Conditional array accumulation inside parfor

I have a situation were I am testing a condition inside a parfor loop, and if true append the results of a computation to an array. A simplified example is as follows

ary = [];
parfor n=1:N
    for m = 1:M
        if (f(m,n)>0) % do some test, this is not easily vectorizable
            ary = [ary; n m];
        end
    end
end

I would like, however, to avoid growing arrays in the loop.

I could estimate an upperbound for the size of ary and try to do it this way,

ary = zeros(ubound,2);
ind = 0;
parfor n=1:N
    for m = 1:M
        if (f(m,n)>0) % do some test, this is not easily vectorizable
            ind = ind + 1;
            ary(ind,:) = [n m]; % such indexing will not work within parfor
        end
    end
end

but that wouldn't work as shown in the comment.

Another idea I had was using a logical array to keep track of the conditional result.

condary = false(N*M);
for k = 1:N*M % flatten the loop
    % get n and m from k; k = (n-1)*M+m, therefore
    m = mod(k,M); if m == 0, m = M; end
    n = (k-m)/M+1;
    if (f(m,n)>0)
        condary(k) = true;
    end
end

The desired array, ary, can then be back-constructed from the logical array in a second loop. In fact, ary, can be preallocated at this point. Or the operations meant to be performed using ary can be performed based on condary in a second loop. But this involves flattening the loop.

I was wondering if there are any better ways to do this.

Answer 1

[Matt J]

map = false(N,M);

parfor k=1:M*N

        [n,m]=ind2sub([N,M],k);
    
        map(k) = ( f(m,n)>0 );

end

[I,J]=find(map);

ary=[I,J];

[Siva]

Matt,

n and m are temporary variables, and can't be used to index arrays inside parfor. When I tried to run it in MATLAB, I got "Valid indices for map are restricted in parfor loops".

I had forgotten about ind2sub, that is better than what I had done.

[Matt J]

Sorry. I fixed it.

        map(k) = ( f(m,n)>0 );

[Siva]

That is clever, to use the single absolute index!

In reference to my question, I am still wondering, might there be a way to build ary directly in the parfor loop, avoiding the logical array altogether?

[Matt J]

I thiink the logical array is the best way, but here is a way to do direct accumulation:

ary=cell(N*M,1);

parfor k=1:M*N

        [n,m]=ind2sub([N,M],k);
    
        if (f(m,n)>0) % do some test, this is not easily vectorizable
            ary{k}=[n,m];
        end

end

ary=cell2mat(ary);

[Siva]

That is pretty cool too. Thank you!

Any implications in terms of the relative memory usage of the two approaches?

[Matt J]

The logical array approach stores memory contiguously, which is one reason I like it better.