H.264 1080p bitrate vs PSNR trajectory

[RyanS]

From this test and analysis "Quick Sync H.264 Encoding Performance and Quality Test" (which uses q264.exe) here: http://www.tetrachromesoftware.com/q264Test1Analysis/q264test_5.html

The PSNR (quality) vs bitrate trajectory falls off for 1080p above about 4000 Kbps, meaning that increasing the bitrate doesn't increase the quality like one would expect. Lower resolutions don't show this trajectory fall-off. This looks to me like something fundamental, but is there some setting q264.exe could be messing up to cause this?

[celli4]

One thing to keep in mind, is I believe how this works is dramatically different between Haswell and Ivy Bridge, Sandy Bridge.

[RyanS]

Yes, thanks, page 4 of the analysis compares those three generations and shows the same trajectory.

[Petter_L_Intel]

Hi Ryan,

can you please provide the specific encoder configuration you are using for the encode workload where you see observe the quality degradation.   You can use the Media SDK tracer tool to fetch the configuration details.

Regards,
Petter

[RyanS]

Here is the tracer output. 

Note that the settings (other than bitrate) are the same for all of the runs that create the bitrate vs PSNR curves.

[Petter_L_Intel]

Thanks for the details Ryan.

We are investigating this and will get back to you shortly.

Regards,
Petter

[Petter_L_Intel]

Hi again Ryan,

We have studied the behavior for the encoder configuration you provided in further detail.

For configurations with high number of B frames (in your case 4) the HW encoder is known to have limitations with regards to quality on high bit rates. So, in essence this behavior is expected.

You should be getting better results using for instance 2 B-frames. You may also explore other rate control modes than AVBR to see if it produced better results.

Regards,
Petter

[RyanS]

q264 uses the normal GopPicSize, GopRefDist, and IdrInterval parameters in mfxInfoMFX to determine encoded frame types. 

Can you clarify what you mean when you say "2 B frames"? Do you mean the number of consecutive B frames? Do you mean reduce the number of consecutive B-frames by setting GopRefDist?

[Petter_L_Intel]

Hi Ryan,

Right. The number of B-frames following a P frame can be controlled by the GopRefDist parameter. If you set GopRefDist to 3 you will generate two consecutive B-frames, such as the following pattern IPBBPBBPBBPBB...

BTW. GopRefDist = 1 means no B-frames.

Regards,
Petter
 

[RyanS]

Okay, thanks. I tried other bitrate control methods and GOP params (including GopRefDist) and saw slight changes in PSNR, but no significant differences.

Ryan