I'm not so good at this, but it would be good for us to incorporate
some rigor into our work.

Yesterday, I made an erroneous statement about the multiplier.  Since,
except in denormalized case (where it doesn't matter), every operand
is going to have a 1 bit on the left, which means the product doesn't
have to be dynamically shifted.  Two 24-bit number, with a 1 on the
left will always beget a 48-bit number on the left, and we can throw
away the lower bits to get another 24-bit number that is already
normalized.

So what I would like is an analysis of the rounding requirements.  Not
practical ones but theoretical, on accordance with IEEE's 3 bits of
extra precision involved in rounding.  (This is actually trivial to
implement in hardware, regardless of how we want to minimize the
hardware.)  But given 1.X * 1.X, what are the POSSIBLE combinations of
bits in the guard bits, and how do those affect our rounding results
when we squeeze the number back into a float32?

(There are better docs, but I found this quickly:
http://www.cs.nmsu.edu/~pfeiffer/classes/473/notes/fp-extras.html)


-- 
Timothy Normand Miller
http://www.cse.ohio-state.edu/~millerti
Open Graphics Project
_______________________________________________
Open-graphics mailing list
[email protected]
http://lists.duskglow.com/mailman/listinfo/open-graphics
List service provided by Duskglow Consulting, LLC (www.duskglow.com)

Reply via email to