access to a gpuArray from a MEX file?

I did a lot of work with CUDA MEX files before gpuArray was an option. Now I would like to simplify my interface to these MEX files by allowing them to access data that the user has stored in gpuArrays -- but I do not want to rewrite all my code.

So: what I need is a way for my MEX code to access a gpuArray's internal pointer to GPU memory, and other metadata. Is there a way to do this?

I suppose I would also need access to the Parallel Computing Toolbox's CUDA context for this to work.

Answer 1

[Edric Ellis]

You're right that you need Parallel Computing Toolbox to access gpuArray functionality. We don't offer a MEX interface with gpuArrays, but we do offer the CUDAKernel which allows you to use gpuArrays together with hand-written CUDA code. Here's the doc for CUDAKernel.

We think that CUDAKernel is easier to use than a MEX layer because you only need to write the "__global__" entry points for CUDA, you don't need to write any of the code to marshal data in and out of mxArrays.

[Jim Mutch]

Thanks for the reply. It may indeed be easier for new code development, but not necessarily for people with existing CUDA MEX files.

[King Fisher]

Edric,

I ve tested the MATLAB .cu code and generate a .ptx file for it. But when I execute it on matlab the following error occurs:

??? Error using ==> iParseToken at 266

Unsupported type in argument specification "fahad.cu".

Error in ==>

/usr/local/Matlab/2011a/toolbox/distcomp/gpu/+parallel/+internal/+gpu/handleKernelArgs.p>iParseCPrototype at 192

Error in ==>

/usr/local/Matlab/2011a/toolbox/distcomp/gpu/+parallel/+internal/+gpu/handleKernelArgs.p>handleKernelArgs at 79

To let you clear that Gpu is funtioning well in my Matlab environment for other functions ,example, gpuarray().

Would you please figure out the problem !

[Walter Roberson]

King Fisher, I recommend that you start a new Question for this.

[Daniel Armyr]

With the release of 2013a, though, you can do what the original poster wanted. See this: http://www.mathworks.se/help/distcomp/create-and-run-mex-files-containing-cuda-code.html

Answer 2

[John Melonakos]

Jacket users have access to the Jacket SDK that has advantages over the other two approaches: http://blog.accelereyes.com/blog/2010/10/29/jacket_sdk_trumps_mex/

[Jim Mutch]

Perhaps, but in my question I said "I do not want to rewrite all my code". So this is off topic.

Do you guys need to pop up in every GPU-related thread?