Evaluation of HEVC decoders (SW, Hybrid and HW)

[huhn]

TEARS OF STEEL remux 720p source http://www.h265files.com/

lav 63.0 with 16 threads 821 fps 66 % CPU usage
lav 63.0 with auto 842 fps 60 % cpu usage
ms mf 732 fps 67 % cpu usage

ShakeNDry_3840x2160_120fps_420_8bit_HEVC_RAW.hevc 300 frames remux to mp4 i got this file on a page with tons of samples in all size and codec even lossless YUV. is a long time ago can't find it right now.

lav auto threads 87 fps 76 % CPU
MS MF 90 FPS 85 % CPU

120 FPS game recoding created by HUHN

lav 63.0 366 FPS 53% CPU
ms MF 363 84%

i don't see a clear winner but huge potential in thread optimization for the decoder in Lavfilter

[NikosD]

I upgraded Win 10 to the latest build 9879 and I got a little performance boost for HEVC decoding:

Microsoft's MFT H.265 20/36/60 CPU Usage: 90%

LAV x64 0.63.5 15/21/23 CPU Usage: 67%

I tried LAV with Auto Threads and 8 threads with no performance difference.

I uploaded the clip tested here, in order more people to check it out.

https://www.sendspace.com/file/1aaex4

[huhn]

lav 63.0 117/159/405 61%
ms mf 150/204/292 72%

something about this file i haven't see something like this with every other file.

[NikosD]

Maybe because it's a 10 bit file.

If you have another one, you could try.

[huhn]

i don't have one time to create one.

[xooyoozoo]

This site has a couple 20+mbps 10-bit 4K HEVC demos: Eutelsat and Astra.

You may need to remux.

[P.J]

Interesting, would you try 10-bit 4K with both MS and LAV?

[huhn]

i try it later. but i haven't found a way to get hevc stream out of the TS. so i may go and simply create a 10 bit video should be faster.

[P.J]

Doesn't WMP play TS HEVC?

[Kurtnoise]

Originally Posted by huhn

i try it later. but i haven't found a way to get hevc stream out of the TS.

did you try with the last version of TsMuxer ?

[huhn]

Originally Posted by Kurtnoise

did you try with the last version of TsMuxer ?

yes it find the stream with Eutelsat.Demo.HEVC.ts but it can't find the fps of the file and an option for 50 is not part of the GUI.
i tried the CLI didn't got it working.
if i remux the stream to m2ts, mmg detects the video stream as AC3/EAC3 X-).

Astra.SES.Demo.HEVC.ts doesn't work at all the stream is not found only audio.

[huhn]

Originally Posted by P.J

Doesn't WMP play TS HEVC?

no and would be pointless in this case. i don't need it to work in a player i need it to work with a benchmark tool and that's not the case.

[xooyoozoo]

It played in a Win10 VM through the Video app after

ffmpeg -i $FILE -map 0:0 -c:v copy -movflags faststart $FILE.mp4

.

Edit: Well, I uploaded a remuxed version of the Astra clip.

[huhn]

astra mp4 remux
ms mf 52/58/66 63%
lav 63.0 57/74/117 77%
that's that i guess.
no i didn't mixed them up

[xooyoozoo]

The results for Astra looks great! But it brings up the question of why LAV does poorer on the 10bit Gravity clip.

Originally Posted by huhn

something about this file i haven't see something like this with every other file.

So I went through and looked at the parameter sets in a few of the clips here, and compared to the others, that clip (Gravity Trailer) has two "unique" features: weighted-prediction and wavefront parallelism.

I hope that either one isn't the sticking point for the LAV decoder, but just to check, can you also try this clip? It's a 90-seconds 10bit 16Mbps 4K encode with WPP and weighted-pred.

[huhn]

lav 63.0 46/56/68 93%
ms mf 36/48/54 85%

i guess it's wavefront parallelism then.

[xooyoozoo]

WPP is shorthand for wavefront parallelism, so that feature is already turned on.

In any case, the two 10bit-4K clips's results are enough to allay my doubts about LAV's decoder optimizations.

[huhn]

for some reason lav is a lot faster with UHD content but why.

maybe the cabac is faster in ms mf?

[NikosD]

I tried every clip posted here, but unfortunately as I have said before, my laptop probably due to graphics card or DXVA Checker's bug can't benchmark 4K HEVC files (not even 4K H.264), at least under Win 10.

But I uploaded for you another non 4K HEVC file to run a benchmark.

Once again for me, MS MFT 9879 is faster than LAV

TOS sample

MS MFT x64 9879 11/23/36 CPU: 91%
LAV x64 0.63.5 12/19/31 CPU: 91%

Same CPU utilization, MS MFT faster than LAV (about 21%).

You can download the clip here:
https://www.sendspace.com/file/uq8osh

[NikosD]

Originally Posted by xooyoozoo

So I went through and looked at the parameter sets in a few of the clips here, and compared to the others, that clip (Gravity Trailer) has two "unique" features: weighted-prediction and wavefront parallelism.

How did you do that ?

How did you find the encoding parameters of the Gravity trailer and the other clips ?

Is there a tool ?

Thanks!