--- On Mon, 5/30/11, Tobias Burnus <bur...@net-b.de> wrote:

> From: Tobias Burnus <bur...@net-b.de>
> Subject: Re: GCC on GPU - graphite OpenCL (was: WHOPR Linux distribution)
> To: "Robert Beeporbop" <rbeepor...@yahoo.com>
> Cc: gcc@gcc.gnu.org, "Sebastian Pop" <seb...@gmail.com>
> Date: Monday, May 30, 2011, 5:11 AM
> On 05/30/2011 12:10 PM, Robert
> Beeporbop wrote:
> > I have some questions for the gcc developers:
> > Also, I had a couple other general development
> questions:
> > 
> >    ~ Has anyone been working on GPU support,
> automatically-utilized or otherwise?
> 
> I have not used it - nor have I seen much documentation,
> but the Graphite branch of GCC supports shovelling loops to
> the GPU via OpenCL.
> 
> I have to admit I have not seen much documentation for it,
> just the talk at 
> http://gcc.gnu.org/wiki/summit2010?action=AttachFile&do=get&target=belevantsev.pdf
> 
> Thus, I cannot comment on how to use it nor how well it
> performs. Others might know better.
> 

Thank you for the link. The paper shows good results, for many cases
where good results would be expected. It looks like the benchmarks
relate mostly to how well general purpose code could be sped up using multiple 
cores or a GPU, while most of the code I deal with that I 
would use a GPU for is readily GPU accelerated.

I turned on all the bells and whistles in gcc 4.6.0, and 
compiled code written in the way it looks nice, rather than 
optimizing the code so that it compiles well. And received 
an order of magnitude speedup versus a few minor releases 
back (!) in some cases. Thank you too all who have made this possible!

From the paper, it looks like the Graphite branch takes advantage 
of the same code analysis techniques that make 
this speedup possible. So I'm wondering: has anyone ran 
benchmarks of code that could obviously (or not so obviously) 
be sped up by GPU acceleration? Maybe some software-defined 
radio code, the x264 codec, or code that already has specific GPU
support, which has been taken out, then compiled with this gcc branch
to see how close it gets?

Rob

> Tobias
>

Reply via email to