[mythtv] Re: [committed] denoise3d "black spots" fix
davatar at comcast.net
Thu Jan 29 21:54:59 EST 2004
From: "Andrew Mahone" <andrewmahone at eml.cc>
> On Thu, 29 Jan 2004 17:47:02 -0800, "Bruce Markey" <bjm at lvcm.com> said:
> > Go figure. The first thing I tried was USA (white logos) PGA Golf
> > (white shirts in the sunshine). No black spots. Maybe a
> > compiler/optimizer error?
> That's my thinking. The big oddity for me is that the three instances
> of this loop are cut-and-paste copies of each other, but only the last
> one, for the line 1+ temporal values, produces invalid output.
> Differing behavior from identical code feels like a compiler bug to me,
> and when I get a chance, I'll see if I can make this go away by fiddling
> with CFLAGS.
Well, the only thing I can see is that you're not using an asm volatile. I'm
not sure what gcc will do with that code, since if you don't indicate
volatile it's free to move it around if it feels it can make an
optimization. That may be your problem. Also why I stuck to doing single
blocks of asm in quickdnr, mmx is already hand optimized, I want the
compiler to do as little as possible to it.
This revision seems to be much better, no spots, lines or other artifacts
More information about the mythtv-dev