Evaluation of HEVC decoders (SW, Hybrid and HW)

[clsid]

2GB is more than enough. A decoded 4K frame is ~32MB. The person that told you it isn't enough probably uses a too high GPU buffer in madVR.

[huhn]

even with an GPU queue of 4 you get somethining like 1700+ mb Vram usages.

for save deinterlacing a GPU queue of 8 is kind of needed which is not possible with 2 GB Vram.

i'm pretty sure i can play the Samsung_UHD_Ride_on_Board.ts with my 960 on my UHD screen i will try it later.

[huhn]

playback with EVR (not EVR-CP) is no issue at all.

madVR playback is doable with bilinear chroma scaling.

Vram usages
EVR ~1100
madVR ~1700 with a GPU queue of 4 and present frames in advance at 3.

[clsid]

crowd_run_2160p50_Main10-127Mbps.mkv

Played on a 1920x1200 screen. But should not make huge difference since most buffered data will be in the full 4k size.

EVR-CP with standard buffer size of 5:
OS (no playback): 180 MB
No HWA: 438 MB
DXVA2 native: 927 MB
DXVA2 cb: 1296 MB

MadVR (GPU queue 4, present 4):
DXVA2 cb: 1072 MB

MadVR (GPU queue 8, present 4):
DXVA2 cb: 1270 MB

MadVR (GPU queue 8, present 8):
DXVA2 cb: 1299 MB

[huhn]

native: http://abload.de/img/vramusagef2bqr.png

copyback: http://abload.de/img/vramusage2kqjp6.png


the OS eats ~650 MB Vram with an open browser and i don't have unusual programs running in the back ground.

you get the Vram usage without browser later.

most likely the 200% scaling is eating a lot.

~350 mb after reboot with and without 200% scaling.

after closing madVR Vram usage was normal.
than i was running madVR with an 3D LUT and after closing i got 521 mb Vram usages.

[Rza11]

I am getting micro stuttering while playing x265 @ 24p.

PC: I7 920 @ 4ghz with GTX 670
TV: 1080p @ 24p
Source file: 1080p

Shouldn't my PC be enough to play these files correctly? As you guys can see i'm using latest lav with madvr and the cpu doesn't seem to be taxed much.

Does Strongene have the same quality as lav for 10 bit ?

http://i.imgur.com/1VDyXmC.jpg

http://i.imgur.com/gTMaN25.jpg

[nevcairiel]

From the madVR screenshot, its not a decoding problem, madVR receives enough frames, but the GPU seems to not handle them fast enough.

[Rza11]

From the madVR screenshot, its not a decoding problem, madVR receives enough frames, but the GPU seems to not handle them fast enough.

Really ok, I will try update drivers (currently 361.91) and all software to absolute latest and report back.

[Rza11]

I installed latest nvidia drivers and checked all codecs and kodi are latest.

I'm not getting any presentation errors but i still get micro stutters.

Anyone know what else should i look into?

Updated Stats:

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

http://i.imgur.com/7mRdqrc.jpgc

[chros]

I installed latest nvidia drivers and checked all codecs and kodi are latest.

I'm not getting any presentation errors but i still get micro stutters.
I think the problem is that the refresh rate of your screen (23.971) is lower than the content's (23.976).
Solution: use 24p in madvr config or in OS not 23p, OR try to tweak the frequency in 23p to get as close to 23.976 from above.

[nevcairiel]

Your GPU just can't keep the GPU queues filled, all GPU queues are nearly empty and you have tons of dropped frames. There are no obvious answers. Reset madVR settings to defaults, try again. See if some particular setting causes it to break again, etc.
D3D11 10-bit output is problematic on some systems, so try to avoid that. In general, greatly simplifiy the setup until it works, then try adding more options.

[Rza11]

I have disabled exclusive mode which disabled 10 bit and everything is fine now, presentation is filling and no dropped frames. I think this will have to do until i find out the real problem.

in general is a 670 gtx enough for 10 bit or is it borderline? maybe i should upgrade gpus

[krille]

I installed latest nvidia drivers and checked all codecs and kodi are latest.

I'm not getting any presentation errors but i still get micro stutters.

Anyone know what else should i look into?

Updated Stats:

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

http://i.imgur.com/7mRdqrc.jpgc

This happens to me too in exclusive mode with my Nvidia 960, the queues doens't fill up. Disabling exclusive mode solves the problem for me.

