Difference in the evaluation of trigonometric function in CVF and IVF

[tlahydrema_com]

I've just bought IVF and is moving my code from CVF 6.6.B to IVF 10.1. The two programs are giving small difference in the output. Is it possible that the two compilers evaluate trigonometric functions different?

i.e. I use asin, cos and acos til calculate three variables, which all is used in a equation containing sin.

I have checked initialisation for all variables, which I have figured out, that IVF is much more sensitiv to initialisation (var = 0.0D0). I use double precision in my code (real(double) :: xx).

Anybody who have experiende same behavior moving code from CVF til IVF?

/Regards Thomas Langer

[Les_Neilson]

It could be due to a difference in optimisation;but it could also be due to your initialisations.
For example x = 0.1 is not the same as x = 0.1d0

(The first one gives an internal representation of x'3FB99999A0000000'

the second is x'3FB999999999999A)

Les

[TimP]

Quoting - tlahydrema.com

I've just bought IVF and is moving my code from CVF 6.6.B to IVF 10.1. The two programs are giving small difference in the output. Is it possible that the two compilers evaluate trigonometric functions different?

i.e. I use asin, cos and acos til calculate three variables, which all is used in a equation containing sin.

I have checked initialisation for all variables, which I have figured out, that IVF is much more sensitiv to initialisation (var = 0.0D0). I use double precision in my code (real(double) :: xx).

Anybody who have experiende same behavior moving code from CVF til IVF?

/Regards Thomas Langer

[TimP]

Math functions in CVF take advantage of x87 extra precision for improved accuracy, while ifort SSE2 libraries would use extra precision only where necessary for stability. To increase likelihood of identical results, you would need to set float consistency option in CVF and /assume:protect_parens and no SSE options in 32-bit ifort.

[g_f_thomas]

IIRC, fp arithmetic is supposed to be done the same way acrossI32 - IA64. Q4 of

http://developer.intel.com/technology/itj/archive/1999.htm

might be of interest.

Gerry