[MPlayer-users] Fresh svn mplayer crashes on many ac3 streams

[Reimar Döffinger]

On Sun, May 29, 2011 at 11:26:47AM +0200, Reimar Döffinger wrote:
> On Sun, May 29, 2011 at 11:10:34AM +0200, Reimar Döffinger wrote:
> > On Sun, May 29, 2011 at 01:01:17AM +0400, Vladimir Mosgalin wrote:
> > > Anyhow it's a bit unfair to blame it on ".0" here, I think, since this
> > > issue is related to avx as well and 4.6.0 was first stable release to
> > > include avx support. It's, like, not gccs fault that avx support
> > > appeared in .0 version and not some other..
> > 
> > In the about 10 years since I've been with MPlayer I don't think
> > I've ever seen a .0 gcc version that didn't miscompile some
> > part of it.
> > 
> > > I can turn off avx or ftree-vectorize, it's np to me - though I expected
> > > some serious speedups from gcc 4.6.0 and avx support, like shown in some
> > > phoronix benchmarks, for example
> > > http://www.phoronix.com/data/img/results/intel_avx_gcc/6.png
> > > http://www.phoronix.com/data/img/results/intel_avx_gcc/7.png
> > 
> > If I was mean I'd ask if they actually checked the results.
> > It's really easy to be fast when you generate incorrect code.
> > Though the real point is that that was a pure-C program for
> > compute-intensive stuff with probably a lot of embarrasingly
> > parallel code.
> > Which is to say, it probably has no relevance to MPlayer speed.
> > Also if you care so much for speed you should probably rather
> > compile with Intel's ICC compiler, for personal use
> > (or something like that) it is free of cost IIRC.
> 
> You also forgot to mention that a lot (if not most) tests
> lost performance when enabling the AVX-math option that
> gave a speed boost in that case.
> Past experience tells that for most, and all relevant,
> formats (some unusual formats are simple but still not
> asm-optimized, those might gain) a reduction in performance
> is exactly what you'll see with MPlayer, too.

Sorry for the many posts, but I just finished the article.
The summary is a joke, my suggestion for a accurate summary would be
"In some artificial tests performance improved significantly when
enabling AVX math.
In most real-world applications though performance changed either
in an insignificant way or lead to a catastrophic performance loss
both when enabling optimizing for the new CPUs and particularly
when enable AVX math.".