Matlab - highly parallel Tasks in CPU workers or GPU Processors

Question

Or how to convert Parfor loops to gpuArray operations ?

Problem at hand : A column-wise operation on large data matrix distributed to multiple CPUs (workers) via parfor.

data = 1000 x 200 matrix
[nrows, ncols] = size(data)
parfor ix = 1:ncols
     workerData = data(:,ix);
     colwiseResult(ix) = function(workerData,params);
end

Now, how could this be efficiently made to utilize GPUs ? especially since the problem needs to scale up to 1000x1000 matrices (& more )

dataGPU = gpuArray(data);

But I don't find an easier way of doing column wise operations on GPU. Functions such as arrayfun or bsxfun operate element-by-element which is not what am interested in. Since these are trivially parallel tasks, exploiting multiple processors in the GPUs could be ideal (& avoids bothering with the parfor etc., )

(Well actually, am performing a likelihood calculation that does matrix diagonalization, exponents & products progressively over each data point on the column vector using a for loop inside the Workers). Have checked that all these ops have gpu overloading )

C.Colden C.Colden · Accepted Answer · 2013-08-05T14:22:19

Have you watch this tutorial?

http://www.mathworks.ch/videos/introduction-to-gpu-computing-with-matlab-68770.html

Please check first whether you can use your GPU, since not all GPUs are supported. Afterwards you can pass variables to your GPU or you can include GPUs into your parfor loop like this:

parfor ix = 1:10
gd=gpuDevice;
end

In the end of day using one GPU is not a big advantage. Please give some more details on your workstation for further information.

Matlab - highly parallel Tasks in CPU workers or GPU Processors

1 Answers