Re: float vs. double?

There have been repeated myths that float is faster than double. In antique machines, the
kind where you could see the transistors without a microscope, this was true. It has not
been true in the Pentium series in many years. A floating-point multiply takes, if I
recall correctly, one CPU clock cycle, that is, about 1/3ns on a 3GHz machine. And it
will happen concurrently with two other integer operations. So there's very little
motivation to use 'float' these days.

Note that the floating-point ALU of Intel chips supports an 80-bit floating point number,
but as of VS 2005 there is no C compiler support for 80-bit floating point. But going to
the MSDN and clicking "long double" will lead to informative articles, such as one on
floating-point precision.

32-bit: sign, 8-bit exponent, 23-bit mantissa
65-bit: sign, 11-bit exponent, 52-bit mantissa
80-bit: sign, 15-bit exponent, 64-bit mantissa (unsupported by C/C++)
joe

On Thu, 21 Dec 2006 07:17:33 -0000, "David Webber" <dave@xxxxxxxxxxxxxxxxxxx> wrote:


"Robert Wong" <robertwong@xxxxxxxxxxx> wrote in message
news:OK%23gpmGJHHA.1252@xxxxxxxxxxxxxxxxxxxxxxx

I wasn't sure if it was a code generation problem. Also kind of weird is
depending on the variable type, is different precision being stored?

Of course - "double" is called "double" because it stores about twice as
many significant figures as "float".

FORTRAN is a bit more verbose "float" = "REAL" and "double" = "DOUBLE
PRECISION".

The same multplication is used.

answer = value * lsb; // debugger - 1004714.7 float
danswer = value * lsb; // debugger - 1004714.7080078125 double

It is also worth noting that multiplication essentially preserves the number
of significant figures - the result will be accurate to the number of
significant figures as the less accurate of the multiplicands. [And note
that integers are first converted to float or double when multiplied by
float or double.]

But lesson number 1, which I learned in FORTRAN IV in my first lesson (in
1968), is "never use single precision" as you're laying yourself open to
errors which can be mitigated just by using double precision. Since then I
don't think I have ever done any floating point computation in single
precision - and these days, as I say, (because of me, obviously, and
possibly the other David too <g>) Mr Intel designs his chips to work most
efficiently in double precision.

If you're going to be doing a lot of floating point computation, I'd
seriously recommend looking at a textbook or two on numerical maths. [I'd
recommend those on my shelves but they've probably been out of print for 30
or 40 years.]

Dave
Joseph M. Newcomer [MVP]
email: newcomer@xxxxxxxxxxxx
Web: http://www.flounder.com
MVP Tips: http://www.flounder.com/mvp_tips.htm
.