Intel® Fortran Compiler
Build applications that can scale for the future with optimized code designed for Intel® Xeon® and compatible processors.
28383 Discussions

Reason for difference in results

Freitas__Henrique
443 Views

Hello,

I have a scientific code that implements a numerical method in Fortran, which is compiled with ifort using "-O2 -mieee-fp" flags, and the results with this compilation are the reference results.

After compiling the code without using "-mieee-fp", intermediate variables present small differences, and as the numerical method iterates and checks for convergence, the final results differ by either hundreds or thousands when compared to the reference results. I'd like to know what is the reason for these differences.

I also used another compiler (pgf90) for comparison, which produces the same results as ifort without using "-mieee-fp", even if explicitly using this type of flag ("-Kieee").

The sample code below illustrates the differences.

System: Intel Core i7 7820x, Linux CentOS 7.5.1804
Compilers: Intel Fortran 19.0.5.281, PGI Fortran 19.10-0

Thanks

Results:
ifort -o prog1.x prog.f90 -O2 -mieee-fp; prog1.x
val: 2914347.25
ifort -o prog2.x prog.f90 -O2; prog2.x / pgf90 -o prog2.x prog.f90 -O2 [with/without] -Kieee; prog2.x
val: 2914347.50

program precision

real x1, x2, x3, val

x1 = 0.00999999977648258209228515625
x2 = 539.693969726562499999999999999999999999
x3 = 150.0

val = x1*x2*x3*3600.0

write(*,'(F30.20)') val

end program precision

 

0 Kudos
1 Reply
mecej4
Honored Contributor III
433 Views

The differences that you report are illusions. In single precision, i.e., default real size, machine epsilon is 1.192093E-07. The implication is that you should not compare more than the seven most significant digits in deciding whether two numbers are equal or different. The two numbers that you reported differ only in the least significant bit; that is, they have the internal representations Z'4A31E0AD' and Z'4A31E0AE'.

Secondly, when you output more digits, you can see misleading (and, indeed, non-existing) differences because of the arithmetic involved in converting from the internal binary format to a decimal string representation.

If you are not familiar with the ways in which floating point arithmetic is inexact and violates familiar rules of algebra (such as A + (B + C) = (A + B) + C), please read some articles on the topic, such as https://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html .

0 Kudos
Reply