18 Nov
2019
18 Nov
'19
12:12 a.m.
On Sat, 16 Nov 2019 21:50:58 -0800 Larry McVoy <lm(a)mcvoy.com> wrote:
Larry McVoy writes:
On Sun, Nov 17, 2019 at 04:30:15PM +1100, Dave
Horsfall wrote:
Many numbers can't be represented perfectly using integers or
rationals (a pair of integers) but can be computed using a
series expansion to arbitrary precision. I thought FP numbers
were a clever & practical compromise that worked quite well.
David Goldberg's "What every computer scientist should know
about floating-point" is worth reading.
https://www.itu.dk/~sestoft/bachelor/IEEE754_article.pdf
Earlier I remember reading the "Numerical Recipes" books by
Press, Teukolsky, Vetterling & Flannery. IIRC, the original
version used Fortran. They also had versions using Pascal and
C (I finally bought the C version in '80s though never used it).
Note that Scheme & CL have a full complement of numeric types:
big nums, rationals, reals and complex numbers. At least some
versions of CL have arbitrary precision FP numbers.
What I really want is a programming language with support for
symbolic manipulation of formulas!
On Sat, 16 Nov 2019, SPC wrote:
I'm pretty sure that was the Fortran I learned. Yeah, it was not C. But
it was math. I spent a bunch of time learning accumulated errors and
more time on floating point numbers. My dad was a theoretical physics
guy so I did some coding for him. I respected Fortran for what it could
do but I developed a hate for floating point. In my mind, floating
point numbers meant you couldn't handle the world you were working in.
It just felt like you could shift the domain you were working in so
integers could work. If you couldn't do that, you were admitting that
you were not accurate. My first FORTRAN textbook was titled "FORTRAN
with WATFOR and WATFIV". It
had a long print run as well.
Now *that* brings back memories (not necessarily pleasant). WATFOR was as
ugly as sin