[Ffmpeg-cvslog] CVS: ffmpeg/libavcodec mpegvideo.c,1.511,1.512

[Michael Niedermayer]

Hi

On Tue, Mar 28, 2006 at 03:24:57PM -0800, Loren Merritt wrote:
> On Thu, 23 Mar 2006, Jeff Clagg wrote:
> 
> >On Fri, Mar 24, 2006 at 12:53:35AM +0200, Jan Knutar wrote:
> >>On Friday 24 March 2006 00:43, Diego Biurrun wrote:
> >>>On Thu, Mar 23, 2006 at 10:54:50PM +0100, Loren Merritt CVS wrote:
> >>>>
> >>>>Modified Files:
> >>>> mpegvideo.c
> >>>>Log Message:
> >>>>prefetch pixels for future motion compensation. 4-10% faster mpeg1/2/4.
> >>>
> >>>Have you benchmarked FFmpeg against libmpeg2 after the change?  Would be
> >>>very interesting...
> >>
> >>As would benchmarks comparing ffmpeg before prefetch, with ffmpeg after
> >>prefetch, on other CPUs than Loren's.
> >
> >I tried on one machine (not my own) and got mixed results, mostly
> >unfavorable.
> >
> >The machine is a dual core 3.2ghz p4. I compared cvs from 3-21 (just
> >prior to Loren's last bunch of mmx commits) with cvs from after the
> >prefetch change, using GCC 4.0.3. In a nutshell, mpeg4 ASP is about 8%
> >slower, mpeg2 is 10%-20% slower, but H.264 speed improved by less than
> >2%. Below is a summary of the results. Maybe sometime later on, I'll try
> >these tests at home, using my Athlon 64 in 32-bit mode, with an older
> >GCC. Though maybe Loren's already done exactly this test.
> 
> I found a P4 2.8GHz with gcc 4.0.3, and compared current cvs vs cvs with 
> the prefetches commented out. Result: with prefetches gave 4% faster 
> mpeg4asp, 1.5% faster h264.

hmm, interrestng, are both of you using the same resolution? maybe theres
some cache line aliasing causing this?

[...]

-- 
Michael