Re: GPU computing for octave

[Ronny Lindner]

Hello,

I can finally present my work to you. I tried to create an Octave
plugin which uses BrookGPU to access the GPU (NVIDIA 7000+, ATI
x1000+). There is a matrix multiplication example with good performance
and some other less used algorithms. The drawback is that everything is
calculated in single precision! The machine that runs the plugin must
have hardware acceleration - so you need to install the 3D driver for
your graphics card.

My diploma thesis is in german, so I only attached the commands to get
everything that's needed.

Before you can use BrookGPU on a regular basis there is a nasty bug
left: Every time you start a brook program in octave, some memory is
not freed, so octave will crash after a while (100+ calls). When
Octave is closed or crashes, the memory is freed!

To start a compiled octave plugin you have to set the environment
variable "BRT_RUNTIME=ogl" (ogl or dx9 for windows), you can do this
for example by calling octave like this: "BRT_RUNTIME=ogl octave"
The maximum size per dimension is set to 4096 in BrookGPU, so the
biggest matrices you can multiply are 4096x4096!

I hope this code helps to speed up some calculations or at least gets
the Octave <-> GPU programming going :)

Bye, Ronny

Am Mon, 2 Feb 2009 08:46:42 +0100
schrieb Andreas Romeyke <address@hidden>:

> Hello
> 
> Am Fri, 30 Jan 2009 22:31:25 -0800 (PST)
> schrieb melonakos <address@hidden>:
> 
> > We recently released Jacket v1.0 - a CUDA-based engine for MATLAB.
> > Jacket uses MEX to interface with MATLAB, so we're wondering if it
> > will work with Octave.  If there is any Octave developers interested
> > in trying this out, please email us and we'd be happy to arrange
> > this with you.
> > 
> 
> A Student of me is finishing his thesis about API for GPU based
> computing and its usage and implementation within Octave.
> 
> He is using CGIS (API of a team around University of Saarbruecken) and
> Brook/Brook+, finally. The intention not to use CUDA was the License
> conflict and because CUDA focuses on a specific hardware manufacturer.
> 
> The work will be released until March 2009 and distributed under terms
> of GPL. Please be patient... :)
> 
> Bye Andreas
> 


-- 
Student of Computer Science
Max Planck Institute for Human Cognitive and Brain Sciences
Department of Psychology
Stephanstr. 1a, 04103 Leipzig, Germany

brookoctave.txt
Description: Text document

signature.asc
Description: PGP signature