Author: Gerd Isenberg
Date: 12:33:59 02/26/06
Go up one level in this thread
On February 26, 2006 at 15:07:09, Keith Evans wrote: >On February 26, 2006 at 07:39:10, Gerd Isenberg wrote: > >>Hi all, >> >>First greetings and good luck to all cct8 participants! >>Second some thoughts and suggestions for improvements of Dann's switch-approach: >> >>Dann's approach, to do a switch of a masked row per square and ray-kind with 128 >>bitboard cases does almost the same as rotated lookups with rotated occupied >>state and square - it may supply appropriate attack bitboards for that ray-kind >>and square - and even other precalculated information such as possible covered >>xray information. Correct me if i'm wrong, Dann. >> >>Despite it was interesing to see how the compiler translates a switch with 128 >>64-bit cases in a binary search manner, Dann's switch-approach covers a lot of >>branch target buffer slots and branch prediction ressources. Also each maked >>move inside the search changes almost 7 occupied rays so that some >>miss-predictions are likely in the compare/conditional-jump chains of the >>switches. > >One thing that I was wondering about the original thread was is there's any way >to approach this the same way for game with boards with more than 64 squares. >Even if you don't worry about efficiency, can you even switch on say 128 bits? >(e.g. Chinese chess, capablanca chess,...) I wonder if your suggestions would be >useful there. I have to admit that it's a bit over my head at the moment... In principle, yes. A 64-bit switch as case of an outer 64-bit switch ;-) De Bruin folding needs a 128*128=128-bit product. Also, lookup tables indexed by occupied states double their size with every new square per ray. I think fill algos are great with 128-bit SSE2 for greater boards.
This page took 0.01 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.