[XviD-devel] Re: OpenMP parallelization

[Christoph Lampert]

On Sun, 9 Feb 2003, Felix von Leitner wrote:

> Thus spake Christoph Lampert (chl at math.uni-bonn.de):
> > what would you think of OpenMP parallelism instead of
> > pthread-Multithreading?
> 
> Why "instead of"?  Is there reason to believe that openmp could scale
> better than the pthread stuff?

No, but pthread stuff requires big modifications and make the source either
completely unreadable code or needs a double code basis for SMP and
non-SMP. 

OpenMP would only be a few pragmas at crucial positions. I don't expect
linear speedup, but on SMP it might help in heavy calculation routines
which are still C-only, like GMC. 


> > Does VisualC++ suport this (yeah, I know, of course not).
> > There is something call Omni OpenMP for cygwin and other projects, I never
> > heard of.
> 
> > icc for Linux does support openMP, so maybe Windows version does, too. 
> > And there might be some modification of gcc that does, but I didn't find
> > it.
> 
> Just a few days ago, a gcc cvs branch was created for openmp, but AFAIK
> no actual code is in it yet.
> 
> I'm not involved, I just saw it on the mailing list.
> 
> > For those who haven't met openMP: It's a "trick" to tell the compiler
> > which parts of a program can be multithreaded by  #pragma  's in the code.
> > If the compiler doesn't know about openMP, it ignores the statements.
> 
> I wonder how this can me more efficient than pthread on any CPU you can
> buy today.  

I'm pretty sure that it isn't more efficient. But it's simpler to do,
because you don't have to deal with all the semaphores and mutex stuff. 

gruel