[gambit-list] Suggestion for configuring gambit on high-end x86 processors.

Bradley Lucier lucier at math.purdue.edu
Wed Feb 16 00:56:59 EST 2005


Marc:

High-end P4, Athlon, etc., as well as Athlon 64, have 64-bit 
floating-point instructions pretty much the same as any other RISC (I 
don't mean the fused multiply-add of IBM's Power).  This is accessed 
through the SSE2 instruction set on Intel, and similar things on AMD's 
Athlon.

Some algorithms in _num.scm (for converting strings to and from 
floating-point numbers, for example) rely on proper 64-bit 
floating-point arithmetic.  Gambit simulates this by setting the 
rounding precision to 64-bits on x86's 80-bit floating-point stack, but 
there is still the possibility of double rounding, especially for 
denormalized numbers, because of the extended exponent range of 80-bit 
numbers on the floating-point stack.

I'm thinking it might be a good thing for Gambit's configure script to 
test for the availability of SSE2 or the similar thing for AMD and to 
set those options when building Gambit with gcc.

Brad



More information about the Gambit-list mailing list