Re: [Xsb-development] Regarding XSB floating point changes

[Kostis S. <ko...@us...>]

Bart wrote:

 > > The very quick benching that we did (e.g. like adding up 1000000
 > > floats in a list) was something like 50-70% slower than the old way.
 > > To me that's not bad at all, seeing as how we're now manipulating 4
 > > words instead of 1 each time.
 > 
 > You should add "and now doing the IEEE doubles [or floats] RIGHT".
 > I know that "Worse is better", but in this case XSB would really win
 > by doing it right. Don't compromise with floats :-)

I am very much with Bart here...

I can of course understand arguments of shortage of developer power
or having other more urgent implementation priorities, but I have
serious trouble interpreting performance arguments when comparing
a serously crippled implementation of floats with an IEEE one.

Also, I do not understand what's the rationale of implementing IEEE
floats as boxed 3-word integers rather than implementing them as a
float-tagged header followed with two words (on 32-bits machines)
or one (on 64).

Cheers,
Kostis