Strange results of IPP H.264 encoding for subblock split = 1 option

andrewk88 · ‎10-18-2005

Hi,

I'd like to report some experimental data from H.264 encoding of test sequences: stefan and akiyo. For reference I've used JM10.0 reference software using Simplified FME algorithm. Basically, it seems like enabling subblock split = 1 option in IPP-based H.264 encoder (which is to enable 8x4, 4x8 and 4x4 subblocks for motion estimation)sometimes hurtsperformance of ipp-based h.264 encoder e.g. for stefan sequence (comparing to JM reference software) as follows:

h264IPP-stefan-nosplit 34 31.27dB 632.73kbps 9.712sec
h264IPP-stefan-split 34 31.27dB 638.45kbps 14.922sec

by the 1% bitrate increase instead of expected decrease at unchanged SNR level. Normally adding smaller sizes to motion estimation with a properly functioning R-D scheme shouldboth decrease bitrateand increase SNR at the same time. In the case of IPP code that only happens for Akiyo sequence and not for Stefan sequence.

Please comment or maybe that is already fixed in upcoming release?

Best regards,

AndrewK

PS. For your reference I'm attaching complete experimental results and configuration files of iPP H.264 encoder used in this experiment merged into a single file

Ying_S_Intel · ‎10-25-2005

Dear Customer,
Please use the latest version of IPP 5.0 gold from
http://www.intel.com/software/products/ipp

and updated sample with 5.0 from
http://www.intel.com/software/products/ipp/samples.htm

to see if the problem still exists in 5.0. The sample you need to use is Media Processing Sample where contains H.264 decoder/encoder componetns. You need to build the video encoder application in order to encoder h.264.

Please check the readme for instructions.
Thanks,
Ying S
Intel Corp.

Vladimir_Dudnik · ‎10-27-2005

Hi, there is developer's comment:

Thank you for your detailed information. There is such an effect in the encoder and it will remain also in the IPP v5.0 final release. Such a behaviour can be explained by underestimation of extra MV coding penalty in a choosing best prediction algorithm. We are working on this issue.

Regards,
Vladimir

andrewk88 · ‎10-27-2005

Hi Vladimir,

Thanks for the update.

Please let us know when fixed.

Best regards,

Andrew

Vladimir_Dudnik · ‎10-27-2005

Sure.

But it is better to submitt issue to Intel Premier Support, so it will be tracked and you will be notified automatically when fix became available.

Regards,
Vladimir