topic Re: dble() changes value in Software Archive

dble() changes value

Intel_C_Intel — Wed, 14 Mar 2001 19:36:25 GMT

Hi,
I'm wondering, how to get exact same real*8 number from real*4 numbers. Example:
real*4 x4
real*8 x8,y8
x4=2.7
x8=dble(x4)
y8=2.7d0
write(*,*) x4, x8, y8
end
gives:
2.700000 2.70000004768372 2.70000000000000

So,shouldnt dble() initialize the rest of the mantissa to zero to have x8=x4 in common sense (y8 shows, this number exist :)

Cite from dble(a) documentaton in CVF:
"If a is of type integer or real, the result has as much precision of the significant part of a as a double precision value can contain."

Any work-around, which stands optimization ?

Thanks
Martin

Re: dble() changes value

Steven_L_Intel1 — Thu, 15 Mar 2001 00:55:53 GMT

It DOES set the rest of the mantissa to zero - binary zero! That's not the same as decimal zero! (Look for an article in our upcoming newsletter that discusses this issue in general - it will be available to registered users only.)

Once you have assigned a REAL*4 value, there's no way to "create" the extra precision that a REAL*8 value would have had.

Steve

Re: dble() changes value

Intel_C_Intel — Thu, 15 Mar 2001 01:22:41 GMT

Thanks for your prompt answer. But ...

The consequence of this unpredictable 'binary' addon: no local change to double precision, e.g. in critical functions wich handle small differences of large numbers. I have constructed a small example, where double precision results in larger error than staying in single precision.

You mentioned a newsletter article, I guess not out yet. Any chance to get an idea how to get rid of that problem ? (I'm registrated, email ok)

Martin

Re: dble() changes value

Steven_L_Intel1 — Thu, 15 Mar 2001 02:34:33 GMT

It's not unpredictable - it's VERY predictable. DBLE will zero-extend the fraction from single-precision to double-precision with binary zeroes. You evidently want decimal zeroes added - that's not the way it works. (And if you had assigned x8=x4, you'd see the same thing.)

My first suggestion would be to use double precision throughout your application, being sure to use the D0 form of constants. You could also do a conversion as follows:


character*13 buff
...
write (buff,'(E13.6)') x4
read (buff,'(E13.6)') x8

That should extend the value with decimal zeroes.

I expect the newsletter to be out in a week or so.  Watch your e-mail for the pointer.

Steve