Re: Gold Hill's floating-point

[coffee@aerospace.aero.org]

Bob Mathis notes, correctly, that my earlier reply to samalone assumed
that you had all seen the original message. Mea culpa. The problem that
_he_ had fixed was Gold Hill 286's habit of returning _weird_-looking
floating-point results (multiple decimal points, etc.) after a period
of use: apparently, this is due to pushing to DOS and leaving the 8087
or 80287 in an odd state that Gold Hill's interpreter does not correct
on returning. The fix is to make the call (SYS::8087-FPP) after returning
from DOS. I have my own (dos) function defined in USERINIT to provide
a "press a key to return to Lisp" prompt if (dos) is called with an argument,
so adding this call to that function has fixed the problem.

The Hummingboard release uses the call (SYS::8087-FPP :emulate) in view
of the absence of a physical numeric chip on that board, and I have seen
no such bizarre behavior -- not this specific variety, I mean -- on it,
so there is no apparent reason to do this on the Hummer. The general problem
of floating-point creativity in GC 386 Developer continues, I'm sad to
say: how about (log 1024 2) --> 9.53638... to double precision? The same
result also occurs with single- or double-precision real arguments. Further,
another user tells me that regular old non-Hummingboard GC386 does the
same thing on her "Inboard" 386 card.

BTW, I have mentioned the substantial enhancements to the editor as a major
advantage of the Gold Hill 386 vs. 286 products; I learned yesterday that
a new, compiled, enhanced editor (sounds exactly like the 386 version) is a
key feature of the GC286 2.2 upgrade, $90. Since we're not seeing nearly the
factor of five speed increase that Gold Hill claims versus "an AT" (maybe
they ran against a 6 MHz machine? We're seeing speed ratios of from one to
four with a Hummingboard in a DeskPro 386 against an 8-MHz DeskPro 286),
consider this in your cost-effectiveness evaluation. A twelve- or sixteen-MHz
80286 with adequate RAM may be quite competitive, especially if (like us)
you're doing pattern recognition in telemetry streams or other number-
crunching things. Of course, the Hummingboard can be dropped right into any
XT clone with a decent hard disk, an important factor if you already
own such a box.

I'm sorry if this extended discussion of microcomputer implementations has
annoyed anyone: in general, I direct such comments to lang.lisp, and will
continue to do so now that the loose ends on this exchange are (I believe)
tied off. I would also like to stress that Gold Hill's people have been
prompt, courteous, and helpful in our discussions of these problems and that
I have every confidence that they will be quickly resolved.

						Regards, Pc^2