Octave and CUDA/GPGPU programming

[John Swensen]

I was just wondering if anyone has tried compiling the mex file provided by NVidia for doing 2D FFTs using a video card GPU?  http://developer.nvidia.com/object/matlab_cuda.html
If there are any in the Octave community familiar with CUDA and GPGPU programming, I can think of a few other operations that would benefit greatly (e.g. conv and conv2, sum, others???)

John Swensen

[John W. Eaton]

On 11-Sep-2008, John Swensen wrote:

| I was just wondering if anyone has tried compiling the mex file  
| provided by NVidia for doing 2D FFTs using a video card GPU?  http://developer.nvidia.com/object/matlab_cuda.html
| If there are any in the Octave community familiar with CUDA and GPGPU  
| programming, I can think of a few other operations that would benefit  
| greatly (e.g. conv and conv2, sum, others???)

Maybe you should ask the help list since that is read by more people.

OTOH, is underlying CUDA toolkit free software?  The Octave lists are
probably not the best place to promote using non-free software...

jwe

[isaacgerg]

They are indeed free.

[John W. Eaton]

On 11-Sep-2008, isaacgerg wrote:

| They are indeed free.

This page: http://www.nvidia.com/object/cuda_get.html, says "CUDA
tools are provided free of charge" but that is not the same as what I
meant by "free software".

Looking at the detailed license terms, I see

  Source Code: Developer shall have the right to modify and create
  derivative works with the Source Code.  [...]

and that's encouraging, but then I also see:

  Object Code: Developer agrees not to disassemble, decompile or reverse
  engineer the Object Code versions of any of the Materials.  Developer
  acknowledges that certain of the Materials provided in Object Code
  version may contain third party components that may be subject to
  restrictions, and expressly agrees not to attempt to modify or
  distribute such Materials without first receiving consent from NVIDIA.

so are there binary-only blobs that are required and that are not
free (as in speech) software?

jwe

[John Swensen]

On Sep 11, 2008, at 4:22 PM, John W. Eaton wrote:

> On 11-Sep-2008, isaacgerg wrote:
>
> | They are indeed free.
>
> This page: http://www.nvidia.com/object/cuda_get.html, says "CUDA
> tools are provided free of charge" but that is not the same as what I
> meant by "free software".
>
> Looking at the detailed license terms, I see
>
>  Source Code: Developer shall have the right to modify and create
>  derivative works with the Source Code.  [...]
>
> and that's encouraging, but then I also see:
>
>  Object Code: Developer agrees not to disassemble, decompile or  
> reverse
>  engineer the Object Code versions of any of the Materials.  Developer
>  acknowledges that certain of the Materials provided in Object Code
>  version may contain third party components that may be subject to
>  restrictions, and expressly agrees not to attempt to modify or
>  distribute such Materials without first receiving consent from  
> NVIDIA.
>
> so are there binary-only blobs that are required and that are not
> free (as in speech) software?
>
> jwe

Yeah, sorry about mentioning it.  In my enthusiasm about fancy new  
things that are on the surface "free of charge", I often overlook the  
limitations of not being "free software".  This would be more  
appropriate for the nonfree section of OctaveForge.

John Swensen

[Andreas Romeyke-3]

Hello,

Am Thu, 11 Sep 2008 14:09:33 -0400
schrieb John Swensen <jpswensen@...>:

> I was just wondering if anyone has tried compiling the mex file  
> provided by NVidia for doing 2D FFTs using a video card GPU?
> http://developer.nvidia.com/object/matlab_cuda.html If there are any
> in the Octave community familiar with CUDA and GPGPU programming, I
> can think of a few other operations that would benefit greatly (e.g.
> conv and conv2, sum, others???)

There is a student supported by me who is developing Octave-plugins
using Brook or CGIS (a GPGPU project of University of Saarbrücken, I
think).

He decided not to use Nvidia's CUDA because license issues. There will
be problems with Brook too, because the Brook development ist still
stalled. The original developers are angry about AMD, because AMD's
Brook+ (extensions) developers does not return code back.

CGIS is a general GPGPU project and will be released under a free
license.

My  student will finished his work between 12-2008 and 1-2009. Please
be patient... :)

Bye Andreas

PS.: If you have algorithms which would benefit from parallel
architectures of GPUs, please let us know. But there are some
limitations. The transfer between GPU and CPU is expensive, random
memory access will be expensive, too. The maximal amount of data is
limited and the precision could be restricted to single float on older
cards. The advantages of GPU programming found in existing literature
is outdated by the progress in CPU development and therefore you should
rethink about your problems.
--
Software Developer / Dipl. Inform. (FH)
Max Planck Institute for Human Cognitive and Brain Sciences
Department of Psychology
Stephanstr. 1a, 04103 Leipzig, Germany