Problems with IFC 8.0 and CSQRT

[Deleted_U_Intel]

I'm having some trouble with the CSQRT in IFC 8.0. It seems to give different results on the same data depending on the subroutine it is contained in. Example:

program test
implicit none
COMPLEX z
z = (-0.5,0.)
write(6,*) 'z =',z,' csqrt =',CSQRT(z*z-1.)
call W(z)
stop
end

SUBROUTINE W(dummy)
IMPLICIT NONE
complex dummy
COMPLEX z

write(6,*) 'Using assigned values for kz and k'
z = (-0.5,0.)
write(6,*) 'z =',z,' csqrt =',CSQRT(z*z-1.)
return
END


This program outputs:

z = (-0.5000000,0.0000000E+00) csqrt = (0.0000000E+00,-0.8660254)
Using assigned values for kz and k
z = (-0.5000000,0.0000000E+00) csqrt = (0.0000000E+00,0.8660254)


Thus the complex part changes sign. g77 handles the code fine and so does IFC 7.0

[TimP]

It looks like this comes under the heading of how to minimize incidental numerical differences. A few suggestions:

Replace z*z-1 by (z+1)*(z-1)

Attempt to have the entire compilation done with SSE (option -xK or -xW), unless you have a very old CPU.

Try the -mp and -mp1 options, and -fp_port if you can't use SSE.

It does look in this case that all possible versions should produce exact results for z**2-1, so I agree that it doesn't make sense for one of the intermediate results to fall slightly below the branch cut.

[Deleted_U_Intel]

It seems that -mp makes the result consistent (both CSQRT returns a negative number)
When re-written to (z+1)*(z-1) both return positive.

Hmm.

[Deleted_U_Intel]

It is clear that the differences has to do with the branch cut implementation.
However, the compiler is not consistent in its exact location of the branch cut.
Two examples:
1) If the call to W (line 7) is commented out by a c in column 1 the maginary part of the result from the main program changes from negative to positive.
2) If the argument of the "subroutine W" (line 13) is changed from dummy to z the maginary part of the result from the subroutine changes from positive to negative.