Using "parfor" in non constant bounds [Parallel Computing]

Hi all!

Recently I started working with Matlab, and until now everything went fine. Because I'm working with giantic files, my run times have gotten out of hand. Luckily I got a lot of processors & RAM at my disposal.

At the moment I'm trying to use the "Parallel Toolbox", but I'm running into trouble. Figuring out that the function parfor could be useful, I found out that the bounds within the loops should be constant.

I need help converting my code in something suitable for parallising. To clearify this, below my code is given. The imporant part is the outer for loop: snap=1:size(A_tensor,3).

Basically I think these loops could run independantly, but still Matlab is having problems. Any suggestions what I should change to be able to parallise my code?

Thanks a bunch!

%%%Step 4: Make adjency matrices depending on snapshot size
n =length(Unique_elements);
adj = sptensor([n n size(A_tensor,3)]);
%adj = zeros(n, n, size(A_tensor,3), 'logical');
tic

for snap=1:size(A_tensor,3) %  snap= snapshots in timeframe specified
    for i=1:length(Unique_elements) % Go over all callers
        index_caller_out = find(A_tensor(:,1,snap)==Unique_elements(i)); %Indexen of calls started
        index_caller_in = find(A_tensor(:,2,snap)==Unique_elements(i));  %Indexen of calls received

          for j=1:length(index_caller_out)
              %J=find(Unique_elements==M.b_number(index_caller_out(j)));
              J=find(Unique_elements==A_tensor(index_caller_out(j),2,snap));
              adj(i,J,snap) = 1;
          end    
          for q=1:length(index_caller_in)
              %Q=find(Unique_elements==M.a_number(index_caller_in(q)));
              Q=find(Unique_elements==A_tensor(index_caller_in(q),1,snap));
              adj(i,Q,snap) = 1 ;
          end
      end
  end

Answer 1

[Walter Roberson]

Using size() as an upper bound is not a problem for parfor.

However, instead of the multiple references to A_tensor(?, ?, snap) you should take a local copy of A_tensor(:,:,snap) and use local indexing within that, writing into a local 2D copy of adj. After your j and q loops have built the local 2D copy of adj, write the local 2D copy into adj(:,:,snap) as a single step.