How to vectorize the loops

Dear all,

Could anyone tell me how to vectorize the follow loops:

Best regards.

Yeping Sun

[Andrei Bobrov]

What is it - dv? Scalar or vector?

[Jan]

The EXP function is very expensive. It seems in your example like you apply it for a constant repeatedly. The inner loop overwrites the value of Hr(i,j) 5596 times, so you can easily get a speedup of factor 5596 if you omit the loop over k.

Is Hr pre-allocated before the loops?

Summary: It is hard to optimize code, which is shown partially only. Better post the real code together with some test data.

[Yeping Sun]

The code I mean should be:

for i=1:1:40
      for j=1:1:40
          for k=1:1:5596
               Hr(i,j)=sum(delta_2(k,i,j)*exp(dv(k));
          end
      end
end

delta is a (5596x40x40) matrix,dv is a 5596-element vector. The attached test.mat file contains these two entities.

What I want to do is to multiple an exponent factor (exp(dv)) to each element of the first dimension of the 3D matrix delta_2 and then sum them up, and the sums are assigned as the element of the result Hr matrix (40x40).

Now I have two questions: (1) it takes over 8000 s for the code to finish the calculation in my computer, which is too long; (2)the elements in dv are too large and exp(dv) cannot be properly calculated in matlab. Would you minding helping me solve these questions?

Warm regards.

Yeping Sun

[Andrei Bobrov]

Your code should be:

ev = exp(dv(:))';
Hr = zeros(40,40);
for ii = 1:40
    for jj = 1:40
        Hr(ii,jj) = ev*delta_2(:,ii,jj);
    end
end

Vectorize form see my answer in part 3.

Answer 1

[Andrei Bobrov]

if dv -scalar:

out = permute(sum(delta*exp(dv)),[2,3,1]);

if dv - vector:

out = permute(sum(...
     reshape(sum(bsxfun(@plus,reshape(delta,[],1),exp(dv(:)')),2),size(delta))...
                                                                        ),[2,3,1]);

part 3

ev = exp(dv)';
[~,n,k] = size(delta_2);
out = reshape(ev*reshape(delta_2,k,[]),n,[]);