For loop in parfor loop

Dear all, I am trying to run the following script

a = linspace(0,10);
parfor ii=1:2
   for jj=1:length(a)
      x(ii,jj) = rand;
   end
end

which fails with an "Error: The variable x in a parfor cannot be classified." Is this a bug in Matlab? I would simply consider the vector "a" a broadcast variable, since "a" and "length(a)" are unmodified by the script.

How would you recommend to fix the script, taking into account that length(a) should be evaluated inside the parfor loop?

Thanks, Niels

Answer 1

[Greig]

I don't know the details (perhaps another answer can get technical), but parfor loops don't like assigning values to things like x(ii, jj) directly, so you have to use temporary variables that are accessed only in the parfor loop. Try something like this....

a = linspace(0,10);
x=NaN(2,length(a)); % pre-allocated for speed

parfor ii=1:2

      tmp_var=NaN(1,length(a)); % pre-allocate a temporary variable

      for jj=1:length(a)
          tmp_var(jj) = rand;
      end

      x(ii,:) = tmp_var; % assign it to x

end

Answer 2

[Edric Ellis]

The limitation you're hitting here is that any for loop nested inside a parfor loop must have constant bounds to allow slicing. In your case, the bounds are constant, so you can simply write:

a = linspace(0,10);
n = length(a); % extract length outside PARFOR
parfor ii=1:2
   for jj=1:n % constant bounds loop inside PARFOR
      x(ii,jj) = rand;
   end
end

This limitation is documented here.

In the case where your inner for loop has non-constant bounds, then @Greig's solution is the way to go.

[Niels]

Great, thank you both for your replies, I will follow your suggestions! :-) Best regards, Niels