Author: Dan Newman
Date: 02:25:05 09/27/99
Go up one level in this thread
On September 25, 1999 at 13:56:29, Vincent Diepeveen wrote: >We should compare speeds. > >What i do after 1.e4,e5 2.d4,d5 is >generate the move list a few million times (n times). >Then i divide the time by (n * #semi legal moves). > >#nodes a second generated at a PII450 for diep is 15.5M > >Haven't seen anyone faster so far, though most are not >reveiling their speed. > >Note that if we have a merced within say 5 years that i >can implement bitboards for a few eval terms within a day by then. > >No need for rotated bb of course as i keep my current generator >and makemove. > >Majority of code is not faster using bitboards even at 64 bits >machine. > I tried this position out on the (non-rotated) bitboard engine that I'm just starting. I compiled with MSVC++5 and get 10.25 M moves generated/s on a P6/200. So, (ignoring cache speed differences) I should get something near 2.25 * 10.25 == 23.6 Mm/s on the PII450. OTOH, my (more developed) 0x88 engine gets 11.3 Mm/s on this position, which is slightly better than the bitboard program (roughly 25 million moves generated per second on the PII450). It seems like if I were to go to a 64 bit machine, the bitboard version would benefit somewhat more and would probably go faster than the 0x88. Of course these measurements, made by repetitively calling the move generator in a tight loop for the same position, may not tell us much because of (among other things) the branch prediction hardware. This will artificially inflate these numbers. In real life when the move generator is called with different positions and is continually getting blown out of cache, it may perform quite differently... Actually, the speed at which moves are generated is not a good reason to choose between bitboards and 0x88, or whatever. The last time I profiled my 0x88 program, the move generation was taking about 20%. The SEE was taking 25% and the eval 4%. If I ever get a decent amount of eval added (say 50% of cpu time) this will greatly depress the move generation figure. So a 10% change in move gen speed might change the overall speed by 1% in a well developed chess program. Now, if my (non-rotated) bitboards can help speed up the SEE or some of the (as yet unwritten) eval terms, it might be a good deal faster than the 0x88--but I don't really have a good feel for this yet. I do like the way bitboards replace the piece-list structures though. It was always a drag trying to figure out what to do with captured pieces, where to put promoted pieces, and so forth. All this goes away with bitboards. I think this is my favorite aspect of bitboards at this point. -Dan.
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.