SPARC2 FLOPS hw? sw?

[Jim Reynolds 37633]

I'm converting a major application from the Mac platform to the Sparc2, and did a
few trivial benchmarks to get an idea of the performance difference I might see.
I didn't spend very much time on these benchmarks, since I can't do anything about
the new platform, and there's  quite a bit of public literature on this subject.
A question arose however, about which I'm somewhat baffled.

Mac IIcx: replaced original 68030 with 40MHz Daystar 68030 accelerator card (comes
          with -new crystal, -40MHz 68030, -cache memory, -support circuitry, 
          -optimized math/sw routines/libraries.

Sparc2: -std. config., -std. AT&T 'C' compiler, ...

1. On 1 million loop test with floating point arithmetic operations only, the
   Sparc2 was approx. 3x as fast as the Mac.

2. Adding into above loop floating point trigonometric functions lowered the
   Sparc2's advantage to 2x over the Mac.

3. Adding in floating point exponential & logarithmic operations into the mix
   resulted in the Sparc2's performance lowering to 1/2! that of the Mac.

OK, these benchmarks were full of holes, but this question arose: Does the Sparc2
architecture have its floating point instructions hardwired into its RISC chip
like the 80486, as a separate co-processor like the 8088 <-> 80386 (unlikely), or
are these FLOPS all or partially emulated in software, or what?? I can't endure
the thought that a Mac is twice the speed of a Sparc2 in this niche area, so am
scrambling by checking out compiler directives, alternative libraries, ... .      


-- 
reynolds at aviary.kodak.com ( Jim Reynolds )