CUDA kernels from Matlab Launching CUDA kernels from matlab
Hi,

So I am working on this large computation project. My departments GPU machine has Matlab with version as Version: 7.14.0.739 (R2012a).

This machine has Parallel Computing Toolbox installed.Few months ago,I read about this nvmex method by which you can launch cuda kernels from matlab.

But is this method still supported in the version of matlab that I am using?

If not then, should I use parallel computing toolbox methods for launching custom kernels ?

If yes, where can I find the comprehensive guide to install the nvmex on linux gpu machine?

I have checked CUDA Zone and MAthworks site but I just couldnt get the answers for nvmex.

Thanks in advance.

Cheers
Hi,



So I am working on this large computation project. My departments GPU machine has Matlab with version as Version: 7.14.0.739 (R2012a).



This machine has Parallel Computing Toolbox installed.Few months ago,I read about this nvmex method by which you can launch cuda kernels from matlab.



But is this method still supported in the version of matlab that I am using?



If not then, should I use parallel computing toolbox methods for launching custom kernels ?



If yes, where can I find the comprehensive guide to install the nvmex on linux gpu machine?



I have checked CUDA Zone and MAthworks site but I just couldnt get the answers for nvmex.



Thanks in advance.



Cheers

#1
Posted 04/27/2012 06:16 PM   
Here is the latest state of things:

1) NVMEX is not actively maintained by NVIDIA since Jacket first went commercial in 2009. However, if you hack at those old scripts, you can get them to work in the newer versions of MATLAB.
2) Parallel Computing Toolbox is slow, unstable, and not very functional: http://accelereyes.com/compare
3) Jacket is available and the [url="http://blog.accelereyes.com/blog/2010/10/29/jacket_sdk_trumps_mex/"]Jacket SDK[/url] is a great way to integrate your own custom CUDA code (that's how most people are doing CUDA integration into MATLAB today).

I'm happy to provide you with a free trial to #3 if that's interesting. Otherwise, you might try your hand at hacking to get #1 to work.

Good luck!
Here is the latest state of things:



1) NVMEX is not actively maintained by NVIDIA since Jacket first went commercial in 2009. However, if you hack at those old scripts, you can get them to work in the newer versions of MATLAB.

2) Parallel Computing Toolbox is slow, unstable, and not very functional: http://accelereyes.com/compare

3) Jacket is available and the Jacket SDK is a great way to integrate your own custom CUDA code (that's how most people are doing CUDA integration into MATLAB today).



I'm happy to provide you with a free trial to #3 if that's interesting. Otherwise, you might try your hand at hacking to get #1 to work.



Good luck!

John Melonakos ([email="john.melonakos@accelereyes.com"]john.melonakos@accelereyes.com[/email])

#2
Posted 04/27/2012 06:24 PM   
Scroll To Top