Code Optimization, for loop

Is it possible to make this code faster?

%
RevPSD = zeros(size(X));
for m = Tmin + 1:size(X,2)
  for i = Tmin:Tmax
      if m - i <= 0
          break
      else
          RevPSD(:,m) = RevPSD(:,m) + exp(-2*delta*i*hop/fs)*(Xpsd(:,m-i));
      end
  end
end

[Christoph F.]

Two suggestions:

The inner loop only loops from Tmin to (m-1). The extra comparison of m and i inside the loop could be removed if the loop condition is changed, e.g.

for i = Tmin:(m-1)
  RevPSD(:,m) = RevPSD(:,m) + exp(-2*delta*i*hop/fs)*(Xpsd(:,m-i));
end

The values of the term

exp(-2*delta*i*hop/fs)

only depend on i can could be pre-calculated for every possibly value of i outside the loop. e.g.

exptable = exp(-2*delta*(Tmin:Tmax)*hop/fs);
for m = Tmin + 1:size(X,2)
  for i = Tmin:(m-1)
    RevPSD(:,m) = RevPSD(:,m) + exptable(i-Tmin+1)*(Xpsd(:,m-i));
  end
end

[Jan]

@Christoph F: Please post this in the answer sections.

Answer 1

[Christoph F.]

Two suggestions:

The inner loop only loops from Tmin to (m-1). The extra comparison of m and i inside the loop could be removed if the loop condition is changed, e.g.

for i = Tmin:(m-1)
  RevPSD(:,m) = RevPSD(:,m) + exp(-2*delta*i*hop/fs)*(Xpsd(:,m-i));
end

The values of the term

exp(-2*delta*i*hop/fs)

only depend on i can could be pre-calculated for every possibly value of i outside the loop. e.g.

exptable = exp(-2*delta*(Tmin:Tmax)*hop/fs);
for m = Tmin + 1:size(X,2)
  for i = Tmin:(m-1)
    RevPSD(:,m) = RevPSD(:,m) + exptable(i-Tmin+1)*(Xpsd(:,m-i));
  end
end

[Yusuke Kadowaki]

Thanks a lot Christoph. The second one worked a bit.

[Christoph F.]

Looking at the expression more closely, the whole term

exp(-2*delta*i*hop/fs)*(Xpsd(:,m-i))

only depends on m and i. I assume that with some creative thinking, the inner loop could be completely removed and replaced with adding a matrix to RevPSD.

The general idea is to use built-in MatLab matrix/vector functions (add, multiply, sum, etc) instead of loops whenever possible, and not to calculate expressions repeatedly that only need to be calculated once.