One basic approach to writing fast MATLAB (or R) code is to understand vectorized operations. Take a look at this PDF, which addresses this particular problem and I guess will be incredibly helpful:

http://www.ee.columbia.edu/~marios/matlab/Writing_Fast_MATLAB_Code.pdf

Best,

José.

One basic approach to writing fast MATLAB (or R) code is to understand vectorized operations. Take a look at this PDF, which addresses this particular problem and I guess will be incredibly helpful:

http://www.ee.columbia.edu/~marios/matlab/Writing_Fast_MATLAB_Code.pdf

Best,

José.

If your code is time cosuming, I think the fastest way to see the results is to use 'matlabpool" and 'parfor' loops.(if your cpu is multiprocessor).

Also a better way, is to convert the m-file code to simulink and to use xpc to calculate. But it's a little time consuming to convert the code.

There are different factors that affect the speed.

First of all, the instruction of your program is very important .. Do you embed all the things in one m-file or in multiple files as functions to one main m-file?

Also, some commands can consume more cpu time than other commands that produce same purposes!

If you can avoid the number of "for" and "while" loops, then you can save huge amount of the total consumed time

In addition to the above, the new matlab program lets you to debug your m-file and export your final work into .exe file

In my usage, I will avoid using too many loops, which is one of the most disadvantages of Matlab. Or maybe you can develop the mixed procedure to take advantage of the superiorities of the different compilers.

Dear Ming He

Thanks for your comments.

How can I develop a mixed procedure? Please explain.

Matlab was built to perform matrix computations.. so as stated above.. avoid loops and if and use vector and logical operators.. then use parallel computing when you can..

To perform mixed procedure you can call from matlab other programs and ask them to to some calculations.. I have never tried this but i know that this can be easily done

Hi,Reza

I don't contact mixed procedure. I just heard that you can write a "mex (Matlab Executable)" file in C language or Fortran Language, and then call it from matlab. If you need more information on mixed procedure, you can see the following webpage from the official website of mathworks.

http://www.mathworks.com/help/matlab/create-mex-files.html?s_tid=gn_loc_drop

Best regards,

Ming He

Hi, if you want I can prepare simple example of mex in C++. All you need is installed compiler on your system. Please let me know if you are interested and I will prepare example for you.

Hi, Gregor

Thanks, It may be useful.

Best regards,

Reza

Hi, please find attached files. There are two examples, one is extremely simple. It just takes two inputs and multiplies them. Second one computes error of supplied solution of 2D diffusion problem. Please let me know if you won't understand the examples.

best regards, Grega

https://dl.dropboxusercontent.com/u/1120061/MexProgarmming.rar

With preallocating u can have faster programs .

I agree with Schrader. After getting used to the basics as in http://www.mathworks.se/help/matlab/matlab_prog/techniques-for-improving-performance.html and José María Mateos Pérez' document, learn to use the profiler. To get a quick start:

profile on

profile viewer

Now you will see the expensive parts of your program.

I think it is not an issue with Matlab code, but rather with how you run it. If you run it interactively within Matlab environment then it it interpreted. If you compile it and then run it outside of Matlab environment, then it will be faster. Back in the old days you could do the same with an interpreted version of FORTRAN 77. BASIC was also the same. The nice part of interpreters were that you could actively bug fix as you run them. Though compiled code runs faster, if you are in the development phase, the cycle time to compile run and then bug fix is more tedious.

You might also consider

http://www.mathworks.com/products/matlab-coder

best regards

I tried matlab coder. It is quite nice. I compiled the most critical parts of my CFD code (loops that cannot be vectorized) into MEX without changing it much. All you have to do is define arguments of the functions and setup the compiler (coder class). Of course the speedup depends on the size of the problem, but for example, driven cavity test computed on 6400 discretization points executes roughly 20 times faster. If you want, I can prepare example .m files, so you can try it.

You can compile the code into mex, dynamic library or executable. For the executable you have to define main.c where you call your matlab functions.

Dear Gregor,

Many thanks for your comments. I am willing to access the example.

Regards

Hi,

I have prepared simple example, you can find it here:

https://dl.dropboxusercontent.com/u/1120061/matlab.rar

In this example the speedup isn't that high since the code is very simple and matLab can optimize script execution. There are also some overheads for calling compiled mex. However, with more complex codes you can expect better speedups.

Regards

Avoid loops. Vectorize.

vectorize

and create m file

here is tutorial

www.mathworks.com/academia/student_version/learnmatlab_sp3.pdf‎

https://www.mccormick.northwestern.edu/docs/efirst/matlab.pdf‎

http://www.ee.columbia.edu/~marios/matlab/Writing_Fast_MATLAB_Code.pdf

hope this will help

Hi,

I have implemented C++ and Matlab Fluid flow solver. Both implementations are quite optimized. In Matlab I vectorized most of the operations and compiled the rest of the code into mexw64. I've compared the performance on lid driven cavity Re=100 on 900 nodes (I am using DAM numerical approach) and the difference in the execution time is quite big. The c++ code took about 3s while matlab was crouching it for more than a minute (83s). If I find some time I will do more thorough test. The Matlab's parallel for did not help improving the performance, while in C++ I could speedup the code with efficiency 0.6 on four cores. I used Intel(R) Core(TM) i7-2630QM CPU @ 2.00GHz.

Best regards, Grega

Other useful links:

http://www.slideshare.net/jbhuang/writing-fast-matlab-code

http://www.matlabtips.com/optimizing-your-code/

For Matlab Code, Visit link

http://dg-algorithm.blogspot.com/

Hi Reza;

a very good way is to transfer your codes to python using NumPy which runs matlab codes almost exactly as it is. there are very few changes needed ! ( might be needed) . this runs much faster than matlab but the same results. also you can share the script on all OS platforms as it is.

http://www.numpy.org/

This is an old thread, and I had not done that in a while, but what makes a huge difference is how Matlab stores arrays/vectors and how you read it in your loops. Say you have a matrix and Matlab stores it row-wise, then you get a huge penalty if you are looping on elements in the same "column". Sorry I can' be more specific, but this is the rough ideea. Using it I was able to speed some matrix calculation code by about a factor of 10.