Author: Robert Hyatt
Date: 20:06:28 12/16/02
Go up one level in this thread
On December 16, 2002 at 20:32:17, Matt Taylor wrote: >On December 16, 2002 at 16:31:21, Robert Hyatt wrote: > >>On December 16, 2002 at 16:03:11, Miguel A. Ballicora wrote: >> >>>On December 16, 2002 at 14:51:18, Robert Hyatt wrote: >>> >>>>On December 16, 2002 at 13:32:46, Dieter Buerssner wrote: >>>> >>>>>I did similar tests. IIRC, I got differences of > 5% faster for, what you called >>>>>"new". I unfortunately do not have the code for the "old" (2 different tables) >>>>>handy anymore, to recheck. You are not probing in quiescence search, so it is >>>>>reasonable to expect, that your differences are smaller. >>>>> >>>>>Anyway, for a very small code change (I would guess not more than 1 hour), >>>>>getting 5% is not so bad. >>>>> >>>>>Regards, >>>>>Dieter >>>> >>>> >>>>No, although for me it was 2% which is not so hot... >>>> >>>>I will probably stick with it since it is done, of course.... >>> >>>Did you make sure that your table is aligned avery 16 bytes and not 8, for >>>instance? >>> >>>Miguel >> >> >>Yes. If you look at init.c you will see where I force the starting address to >>be a multiple >>of 16 regardless of what malloc() gives me. >> >>However, my entries are 16 bytes long, and have three in a "set". No way to map >>a "set" to >>the front of a cache line every time. It is going to be staggered no matter >>what... > >You can if you don't mind bloating the size of the hash by 33%... > >-Matt There is a problem. That reduces the hash size by a factor of two. With the current three entry approach, I can use 3/4 of memory for hashing. If I go to four entries, I have to back up to 1/2 of memory. Also there is a problem mapping to one of three (four entries, one depth-preferred and three always store) entries vs the current approach that maps to one entry to get all three then uses one more bit in the signature to choose between the two always-store entries...
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.