[FFmpeg-devel] MPEG-2 FLAGS2_FAST benchmarks

Kieran Kunhya kierank at obe.tv
Sun Jun 14 23:32:12 EEST 2020


>
> > Also the speed benefits are tiny overall, sub-1%.
>
> This claim is not supported by the statistics you posted
>

For the overall decoding of MPEG-2 the VLC/dequant functions take up
roughly 30% of decoding time.
FAST vs normal is about 2% faster on the VLC/dequant functions, therefore
the total speed benefit of FAST is sub-1% for MPEG-2.


> > Removing it and some
> > other small changes to the existing code will make the existing code run
> at
> > the same speed.
>
> its a long time since we had mpeg1/2 optimizations
> iam very happy about this being speed up more.
> If you manage to make it as fast as fast mode i surely would be happy
> If you manage to make the difference shrink by half i too would be happy.
> Faster decoding of mainstream codecs is a good thing. It has many
> advantages ...
>

Part of the speed difference is the if(FLAGS2 & FAST) branch being gone on
each macroblock.
However, if you are not willing to remove FAST I don't see the point of
continuing.
You have free reign to "optimise" FAST to justify its existence as it
doesn't have to conform to any spec.

Kieran


More information about the ffmpeg-devel mailing list