"Steve Myers" <noone@[EMAIL PROTECTED]
> wrote in message
news:5cc42$4481b58a$d14739a1$3604@[EMAIL PROTECTED]
> Something that may be immaterial, or nearly so. I suspect the average
> bit ****ft was <= 2. Of course, in "hexadecimal" floating point, it was
> probably closer to 0
Although on average this would be so for hex,
the quoted times were fixed, and did not depend on
the number of hex ****fts.
As for binary, that's a different kettle of fish altogether,
as it would be rare for the exponents to be equal
(e.g., add 1 and 2).
A ****ft of 2 wouldn't even cover addition of 1 and 8.
In practice, the range of exponents of typical values
would differ much more than this.
> robin wrote:
> > Further to my note about hex float times,
> > you will observe from the times of
> > AE of 19.20µS and AU of 18.96µS
> > that the time for post-normalization was 0.24µS
> > Compare that with single bit ****fts using SLL
> > of up to 21 bits, say, of 0.48µS per bit --
> > Equals 10µS.
> > Thus, we could expect that single-bit ****fts for AE
> > would add 20µS (recall that pre-normalization
> > and post normalization are required for add and subtract)
> > to the instruction time. Thus, instead of AE taking
> > 19.20µS, it would take
> >
> > 19.20 - 0.48 + 20 = 38.82µS
> >
> > which is nearly double the time taken for hex of 19.20µS.
|
