Subject:
|
Re: FW: Something else is needed, I think...
|
Newsgroups:
|
lugnet.robotics
|
Date:
|
Thu, 6 May 1999 07:48:23 GMT
|
Viewed:
|
1079 times
|
| |
|
|
In lugnet.robotics, Todd Lehman writes:
> 28.4 is a pretty good general-purpose fixed-point representation. It gives
> a domain of -134,217,728 to 134,217,727.9375, and you can square numbers up
> to 11,585 without overflow. 28.4 is also well-suited to vector graphics on
> relatively low-resolution devices, for example, where you have oodles of dot
> products and sub-pixel coordinates.
>
> A couple of other popular representations are 24.8, 16.16, 8.24, and 2.30,
> but these are progressively limited in the magnitude of domains they can
> safely handle.
Thanks Todd for your suggestions. 16.16 would probably meet my requirements
for the particular task I am working at present moment. Anyway I suppose I can
write general-purpose code to be used with different radix points.
> In choosing where to put the radix point, it's important to consider what
> sorts of operations can occur in the problem domain. Will there be lots of
> adding or lots of mulitiplying? Multiplying large numbers by small numbers
> only, or multiplying large numbers by large numbers? Summing small
> quantities of large numbers or large quantities of small numbers?
>
> For example, 16.16 is probably excellent for neural networks, but 8.24 --
> though it would provide much more detail to the all-important numbers inside
> the unit interval -- would likely fail from overflow when summing complex
> columns of weight vectors.
I made all these considerations in writing my code (I'm not completely new to
numerical analysis). Simply took the wrong way using decimal instead of binary
radix and this made the implementation a little bit more complex and slower.
I'm going to rewrite my project with radix 2 fixed point precision.
Thanks for your help
Mario
|
|
Message is in Reply To:
 | | Re: FW: Something else is needed, I think...
|
| (...) 28.4 is a pretty good general-purpose fixed-point representation. It gives a domain of -134,217,728 to 134,217,727.9375, and you can square numbers up to 11,585 without overflow. 28.4 is also well-suited to vector graphics on relatively (...) (25 years ago, 6-May-99, to lugnet.robotics)
|
32 Messages in This Thread:
 
- Entire Thread on One Page:
- Nested:
All | Brief | Compact | Dots
Linear:
All | Brief | Compact
|
|
|
|
