Performances of nested functions

Hi,

I'm new to matlab, so I hope I'm doing this right. I tried to find a documentation about performances of nested functions, but I haven't found something satisfying for my concern.

I ran this simple code on my system (winXP, MATLAB R2010a):

function [] = testNestedFunctions()

tic;
A = 0;
for i = 1:1000000
    A = i * i;
end
toc;

tic;
A = 0;
for i = 1:1000000
    A = nestedFun(i);
end
toc;

      function [res] = nestedFun (i)
          res = i * i;
      end
  end

The non nested code runs in 0.002511 seconds. The code with nested functions runs in 0.198646 seconds.

Is there a reason about this significant time difference? I thought it may be related to the use of loops, which may be not optimized using the nested function, but I didn't find any information about that.

Thanks!

Answer 1

[Jan]

Matlab's JIT accelerator can optimize the first version, but has less power, when the code inside the loop contains calls to not built-in functions like your nested function here. The JIT is not documented and the timings might change with the Matlab release. So I guess that the JIT can recognize, that A is overwritten in each iteration in the first version, but in the 2nd case, the called function could contain side-effects.

Timings are not meaningful for such minimal examples, because you cannot draw much conclusions. I'd compare it to a subfunction and a function written to another M-file. And because optimizing experimental code is not useful, I'd compare the nested function versus inlined code for the real application.

[Nicolas]

Thank you for your answer!

Actually I showed here a minimal example for clarity but in my code, I have something like that:

function [] = foo()
    A = zeros(512, 512, 'double');
    for i = 1:size(A, 1)
        for j = 1:size(A, 2)
            A(i, j) = aDistanceCenteredAt(i-1,j) + aDistanceCenteredAt(i+1,j);
        end
    end
    function [res] = aDistanceCenteredAt(i,j)
        % distance between two histograms computed on 2 windows
    end
end

In this case, the non nested version runs in 0.97 second, while the nested version runs in 2.8 seconds.