Evaluation of HEVC decoders (SW, Hybrid and HW)

[CruNcher]

It seems that Nvidias Hybrid H.265 Decoder is more efficient with CUVID Decoding especially on intra frames DXVA Native has the tendency to cause latency spikes (no matter which decoder used) on edge decoding cases i guess it has todo with DXVA Natives Frequency Scaling issues in conjunction with Nvidias Boost Power Management System.

I frames seem to heavy fluctuate in my tests with Nvidias Hybrid Decoder GM204 and a DXVA frequency of around 759 MHz @ 0,862v Real 4K 23.976 Fps

CUVID doesn't show these Latency I frame fluctuations @ 1126 MHz @ 1.025v


NVAPI Hook readouts (MSI Afterburner hooked NVAPI results gained after test)

Though Unwinders code is super nice it itself creates almost 0 fluctuations when polling of Nvapi is kept in sane ranges

Im mainly using as seen MPC-BE Sync Graph which is awesome fine grained for a almost Realtime Graph especially in Analyze Mode 2 when CPU overhead is way lower then with the Full OSD and additional timer polling (NVAPI ?) eliminated

@V0lt

Please add a Sync Graph only Analyze Mode without any additional OSD timer running maybe it can further lower the Overhead and it's caused EVR Latency Fluctuations that become visible in the Graph itself increasing the Audio and Video distance from the middle with DWMs tripple buffering
same happens with Energy Saving Modes the lines move away when timer precision is lowered and CPU overhead gets higher with lower Frequency


PS: About the Intel vs AMD thing Intel had a major advantage with shrinking up to 22nm before AMD could do it thus is the only reason they won the Mobile Space with Baytrail and Cherrytrail adding more CUs each time this advantage is slowly over, but it gave Intel the time to improve the GPU each time and close the gap to AMD and Intels GPU Core being now more feature advanced then all of them including the Video Asic part, where we will see surely another interesting development but i guess not with Kabylake in terms of Intels capability with their Video Decode/Encode time to market updates in the future

At least with 14nm we reach pretty much parity for the first time between virtualy all of them for a longer time period.

Intel = 22nm/14nm/ in the future targeting 11nm but the shrinking advantage will be lower as ever before
Nvidia =28nm/16nm/14nm
AMD = 28nm/14nm

so 14nm is the common denominator on that a lot architecturally will happen finally and no shrinking race advantage anymore for a longer period

[aufkrawall]

Who cares about minor fluctuations when you typically have more than a dozen frames queued ahead by decoder and even can have madVR render them ahead?

[Paul Tronc]

Originally Posted by CruNcher

It seems that Nvidias Hybrid H.265 Decoder is more efficient with CUVID Decoding especially on intra frames DXVA Native has the tendency to cause latency spikes ...

Indeed : I switched to CUVID, the Netflix_TunnelFlag is almost playable. Globally smooth, only 170 frames dropped for the entire sequence.

[Paul Tronc]

Originally Posted by CruNcher

Im mainly using as seen MPC-BE Sync Graph which is awesome fine grained for a almost Realtime Graph especially in Analyze Mode 2 when CPU overhead is way lower then with the Full OSD and additional timer polling (NVAPI ?) eliminated

@V0lt

Please add a Sync Graph only Analyze Mode without any additional OSD timer running maybe it can further lower the Overhead and it's caused EVR Latency Fluctuations that become visible in the Graph itself increasing the Audio and Video distance from the middle with DWMs tripple buffering ...

I'll have to dig into this 'Sync Graph', for now I don't understand what it is.

[CruNcher]

It's pretty simple a Graph that shows you the current Render Jitter and this is dependent on many System factors the lower the jitter the smoother the Rendering playback

And the Graph is very sensitive to System Problems and fluctuations in the whole WDDM Playback Chain because it's own Render Overhead isn't really high though it works only in the Render context of D3D9 in MPC-BE for now it's though muich nicer then some static numbers you can't bring into context over time of what you seeing even if it's slightly delayed

Push for example the Print Screen button and see if you get a spike

Originally Posted by Paul Tronc

Indeed : I switched to CUVID, the Netflix_TunnelFlag is almost playable. Globally smooth, only 170 frames dropped for the entire sequence.

Not bad if you take into account that you have only 2 GB

Though you wouldn't really expect anything else from VP7/8

Nvidia is very picky about memory see the last AV1 Optimizations they proposed for the De-ringing Filter or their latest AGGA R&D

Nvidias Hybrid Decoder failing not even Broadcast Complexity but high enough to get it performing very unstable



13 sm vs 4 Intel Cores to the rescue




PS: Did some RX 460,470 480 user tried if Strongene Lentoids Hybrid OpenCL Decoder still works, as each IDCT,MC and PP part is generation individually optimized and chosen at runtime i wonder if it still works @ all on Polaris GCN 4.0 supporting the Decoding or fails or if it's going to fallback to the Tonga\Hawai or some Generic GPU/CPU Path ?

Lentoid can even partly outrun lav video in complex bitstream parts when the CPU starts to starving under heavy load and push out more stable results then but therefore it also uses lot more System Ram and it performs equal in 32 bit as in 64 bit no slowdown like lav video



Fast Scenechange sudden Bitrate Peak test

Lav Video CPU




Lentoid CPU

[Paul Tronc]

I'm back, to share my findings. I'm trying to optimize my CPU and GPU power in order to get the maximum HEVC decoding result.
For now, the best thing I can achieve on a 4K 40Mbps HEVC sample, using a 2600K@4.4 and GTX960, on a 1080p display :

- GPU HEVC decoding : OK.
- Madvr : Jinc and other "rather good" algoritms
- FRC (Frame Rate Conversion) using SVP, from 24p to 60p

The GPU assumes most Madvr + LAV work, while the CPU is used for frame interpolation.

I'm currently tuning the settings to get a rock stable smooth experience.
The 1080p output helps a lot, on my 1440p the rendering time grows quickly.

[wanezhiling]

http://i.imgur.com/CYpKdmp.jpg
http://i.imgur.com/opaTGVd.jpg
Pascal 8K VP9

[NikosD]

And 10bit VP9 ?

[wanezhiling]

Do you have 10-bit VP9 clips?

[NikosD]

No and I haven't found an encoder

[aufkrawall]

Any VP9 decoding still not enabled with Polaris?

[huhn]

here a 3 sec VP9 10 bit file: http://filehorst.de/download.php?file=bDyyChti
i encoded it with VPXENC

edit: i will check out 16.10.1 for polaris VP9 support

[NikosD]

You own a polaris card ?

[huhn]

yes and no VP9 support in 16.10.1.

[NikosD]

It would be useful if you could post H.264 & H.265 (8bit & 10bit) benchmarks of the usual clips of this thread.

Which one by the way ?

[huhn]

RX 480 4Gb.

it's broken or at least i have major problem with it to even get an Constant image have to replace before i can do any proper test.

i will do some test with the card in some of my other PC this weekend to make sure it is really the card. even though it is pretty clearly the card.

in short no performance tests for now. and most likely not any time soon.

[NikosD]

Do you have problems in 2D and 3D besides video ?

[huhn]

i get the problem in madVR and in browser. so it is generally.

nothing special just bad luck.

i could work around it and do some tests but testing this card it self is just more important right now.

trust me i'm very very upset about this.

[NikosD]

Brand ?

[huhn]

XFX.

i don't know why this should matter.