Re: Floating point types in MCL...

[gat@forsight.jpl.nasa.gov (Erann Gat)]

In article <43702@mimsy.umd.edu> waander@cs.umd.edu (Bill Andersen) writes:
>
>   Recently I built a toy neural net program using MCL.  To my dismay,
>I discovered that all floats are of type double-float in MCL, no
>matter what you declare them.  So, every time I run my net, doing many
>floating point additions and multiplications, I *CONS LIKE THERE'S NO
>TOMORROW*!!!  For example:
>
>  (time (* 1.0 1.0)) returns something about 16 bytes being allocated.
>
>  This is unacceptable.

Check out FPC, a floating point accelerator program I wrote for MCL.
FPC speeds up floating point math, doesn't cons intermediate results,
and allows destructive floating point operations so you can do float
math without consing at all.  FPC is available for anonymous FTP
in the contrib directory at brazil.cambridge.apple.com.

Erann Gat
gat@robotics.jpl.nasa.gov