What's wrong with this parfor loop?

When I try to run the following parfor loop, Matlab returns the error claiming that A cannot be classified.

N = 10;
T = 3;
A = zeros(N,T);

parfor i = 1:1:N
    for t = 1:1:T
        A(i,t) = rand(1,1);
    end
end

Why isn't A considered to be a sliced output variable? The following modification works just fine.

N = 10;
T = 3;
A = zeros(N,T);

parfor i = 1:1:N
    A(i,1) = rand(1,1);
end

So it seems to be something to do with the inner loop over t. From reading the help document on sliced variables, I don't understand why this inner loop would make a difference.

[even coil]

Is there a workaround besides using the vectorized version of rand? (Doing the analgous would be difficult in the application I have in mind.)

[Matt J]

If you interchange the order of the loops you've shown, it will run. Whether that's the best solution is hard to tell, though, without seeing an example closer to the application you have in mind.

Answer 1

[Edric Ellis]

In this case, the fix is simply to remove the increment from the inner for loop, like so:

N = 10;
T = 3;
A = zeros(N,T);
parfor i = 1:1:N
    for t = 1:T
        A(i,t) = rand(1,1);
    end
end

(I'm not sure why this limitation exists, it doesn't appear to be explicitly mentioned in the documentation)

[Matt J]

(I'm not sure why this limitation exists, it doesn't appear to be explicitly mentioned in the documentation)

Maybe for the same reasons that this isn't supported:

parfor i = 1:1:N
  for t = 1:i:T
      A(i,t) = rand(1,1);
  end
end

If A isn't pre-allocated, I can see parfor having a hard time figuring out what size(A) should ultimately be.

[even coil]

Why isn't this considered a bug? Isn't "1:T" just shorthand for "1:1:T"?

[Matt J]

Hard to say if it can be called a bug because it's not yet clear what level of support is intended for nested loops that interact with sliced variables. Note, for example, that the following also produces a classification error,

    parfor i = 1:1:N
        for t = 1:2:T
            A(i,t) = rand(1,1);
        end
    end

Answer 2

[Walter Roberson]

N = 10;
T = 3;
A = zeros(N,T);

parfor i = 1:1:N
    v = zeros(1,T);
    for t = 1:1:T
        v(1,t) = rand(1,1);
    end
    A(i, :) = v;
end

[even coil]

Thank you, this is a good workaround.