Author: Dan Newman
Date: 19:30:13 01/29/01
Go up one level in this thread
For comparison with the bitboard data, here's what Skyblue
(my old 0x88 program) gets:
On a P3/600 (non-coppermine):
position captures non-captures both together
------------------------------------------------------
Vincent's 1.8 33.4 17.7
Dann's 0.0 57.1 30.3
On a P3 (coppermine) at 980 MHz:
position captures non-captures both together
------------------------------------------------------
Vincent's 3.1 56.1 29.7
Dann's 0.0 95.2 51.3
Hmm. Very interesting. Here is the P3/980 data from above
compared to the P3/980 bitboard data from my previous post:
Vincent's position:
movegen type captures non-captures both together
-----------------------------------------------------
bitboard 6.1 62.3 43.6
0x88 3.1 56.1 29.7
Dann's position:
position captures non-captures both together
------------------------------------------------------
bitboard 0.0 72.6 53.7
0x88 0.0 95.2 51.3
The bitboard ends up edging out the 0x88 because its capture
generator is much faster. Given a sparsely occupied board
(Dann's position), the 0x88 is a good deal faster generating
non-captures, but the bitboard capture generator gets done a lot
faster--and that's the generator that really counts.
Move generation speed is not really why I favor bitboards though.
I find it a lot easier to deal with bitboards than piece lists.
With piece lists you have to make a lot of decisions about whether
to leave captured pieces in the list and just mark them as
captured, or to contract the piece list and whether to do this
on-the-fly while searching or do it only at the root. Bleh.
Bitboards are a lot simpler in this respect.
I think my bitboard SEE is much faster than the 0x88's too, though
I haven't profiled either recently.
On the P3/980 my 0x88 program currently gets 946 knps on WAC at
1s/posn. The bitboard program gets 1044 knps, and that's with
significantly more evaluation code.
The 0x88 program hasn't got much beyond material eval, just a little
pawn shield code, bishop pair, and some eval for pieces bearing on
the king's position. The bitboard program has all that in addition
to piece square tables, a lot more pawn eval, bonuses for rooks on
7th, rook behind passed pawns, bishop mobility, trapped knight,
outposts, and so forth...
So, for me at least, bitboards are a fairly big win over 0x88 and
over anything else I've tried.
-Dan.
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.