How to parallel this code?

function d = compute_d(vertex)

%vertex is nvert by 3 matrix;
nvert= max(size(vertex));
d = [];
for i = 1:nvert-1
    for j = i+1:nvert
        x_ij = vertex(i,:) - vertex(j,:);
        d = [d; x_ij/norm(x_ij)];
    end
end
save('d', 'd');

[tao jiang]

I try to use the code below to parallel, but it said the variable isn't sliced. How should I do to deal with this problem? thx

 function d = compute_d_pal(vertex)
  %vertex = cat(2,vertex1(:,1),vertex1(:,2),zeros(size(vertex1,1),1));
  nvert= max(size(vertex));
  n = nvert*(nvert-1)/2;
  d = zeros(nvert*(nvert-1)/2,3);

parfor i = 1:nvert-1
    for j = i+1:nvert
        x_ij = vertex(i,:) - vertex(j,:);
        %d = [d; x_ij/norm(x_ij)];
        index = nvert*(nvert-1)/2 - (nvert-i)*(nvert-i+1)/2 + j - i;
        d( index,: ) = x_ij/norm(x_ij);

      end
  end
  save('d', 'd');

[Walter Roberson]

The object your store into, d, cannot have an index that complicated. You should switch to a linear index that you then break up into i and j.

Answer 1

[Walter Roberson]

The biggest speed up would be to pre-allocate d. It will run (nvert) * (nvert+1) / 2 times, it appears, each size(vertex,2) rows, so you can figure out the needed memory ahead of time, and store into the right place in the matrix.

You can use a linear index instead of a double-nested loop, decoding the linear index into the proper i/j pair. And that allows you to use a something-by-3 output matrix with its first index being the linear index. If you had not sped the code up enough already by that point, you could parfor that linear index -- but I suspect that would end up slowing things down.