[chros]

Your GPU just can't keep the GPU queues filled, all GPU queues are nearly empty and you have tons of dropped frames. There are no obvious answers. Reset madVR settings to defaults, try again. See if some particular setting causes it to break again, etc.
D3D11 10-bit output is problematic on some systems, so try to avoid that. In general, greatly simplifiy the setup until it works, then try adding more options.
Oh, I missed that part completely :)
I have disabled exclusive mode which disabled 10 bit and everything is fine now, presentation is filling and no dropped frames. I think this will have to do until i find out the real problem.
in general is a 670 gtx enough for 10 bit or is it borderline? maybe i should upgrade gpus

This happens to me too in exclusive mode with my Nvidia 960, the queues doens't fill up. Disabling exclusive mode solves the problem for me.
That's interesting. With my GT650m (optimus system! , using DXVA-copyback in LAVvideo) I CAN use D3D10 10bit FSE output up to 1080p@30p material (4K is a different beast, and 50p/60p content (like sports caps) are more demanding). So that means GPU isn't the problem, it has 2GB GDDR5 VRAM.
If you use dxva-copyback hardware acceleration, then try out software decoding, maybe it helps.

[Rza11]

Interesting,

With my 670 gtx most stuff >90% plays correctly @ 24p 10 bit but i have a few action vids that stutter.

Maybe it has something to do with the way there encoded.

I tested all the lav accelerator settings but there all the same.

I may try overclock the evga 670 and see if it makes a difference.

very interesting it does that too with the 960 i thought the hardware decoding would of taken care of all the x265 issues

[chros]

very interesting it does that too with the 960 i thought the hardware decoding would of taken care of all the x265 issues
You can upload a sample file of the problematic part somewhere so we can check it.

[Rza11]

I might do, I'll look into making a sample

[DragonQ]

Which cards support full hardware HEVC decoding nowadays? Is it just the GTX 750 Ti, 950, and 960? Thinking of upgrading my GT 430 since it's useless for anything in MadVR beyond Bicubic and Lanczos, and I may as well ensure I can play UHD video encoded as HEVC whilst I'm at it.

Also, I notice some people in this thread using Kodi DSPlayer with MadVR. I am testing that on my setup but have an issue with TV - when switching from an SD to an HD channel, the screen goes entirely black. I can still Alt-F4 or even change channel using keyboard shortcuts and if I change back to an SD channel, the video and UI reappear. Anyone else see this issue?

[NikosD]

10bit HEVC decoding in fixed function hardware (full) is supported only by Nvidia cards - 950 and 960 (+2nd edition of 750 ti).

But 8 bit HEVC full hardware decoding is supported by Skylake integrated iGPU and latest iGPUs and discrete cards from AMD.

Personally, I'm waiting for the new cards of Nvidia and AMD which are just around the corner.

They will support 8&10 bit HEVC and probably VP9 full fixed function HW decoding.

[clsid]

Yes, those are the only ones with 10bit. For 8-bit there is also AMD Fury/Nano and Intel Skylake.

[P.J]

... and another one collegue with Braswell NUC (Pentium N3700 - HD Graphics). Seems to be better, like GTX960.

http://i60.tinypic.com/nej7yq.jpg

No benchmark for Braswell/Cherry Trail? :rolleyes:

[wanezhiling]

http://www.anandtech.com/show/10256/intel-unveils-apollo-lake-14nm-goldmont
The new graphics core is listed as being more powerful (most likely due to both better architecture and a higher count of execution units), but will also integrate more codecs, enabling hardware-accelerated playback of 4K video from hardware decoding of HEVC and VP9 codecs.

[wanezhiling]

It's Gen9 GPU though, same as Skylake so no HEVC Main10 hardware decoding, only hybrid/partial VP9 hardware decoding, no HDMI 2.0/2.0a support, no HDCP 2.2 support.
This Gen9 is probably a little different from skylake's.
Because braswell/Cherry Trail's Gen8 has totally same decoding ability with skylake's Gen9 (8-bit HEVC).;)

[P.J]

Skylake's iGPU can decode 10-bit HEVC in hybrid mode but Braswell/Cherry Trail =/

[nevcairiel]

It's Gen9 GPU though, same as Skylake so no HEVC Main10 hardware decoding, only hybrid/partial VP9 hardware decoding, no HDMI 2.0/2.0a support, no HDCP 2.2 support.

The media support of the Atom platform typically is a generation ahead of the desktop platform. Braswell has a Gen8 GPU but still VP9-Hybrid support, which on the desktop is a Skylake feature.
Its important not to confuse the actual GPU with the media capabilities, which are not directly tied together.

[wanezhiling]

Yes, as nevcairiel said, I'm sure that apollo lake already supports hevc 10-bit & vp9 8-bit(maybe also 10-bit?) hardware decoding, not hybrid any more.
Of course the next desktop gen10 gpu(Kabylake) will has same supports like apollo lake.

