Author: Vincent Diepeveen
Date: 11:00:17 06/30/04
Go up one level in this thread
On June 30, 2004 at 13:22:31, Matthew Hull wrote: >On June 30, 2004 at 13:15:02, Vincent Diepeveen wrote: > >>On June 30, 2004 at 13:13:19, Matthew Hull wrote: >> >>>On June 30, 2004 at 13:08:48, Gian-Carlo Pascutto wrote: >>> >>>>On June 30, 2004 at 13:06:24, Matthew Hull wrote: >>>> >>>>>On June 30, 2004 at 12:57:26, Vincent Diepeveen wrote: >>>>> >>>>>>Hello, >>>>>> >>>>>>For diep moving from 32 to 64 bits is not so interesting. But the 8==>16 >>>>>>registers is. >>>>>> >>>>>>All tested with same diep version : >>>>>> >>>>>>at windows 64 bits : >>>>>> >>>>>>net2003 32 bits : 135k nps >>>>>>amd SDK 64 bits : 130k nps >>>>>>gcc 3.4 32 bits : 130k nps (with pgo) >>>>>> >>>>>>at linux (gentoo) 64 bits : >>>>>>gcc 3.3.3 64 bits : 140k nps >>>>>> >>>>>>So it's trivial that moving from 8 to 16 registers increased it by 10k >nps, >>>>>despite also losing to larger instruction sizes. >>>>> >>>>> >>>>> >>>>>If you had gone with bitboards, you would have seen a 47% increase. >>>> >>>>Which would have been no help if it had slowed him down 47% in 32 bit mode >to >>>begin with. >>> >>> >>>Do you think bitboards are that much slower on 32 bit machines? >> >>As proven 2 years ago clearly by me it is 2 times slower. >> >>Diep generates at 32 bits exactly 2 times faster than crafty. > > > > >But "generates" is not a very big part of the program. The program as a whole >is 47% faster, not just move generation. the invariants are used everywhere in evaluation to loop scan or whatever. it's the basic principle simply that scanning in a simple 32 bits way goes faster than getting first a bit out of bitboard and combining a lot of bitboards with each other. In 32 bits that's a lot slower. Factor 2.2 initially, then some clever guys managed to speedup crafties bitboard routines and it was factor 2.0 then. > > > > >>Tested at K7. At P4 difference is bigger though. >> >>> >>>> >>>>-- >>>>GCP
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.