Mex and Matlab Code

Using the Concept from

http://in.mathworks.com/help/coder/examples/c-code-generation-for-a-matlab-kalman-filtering-algorithm.html?prodcode=ME&language=en

I have done :

function out = sum_it(a,b)

out = a + b;

>> codegen -args {2.0 3.0} sum_it.m

My Test Function :

ip1 = rand(100,1);

ip2 = rand(100,1);

for j = 1:20

    tic()
    for i = 1:len
        Pa(i) = sum_it(ip1(i), ip2(i));
    end
    a(j) = toc();

    tic()
    for i = 1:len
        Pb(i) = sum_it_mex(ip1(i), ip2(i));
    end
    b(j) = toc();
    fprintf('%f   %f\n',mean(a),mean(b))
end

Unfortunately a < b. How can it be so ??? I was expecting b <<< a by big margin ! What am I doing wrong?

0.004559 0.004008

0.002347 0.003053

0.001587 0.002737

0.001214 0.002609

0.000989 0.002450

0.000834 0.002305

0.000728 0.002175

0.000645 0.002064

0.000580 0.002015

0.000534 0.001983

0.000491 0.002014

0.000458 0.002048

0.000430 0.002079

0.000406 0.002092

0.000386 0.002094

0.000369 0.002105

0.000353 0.002124

0.000338 0.002144

0.000326 0.002146

0.000313 0.002114

[Ryan Livingston]

Could you please format your code to make the question a bit easier to read?

Answer 1

[Ryan Livingston]

Generally what is happening here, is that the body of sum_it calls a MATLAB built-in function plus. Such functions are already implemented in compiled code in MATLAB and generating code for them is not expected to accelerate them significantly. The documentation for acceleration describes this more under the heading Avoid Generating MEX Functions if Built-In MATLAB Functions Dominate Run Time.

The likely reason for the small slowdown is that calling a MEX function has some overhead. Because the body of sum_it is so simple, that overhead is dominating the small execution time.

If you have a larger application that you would like to accelerate, then I would suggest following the steps on that documentation page that describe using the profiler to identify the bottlenecks in your code. Once those are identified, then you can determine if they are good candidates for acceleration via MATLAB Coder and follow the steps described to generate code for those small portions.