Matlab -- how to convert this code so that it can run using parfor?

Question

I have the following code in Matlab (simplified, of course, but this shows the root of the problem), which used for previously; I want to change it to use parfor instead:

isForeground = ones(1, size(In, 1));

% this used to be a simple for
parfor i=1:X
   data = output_of_some_expensive_function();

   % do some other stuff with data here

   % the part below is the problem; isForeground basically keeps track of all the places where data was never a zero
   isForeground(data == 0) = 0; 
end

Matlab is complaining saying that Valid indices for 'isForeground' are restricted in PARFOR loops. Is there a way to get around this problem? I thought about just saving all the data outputs, and then running a separate traditional for loop where I would do the isForeground part, but the problem is that X is very large and saving all the data outputs would be very memory-intensive.

Is there another way of doing this?

Answer 1

How dense are the 1s in the results? If they're sparse (i.e. mostly 0), find() could shrink the logical masks down to numeric index lists. If you need further compression, depending on the size of X, you could convert the stored masks to uint32 or uint16, which would take 1/2 or 1/4 the space of double.

Then stash the results in a temporary variable that is still sliced by the loop index, and then combine them outside the parfor loop, like you say.

masks = cell(1, X);
parfor i=1:X
   data = output_of_some_expensive_function();

   % do some other stuff with data here

   % stash the loop results in a non-combining manner
   masks{i} = uint32(find(data == 0));
end

% Combine in regular for loop

for i=1:X
   isForeground(double(masks{i})) = 0; 
end

If that's not enough, maybe you could write each result mask out to disk in a .mat file on a commonly-visible drive space or other persistent object store. If the workers for your parfor are on the local machine, any drive will work, and should be fast because you may just go through your OS's disk cache.

You could also try transforming the output mask to a compressible form. E.g. run diff() on it and zip the results in to a byte[] "blob", and then do the inverse right before applying the mask in the regular for loop. Depending on the structure of your masks, this may save additional space.

Note: logicals take more storage space than you might think; even though each element only has one bit of meaningful information, it takes a whole byte of memory, because that's the smallest addressable unit of memory.

Answer 2

This code now uses struct to pass the data out of parfor:

In = eye(3); % some input
X = size(In, 1);

isForeground = ones(1, X);
isForeground_struct = repmat(struct('data', ones(1, X)), [1, X]);

% is used to be a simple for
parfor i=1:X
   data = ones(1, X);
   data(randi(X)) = 0; % a random operator just to make an example for data

   isForeground_struct(i).data = data; 
end

isForeground = prod(cat(1, isForeground_struct.data)); % assuming you want AND combination