Evaluation of HEVC decoders (SW, Hybrid and HW)

[NikosD]

On my i5-4440/ HD4600 (v.4703), GPU monitor of Fall Creators Update only shows 3D chart, the rest of charts (video decoding, processing etc.) shows nothing.

Make sure you are using HW acceleration for compatible video formats.

I tried MPC-HC with latest nightly LAV and DXVA native using H.264 clips.

Next time, I'll try also MPEG2/VC-1/WMV3 and the hybrid H.265 video.

[aufkrawall]

The VPU of the GTX 1050/Ti or 1030 was not just able to decode VP9 10, it was also faster in general than of other Pascal GPUs, right?
YouTube 8k HFR VP9 8 bit almost is too much for the VPU of my 1070, it gets up to ~95% usage and the decoding queue reported by madVR sometimes gets lower by 1 frame (playback stable though).

It also reveals DXVA2 as slow crap: Contrary to D3D11 it gives me lots of dropped frames, even with EVR.
Not to speak of copyback modes, which are totally out of the game.

[el Filou]

I don't think it's faster, I also get just around 60 fps with my 1050 Ti on VP90 8K with DXVA native. It's good enough to play 8K 24/30p but not HFR smoothly.

[aufkrawall]

It can play HFR smoothly here.
Don't underestimate chroma scaling for 8k video, probably set it to bilinear (or even nearest with a slow GPU) or enable the "scale chroma separately" checkbox. It saves tons of performance when doing much downscaling.

[NikosD]

GPU Monitor tool of FCU is a great tool.

It works for my HD 4400 for all codecs supported in HW (MPEG2, VC-1, WMV3, H.264) and the interlaced ones (MPEG2, VC-1, H.264)

It doesn't show any HW resource used when decoding H.265 and that proves for good that hybrid H.265 decoder of Haswell is a pure GPU based decoder (EU only)

[aufkrawall]

It also works with a Gen 8.5 Atom IGP while it has only a WDDM 2.0 driver.

[wanezhiling]

https://i.imgur.com/EIefn9l.jpg
https://i.imgur.com/fRQVzXm.jpg
8th APU

[mrcorbo]

https://i.imgur.com/EIefn9l.jpg
https://i.imgur.com/fRQVzXm.jpg
8th APU

Thanks for this. Finding it really annoying that no site I've seen seems to go even slightly out of their way to provide this information anymore.

[NikosD]

Thanks for this. Finding it really annoying that no site I've seen seems to go even slightly out of their way to provide this information anymore.

Desktop VEGA as we know, doesn't have fixed-function (pure hardware) VP9 decoder.

I think is the same for APUs using VEGA GPUs, so those figures for VP9 8bit/10bit have to be the decoding performance of hybrid VP9 decoder.

Unless APUs are using a 2nd edition of VEGA video decoding hardware.

[huhn]

the hardware decoder functionality of vega 56 and 64 is not a high priority. people that buy such a card don't do that for hardware decoder.
this totally changes with APUs these system not dedecated to pure gamer or GPGPU application adding hardware decoder to these iGPU makes a lot of sense and it is about time.

it would be disappointing if the VP9 decoder are hybrid

[NikosD]

Just to make things even more complicated, I found this official image:
https://s1.postimg.org/632gctw4bz/battery.png

Ryzen APUs double the efficiency of VP9 playback compared to a previous generation, but still that could be achieved by a transition from CPU to GPU (hybrid) decoding, not necessarily to fixed-function VP9 decoding.

If someone doesn't test it, especially under Fall Creators Update in order to see the Video Decoding Engine utilization, I can't be sure.

[darkrenata]

What would be the consensus of i7-8700k working for 4k 60fps HEVC 10bit? I've tried with mpc-hc LAV+madVR, but I was getting frame drops or stuttering with dxva copyback while using the display out from a 970 (with the copyback I have the iGPU decoding). Don't think I was using x64 so I might have some stuff I can change around to test and see if that's my issue, it looks like it decodes fine, but there is playback issues.

iGPU usage is like 60% max, so if it couldn't handle decoding it shouldn't it hit 100%?

[nevcairiel]

The memory bandwidth of copyback for 4k 10-bit 60fps might be putting quite a strain on the system.

[aufkrawall]

Also give D3D11VA of LAV Filters nightlies a try.

[darkrenata]

The memory bandwidth of copyback for 4k 10-bit 60fps might be putting quite a strain on the system.

Yeah it seems that way, playback with dxva2 native works flawlessly. Problem is that my motherboard displayport out to and hdmi AVR for my projector apparently doesn't do 4K ????? I think it's supposed work via LSPCon, but I'd have to do research to consider using that, so it seems the Intel HD 630 is pretty useless to me since I'll probably just buy a 1080Ti unless Nvidia announces something new soon.

Also give D3D11VA of LAV Filters nightlies a try.

Can you explain exactly what I'm looking for? I'll try to search, but do you think it would make any difference for dxva2 copyback decoding?

For those curious my results for decode benchmark on Jellyfish 400Mbps
are average 85fps with range [67-112].
Playback benchmark is average 84fps with range [68-110]

This is with dxva2 native.

