MEX/Mkoctfile questions about matrices algorithms...why slower?

[钟 伟生]

Hi,

I really appreciate your reply. ATLAS project maybe an efficient way to do 
the improvement on matrix math. But I am still wondering to know the reason 
why the run time increased.

Theoretically speaking, while using the interface functions (MEX/DEDUN-DLD) 
combining matlab/octave with C/C++ language, the matrix math should be 
faster than before. So, what is the key problem which makes it slower? The 
algorithms or the overhead generating by the MEX? If the front one, I will 
try to solve it. Thank you!

Best Regard,

Weisheng


> Weisheng wrote:
> > I am trying to use the interface functions between MATLAB/OCTAVE and C
> > to compute the matrices algorithms( like matrix-matrix multiplication
> > and additon, matrix-vector multiplication ), such as MEX and
> > DEFUN_DLD. However, after having compared with Matlab, I found the
> > implemented with C/C++ are not faster than the same on in Matlab,
> > moreover, it costs more about 30% time. I compared them both in big
> > and small matrices multiplications.
> > The results are the same as before.
> >
> > Are there good and bad ways of implementing matrix multiplication
> > combining MATLAB/OCTAVE and C?
> >
> > Thank you very much!
>
> Both Octave and Matlab link to external libraries to perform their
> matrix and vector computations, and those libraries are already highly
> optimized and probably faster than anything you could using a simple Mex
> file. If you want faster matrix math, the ATLAS project has made this an
> area of ongoing research. If you are not already using ATLAS libraries
> with octave, I suggest downloading them and learning how to optimize
> them for your computer.
>
> Quentin

_________________________________________________________________
享用世界上最大的电子邮件系统― MSN Hotmail。 http://www.hotmail.com