[P.J]

http://www.tweaktown.com/news/51746/amds-official-2016-2018-gpu-roadmap-released-vega-hbm2-2017/index.html

http://imagescdn.tweaktown.com/news/5/1/51746_01_amds-official-2016-2018-gpu-roadmap-released-vega-hbm2-2017.png

[Yups]

https://lists.freedesktop.org/archives/intel-gfx/2015-October/077255.html



Kaby Lake is Gen9.5. Cannonlake should be Gen10.


Kabylake is also often referred to Gen9 from Intel. There are some hints that Broxton supports HEVC 10 decoding in hardware.

https://lists.freedesktop.org/archives/libva/2015-December/003756.html

Broxton being Gen9 doesn't mean it can't have HEVC 10 bit decoding in hardware. Intel wrote that Broxton Gen9 differs slightly to Gen9 from Skylake. I can't find the link at the moment but found a similar one.

http://permalink.gmane.org/gmane.comp.freedesktop.xorg.drivers.intel/55947/gmane.comp.freedesktop.xorg.drivers.intel/55947

Very similar means it's not identical.


Here is a good overview

https://build.opensuse.org/package/view_file/X11:XOrg/vaapi-intel-driver/vaapi-intel-driver.changes?expand=1

+ Add support for Skylake.
- Decoding: H.264/MPEG-2/VC-1/JPEG/VP8/HEVC

+ Add support for Broxton:
- Decoding: H.264/MPEG-2/VC-1/JPEG/VP8/HEVC/HEVC 10-bit/VP9.

* Add support for Kabylake
- Decoding: H.264/MPEG-2/VC-1/JPEG/VP8/HEVC/HEVC 10-bit/VP9/VP9 10-bit

[leonccyiu]

I can't wait to find out what the hardware decode capabilities of the new Nvidia Pascal cards are. Since they have displayport 1.4 I wonder if they'll do Hevc 10bit at 8k. I assume that they'll have VP9, in 10bit as well, although I doubt google chrome will able to take advantage of that.

I believe the reviews are coming on May 17th.

Hopefully we'll hear something about Kabylake at computex.

[Yups]

http://fs5.directupload.net/images/160515/o9qx4kyv.png
http://videocardz.com/59962/nvidia-geforce-gtx-1080-final-specifications-and-launch-presentation


Even 12 bit decoding.

[nevcairiel]

Even 12 bit decoding.

I doubt that until I see it. The DXVA2 HEVC spec includes nothing for it. Leaks will be leaks.

[aufkrawall]

Maybe Microsoft will again silently edit some knowledge articles? :D

[nevcairiel]

Maybe Microsoft will again silently edit some knowledge articles? :D

Perhaps, but 12-bit is not part of the Main/Main10 profile, so adding that would involve quite some additional changes, and not just add a new profile specific for this mode.
So far they did not, anyway.

But I'll be getting a 1080 myself, so I can see whats what.

[Yups]

I doubt that until I see it. The DXVA2 HEVC spec includes nothing for it. Leaks will be leaks.


Who is talking about DXVA2 HEVC specs? It doesn't say anything about the hardware capability of Pascal. If DXVA2 doesn't support HEVC 2.0 Main 12 I'm sure it will be supported sooner or later.

[nevcairiel]

Who is talking about DXVA2 HEVC specs? It doesn't say anything about the hardware capability of Pascal. If DXVA2 doesn't support HEVC 2.0 Main 12 I'm sure it will be supported sooner or later.

It can have whatever hardware capabilities it wants, but if its not exposed through documented interfaces, we can neither verify it actually has them, nor use them.
But we'll see.

[nevcairiel]

Date Published:
2/17/2016

Here we go again with the silent updates? Or nothing actually changed?

Thats 3 month old and only contained some extra clarifications, no new things.

[JohnLai]

http://fs5.directupload.net/images/160515/o9qx4kyv.png
http://videocardz.com/59962/nvidia-geforce-gtx-1080-final-specifications-and-launch-presentation


Even 12 bit decoding.

Hmmm? 10bit HEVC encode? I wonder how many 'specification' being skipped for the encode again.......

Last time maxwell v2 hevc is missing SAO, 64 CTU, B-frame, what else?

[leonccyiu]

http://fs5.directupload.net/images/160515/o9qx4kyv.png
http://videocardz.com/59962/nvidia-geforce-gtx-1080-final-specifications-and-launch-presentation


Even 12 bit decoding.

Wow I didn't expect the specifications to leak tonight
I can't remember from several posts back, someone testing Intel's Skylake decoder for 8k HEVC benchmarks, did it manage 60?

A concern of mine is regarding VP9 decode, it lists 4k and 2x @120hz, the same as HEVC 8bit, but no mention of 8k. Youtube is going to be our prime source of 8k content for the foreseeable future, and with google's lackluster libvpx codec, at least a 5820k will be required for playback.

