🚀 go-pugleaf

schmid@supernova.commodore.com (Bob Schmid) writes:
>Floating point benchmarks we have done with gcc 2.3.3 on a Sparc 10/41
>have been disappointing, about 30% faster than a Sparcstation 2.  We
>were hoping for 2.5x-3x speedups.

You were, perhaps, hoping for too much.  I'm puzzled why you're blaming
gcc for this; I think perhaps you were too gullible when listening to Sun
marketing people.

>Does the latest version of gcc implement SuperSparc optimizations?

By default, gcc generates code for version 7 of the Sparc architecture.
You can get an improvement on some code (especially code that does a
lot of integer multiplication) by giving the -mv8 switch, which tells
the compiler to use the Version 8 instructions (basically, this adds
integer multiply and divide instructions, instead of using subroutines
to do these).  I don't think this will help you on code dominated by
floating point computation.

>The benchmarks were done with OS 4.1.3, and gcc -O.

If the benchmarks access a lot of memory, there are situations where the
Sparc 10 is actually worse than the Sparc 2, because a secondary cache
miss on several models of the Sparc 10 architecture is so expensive.
This can happen for numerical applications that go sequentially through an
array in memory that exceeds the size of the secondary cache; people have
posted benchmarks to comp.arch that make the Sparc 10 look really bad on
this kind of code.





--
Joe Buck	jbuck@ohm.EECS.Berkeley.EDU