Some peculiar problem with Intel Linux Fortran compiler

[praveen_korukonda]

Compiler Used: Intel Fortran 10.1

Operating system: Fedora core 3.0

Problem :

At some places in the code where there are multiple additions and multiplications we get a NAN even though all of its arguments are well defined numbers.

The above problem was noticed with the above operations for both real and complex numbers.

Example:

----------------------------------------------------------------------------------------------------------------------------------------

SUBROUTINE CQCSUM(RESP,DAMP,FREQ,SDVA,AMODS,TEMP,COEF,NMODES,NRSP,NCQC,IDRKEY)

IMPLICIT REAL*8(A-H,O-Z)

DIMENSION RESP(NRSP),SDVA(NMODES,NCQC),AMODS(NRSP,NMODES),

* COEF(NMODES),DAMP(NMODES),FREQ(NMODES),TEMP(NRSP)

DO 50 I=1,NMODES

COEF(I)=0.0D0

50 CONTINUE

DO 100 I=1,NRSP

TEMP(I)=0.0D0

100 CONTINUE

DO 1000 JCQC=1,NCQC

DO 200 I=1,NRSP

RESP(I)=0.0D0

200 CONTINUE

DO 700 IMODE=1,NMODES

DAMPI=DAMP(IMODE)

FREQI=FREQ(IMODE)

DO 300 J=IMODE+1,NMODES

DAMPJ=DAMP(J)

FREQJ=FREQ(J)

IF(FREQI.GT.0.0D0)THEN

R = FREQJ/F REQI

S2 = DAMPI*DAMPJ

R2 = R*R

TEMP1=8.0D0*DSQRT(S2)*(DAMPJ+DAMPI*R)*DSQRT(R2*R)

CON1=(1.0D0-R2)*(1.0D0-R2)+4.0D0*S2*R*(1.0D0+R2)

TEMP2=CON1+4.0D0*(DAMPI*DAMPI+DAMPJ*DAMPJ)*R2

COEF(J)=TEMP1/TEMP2

ENDIF

300 CONTINUE

700 CONTINUE

1000 CONTINUE

RETURN

END

-----------------------------------------------------------------------------------------------------------------------------------

In this subroutine CON1 and TEMP2 are getting NaN.

We are unable to trace the source of the problem. The above code works fine when compiled with a GNU compiler.

Other Observations:

Breaking up the operations into series of smaller operations seems to get around the problem until the same problem surfaces at some other part of the code.

The above problems are not noticed in Windows environment (XP)

Example:

In the same subroutine if I broke up the computation as below then it works.

----------------------------------------------------------------------------------------------------------------------------------------

TEMP1=8.0D0*DSQRT(S2)*(DAMPJ+DAMPI*R)*DSQRT(R2*R)

omr2=(1.0D0-R2)

opr2=(1.0D0+R2)

CON1=(omr2)*(omr2)+4.0D0*S2*R*(opr2)

DIDJ = ((DAMPI*DAMPI)+(DAMPJ*DAMPJ))

&n bsp; TEMP2=CON1+(4.0D0*DIDJ*R2)

COEF(J)=TEMP1/TEMP2

-----------------------------------------------------------------------------------------------------------------------------------------------

It would be helpful if you could help us throw some light as to what could be the source of the problem and a possible procedure to fix it.

[TimP]

You may have to provide more complete information, such as:
is this the 32-bit compiler? Which options? Which specific version of ifort 10.1? A full test case?
I would have thought that
R*SQRT
would be more stable than
DSQRT(R2*R).
It is also possible that
(1+R)*(1-R)
may be more stable than
1.0D0-R2.
If it is necessary to perform operations in the order you have specified, rather than permitting standard optimizations, you should try -fp-model precise -assume protect_parens, if you haven't done so.

[joseph-krahn]

This URL explains the different precision defaults for Linux/Windows:
http://www.intel.com/support/performancetools/fortran/sb/cs-007783.htm

If numbers are close to becoming NaNs, different math options can change the result either way for a given set of data.

[praveen_korukonda]

Thanks for the reply i will try the options provided by you people.