[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: GPU computing for octave
From: |
Xavier Delacour |
Subject: |
Re: GPU computing for octave |
Date: |
Fri, 14 Mar 2008 11:57:47 -0400 |
A more interesting (and more ambitious) approach might be to generate
bindings for cuda/brook/sh frameworks themselves, such that users
could compile/prepare and then invoke a cg routine from within Octave.
eg,
mkcudafile("cuda_fft.cu");
cuda_fft(...);
There is definitely value in parallelizing common routines like FFT,
GEMM, etc. But quite a bit of research depends on custom cg coding.
Lowering the complexity barrier to integrate that custom work into
Octave would be useful I think.
Xavier
On Fri, Mar 14, 2008 at 9:55 AM, David Bateman
<address@hidden> wrote:
> Ronny Lindner wrote:
> > hello,
> >
> > I recently started my diploma thesis on GPU computing and I want to
> > develop plugins for some octave functions so that they run on the
> > graphics card. It might be possible to achieve a speedup of up to 20!
> >
> > I know there is a CUDA (GPU computing) plugin for matlab, but it only
> > works on nvidia cards and my aim is to develop an octave plugin for
> > amd+nvidia, windows+linux.
> > Has anyone tried something similar before?
>
> However, wouldn't trying to use the CUDA interface with Octave already
> be a good start.. Octave can use matlab mex file in source code form and
> CUDA mex files are supplied with their source.. Also it seems that the
> NVIDIA GPU works on C99 complex values rather than their real and
> imaginary parts as matlab doe, so there will be speed improvements if
> these files were converted directly to oct-files
>
> >
> > Do you know any functions that are good candidates for parallelization?
> What do you mean? I'd probably wouldn't base an answer on what is most
> easily parallelized, but rather what function being parallelized will
> give the most benefit to my (or rather your) programs.. I'd probably
> target a xGEMM blas replacement and an FFT as the first targets as the
> most useful code to have running fast.
>
> D.
>
> --
> David Bateman address@hidden
> Motorola Labs - Paris +33 1 69 35 48 04 (Ph)
> Parc Les Algorithmes, Commune de St Aubin +33 6 72 01 06 33 (Mob)
> 91193 Gif-Sur-Yvette FRANCE +33 1 69 35 77 01 (Fax)
>
> The information contained in this communication has been classified as:
>
> [x] General Business Information
> [ ] Motorola Internal Use Only
> [ ] Motorola Confidential Proprietary
>
>