Floating point hardware/software

[David F. Carlson]

In article <1397 at qiclab.UUCP>, neighorn at qiclab.UUCP (Steve Neighorn) writes:
> In article <682 at amelia.nas.nasa.gov> wgreene at amelia.nas.nasa.gov (William H. Greene) writes:
> >
> >                      Weitek 1167
> >
> >     A lot of companies are talking about this coprocessor board. Is
> >there anyone out there actually using it under Unix/Xenix ? Are there
> >versions of Unix/Xenix that support the 1167 available now ? I know
> >top, I tend to wonder more. Anyone have any experiences with this ?
> 
> -X143 specifies the generation of Weitek floating point code instead
> of 80387 code.
> 
> -X213 specifies using the Weitek software emulation library, and
> requires -X143 to work. 
> Steven C. Neighorn            !tektronix!{psu-cs,reed,ogcvax}!qiclab!neighorn

There is no kernel support in 2.2 SV/386 for Weitek.  There is a module but it
contains stubs and is only 888 bytes long!  A beta kernel for an unreleased
version (2.3!!! ~fall 88) contains a whole weitek driver (14k bytes).  The 
GHC *will not* work without this kernel as floating point is protected in
this 386 version of UNIX (for right or wrong).  Actually, the "big" win
of protected floating point is that weitek, 287, 387, etc. or emulation can
be used without modifying executables (ie changing to -lm68881 as our
Motorola cousins have to do.)

Look for full weitek support real-soon-now.  Seriously!

One problem I haven't seen discussed much is that the weitek costs 2X a 387
and performs add/sub/mlt/div ~4X but floating point conversions, trig and
other transcendentials much slower that the 387.  What is the "real" world
speed up over its Intel competitors?  Not as much as one might think I posit.



-- 
David F. Carlson, Micropen, Inc.
...!{ames|harvard|rutgers|topaz|...}!rochester!ur-valhalla!micropen!dave

"The faster I go, the behinder I get." --Lewis Carroll