[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: I'll try again: oct and arrays
From: |
Martin Helm |
Subject: |
Re: I'll try again: oct and arrays |
Date: |
Thu, 30 Dec 2010 14:14:43 +0100 |
User-agent: |
KMail/1.13.5 (Linux/2.6.34.7-0.5-desktop; KDE/4.5.4; x86_64; ; ) |
Am Donnerstag, 30. Dezember 2010, 12:52:21 schrieb Jesper Schmidt Hansen:
>
> I have tried to use fortran_vec() as Martin and Søren suggested: this
> gives a speed-up of two compared to using the ()-operator, but is
> still slower than the corresponding MEX file. On the current machine
> I get (different from the one I used in the original question), the
> numbers in parenthesis denote the standard deviation
>
> OCT using (): 0.0362 (0.0006)
> OCT using fortran_vec(): 0.0143 (0.0006)
> MEX : 0.0078 (0.0005)
>
> Jesper
Just out of curiousity, did you also try how it scales when you make your
problem larger with the Gauss-Seidel?
I have tested your original example with () and fortran_vec and with a
1000x1000 matrix I had a speedup for the latter by about a factor 2 but for
10000x10000 by a factor about 10 (i repeated every calculation 10 times to
avoid measuring some startup effect).
So it looks to me that there is some (constant?) overhead in using the oct file
which leverages when the size of the problem/raw computing time grows.