely slower right now if you use it without madVR, because that results in D3D11 copy-back, which is for some reason a tad bit slower then DXVA2 copy-back even. There are some technical reasons for it being slower then DXVA2 copy-back, but I didn't expect them to have such a huge impact - but I guess an Intel iGPU which uses system memory feels it the most.

[mariner]

A fully optimized D3D11 video playback chain is the most efficient you can get on a PC. Unfortunately DirectShow is too old to really support that out of the box. You might be able to achieve some improvements using D3D11 native in LAV Video with madVR, but madVR internally still uses D3D9 a lot, so that loses a bit of the efficiency again there, because the image needs to be transfered from D3D11 to D3D9 again.

Greetings nevcairel.

Been testing hevd clips with HD620 in w10:

1. When D3D11 option is selected in LAV filters, mpc reports dxva not in use, and cpu usage hits 100%.

2. The D3D11 option is not available in mpc's internal LAV filters.

Are these expected behavior?

Many thanks and best regards,

[huhn]

try it again with this version: https://github.com/clsid2/mpc-hc/releases

[alex1399]

The win10 Movies and Tv app isn't using D3D11 mft filter but some sort of mvd filter stuff called msmpeg2vdec.dll as shown in the filter chain by GraphStudioNext. The image quality is terrible. That's why it is so fast.

No, the crappy rendering stuff is out of the question to affect the decode stuff. MS made something too cheap to make it fast.

[huhn]

a decoder for a modern codec should not show any difference in image quality what give you the idea the decoder is the reason for the lowered quality?

[littleD]

The win10 Movies and Tv app isn't using D3D11 mft filter but some sort of mvd filter stuff called msmpeg2vdec.dll as shown in the filter chain by GraphStudioNext. The image quality is terrible. That's why it is so fast.

No, the crappy rendering stuff is out of the question to affect the decode stuff. MS made something too cheap to make it fast.C'mon GSN cannot use media transform, can it?

Edit:
Dxvachecker shows some hints. In d3d11 decoding benchmark, both lavfilters and windows decoder are equally fast. But in playback benchmark, lav filters use 20% video engine while mt decoder over 80%. Such situation gives lower framerate in mphc than in windows app. Im sure lav outputs best quality possible. But comparing things and drawing conclusions aren't bad right? According to new hardware monitor in newest win10, the shared graphic memory is saturated when d3d11 enabled and 8k video played. Shared Graphic ram is full filled with some data. Maybe transfer between small GPU Ram and shared System memory is very costly. Maybe nevcairiel can lower memory requirements or use some texture compression to speed up things at 8k? I dont know where it comes from but win10 shows shared 4GB gpu memory on intel. During mft 8k decoding its half filled.

[foxyshadis]

Like NikosD already pointed out, we cannot open your files.
Do it like that:
I created a screenshot (https://drive.google.com/file/d/1B-R1h4cfhP8HFpmqqc-wjq8UzggMKCyy/view?usp=sharing), here you can see I can not open your files

I really wish someone would report these attachment posts, even if it's not the OP. It's not the workflow I would have chosen, but it's the only one I have until I have time to browse over the weekends.

[pirlouy]

Maybe this subforum needs another moderator, someone well-recognised and quite active like nevcairiel or clsid for example ?

[nevcairiel]

No thanks :p

[NikosD]

Actually, this thread doesn't need a moderator.

It's self moderated by the members writing here and the OP does a little moderation, like in all threads.

Please, don't try to make things look like complicated.

[pirlouy]

Like if your thread was the subforum. :rolleyes:
It was a general statement about screenshot approval.

[mariner]

8K playback on w10: MTV vs DXVA Checker vs MPC

Test system: Stock i7 7500u/HD620/16GB DDR3 1600 RAM/2k display
OS: w10.1803.285, pagefile disabled.

1. While HD620 can decode 8K HEVC10bit in hardware, MTV plays the 59fps 4in1 clip in slow motion, while MPC and Potplayer cannot display the video. What does the creator of the clip use for playback?

2. The 8K HEVC8bit LondonEye test clip is too short to allow any meaningful conclusion drawn. Longer samples most welcome.

3. 8K VP9 8bits: Test sample used Yellowstone is a randomly selected 8k 59fps yt clip (those numbered 272).
(a) MTV: Plays smoothly, but there could be a hint of video lag at the end. Clip too short to be conclusive. Is there a way to check frame rate in MTV?

(b) DXVA Checker:
(i) Playback performance benchmark comes in lower that playback benchmark. It is also 1/3 slower than w7's benchmark.
(ii) While playback is smooth for both LAV filters and MF decoder, video soon lags behind audio. LAV actually outperforms MF here.
(iii) Does this mean a different decoder unidentified by DXVA Checker is used by MTV for playback?

(c) MPC, POT and VLC; Can't handle any 8K clip. VLC suffers from decoding artifacts.

4. Suggestions to tweak the system for better performance most welcome.

[Klaus1189]

I really wish someone would report these attachment posts, even if it's not the OP. It's not the workflow I would have chosen, but it's the only one I have until I have time to browse over the weekends.

I don't understand, did I something wrong? :scared:
But I can't watch the attachments, and
again from @mariner, still pending ...

@mariner
we can't open your uploaded files, use any other file hoster, like google drive or whatever ...

Like NikosD already pointed out, we cannot open your files.
Do it like that:
I created a screenshot (https://drive.google.com/file/d/1B-R1h4cfhP8HFpmqqc-wjq8UzggMKCyy/view?usp=sharing), here you can see I can not open your files

click on red "screenshot"