glennrp@BRL.ARPA (Glenn Randers-Pehrson, WMB) (07/07/88)
The behaviour of the EXP() function for large negative numbers on the
IRIS is shameful. I've tried the following program on a number of
machines (VAX f77, Gould f77, Alliant FORTRAN, Cray-2 cft77) and all
return 0.0 for out-of-range negative numbers. The IRIS returns
not-a-number if the floating point board isn't used. Worse, if the
floating point board is used it returns wrong numbers. This on a 3130
running 3.5r2 and a 2500T running 3.6.
program bugexp
dimension trial(10)
data trial/-1., -20., -50., -80., -85., -88.02968, -88.02970,
& -100., -1000., -10000/
do 10 i=1,10
x=trial(i)
xx=exp(x)
write(*,'('' exp('',f13.6,'')='',g15.8)')x,xx
10 continue
stop
end
Results with f77 -Zf:
exp( -1.000000)= .36787950
exp( -20.000000)= .20611540E-08
exp( -50.000000)= .19287500E-21
exp( -80.000000)= .18048510E-34
exp( -85.000000)= .12160990E-36
exp( -88.029680)= .15974800E-42
exp( -88.029700)=???????????????
exp( -100.000000)= -.43075530E+34
exp( -1000.000000)= .12234650E+29
exp(-10000.000000)= .41732680E-27
Programmed STOP mccalpin@MASIG1.OCEAN.FSU.EDU ("John D. McCalpin") (07/08/88)
Silicon Graphics has known about the pathetic behavior of the exp function for a long time, but they do not seem at all interested in doing anything about it.... As I understand it, they bought the compiler from Silicon Valley Software, and did an in-house rewrite of the intrinsic transcendental functions. The behaviour can be modified (though not corrected) by initializing the floating-point exception handler. You will need to add the following lines to the example code: #include <fortfpe.h> ! somewhere in declarations call setfpe (NOCOREDUMP, ZEROVALS) ! first executable statement This makes MOST underflows go gently to zero, but it does not fix the exp() problem. I don't know why the floating-point exception handler is not active by default.... I have to uses a fix to the exp() problem, I use x = min(x,88.0) x = max(-88.0,x) result=exp(x) or, in one line: result = exp(max(-88.0,min(x,88.0))) The constant 88.0 is the largest value which returns a correct result with 32-bit precision. I can't recall the number for 64-bit precision --- I think it is about 300. Now that the 4D's are the SGI's main interest, I would not count on getting any improvements to the SVS Fortran on the IRIS 3000's.... especially since they did not respond when the IRIS 3000 was their only entry. I suggest that you try a SUN machine. They have troubles, but at least the Fortran numerics are quite robust. John D. McCalpin mccalpin@nu.cs.fsu.edu