Author: Andreas Guettinger
Date: 10:21:39 01/13/03
Go up one level in this thread
On January 13, 2003 at 04:22:06, Bas Hamstra wrote: >On January 12, 2003 at 21:01:01, Russell Reagan wrote: >>On January 12, 2003 at 20:25:01, Bas Hamstra wrote: > >>>I test it with all pieces being still on the board, while the square in >>>question is not attacked. So a worst case scenario. >> >>In that case, the bitboards should perform about equally fast in any position >>whether there is an attack or not. The 0x88 isn't the same though. It will be >>slower if there is no attack, and faster if there is. So if you used more >>positions (not just a few, but a lot, like during a search) you might find it\ >>to be faster overall. Gerbil is faster than Crafty, sometimes almost twice as >>fast, and so that speed has to come from somewhere. > >Thanks for your data about Gerbil, this motivates me again :-) It is difficult >to predict how fast a program is going to be. I remember the GNU guys tested all >critical functions when they switched to bitboards, and they were faster than >the old versions. But when the rewrite was finished, the nps went *down* >somewhat. > >Yes, where does gerbil's high nps come from? I have a few ideas without knowing >Gerbil too well. > >- it probably does not do SEE sorting in the qsearch (?), which Crafty does do, >this really kills nps but not overall searchdepth. In Tao I assign SEE scores to >ALL captures which is even worse than Bob does (he does not calculate SEE if >Defender > Attacker or a capture is considered "futile"). Don't assign all SEE values in qsearch. This will kill your 0x88 generator, maybe it works better with bitboards. I prune bad captures in qsearch similar to crafty _before_ calculating an SEE value for them. I may have not the best SEE function (it's not so easy for 0x88), but this alone gave me a doubling of NPS. I search you can savely calculate all SEE values, because the slightly better move ordering compenses for the time. regards Andy
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.