precision difference in matlab and gnu g++

Hi all,

I have two short code examples which do the same. One is written in matlab the other in c++.

The result in matlab code is -2.168404344971009e-19.

The result in c++ code is -3.01755e-20

The difference in results is much toooo huge! What can I do?

Thanks in advance!

--- matlab code start ---

w = -0.00161579
rho = -3.23158e-05
gamma = 0.02
Igrad = 0.0004

w = w - rho * gamma / Igrad

--- matlab code end ---

--- c++ code start ---

#include <iostream>

int main() {
  double w = -0.00161579;
  double rho = -3.23158e-05;
  double Igrad = 0.0004;
  const double gamma = 0.02;

    w = w - rho * gamma / Igrad;

    std::cout << w << std::endl;
  }

--- c++ code end ---

Answer 1

[Titus Edelhofer]

Hi,

not really: both answers are equally correct up to round off: the roundoff for the last difference is

eps(rho*gamma/Igrad)

which is in fact the answer MATLAB gives. Every answer that deviates from this by a maximum of 2.168e-19 is "correct".

Titus

Answer 2

[DavidK]

I'm confused as the mex-compiler of matlab is giving the correct result but g++ not.

C++ code "test.cc":

#include "mex.h"
#include <iostream>

void mexFunction(int nlhs, mxArray *plhs[],          /*  output  */
                 int nrhs, const mxArray *prhs[])    /*  input  */
{
  double w = -0.00161579;
  double rho = -3.23158e-05;
  double Igrad = 0.0004;
  const double gamma = 0.02;

    w = w - rho * gamma / Igrad;

    std::cout << w << std::endl;

}

and matlab code main.m:

test();

Why is mex-c-compiler better than g++? Can I tune g++?

[Titus Edelhofer]

Again: everything in the interval [2* -2.168404344971009e-19 0] is equally correct.

[Walter Roberson]

c++ does not use fixed-point arithmetic. c++ has the same roundoff problems that MATLAB has. It is possible, however, that in C++ the rounding mode is initialized differently.