[aufkrawall]

Can you explain exactly what I'm looking for? I'll try to search, but do you think it would make any difference for dxva2 copyback decoding?

Just install the latest nightly build and select D3D11 in the video decoder settings instead of DXVA2.
DXVA2 is slower on the GPU than D3D11VA, perhaps this is also true in your case when using copyback. Doesn't hurt trying.

[littleD]

I see no help for the time being. Decoding on intel in dxva CB or even dxv11va gives slowdown on higher fps, resolution, bitness. Dxva native always smooth.
Decoding benchmark is always much faster in dxvachecker than playback benchmark. Looking on scores in cba mode - if not bus or video memory quantity? limitation, all files would be playable as in native dxva.

Wonder if intel implemented lossless color compression, if not it would be handy here as nvidia and amd have it.

[aufkrawall]

That's what I was arguing with huhn a few pages ago: Copyback can put lots of additional load on the GPU or slows it down.
When I put the the GTX 1070 to the limit with high madVR settings, playback is still smooth with native decoding, but drops frames when using copyback.
Copying lots of data from the GPU into one direction or another should be avoided if possible, it's just ugly. DXVA2 might be slower than D3D11VA though, at least that's the case when using native decoding (I commented about 8k VP9 playback a few posts back).

[wanezhiling]

I never use any copyback mode cuz it's much slower and costs too much cpu/gpu stresses.

[littleD]

Is it measurable property? I mean, can we put some values of - when given system would be limited by internals for CB? Like video memory, bus speed, system ram speed etc.

I have somewhere on the disc test programs one in d3d one in opencl that measure speed cpu<>GPu, not sure if scores would be relevant.

[nevcairiel]

Memory bandwidth hasn't really increased much over the last generations, consumer CPUs are still dual channel and the iGPU still uses that same system memory. At the same time 4K 10-bit 60fps uses around 20 times the raw memory bandwidth that your typical 1080p24 content would use.
On the other hand, running copy-back from an NVIDIA card and uploading back to that same NVIDIA card is still quite fast, even with 4K 10-bit content. So one can only conclude that the Intel GPUs are just not very good at this.

[huhn]

it's not easy to measure i personally never noticed anything worthwhile using CP with any of my video cards.

native is disabling a lot of possibilities which should not be generally ignored.

[aufkrawall]

Just look at the GPU usage reported by tools, it's very noticeable here with GTX 1070 and not too expensive scaling/processing in madVR/mpv.

[huhn]

you know my reports it's meaningless on all my cards in the past 4 years.

[darkrenata]

Just install the latest nightly build and select D3D11 in the video decoder settings instead of DXVA2.
DXVA2 is slower on the GPU than D3D11VA, perhaps this is also true in your case when using copyback. Doesn't hurt trying.

I tried this. Happy to report that for the 4k 60FPS content I have it playbacks flawlessly on my 1080p monitor out of my 970 decoding on the i7-8700k iGPU. I'll try my 4K display next time I turn it on but this seems pretty good so far.

Thanks for the suggestion.

[gagmani]

1070 Ti does not have VP9 Profile 2 just like 1070 and 1080.

https://imgur.com/a/4b4c8

[Razoola]

1070 Ti does not have VP9 Profile 2 just like 1070 and 1080.

https://imgur.com/a/4b4c8

I was thinking of getting a 1070TI, what does the loss of VP9 profile 2 mean in real terms? Does any nvidia GFX card have it?

[huhn]

you can't decode VP9 profile which is used on youtube HDR.
1050 ti and slower 10xx cards and there are information that state the 1080 ti can do it too

[gagmani]

I was thinking of getting a 1070TI, what does the loss of VP9 profile 2 mean in real terms? Does any nvidia GFX card have it?

99% youtube vp9 videos would play fine. Only 10 bit(HDR) ones would fail. Netflix playback can be problematic on HDR content.

[el Filou]

99% youtube vp9 videos would play fine. Only 10 bit(HDR) ones would fail. Netflix playback can be problematic on HDR content.Netflix also supports HEVC so that's not a problem.
If Razoola has a powerful CPU maybe they could decode VP9 profile 2 in software?

[aufkrawall]

Bitrate of YT VP9 10 bit is lower than of 8 bit, so it might work with software decoding.

[Razoola]

I have an old i7 920 at 4Ghz and its not fast enough hence I'm looking for a new gfx card. Can the 950 do it? I understand this card has some HEVC and VP9 support but I'm not sure that includes 10bit.

[huhn]

as far as i remember the first card with support is the 1050 ti.

[sneaker_ger]

Yes, GTX 950 does HEVC 8 bit + 10 bit and VP9 8 bit but not VP9 10 bit.


Note: AFAIK it does not have the DRM features the Pascal series has. So don't expect 4K Netflix or Blu-ray to work through non-pirate/non-ripped means. I also don't know about HDR even if pirated. Power and RAM are also limited in case you want any advanced madvr features or similar.

[nevcairiel]

there are information that state the 1080 ti can do it too

I can confirm this, the 1080Ti indeed has VP9 10-bit support.

