Author: Robert Hyatt
Date: 09:14:12 03/18/04
Go up one level in this thread
On March 18, 2004 at 11:02:23, Sune Fischer wrote: >On March 18, 2004 at 10:41:44, Robert Hyatt wrote: > >>On March 18, 2004 at 04:32:12, Sergei S. Markoff wrote: >> >>>Hello! >>> >>>First try to index history depending of piece that moves. It's a "free" >>>improvement. I don't know why Crafty not doing it. See me code (it complicated a >>>little, but works ok): >>> >> >>It didn't help at all when I tried this. Andm in fact, it hurt because it >>increased the "cache footprint" significantly... > >Do you use a unsigned int[64*64] size table? >That's a 16k table, which gets probed constantly. > >Maybe it will be good for nps (and nodes??) to try with >unsigned short[64x7x2] and age only to control overflow? > >:) > >-S. First, I am not sure everyone read Jonathan's paper. But he mentioned the possibilities of history[piece][to] and history[piece][from][to] in addition to what I am using history[from][to] I tried them all. from/to worked best for me. I have not tried the tests recently, but I can't imagine it changing. I did try short, long and long long. Short caused a few problems with overflow. I settled on long using depth*depth, where Schaeffer used 1<<depth for the increment. That overflowed _badly_ where it would not have at early 80's search speeds... long long didn't seem to help much and was definitely a bit slower.
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.