[FFmpeg-devel] h264 speed regression after PAFF

[Jeff Downs]

On Fri, 12 Oct 2007, Andreas ?man wrote:

> Hi
> 
> After the PAFF code got in h264 is about 3% slower for non-interlaced
> content. See test below.
> 
> The offending diffs are
> svn diff -r10679:10690 (changes in deblock_mb_fast())
> svn diff -r10691:10696 (The "enable" patch)
> 
> The "enable" patch can obviously not be blamed on its own.

Yes.. the other diffs are near completely disabled until the enable is 
applied.  Would be interesting to rerun these with FIELD_PICTURE properly 
defined at each step.

If anyone doesn't need interlace support and wants speedup, they can undef 
interlaced support for now (not terribly tested, but should behave as 
before).

> Either way, I find this a little too much. Michael, whats your opinion?
> 
> I haven't tried to fix this (yet), i just want the share
> the output from the test.
> 
> Jeff, you mentioned earlier that you was going to address the
> (MB_MBAFF || FIELD_PICTURE) -stuff. Are you working on this
> or are there still remaining bugs that needed to be squashed?

I was looking into the various reports so far (seeking, spatial direct,
misordered frames, etc.) but actually got side-tracked on some other
things, too.

I can make the changes which will get rid of one of the tests in most of
the macroblock decoding levels.

I can also put up some of the cosmetic cleanups.

 
> Running rev r10714
> User: avg: 10.033  stddev: 0.027  med: 10.033

Is there an easy way to run this test/benchmark?

	-Jeff