parallel for loop in mex

The outer loop in this code should be parallelizable

#include "mex.h"
#include "matrix.h"
#include <math.h>

//[reX imX]=cfrft(rectpuls(t),0.5,linspace(-10,10,2^12),-5:0.01:5);
void mexFunction(int nlhs, mxArray *plhs[],
        int nrhs, const mxArray *prhs[]) {

      nlhs=2;

      const double pi = 3.1415926535897;
      double *f= mxGetPr(prhs[0]);
      size_t flen=mxGetNumberOfElements(prhs[0]);
      double *ap= mxGetPr(prhs[1]);
      double p =  (*ap)*pi/2; double cscp = 2/sin(p); double cotp = 1/tan(p);
      double *t= mxGetPr(prhs[2]);
  //     size_t tlen=mxGetNumberOfElements(prhs[2]);
      double *w= mxGetPr(prhs[3]);
      size_t wlen=mxGetNumberOfElements(prhs[3]);
      plhs[0]=mxCreateDoubleMatrix(1, static_cast<int>(wlen), mxREAL);
      plhs[1]=mxCreateDoubleMatrix(1, static_cast<int>(wlen), mxREAL);
      double *reF = mxGetPr(plhs[0]);
      double *imF = mxGetPr(plhs[1]);

      double dt=t[1]-t[0];
      for (int i = 0;i<wlen;i++){ //<<<<< THIS LOOP

          double re=0;
          double im=0;
          double wc=w[i];
          double wc2=wc*wc;

          for(int j=0;j<flen;j++){
              re += cos(pi*(cotp*wc2-cscp*t[j]*wc+cotp*t[j]*t[j]))*f[j];
              im += sin(pi*(cotp*wc2-cscp*t[j]*wc+cotp*t[j]*t[j]))*f[j];
          }
          reF[i]=re*dt;
          imF[i]=im*dt;
      }

}

How do I use parallel loops in mex? My compiler is Building with 'Microsoft Windows SDK 7.1 (C++)'.

Answer 1

[James Tursa]

Some Microsoft compilers support OpenMP, but I don't think the vanilla SDK compiler is one of them. Did you mean using OpenMP, or something else?

[htrh5]

Okay, what must I do? matlab itself can run parfor loops. Is this OpenMP something I can simply install? Where would I learn how to make matlab to use that instead of sdk 7.1?

[James Tursa]

To run MATLAB parfor loops you need to purchase the Parallel Computing Toolbox. But parfor is not something you would use in a mex routine. To use OpenMP in a mex routine, you need to purchase a compiler with that capability. E.g., one of the Microsoft Professional/Ultimate/Etc editions.

Are you simply looking for speed improvements for some critical part of your code?

[htrh5]

There is no way to write parallel for loops in C++ without paying money for compilers? That doesn't sound right.

[James Tursa]

OpenMP is a relatively easy way to do it, but it is not the only way. Another way is to write your own threading code. For that I would suggest looking at some examples in the FEX to see what others have done. E.g., here is one example:

http://www.mathworks.com/matlabcentral/fileexchange/37515-mmx-multithreaded-matrix-operations-on-n-d-matrices

[htrh5]

if this table is true http://openmp.org/wp/openmp-compilers/

maybe I have access to intel c++ compiler, if I installed that and changed matlab's compiler to that one, would OpenMP work then?

[htrh5]

I did try writing that code in matlab (and I vectorialized the inner loop using sum()). The function runs about 30% faster when it uses 4 cores. That much difference is not worth keeping 4 extra matlab processes running and waiting about a minute for the first function call to me.