Hi Chris,

On  Chris Mcbrien to All:

 CM>         float temp=26.4;

 CM> My system tells me the value is 2.639996E+01.  

Borland (1.0) says: 2.63999996185303e+01
IBM Cset 2.01 says: 2.63999996185303e+01
Watcom Fortran 9.5: 2.63999996185303E+01

They all agree. Isn't that nice! So it is originated by your system, not
your compiler. If you would have been using doubles, you would have been
obtaining a better result: 2.64000000000000e+01 or in 40 digits:
2.6399999999999998580000000000000000e+01

The problem is the sheer number: 26.4. This can be written as 32 * (33 /
40). Unfortunately 33/40 can not be written as a fraction p/q where p is
integer and q is a power of 2. 

The solution for your problem is well known: if you compare one temperature
T with your criterion temp (26.4) it will be never the same, so ask whether
it is nearly the same. 

if (abs(1-T/temp) < 1.E-6)
        stop;
else
        continue;

if 1 ppm is still to large, use 1 ppb (1.E-9). 

A rather crude trick, a 4 byte floating point has a mantissa of 23 bits, so
it is relatively accurate to roughly 2^-23 appx. 1.E-7. An 8 byte floating
point has a mantissa of 52 bits, roughly 2^-52, so it is relatively
accurate to appx 1.E-16. The figures 26.4 in float and double confirm these
crude estimations.


73 es cuagn,
 Henk

--- The-Box Point 0.15+ PC