> In article <email@example.com> Raymond Toy <firstname.lastname@example.org> writes:
> > But I do still think we when we compute estimates of costs, MIPS,
> > cycles, whatever, we tend to ignore things like table lookups,
> > additions, etc, and assume they are free, compared to multiplication,
> > when in fact they can be even more expensive.
> Well, if somebody does that he is wrong. There are very many processors
> (and not only DSP's) where a floating point multiply takes as many
> cycles as a floating point addition. Vector processors are of this
> kind. And the first one I did use with this feature was a CDC Cyber,
> back in the seventies.
Floating-point addition is very expensive. Before they are added, one of
the numbers must be adjusted to have the same exponent as the other.
After, the result must be renormalized. For multiplication, only the
post-operation renormalizing is needed.
Engineering is the art of making what you want from things you can get.