Author: Robert Hyatt
Date: 12:49:21 07/01/04
Go up one level in this thread
On July 01, 2004 at 15:02:34, Gerd Isenberg wrote: >On July 01, 2004 at 11:13:14, Robert Hyatt wrote: > >>On July 01, 2004 at 02:50:35, Tony Werten wrote: >> >>>Hi all, >>> >>>although I like the principle of bitboards, it really bothers me that I can't >>>seem to find a decent/fast way to evaluate weighted safe squares. >>> >>>Suppose I want to (simple) evaluate a rook, I generate a bitboard with all >>>reachable squares and mask off the squares attacked by lower pieces (that's no >>>problem). >>> >>>(This doesn't exacly generate safe squares, only the ones that aren't attacked >>>at all by opponents pieces are, for the remaining squares one would need a SEE, >>>but that's not the point ) >>> >>>Now I can use this bitboard ( say rook on e4 ), mask the rank state, and look in >>>a precomputed table how this rankstate scores on an e rank. No problem. >>> >>>But how to do the files ? If I use the rotated board, I need to have the >>>opponents attackboard in this rotated board as well, wich would be very costly >>>to compute (ie also for the bishops,queens ) and very complicated. >>> >>>Any ideas ? Am I missing something ? >>> >>>BTW, doing a popcount isn't a solution, since it violates the elegance of >>>bitboards ( and is slow ?) >>> >>>Tony >> >> >>On the Cray there is an elegant solution, but not on X86 so far... >> >>You can create a 64-word vector of "weights". How you compute these is up to >>you. In Cray Blitz I did this as I did the evaluation, figuring out which >>squares were weak, unimportant, strong, useful, painful for opponent, etc. >>After the normal eval, I had a vector of values, one per square for all squares >>on the board. Now I computed the "attack bitmap" for a piece, and stuck that in >>the vector mask register. Now when I sum up the square value vector, it only >>sums the values with a corresponding bit mask of 1, meaning this piece attacks >>that square safely. > >Wow great, a scalar product 64word*64bit. >Was it implemented in hardware or a kind of micro-program? Took a couple of instructions. "vector mask" selects the words you want, you pipe them into a "reduction" operation successively to collapse N words to 1 final sum. This "chains" so it takes essentially no extra time to do, which is cute.. :) But no similar facility on non-cray cpus to date... > >> >>I obviously don't do that at present, since X86 has no such direct capability >>and the software approach is expensive... > >Thinking about some oppropriate SSE2-instructions for that scalar product, eg. >64 bytes * 64 bit. Four 128-bit (16Byte) xmmm registers where each byte is >associated with one bit of the other operand. > >One subtask, may be the most expensive, is to expand each bit to one byte, so >that 1 becomes 0xff. From 64-bit word to four times 128-bit words. > I hate corresponding with you. I end up with a _headache_ every last time you start that stuff. :) >Four parallel "and" and some final horicontal (3DNow psadbw is fine for that)and >vertical (over the four xmm-registers) parallel adds. > >Ok it will take some cycles, anyway... > >> >>In any case, I believe there are plenty of viable alternatives to this kind of >>mobility concept that it isn't that important... IE mobility has many forms of >>expression. A rook on an open file is one such idea...
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.