How to use multiple GPUs asynchronously

I have two GPUs connected to my computer. I've explored on Mathworks, but can't find an easy way to use both GPUs in my code (there seems to be a 'parfor' and 'spmd' method, but neither seem to work for me). I would like to have each GPU run asynchronously as in:

x = gpuArray();

y1 = gpuFunction1(x);
y2 = gpuFunction2(x);

where gpuFunction1 runs on the first GPU and gpuFunction2 runs on the second GPU. Is there an easy way to do this? Thank you.

[Srinidhi Ganeshan]

https://www.mathworks.com/matlabcentral/answers/162421-how-to-use-multiple-gpus-asynchronously#answer_158769

could you check the comment below and could you confirm if the below answer as stated by @edric ellis is right ?

Answer 1

[Edric Ellis]

If you have 2 GPUs, you could use SPMD like this:

parpool('local', 2); % ensure the pool is of size 2
spmd
  % set each worker to use a GPU device based
  % on their "labindex"
  gpuDevice(labindex);
  x = gpuArray(...);
  if labindex == 1
    y1 = gpuFunction1(x);
  elseif labindex == 2
    y2 = gpuFunction2(x);
  end
end
% Note that y1 and y2 are 'Composite', so we must
% use indexing to retrieve the values
y1 = y1{1};
y2 = y2{2};

Note that the SPMD block approach shown is still synchronous with respect to the client. If you want the GPU work to be asynchronous with respect to the client, you need to use either batch (without a parallel pool open), or parfeval (with a parallel pool open).

Here's how you might use parfeval:

% setup as before
parpool('local', 2);
spmd, gpuDevice(labindex); end

% now invoke asynchronous work
x = gpuArray(...);
% Here the "1" is the number of outputs from gpuFunction#
f1 = parfeval(@gpuFunction1, 1, x);
f2 = parfeval(@gpuFunction2, 1, x);

% The MATLAB client is now idle, you can check to see if the
% work is done by examining the state:
disp(f1.State)

% Or, you can wait for the results - "fetchOutputs" blocks
% until the results are ready.
result1 = fetchOutputs(f1);
result2 = fetchOutputs(f2);

[Jonathan]

When using parfeval in this way, are f1 and f2 gpuArrays or regular variables stored in system memory? Is it possible to have gpuArrays residing on different GPUs accessible to the client?

[Edric Ellis]

In the example above, f1 and f2 are parallel.Future objects returned by parfeval. The results that they return from fetchOutputs can be gpuArray if you wish (i.e. if that's what gpuFunction1 returns), or you could arrange for your functions to gather things before returning.

[Srinidhi Ganeshan]

% setup as before
parpool('local', 2);
spmd, gpuDevice(labindex); disp(gpuDevice); end
% now invoke asynchronous work
x = gpuArray(...);
    disp(gpuDevice);
% Here the "1" is the number of outputs from gpuFunction#
f1 = parfeval(@gpuFunction1, 1, x);
    disp(gpuDevice);
f2 = parfeval(@gpuFunction2, 1, x);
   disp(gpuDevice);
% The MATLAB client is now idle, you can check to see if the
% work is done by examining the state:

% Or, you can wait for the results - "fetchOutputs" blocks
% until the results are ready.
result1 = fetchOutputs(f1);
result2 = fetchOutputs(f2);

Inside spmd, I tried checking gpuDevice, it shows labindex=1,labindex 2, which is gpuDevice1 and gpuDevice2. But when I try, printing gpuDevice inbetween the lines to check if both gpuDevices are working, it always shows 2, which means gpuDevice 2 is only ON throughout the code and gpuDevice is idle. Why is that gpuDevice 1 is idle ? @Edric Ellis

[Andrea Picciau]

There's a problem with your code. It should be:

spmd, currGpu = gpuDevice(labindex); disp(currGpu); end