braner@batcomputer.UUCP (06/08/87)
[reposting after the crash at SCORE]
Finally I got around to linking the NS32081 floating-point chip I have on my
ST to the Absoft FORTRAN77 run-time library. (By "linking" I mean patching
the library so that standard FORTRAN code uses the FP chip, transparently.)
Here are _approximate_ timings, all in microseconds: (32081 clock: 4 MHz).
op single precision double precision
SWFP HWFP SWFP HWFP
add 100 60 160 75
mul 130 60 300 80
div 330 70 900 90
And here, for what they're worth, are the Savage benchmark results:
System CPU / FPP CLOCK LANGUAGE TIME ERROR
(MHz) (Sec) Abs(a-2500)
Atari ST (68000/-----) 8.00 Absoft F77 V2.2 67.6 1.7 E-07
Atari ST (68000/32081) 8.00 Absoft F77 V2.2 16.3 3.1 E-07
That's faster than a 68020 at 16 MHz with no FP chip, and about twice slower
than the 8088/8087 combo at 5 MHz. But note:
(1) - The 32081 is here used as a peripheral, not a coprocessor. Most of the
time is spent on the communications, not the calculations! (And that's
after I hacked the cartridge port to be able to write to it. The usual
trick of writing to that port via the address lines would have resulted
in much slower performance.)
(2) - The 32081 does NOT have the transcendental functions built-in. The
above result is with the Absoft software library calling my hardware
drivers every time it needs an addition, subtraction, multiplication
or division. Optimized code could double the speed (I have tried that
with log() and sqrt()), and a 68881 chip would do a _lot_ better.
Note: I had to cut pins, solder wires and generally hack the ST hardware to
attach the 32081. This is not a commercial product or set-up. The
32081 is cheap, but the 68881 is now (finally!) affordable.
- Moshe Braner