Re: Extended binary numbers

From: William Kempf (sirwillard_at_[hidden])
Date: 2000-09-21 08:42:48

--- In boost_at_[hidden], "Paul Moore" <gustav_at_m...> wrote:
> From: William Kempf [mailto:sirwillard_at_m...]
> > Also, I'd *really* like to see the related class bigint (or some
> > other name), which is an arbitrarily sized integer type that 
scales
> > to larger sizes as needed.  The bitint class is easier to deal 
with,
> > and will work for 90% of the cases where you need large integral
> > types, but those 10% cases would be nice to cater to as well.  
Along
> > these same lines we could use a bigfloat as well.  This type might
> > use a fixed point representation and bigint in order to avoid the
> > normal binary to decimal conversion error.
> 
> I've been thinking about these classes for a long time, since I did 
the
> rational class. I haven't got round to it because of (a) lack of 
time, and
> (b) lack of certainty as to what the most efficient algorithms are. 
I have
> Knuth's book, but it's quite hard going...

Been there, done that.  Knuth's algorithms are complex enough to make 
for slow going if you want to fine tune it.  Once I did so, I still 
found it to be awfully slow in comparison to say, GNU MP.  
Unfortunately, that library is very *nix based as well as very fine 
tuned for specific architectures, making it very difficult to port.  
Not to mention it's a C library.

Knuth's algorithms would be good enough for a start, IMHO.  For 
Boost, the interface is more important than the implementation.  We 
can always improve the implementation over time.

Bill Kempf