Pentium and Repentium

Torbjorn Granlund tg at
Sat Mar 20 07:36:02 CET 2004

I've run some timing tests of the assembly loops on Precostt, and it is
slower than Willamette/Northwood.  No big surprise there.  The timings
fluctuate just as for Willamette/Northwood, only slightly less.

I'd expect we could get things to run well on this processor with
some software pipelining.  Our current code relies on out-of-order
execution, but Prescott's out-of-order capability seems
insufficient for its over 30 steps deep pipeline.

                Northwood       Prescott
mpn_add_n       4.25            4.5
mpn_sub_n       4.25            4.5
mpn_mul_1       4.0             5.0
mpn_addmul_1    5.5             6.0
mpn_submul_1    7.5             8
mpn_lshift      1.87            2.12


More information about the gmp-devel mailing list