[aufkrawall]

Too bad your VP9 10 bit HW decoding patches haven't got mainlined yet in ffmpeg. :(

[nevcairiel]

Thats because they're not actually complete. LAV uses a bunch of its own HW functions, so it wasn't really required to add the extra stuff. But maybe I should.

[Razoola]

I can confirm this, the 1080Ti indeed has VP9 10-bit support.

I'm a little confused now. So the 1080TI has VP9 10 bit support but the 1080 and 1070TI does not? I somehow got the impression these used the same GPU so I though they would all have the same HW decoding spec.

[nevcairiel]

Indeed, the underlying chip variant is what matters, and only GP102 and GP107/108 seem to support it, since both are relatively new. So that is 1080Ti (and possibly the Titan Xp) based on GP102, and the 1050 Ti and below on GP107 or 108.
The 1070 Ti, while a "new" card, is based on the exact same chip as the 1080 and 1070.

[Razoola]

Thanks for the clarification.

[el Filou]

I'm not sure those VDPAU Feature Sets still provide valuable information, for example all Pascal GPUs are supposed to be Feature Set H but GP107, GP108, and GP102-350 support 10-bit VP9 while the rest don't.

EDIT: just out of curiosity, has anyone compared VDPAU deinterlacing to Windows DXVA deinterlacing on NVIDIA?

[aufkrawall]

mpv offers first-field bob temporal temporal-spatial (default: temporal) for vdpaupp. Just like the available D3D11VA deinterlacers, they all show that aliased look compared to YADIF. So I don't think VDPAU deinterlacing is much different.
I guess it at least doesn't show that idiotic DXVA2 chroma shift, just like VDPAU decoding doesn't show that idiotic chroma blur problem with OpenGL.

There's btw. no deinterlacing for Nvidia's new proprietary decoder NVDEC, and CUVID deinterlacing seems broken on Linux.

[el Filou]

Has anybody ever seen a 4K MPEG2 video in the wild?
I've always found it weird NVIDIA would support decoding that codec at 4K.
I know 4K MJPEG is actually used (some Canon cameras, I think) but I thought everybody else had moved to AVC-Intra or something else H.264 based instead of MPEG2 for 4K.

[gagmani]

I bought a Freesync monitor and plan to buy an RX 580. Is VP9 hardware acceleration still broken on AMD? I have an 1030 which works flawlessly with 4k vp9/hevc.

[Yups]

The new 15.60.1.4877 driver comes with some Media improvements for Kabylake according to the release notes.


- Enhanced performance in HEVC/H.265 video playback on Intel 7th Gen Core processors or higher
- Improvements in the Intel® Media SDK (MSDK), leading to better
encoding quality across dependent media apps such as PowerDirector*, MAGIX*, etc.
https://downloadmirror.intel.com/27355/eng/ReleaseNotes_GFX_15.60.01.4877.pdf

[NikosD]

I bought a Freesync monitor and plan to buy an RX 580. Is VP9 hardware acceleration still broken on AMD? I have an 1030 which works flawlessly with 4k vp9/hevc.VP9 hardware decoding works only for YouTube officially, using Windows 10 and some modern browser in latest versions (like Chrome, Firefox)

Of course, it's not pure hardware decoding but actually it's hybrid (using GPU processing)

[senseipiledrive2]

Hey everyone, I could use help with my video settings. Installing newest klite standard with mpc-hc + madvr, I use a 6770k CPU and 980ti vid card.

After reading, it seems D3D11VA hardware acceleration is a good choice for my setup. However, I left VP9 and HEVC acceleration unchecked, because it will use hybrid GPU (980ti) acceleration, which will probably lower the performance over software for those choices. Is this correct? I also checked the boxes in the Lav Video Decoder settings for Hardware Decoder to use: D3D11, Codecs for HW: All except Vp9, HEVC, HEVC 10. Under hardware device to use: Automatic (Native) is selected. Is this right?

(Edit): Also...I use a 1080p TV that uses limited colors, not full range. I've set my Nvidia panel to full range (new drivers), MadVR to limited range...and I just found out I can change the output to full, limited, or native in LAV Video Decoder settings. I have it set to "Limited." Does this even have an effect since I use MadVR and have it selected there?

Random last question: I unchecked 8-bit audio integer under LAV audio output options: I don't think my computer or receiver support this, nor do I have 8-bit audio files (I think). Should I leave 8-bit checked?

[clsid]

Yes, hybrid decoding uses a significant amount of GPU resources. If software decoding works smoothly (which it should with your CPU) then it best to have all GPU resources available to madVR.

The range setting in LAV should be left at default. Audio output as well. LAV will take care over everything.

[aufkrawall]

The hybrid VP9 decoder of Polaris seems to be working quite well in Chrome, playback of 4k 60fps is smooth and CPU usage is much lower than with software decoding.
Reported GPU usage is low too, but I'd add a questionmark to that since it allegedly even is lower than when using software decoding and the reported GPU usage on AMD is still an unreliable mess compared to Nvidia.

[Cary Knoop]

There's btw. no deinterlacing for Nvidia's new proprietary decoder NVDEC
So for instance -c:v hevc_cuvid is not part of NVDEC? :confused: ?