Speed up Matlab code without changing it?

[yjiang]

Hi, I came across a statement in Tesla Technical Brief regarding speeding up Matlab matrix computation with CUDA without changing Matlab code.

It says “… MATLAB applications can be accelerated by the NVIDIA GPU using two methods. The first method does not require changes to the MATLAB code. By simply plugging in the CUDA FFT libraries underneath the MATLAB application, any calls to FFT or BLAS functions are simply executed on the GPU. …”

I like to try it out but could not find information on how to do it. Has anyone had experience? Please help.

Thanks.

[cascadingbits]

Hi, I came across a statement in Tesla Technical Brief regarding speeding up Matlab matrix computation with CUDA without changing Matlab code.

[…]

I like to try it out but could not find information on how to do it.  Has anyone had experience? Please help.

[snapback]392599[/snapback]

I don’t have experience with it yet, but I put “nvidia cuda matlab” into google and got this: MATLAB plug-in for CUDA.

When MATLAB makes 2D FFT calls of any size, the NVIDIA plug-in intercepts them and handles them with a MEX file that in-turn utilizes an optimized CUDA FFT implementation on the GPU.

There are downloads for Linux and Windows.

Have fun.

[melonakos]

Hi yjiang,

We have just released a free beta version of Jacket: A GPU Engine for MATLAB. It calls the CUDA functions underneath the hood in a highly optimized runtime system, including optimized memory transfers and a compile on-the-fly systems to increase performance.

Jacket overloads standard MATLAB functions so that you can get the transparency you are looking for. All you have to do is cast data to ‘gsingle’ data types.

Anyway hope this helps. You can grab the free version from:

http://www.accelereyes.com

Best,

John Melonakos