Author: Joost Buijs
Date: 07:14:02 08/08/03
Go up one level in this thread
On August 08, 2003 at 09:58:23, Vincent Diepeveen wrote: >On August 08, 2003 at 09:28:58, Joost Buijs wrote: > >If evaluation is only thing you care for, you >should really use a different concept that you previously used. > >Please talk to some very strong ex-bitboard programmers who >did pretty well in world champs and they all threw out the >bitboard code for evaluation. Who was that if I may ask? > >As i already said 5 years ago, i can only see the use of a >bitboard for pawns. And then not the crafty definition which uses > >whitepawns and blackpawns > >but a > >pawns[2] I use two structures containing the pieces for white and for black. Adressing them goes like side[white].pawns or side[wtm].pawns, this is easier then some other progs use and I even found it is somewhat faster with most compilers. I can also adress them with a pointer like own = &side[wtm]; and then own->pawns or opp->pawns works very convenient. This is better then the stuff like if (wtm) then {...} else {...}. > >which is a better form of programming than writing out everything for >black and white. > >Also is better for trace cache. > >So it is no mystery why each new CPU diep speeds up more relatively than other >programs. Itanium2 i get more out of than crafty, despite that it goes from 32 >bits to 64 bits for crafty. opteron i profit more than crafty from and so on. > >Not to mention when the prescott gets released. of course P4 was first 2 steps >back before moving a bit forward, but just consider the posted code here. > >Very P4 friendly! > >>On August 08, 2003 at 07:38:41, Vincent Diepeveen wrote: >> >>>On August 08, 2003 at 02:35:00, Joost Buijs wrote: >>> >>>>Vincent, >>>> >>>>Have you ever tried bitboards? >>> >>>why don't you try this code and find out it is 2.2 times faster than what crafty >> >>I really think it doesn't matter much because move generation takes only a very >>small fraction of the total time involved. Speeding up your move generator by >>100% gives you an overall gain of maybe 1.5%. Why on earth do you want to >>generate 70 million moves/sec. If you can only evaluate at a rate of lets say 50 >>thousand moves/sec.??? I've been making this mistake for years, trying to speed >>things up more and more. But in the end the most important thing is the >>evaluation function. 10 or 20% difference in speed is hardly noticeable in >>playing strength. >> >>>has. i measured this against crafty. >>> >>>please don't complain about L2 cache needs of crafty's move generator >>>because when you generate semi legal move list for a few million times then >>>crafty's move generator is within L1 cache even. >>> >>>In fact this is 0.6% system time so you would save out 2.4% directly. >>> >>>But if you remove the 'if' condition to put the move in the movelist, >>>then you will notice that this is completely branchless code for 1 piece with >>>exception of the loop. >>> >>>So where 0.6% system time in move generator already saves me out 0.6% system >>>time which is worth the effort of some months of fulltime programming as you >>>know, even more important is that all scans in my evaluation that involve semi >>>legal moves, is completely branchless. >>> >>>And there you have a major winner in system time as you already noticed. >>> >>>I can scan more and i can scan it faster than most other software. >>> >>>And there is your answer. >>> >>>Such code like this is a lot cleaner. Bitboards is not easier to evaluate. It is >>>much harder. Also my code is trace cache friendly so to speak. If i evaluate a >>>pattern i prefer general code which works both for white and black. >>> >> >>Evaluation is a lot easier with bitboards. I went really crazy about al these >>'if then else' constructs in my old engine. Now things can be done very easely >>with just a few shift and mask operations. >> >>>That is not always easy to write and in not all cases it gets done, but which >>>single bitboard program is using such general code? >>> >>>the only use i see in bitboards is a pawn bitboard so that complex pawn patterns >>>require just one 'if' instead of a bunch. >> >>Here you already begin seeing the light! >> >>> >>>I feel your decision was wrong. With this code you can get 2 times faster in 32 >>>bits. >> >>I agree on this, but soon everything will be 64 bits, one or two years from now >>32 bit machines will be history. >> >>> >>>Also the advantage is you might get again interested in improving the evaluation >>>function. >> >>Since I've impemented bitboards i'm more interested in the evaluation function >>than ever before. So beware! >> >>> >>>Best regards, >>>Vincent >>> >>>>I agree that on a 32 bit machine it takes a tiny bit longer to generate all the >>>>moves with a bitboard engine. The whole move generation process takes ~3% from >>>>the time the engine needs to determine a move, so thats is not important at all. >>>> >>>>The evaluation fuction is much cleaner and more convenient to handle when using >>>>bitboards, probably you can gain a lot here. >>>> >>>>About two years ago I rewrote my old engine completely, and since that time it >>>>is using bitboards. I still think I made a good decision by doing so. >>> >>>>In fact my program runs even faster after the conversion to bitboards. All other >>>>things like the search and evaluation parameters are exactly the same as in the >>>>old 1991 engine, so I am really comparing apples to apples here. >>>> >>>>B.T.W. I think the move generation is not that slow at all, for instance a perft >>>>6 from the starting position takes here 6.24 sec. on a fast Athlon-XP, that's >>>>about 8.5 million moves/s. Can you tell me what your numbers are with respect to >>>>this? I'm really curious about this. >>>> >>>>Next year I will certainly get a dual or quad Opteron, I think my bitboard >>>>engine will really scream on such a machine. >>>> >>>>The development of my engine has been frozen for 10 year or so, simply because I >>>>was very busy doing other things. But recently I decided to start developing it >>>>further. This year I won't attend the CSVN tourney, but I certainly want to be >>>>there next spring. >>>> >>>>Groetjes, >>>>Joost
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.