time of execution of a code in matlab

Hello, To compare the time of execution of two functions, i wrote a code where i measure the time elapsed for each function and for each index ( with the command clock /etime). However, on my computer the first function is faster than the second, and on an other computer, it's the opposite. ( I tried several time with different indexes and with clearing for each time the data ,but it is still the same result).

Can you explain me why there is such difference?

On both of them, Matlab version is r2017a , the only difference is that on one computer, the compiler is MinGW 4.9.2 ,and on the other one it's Microsoft Windows SDK 7.1 .

Thanks in advance

[Adam]

I imagine different machines are optimised in different ways so one may find one function more optimal the other the other way round.

What kind of runtime are you talking anyway? Tiny or significant? I don't know how accurate clock/etime is at measuring function time. I always use

doc timeit

myself.

[emar]

Hello,

Thank you for the answer. I wanted to compare execution time for a convolution ( optimized code) vs fft. The first code is conv2fft ( I found it on internet . It's based on mex code).

The second one is calculation of fft. Here is the second code :

   function xcorr_ab = freqxcorr(T,A)
T_size = size(T);
A_size = size(A);
outsize = A_size + T_size - 1;
% calculate correlation in frequency domain
Fa = fft2(rot90(T,2),outsize(1),outsize(2));
Fb = fft2(A,outsize(1),outsize(2));
xcorr_ab = ifft2(Fa .* Fb,'symmetric');
end

The code to compare their speed is :

T=rand (800,1280);
j=1;
profile on;
for i=2:0.5:12
  ii=i*10;
  C{j}=rand(ii,ii);
  conv2fft(C{j},T);
  freqxcorr(C{j},T);  
  j=j+1;
end
profile viewer;

On the first computer ( matlab r2017a ,a fast computer), the profiler shows :

On the second one ( slow computer , matlab r2017a ):

It's normal that on the second computer , it will be slower, but it found strange that the fastest function isn't the same on the two computers.

Here is attached all the codes( if you want to test), the test function is named ' test'.

I want also to mention that even if i test with rand matrix's, sometimes, i save the data to compare between the same inputs, but it still the same result .

[Adam]

You forgot to attach anything. Some builtin Matlab code uses multithreading and things that use FFT are especially susceptible to changes in problem definition or resources when compared against non-FFT methods given the usual situation of an FFT and IFFT overhead combined with a (usually) faster algorithm in the frequency domain vs an slower algorithm in the time domain, but which may still end up faster given the FFT overhead.

I have always found it very hard to find any reliable rules for my fft-based convolution function vs non-FFT based equivalent.

[emar]

Sorry i attached the incorrect zip. Here is the zip

[Stephen23]

Better options for timing code would be to use the inbuilt timeit, or the FEX submission timeit.

[emar]

Even with timeit, i still have the same results.

Answer 1

[Jan]

Can you find the code lines inside freqxorr which are accelereated on the faster computer. It could either be multi-threading and a higher number of cores, or the data might match into a larger CPU cache, or the internally used library can use the SSE or AVX features fo the younger processor. Nevertheless, it is not surprising, that different functions can profit differently on different CPUs.

[emar]

Exactly, in the first computer there is 44 cores so that's why fft2 is faster than the other computer (12 cores). Thank you for your help !