How do I convert a for-Loop Into parfor-Loops?

Im trying to use parfor-Loop to speed up the computation. Here is the code (a kind of customized pdist function):

for i = 1:size1-1
    for j = i+1:size1
        index = index + 1;
        Y(index) = distance(dataSet(:,:,i),dataSet(:,:,j));
    end
end

I'm using parfor for the outer loop, but I got an error message on parfor:

• "The PARFOR cannot run due the way variable index is used"
• "The PARFOR cannot run due the way variable Y is used"

and, in the inner loop:

• "Valid indices for Y are restricted in PARFOR loops"

Both the errors occurred because of the use of the indexes in the inner loop, I can't figure out how to change the code for avoiding this problem. I've read the documentation, but i can't figured out how to convert into a parfor loop. How should I use parfor in this context?

Answer 1

[Edric Ellis]

The outputs from parfor loops need to be "sliced" - i.e. one of the subscripts must be the parfor loop index, and the others must be a fixed list of constant subscripts.

In your case, this is a bit tricky because of the structure of your nested loops. Here's one way - I'm converting the doubly-nested loop into a single loop by flattening out the indexing operations, and then using ind2sub to convert back from the flattened index to the co-ordinates.

% linIdxs starts out as a matrix of linear indices
linIdxs = reshape(1:(size1*size1), size1, size1);
% We need the lower triangle of these, and we need the transposed version
% to match the original code's ordering of subscripts
linIdxs = tril(linIdxs', -1);
% Discard the indices that were set to 0 by TRIL
linIdxs(linIdxs == 0) = [];

parfor idx = 1:numel(linIdxs)
    % Convert back from a linear index to [i,j] co-ordinate pair
    [i,j] = ind2sub([size1, size1], linIdxs(idx));
    Y(idx) = distance(dataSet(:,:,i),dataSet(:,:,j));
end

[giovanni grieco]

Thank you! This seem to really solve my problem but the parfor statement throw this error:

Error using parseDistAzInputs (line 46)
Incorrect latitude and longitude data matrices

What does it mean? and how could i avoiding the warning:

The entire array or struct 'dataSet' is a broadcast variable. This might result in unnecessary communication overhead.

[giovanni grieco]

Sorry! No error message, my fault. Thank you again!

[Edric Ellis]

This is simply an information message telling you that dataSet is a parfor broadcast variable. This means that the whole contents of dataSet are sent to each worker, and that is potentially damaging to performance. I think this is inevitable because of the way your loop is structured (sometimes it is possible to avoid "accidentally" broadcasting data, but this is not one of those cases).