ccode atan2 generates nan

Hi!

I have the following question related to the ccode function of the symbolic toolbox.

Let us consider the following piece of code:

    syms a b real
    ccode(atan2(a,b))

the result is:

'  t0 = log((a*sqrt(-1.0)+b)/fabs(a*sqrt(-1.0)+b))*-sqrt(-1.0);'

Questions for the MathWorks:

1. Why ccode replace the atan2 function (which is available in c) with such log implementation?

2. There is a way to prevent such replacement (some additional optional parameter?)

The following Matlab script has been written to generate I/O pairs for the atan2 function to compare the output of the Matlab atan2 vs the C vs the ccode output:

p       = -2:1:2;
[X,Y]   = meshgrid(p,p);
Z       = zeros(size(X));
for i=1:size(X,1)
    for j=1:size(X,2)
        Z(i,j) = atan2(Y(i,j),X(i,j));
    end
end

And with those data I ran the following C code (Visual Studio)

#include "stdafx.h"
#include "math.h"
#include "complex.h"

double Atan2(double a, double b) {
  return log((a*sqrt(-1.0) + b) / fabs(a*sqrt(-1.0) + b))*-sqrt(-1.0);
}

int main()
{
  
  const  unsigned int  nTest = 25;
  const  double      adX[nTest] = { -2.000000,  -2.000000,  -2.000000,  -2.000000,  -2.000000,  -1.000000,  -1.000000,  -1.000000,  -1.000000,  -1.000000,  0.000000,  0.000000,  0.000000,  0.000000,  0.000000,  1.000000,  1.000000,  1.000000,  1.000000,  1.000000,  2.000000,  2.000000,  2.000000,  2.000000,  2.000000, };
  const  double      adY[nTest] = { -2.000000,  -1.000000,  0.000000,  1.000000,  2.000000,  -2.000000,  -1.000000,  0.000000,  1.000000,  2.000000,  -2.000000,  -1.000000,  0.000000,  1.000000,  2.000000,  -2.000000,  -1.000000,  0.000000,  1.000000,  2.000000,  -2.000000,  -1.000000,  0.000000,  1.000000,  2.000000, };
  const  double      adZ[nTest] = { -2.356194,  -2.677945,  3.141593,  2.677945,  2.356194,  -2.034444,  -2.356194,  3.141593,  2.356194,  2.034444,  -1.570796,  -1.570796,  0.000000,  1.570796,  1.570796,  -1.107149,  -0.785398,  0.000000,  0.785398,  1.107149,  -0.785398,  -0.463648,  0.000000,  0.463648,  0.785398, };

      unsigned int  nIdx    = 0;

  for (nIdx = 0; nIdx < nTest; nIdx++) {
    printf("atan2(%+05.5f,%+05.5f)= Matlab(%+05.5f) C(%+05.5f) CCODE(%+05.5f)\n", 
      adY[nIdx], adX[nIdx], 
      adZ[nIdx],
      atan2(adY[nIdx], adX[nIdx]), 
      Atan2(adY[nIdx], adX[nIdx]));
    
  }

  scanf("%d", &nIdx);
    return 0;
}

The results are presented here:

atan2(-2.00000,-2.00000)= Matlab(-2.35619) C(-2.35619) CCODE(-nan(ind))
atan2(-1.00000,-2.00000)= Matlab(-2.67794) C(-2.67795) CCODE(-nan(ind))
atan2(+0.00000,-2.00000)= Matlab(+3.14159) C(+3.14159) CCODE(-nan(ind))
atan2(+1.00000,-2.00000)= Matlab(+2.67794) C(+2.67795) CCODE(-nan(ind))
atan2(+2.00000,-2.00000)= Matlab(+2.35619) C(+2.35619) CCODE(-nan(ind))
atan2(-2.00000,-1.00000)= Matlab(-2.03444) C(-2.03444) CCODE(-nan(ind))
atan2(-1.00000,-1.00000)= Matlab(-2.35619) C(-2.35619) CCODE(-nan(ind))
atan2(+0.00000,-1.00000)= Matlab(+3.14159) C(+3.14159) CCODE(-nan(ind))
atan2(+1.00000,-1.00000)= Matlab(+2.35619) C(+2.35619) CCODE(-nan(ind))
atan2(+2.00000,-1.00000)= Matlab(+2.03444) C(+2.03444) CCODE(-nan(ind))
atan2(-2.00000,+0.00000)= Matlab(-1.57080) C(-1.57080) CCODE(-nan(ind))
atan2(-1.00000,+0.00000)= Matlab(-1.57080) C(-1.57080) CCODE(-nan(ind))
atan2(+0.00000,+0.00000)= Matlab(+0.00000) C(+0.00000) CCODE(-nan(ind))
atan2(+1.00000,+0.00000)= Matlab(+1.57080) C(+1.57080) CCODE(-nan(ind))
atan2(+2.00000,+0.00000)= Matlab(+1.57080) C(+1.57080) CCODE(-nan(ind))
atan2(-2.00000,+1.00000)= Matlab(-1.10715) C(-1.10715) CCODE(-nan(ind))
atan2(-1.00000,+1.00000)= Matlab(-0.78540) C(-0.78540) CCODE(-nan(ind))
atan2(+0.00000,+1.00000)= Matlab(+0.00000) C(+0.00000) CCODE(-nan(ind))
atan2(+1.00000,+1.00000)= Matlab(+0.78540) C(+0.78540) CCODE(-nan(ind))
atan2(+2.00000,+1.00000)= Matlab(+1.10715) C(+1.10715) CCODE(-nan(ind))
atan2(-2.00000,+2.00000)= Matlab(-0.78540) C(-0.78540) CCODE(-nan(ind))
atan2(-1.00000,+2.00000)= Matlab(-0.46365) C(-0.46365) CCODE(-nan(ind))
atan2(+0.00000,+2.00000)= Matlab(+0.00000) C(+0.00000) CCODE(-nan(ind))
atan2(+1.00000,+2.00000)= Matlab(+0.46365) C(+0.46365) CCODE(-nan(ind))
atan2(+2.00000,+2.00000)= Matlab(+0.78540) C(+0.78540) CCODE(-nan(ind))

How can I fix this? Hoping to receive a valid answer asap.

Vincenzo

Answer 1

[Jan]

There is no imaginary value for doubles in C. Then sqrt(-1.0) cannot be resolved with anything but NaN.

You are asking for atan2 in the title, but the symbolic expression contains atan. Is this a typo?

 Questions for the MathWorks:

This is the public forum. If you want to ask MathWorks, use the "Contact Us" button on this page.

Answer 2

[Vincenzo Calabro]

Thanks to Jan Simon to highlight a typo.

The problem is still there and no valuable answer have been provided so far.

I hope someone will come forward with a solution to this bug.

[Karan Gill]

As Jan mentioned, to contact MathWorks, please use the "Contact Us" button at the top right of the page.

[Jan]

@Vincenzo: At least I can confirm that

log((a * 1i + b) / abs(a * 1i + b)) * -1i

calculates atan2, but only if complex numbers are used. Now you know, why the created C code fails. I consider this information and the hint to contact MathWorks directly as useful. Please post here, what they reply. Thanks.

Answer 3

[Vincenzo Calabro]

Matlab confirmed that is actually a bug and no estimate of time to fix has been provided during last contact I had.

Btw, I made an add-on for the symbolic toolbox (sym2code) in which I solved several bugs and I provided support for C/Mex/C#/Java/Python/LabVIEW targets as well ;)

Answer 4

[Christophe Glinel]

I've the same issue with Matlab R2015b, is there is a fix now with the latest version ?

Answer 5

[Vincenzo Calabrò]

to the best of my knowledge the bug is still there. you should use a different code generator to avoid this issue. check for example this third party addon:

Sym2Code

[Christophe Glinel]

Thanks

Is it a commercial addon ?

Answer 6

[Vincenzo Calabrò]

yes it is a cloud service to generate code from symbolic toolbox into C/C++/C#/Java/Python/Labview.

go to: cybertronics.cloud

to get more info and examples.

[Christophe Glinel]

I got an answer from mathworks support and it is fixed in matlab 2018a

Answer 7

[Vincenzo Calabrò]

oh finally :)

good to know. btw sym2code offers more possibilities ;)