Author: Robert Hyatt
Date: 11:07:48 03/03/06
Go up one level in this thread
On March 03, 2006 at 02:48:04, Tony Werten wrote: >On March 02, 2006 at 16:35:17, Robert Hyatt wrote: > >>On March 02, 2006 at 15:00:42, Dezhi Zhao wrote: >> >>>On March 02, 2006 at 14:50:55, Robert Hyatt wrote: >>> >>>>On March 02, 2006 at 13:08:00, Alessandro Scotti wrote: >>>> >>>>>On March 02, 2006 at 11:45:14, Robert Hyatt wrote: >>>>> >>>>>>notice the index calculation above... from * 64 + to, which certainly gives a >>>>>>12 bit index into a 12 * 64 table. That has to be a bug, or else I am missing >>>>>>something. Note that someone just sent me the pieces above, not the entire >>>>>>fruit source so I have not looked at the surrounding code to see if there is >>>>>>more to the story... >>>>> >>>>>I haven't checked Fruit code on that, but I use two separate tables in Kiwi, a >>>>>2*64*64 table (side/from/to) for move ordering, and a 12*64 (piece/to) table for >>>>>reductions. >>>> >>>> >>>>OK. what is the justification? Have you compared them? (I have not tried the >>>>piece/to approach so have no data). Real question has to be is piece-to better >>>>than from-to in any sense other than being a bit more cache-friendly??? >>> >>>I tried. Not worse than from/to at least. >> >>I'll give it a whirl. Less data to copy on parallel splits since each thread >>needs its own history stuff... > >Does it ? > >From a clock cycle performance point of view I would agree, locking is to >expensive. From a search performance pov, I would rather use a "global" table. > >Tony You could be right. But there is "local" data in the history values. I'll try a global table to see what impact it has, as it certainly simplifies things a bit size-wise...
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.