Author: Gerd Isenberg
Date: 01:32:09 03/03/06
Go up one level in this thread
On March 03, 2006 at 03:12:26, Tony Werten wrote:
>On February 26, 2006 at 16:04:13, Gerd Isenberg wrote:
>
>>Applying the old trick - not using the outer squares of the masked occupied ray
>>- makes perfect 6-bit range hashing possible. Most often there are only 32
>>occupied states per ray - only if the source is an outer square, there are 64
>>states on a eight square ray:
>>
>> BitBoard preCalulatedAttacks[64][4][64];
>>
>> struct SMagicMask {
>> BitBoard mask;
>> BitBoard magic;
>> };
>>
>> // sq ::= 0..63 square index of a sliding piece
>> // dir ::= 0..3 kind of ray (two diagonals, horicontal, vertical)
>> occIdx64 = (occupiedBB & mm[sq][dir].mask) * mm[sq][dir].magic) >> 58;
>> attacks = preCalulatedAttacks[sq][dir][occIdx64];
>>
>>Generated mm-values below in a separate longer post...
>
>I can confirm your findings. There are a few points however.
>
>I couldn't get your values to work, in a few cases the magic number seemed to
>map different bitboards to the same index.
oups - really!?
Can you post or mail some of the values so that i can have a closer look and to
eventually find the bug in my checkMagicNumber(). I use a boolean array of bits
to lock found indices to ensure indices all are unique.
>
>I then run the deBruijn generator myself, to get stuck on 40 values or so, for
>wich it couldn't get a magic number.
>
>Then I just pumped random number into my checkMagicNumber() and within 10 milion
>tries, I got a complete set of magic numbers ?!
Ah yes, great idea.
>
>I have implemented them in XiniX, ran a couple of searches with a check wether
>the same moveBB was generated, and it works.
Nice to hear.
>
>It's a bit faster than my (strangly) rotated bitboards on a 32 bit machine. On
>64bit it will be way faster, since a lot of time is now spend in
>_allmull(BB,BB) which isn't needed on 64bit.
Yes, 64bit mul is only 4 cycles direct path.
Thanks for the feedback,
Gerd
>
>Tony
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.