Vectorize this?

yidx = [false(18144,1);true(24,1);false(36882,1);true(64,1);false(6,1);true(198,1)];
temp = find(yidx);
for i=1:1:size(temp,1)
    i
    for j = 1:1:size(cplex.Model.A,1)
        if cplex.Model.A(j,temp(i)) ~= 0
            temp_row = cplex.Model.A(j,:);
            temp2 = find(temp_row ~= 0);
            temp3 = find(temp2 == temp(i));
            temp2(temp3) = [];
            bin_coefficient = temp_row(1,temp(i));
            for k = 1:1:size(temp2,1)
                var = [j,temp2(k)];
                b = cplex.Model.rhs(j,1) - bin_coefficient;
                u = cplex.Model.rhs(j,1);
                new_row = sparse(1,size(cplex.Model.A,2));
                new_row(1,temp2(k)) = 1;
                new_row(1,temp(i)) = (u-b);
                cplex.Model.rhs = [cplex.Model.rhs;u];
                cplex.Model.lhs = [cplex.Model.lhs;-Inf];
                cplex.Model.A = [cplex.Model.A;new_row];

              end
          end
      end
  end

Answer 1

[Stefan]

Hi Edward,

which structure has your cplex.Model? Without this information it's more difficult to understand the vectoroizing problem.

greetz

Answer 2

[Jan]

Do you really need a vectorization or do you want to accelerate the function only?

For the later, a pre-allocation is strongly recommended. For some fields this is even identical to the "vectorizatzion":

yidx = [false(18144,1); true(24,1); false(36882,1); true(64,1); false(6,1); true(198,1)];
temp = find(yidx);
A    = cplex.Model.A;  % Avoid repeated accessing of substructs
for i = 1:size(temp, 1)
  for j = 1:size(A, 1)
      if A(j,temp(i)) ~= 0
          temp_row = A(j,:);
          temp2 = find(temp_row ~= 0);
          temp2(temp2 == temp(i)) = [];
          bin_coefficient = temp_row(1,temp(i));
          cplex.Model.lhs = [cplex.Model.lhs, repmat(-Inf, size(temp2,1), 1)];

            % b and u do not depend on k loop, so move it outside for speed:
            b = cplex.Model.rhs(j,1) - bin_coefficient;
            u = cplex.Model.rhs(j,1);
            cplex.Model.rhs = [cplex.Model.rhs; repmat(u, size(temp2,1), 1)];
            for k = 1:size(temp2, 1)
                % Not used, do not shadow the built-in VAR!
                % var = [j, temp2(k)];
                new_row = sparse(1,size(A,2));
                new_row(1, temp2(k)) = 1;
                new_row(1, temp(i))  = (u-b);
                A = [A; new_row];
              end
          end
      end
  end
  cplex.Model.A = A;

Is this faster already?

Answer 3

[Stefan]

I think you can pre-allocate teh varibale new_row also (command 'ones()' )

Variable A is changig it's size in every loop. To accelerate pre-allocate A and use zour index k. Afterwards you can connect the single values with combine or cell2mat