Youtube's 8k VP9 is monstrous, the bit-rates range from 60-120, and potplayer uses >2gb ram during playback.

This is going off topic but here are some videos that are encoded in 8kvp9 and not the low bit-rate h264 they originally used.

https://www.youtube.com/watch?v=QPdWJeybMo8
https://www.youtube.com/watch?v=VXUdY5P0VJA&authuser=0
https://www.youtube.com/watch?v=_KoGjIDAFk8&authuser=0

if anyone wants to try playing them back by CPU or downloading them for MPC and ffvp9.

[nevcairiel]

Hmmm? 10bit HEVC encode? I wonder how many 'specification' being skipped for the encode again.......

Last time maxwell v2 hevc is missing SAO, 64 CTU, B-frame, what else?

All these features are optional for an encoder, there is nothing saying it has to use them. Their primary purpose for these encoders is high-bitrate realtime encoding of you gaming, it doesn't need all sorts of fancy coding tools to do that.
In fact, for the purpose of streaming to a shield tablet or a TV, you wouldn't want some of those either way to reduce latency.

Such hardware encoders will never really be able to produce high quality encodes on low to medium bitrates. x264 and x265 will always be the best choices there.
The only thing hardware has going for it is that it is fast. To be so fast, it doesn't use all fancy tools.

[Yups]

http://www.computerbase.de/2016-05/geforce-gtx-1080-test/8/#diagramm-videos-abspielen-h264-codec-lav-filters


Much faster decoding.

[NikosD]

Looks very fast, but from the comparison tables Skylake is missing.

Of course, Polaris (10 & 11) and Kabylake are the real rivals for Pascal, regarding video decoding/ encoding.

[huhn]

even if polaris is only half as fast as this new pascal decoder it doesn't matter. it is so fast you are not run in real work issues with it.

the comparison is pretty pointless anyway there are clearly using an 8 bit encode.

[NikosD]

Maybe DXVA checker version of the screen shot is old and represents 8K as 4KUHD.

You need a newer version or even better the latest.

I'm not in front of my PC to check it out.

Also, DXVA checker displays Microsoft's or sometimes OEM's (rarely) device decoders which obviously are not available for HEVC Main12 by both of them (I mean Microsoft and Nvidia)

[nevcairiel]

Like I said on the previous page, MS needs to release an updated specification to allow Main12 decoding through "standard" means, which would mean it gets new GUID, and would not show up as a "named" device in DXVA Checker unless this GUID is known and DXVA Checker updated to support it (or NVIDIA cheats and allows 12-bit decoding through the Main10 device, but that would be evil and hard to detect in software).

We'll have to wait for MS, or NVIDIA, to make a move there, and more information to surface.

For that matter, their DXVAChecker does seem really old. 33FCFE41 is MJPEG_VLD_PureVideo and 463707F8 is VP9_VLD_Profile0.
The two entries below HEVC_VLD_Main10 are new however, so maybe one of those is supposed to be HEVC Main12.

[v0lt]

http://www.computerbase.de/2016-05/geforce-gtx-1080-test/8/#diagramm-videos-abspielen-h265-codec-lav-filters

[baii]

http://www.computerbase.de/2016-05/geforce-gtx-1080-test/8/#diagramm-videos-abspielen-h265-codec-lav-filters
Hmm, all the amd and old Maxwell, can I assume they are h.265 8 bit?
It is good that the performance is better but decoding is not exactly a how fast question but more a can it do it type of thing, which a 960 is enough for now.

Sent from my 306SH using Tapatalk

[ashlar42]

Such hardware encoders will never really be able to produce high quality encodes on low to medium bitrates. x264 and x265 will always be the best choices there.
The only thing hardware has going for it is that it is fast. To be so fast, it doesn't use all fancy tools.
Maybe sooner or later we will be able to have different "profiles" for the hardware encoders. Or are you saying there would not be spped benefits if encoding with the same quality through GPU?

I would say there would but I am not sure. There seems to be a lot of headroom for raising quality while mantaining a speed advantage.

[huhn]

they have different profiles already.

these encoder are missing way to many well tweak feature that x265 or x264 have.

[NikosD]

For those interested, I uploaded a 8K 19s clip from Youtube in two versions.

Original 8K VP9 and converted 8K HEVC (8bit)

Skylake owners and maybe Pascal or probably even future Polaris owners are welcomed to post their pure decode benchmark (or even playback) results, using LAV x64 0.68.1 and their HW decoders in DXVA native mode for both HEVC and VP9 formats.


8K HEVC - 24fps@38Mbps
https://www.sendspace.com/file/80sclj

8K VP9 - 24fps@66Mbps
https://www.sendspace.com/file/qlujz9

[Yups]

Only 8 bit support in hardware for Bristol Bridge?