Precision discrepancy between Fortran and Python (sin function)

问题

I see a discrepancy between python and Fortran when using the sinus function. Could anyone shed light on this, please?

in python:

      import math
      print(math.sin(6.28318530717959))
      >> 3.3077843189710302e-15

in fortran90:

      print*, sin(6.28318530717959d0)
      >> 3.3077720792452914E-15

---

EDIT: As it seems to be a Fortran compiler issue, I used g95 with

       g95 -O3 test.f90 -o test.exe

回答1:

According to IEEE 754 for float representation:

In [7]: bin(3.3077720792452914e-15.view(np.uint64))
Out[7]: '0b11110011101101110010110011010000000000000000000000000000000000'

shows a truncated mantissa, when

In [9]: bin(3.3077843189710302e-15.view(np.uint64))
Out[9]: '0b11110011101101110010110011101100111001100111010111010001111111'

shows a plain one.

Probably a type issue, with a float32 in the process, even the origin is mysterious.

来源：https://stackoverflow.com/questions/42462344/precision-discrepancy-between-fortran-and-python-sin-function