Why is for-loop faster than range index?

I am very confused by the performance difference. fun_A is much faster than fun_B. Anybody has any ideas as to why? I am running R2013a. (I looked into a similar question that concerned matrix multiplication, but this should be much simpler: it only uses ranged indexing!)

function [out] = fun_A(permutation, in)
  NoU = length(permutation);  
  NoF = length(in) / NoU;     
  out = zeros(1, length(in));
  for i = 1:NoU
    lhs_start  = (i-1)*NoF;
    rhs_start = (permutation(i)-1)*NoF
    for j = 1:NoF                            % these 3 lines run 
      out(lhs_start+j) = in(rhs_start+j);    % so much faster
    end                                      % than the equivalent 
  end  
end

function[out] = fun_B(permutation, in)
  NoU = length(permutation);
  NoF = length(in) / NoU;
  out = zeros(1, length(in));
  for i = 1:NoU
    lhs_start = (i-1)*NoF + 1;
    lhs_end   = i*NoF;
    rhs_start = (permutation(i)-1)*NoF + 1;
    rhs_end   = permutation(i)*NoF;
    out(lhs_start:lhs_end) = in(rhs_start:rhs_end);     % this is much slower!!
  end
end

If you need an understanding of what happens here, assume that permutation = [3 1 2] and that in = [a1 a2 b1 b2 c1 c2], i.e. that we have 3 users (call them a,b,c) with two features each (hence a1,a2 and so on). I want to create and return the following: out = [c1 c2 a1 a2 b1 b2], i.e. to rearrange my users according to the permutation but of course have their corresponding features in succession.

• Typical values are : NoU = 6, NoF = 5.
• fun_A (or fun_B) is supposed to be called more than a million times
• I only care about the performance of the lines that have comments next to them. Please don't point out inefficiencies that are relevant to the calculations of NoU, NoF, the temp variables such as lhs_start etc. These appear here just to make this a standalone example

Answer 1

[Prateekshya]

Hello Konstantinos,

The difference in performance between fun_A and fun_B can be attributed to how MATLAB handles array indexing and memory operations.

Range-based assignment involves complex memory operations which makes it inefficient compared to for-loop. It might also not be optimized for cache usage due to less predictable memory access patterns. The overhead of computing and managing the indices in case of range-based assignment is more significant when the function is called multiple times.

I hope this helps!