New approach to floating point

[Mark SiblyKeymaster]

Hi,

Though this paper re: a new float point system was pretty interesting:

http://www.johngustafson.net/presentations/Multicore2016-JLG.pdf

Not 100% sure how it all works, esp. how you perform math, but I think the basic idea is very cool, esp. if it eliminates FP rounding problems.

Bye!
Mark

 

[impixiParticipant]

 

Here’s a link to Gustafson’s book on the subject:

Not read it yet myself, as I suspect it’s beyond my comprehension level.

Another thing that needs “redoing”, in my opinion, is character set representation, but that’s another topic…

EDIT: I’ve read the first few chapters of that book now and I’m actually finding it easily “digestible” and interesting…

 

[Mark SiblyKeymaster]

Cool, might have to buy that! Pricey though…

Going by this (p30)…

http://sites.ieee.org/scv-cs/files/2013/03/Right-SizingPrecision1.pdf

…the key trick seems to be the addition of the ‘uncertainty’ bit and the addition of ‘exponent length’ and ‘fraction length’ fields.

I assume there are 2 length fields ‘coz he’s talking about variable lengthed vars, but with fixed length vars surely you only need one, eg: it could contain the length of the exponent and the length of the fraction would just be some constant minus that? Or is this variable length thing part of what makes it useful.

 

[Author]

Posts