Hi,

I was just jotting some notes on my recollections of Pat Hanrahan's
talk today at school, and some are relevant to Squeak, so here's the
lot

- Talk was entitled "Why are GPUs so fast?"

- GPUs 2.4x faster each year; trend dates back to SGIs in mid-80s

- STREAM PROCESSING!!  Put 1000s of floating-point units on chip, and
keep them fed

- GPUs would need "surprisingly few" changes to be ideal for
general-purpose scientific computation

- Although many problems are not parallelizable, many of the ones that
max out CPUs today can be re-cast as stream algorithms
	- LAPACK-style matrix operations
	- Dynamics (cloth, fluid simulation)
	- Low-level computer vision (eg: edge detection)
	- Statistical pattern recognition (Bayesian/neural nets)

- "Brook": a stream processing language that they will be making an
open-source release of.  Compiles down to Cg, GLSlang, or sometimes 
GPU assembly language.  Hides graphic-centricity of OpenGL, Direct3D.


This last one got me the most excited.  Up until now, Squeak had to
jump through hoops to write performance-critical code essentially in C
via Slang and plugins.  C/C++, on the other hand, wrote performance
code in their crappy own selves.

If we write an interface to Brook (I wouldn't imagine it would be tougher
than the Cg interface I wrote), then Squeak will be on equal footing with
for using stream hardware to do the heavy computational lifting.

Good stuff!

Joshua




On Tue, Nov 18, 2003 at 03:25:59PM -0500, Joshua 'Schwa' Gargus wrote:
> On Tue, Nov 18, 2003 at 08:52:07PM +0100, Bert Freudenberg wrote:
> > Bert Freudenberg wrote:
> > 
> > >Jecel Assumpcao Jr wrote:
> > >
> > >>So we might extrapolate and consider that it is only a matter of time 
> > >>before REYES starts invading OpenGL's turf.
> > >
> > >
> > >Indeed. You might want to take a look at Ashli, if you haven't done so 
> > >yet. See http://ati.com/developer/SIGGRAPH03/AshliOverview.pdf
> > 
> > Oops, that were only the slides. This is the text version:
> > 
> > http://ati.com/developer/SIGGRAPH03/AshliNotes.pdf
> 
> So much reading to do!
> 
> Thanks for the link,
> Joshua
> 
> 
> > 
> > 
> > -- 
> >    Bert
> >