Re: Performance test

[Carlo de Falco]

On 11 Mar 2009, at 23:57, DaNiMoTh wrote:

> Hello to all.
>
> I've tried to execute the same routine (see attachment) with matlab,
> octave and a C program.
>
> There are the results ( both of programs under Linux )
>
> MatLab:
> > > tic ; kernel(magic(100)) ; toc
> Elapsed time is 0.052000 seconds.
>
> Octave:
> octave:1> tic ; kernel(magic(100)) ; toc
> Elapsed time is 56.9406 seconds.
>
> C:
> time ./a.out
> real    0m0.075s
>
> Why this difference?
> Where is the bottleneck of Octave?

In the 'for' loop,

it is well known that writing vectorized code gives a dramatic speed- 
up in Octave.
For exmple on my machine, by rewriting your function as

---------------
function G = kernel2(A)

colonne_m = columns(A);
righe_n = rows(A);
G = zeros(righe_n,righe_n);
for i = 1:righe_n
      for j = 1:colonne_m
        G(i,j) = (A(i,:)*A(j,:)'+1).^3;
      endfor
endfor
---------------

I get

---------------
>> tic ; B2=kernel2(magic(100)) ; toc
Elapsed time is 0.60966 seconds.
---------------

whereas your original function gives

---------------
>> tic ; B=kernel(magic(100)) ; toc
Elapsed time is 48.349 seconds.
---------------

and of course

---------------
>> norm(B-B2)
ans = 0
---------------

I'm sure this can be improved even further

> There is some documentation about this?

yes, this is mentioned in the FAQ:
http://www.gnu.org/software/octave/FAQ.html#MATLAB-compatibility

> Many Thanks. ( Please CC'd me in reply as I'm not a follower of this  
> list )

c.