[vox-tech] more on GPGPU

Norm Matloff matloff at cs.ucdavis.edu
Fri Jun 16 10:18:04 PDT 2006


> Date: Tue, 13 Jun 2006 13:28:48 -0400
> From: p at dirac.org (Peter Jay Salzman)
> Subject: Re: [vox-tech] GPU calculations
 
> On Tue 13 Jun 06, 10:20 AM, Richard Harke <rharke at earthlink.net> said:

> > Why don't you look at   www.gpgpu.org
> > GPGPU -> General Purpose computing on a GPU

> Thanks, Richard.  Good find!  I think you do numerical computing as well --
> have you done any of this?  I've seen GPU implementations for solving sparse

Pete, as I mentioned to you recently, I've been looking into GPGPU too,
and though I am at the VERY beginning stages of this, I thought might
add a little bit to the discussion here.  

First, there are two main APIs that have been developed, brook and Sh
(as well as an "assembly language"), used from C/C++ code.  At this time
it appears that brook is the more popular of the two.  Both are stream
processing languages, basically vector operations.  Anyone who is
familiar with MPI or Python or ... will feel comfortable here.  There is
a reliable rumor that an interface from OpenMP is being developed, which
would be really nice.  All of these seem to work via preprocessors, a
mechanism I've always found to be undesirable, but hey, one can't have
everything.

I've found that brook is quite easy to learn--though less easy to build
on Linux :-) an apparent consequence of the Microsoft orientation of
much of the current GPGPU community.  (They do apparently use
GCC/Cygwin, though.) I've not yet succeeded in building it under Fedora
5, though I haven't put much time into it yet.  All I can tell you so
far is that it didn't build right out of the box.  If you or anyone else
succeeds in building it under any form of Linux, I'd appreciate hearing
how it went.

One of the main movers and shakers in the GPGPU community is John Owens,
an ECE professor at UCD, jowens at ucdavis.edu  He tells me that the brook
discussion group at http://www.gpgpu.org/forums/ is very helpful for
build or other problems.

Another problem is that one apparently needs at least some knowledge of
graphics programming in order to really take advantage of all this.
That of course is not surprising, and it does appear that one can get
pretty far without knowing any graphics, but to really take full
advantage of the graphics hardware one needs to know what the hardware
does, which in turn means knowing a bit of graphics.

There are a number of good tutorials on brook on the Web.

Norm



More information about the vox-tech mailing list