Converting parfor operations to gpuArray

I have a working parallel version of a code that does some likelihood calculations on a reasonably large matrix in parallel (using parfar) It is a trivially parallel operation as the calculation is performed column-wise & the parfor is employed to operate on the columns of data (one worker per column)

How could I achieve the same thing using a GPU (since the matrix is quite big & I have limited number of workers). All the operations are all GPU supported functions (matrix algebra ones like eig, diag & matrix multiplications only )

ie.,

data = 1000 by 200 (1000 rows by 200 cols matrix) 
[nrows, ncols] = size(data); 
parfor ix = 1:ncols
  workerData = data(:,ix);
  likelihood(ix) = funcCalcLikelihood(workerData, params);
end

This is fast enough. But i need to repeat such calculations many times so as to do a parameter sweep, so any speed increment will be good. Also, since my dataset is getting bigger (ncols = 1500 & I only have 144 max workers)

I have 2 Tesla (c2050) GPUS and was wondering if I could convert this into a gpuArray operation.

Thanks for your inputs.

[Edric Ellis]

Does it work simply to convert put 'data' on the gpu by calling gpuArray? gpuArrays can automatically flow in and out from PARFOR loops (providing a GPU is available on the client machine and the workers). Whether or not that is the most efficient thing you could do depends rather on the contents of 'funcCalcLikelihood'.

[nah]

Thanks +Edric Ellis or your comment. I didn't quite get what you mean by converting put data though. Calling gpuArray automatically slices the big matrix by columns you mean ?

[nah]

Any updates on this, anyone ?