On Wednesday 19 September 2007 14:20, Timothy Normand Miller wrote:
> We've talked a lot about multipliers recently, and in the past, we've
> talked about floating point adders, multipliers, and reciprocal.
>
> Another interesting function that might be good to implement is the
> logarithm.  In fact, I think we may actually need that to do
> MIP-mapped textures.
>
> But there are more general applications.  I've studied various
> problems in AI where numberous probabilities must be multiplied
> together, for instance in hidden markov models. (HMMs).  To avoid
> having to multiply smaller and smaller numbers, people commonly switch
> to log space.  Instead of multiplying small positive numbers, you add
> large negative numbers.  Logs of probabilities are commonly referred
> to as "likelihoods" or "log likelihoods."  I don't remember enough
> about the math to explain the advantage, but apparently, there is one.


Am I too old? Adding logs instead of multiplying (And subtracting instead of 
dividing) was how I learnt to do multiplication adn division before you werte 
allowed to use calculators in school exams (Early 1980's in NZ :)

In the mid-late 80's we also used methods like this for steam calculations 
(Water table calculations were quite unweildy things to work with in those 
days. I wonder whether they still use tables or calculate that nowadays).

Didn't some early processors (e..g Pentium floating point bug) use a similiar 
method for multiplication and division by keeping what was essentially a log 
table in ROM and the microcode would lookup a log, add/subtract & convert 
back instead of trying to do multiplication/division longhand?

Come to think of it, would it be more efficient (Thinking of time & logic vs 
table lookup space) to use a lookup table to caluclate log/alog like using 
Eton tables in the old days? Or actually calculating it?


Hamish.
_______________________________________________
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