Run a function using gpu inside a parfor loop or a for loop

I have a code that work perfectly on parfor, and it significantly improve the speed of the simulation. Then the next step for me is to try the process in the GPU. For example, I have a parfor loop or a for loop trying to execute 50 iteration, and the huge processing is done in a function "signalProcessController" to return the output. I would need to speed up the function execution using a gpu not sure how it is done or is this even possible? The only thing I found is the gpuArray.

%Sample code
iCellNum = 1;
activeUeNums = 1;
for i = 1 : 50
%      parfor ttiNum = 1 : 50    
        [timeDomainSignal{i}] = signalProcessController(iCellNum, i, activeUeNums);
end

Answer 1

[Walter Roberson]

See https://www.mathworks.com/help/parallel-computing/run-matlab-functions-on-multiple-gpus.html#MultiGPUExample-2

Notice in particular "Create a parallel pool with as many workers as GPUs available". Provided that you do not have more workers than GPUs then each worker will automatically be allocated a different GPU. Avoid manually selecting a GPU device -- or at least avoid selecting the same GPU on multiple workers, as the communications with the GPU has to be reset each time the device is selected.

It is possible to mix workers that use the GPU with workers that do not use the GPU, so that you can keep your other cores busy even though they have no access to GPU. You will need to figure out whether the worker has access to a GPU and if not then refrain from using gpu functions and variables.

[Yi Xien Yap]

My CPU has 4 cores meaning I have 4 workers, and I only have one GPU. If I understand you correctly, for example, I can have one GPU assign for worker one, and the rest of the 3 workers will process on CPU?

[Walter Roberson]

You would have to specifically program that.

IDs = fetchOutput(parfevalOnAll(@()getCurrentTask().Id,1));
ID_for_GPU = IDs(1);

parfor ttiNum = 1 : 50
    thisID = getCurrentTask().Id;
    if thisID == ID_for_GPU
        proceed with GPU code here
    else
        proceed with non-GPU code here
    end
end

the GPU code would use gpuArray and so on, and the non-GPU code would not.

In some cases you might be able to do something like

parfor ttiNum = 1 : 50
    thisID = getCurrentTask().Id;
    if thisID == ID_for_GPU
        initialize some variables as gpuArray
    else
        initialize variables non-gpuArray
    end
    common code that uses the variables
    if thisID == ID_for_GPU
      gather() results back into the workspace
    end
    common code that creates final outputs
end