Author: Gerd Isenberg
Date: 07:19:25 12/20/02
Go up one level in this thread
On December 20, 2002 at 08:22:15, Sune Fischer wrote: >On December 20, 2002 at 08:00:31, Gerd Isenberg wrote: > >>>The overhead of making/unmaking pinned pieces is extremely small. >>>I count about 0.5-1.5% pinns, the make/unmake move rutines take maybe 20% (in my >>>case) but I only need to do half a move to see if the move is legal, in total >>>that's about 0.075% overhead because of illegal pinned moves!! >>> >>>I'm pretty sure there is _no way_ you can detect pinns faster and improve on >>>this overhead. >> >>Hi Sune, >> >>I'm not sure, specially if we consider hammer or SSE2. I use dumb6fill for 2 or >>3 directions simultaniously (8 mmx-registers). With SSE2 or hammer it can be >>done for all 8-directions in parallel, with sliders and king in one 128-bit-xmm >>register. > >A lot of work to save 0.1%, at best :) Ok, i use two cheap conditions before and the whole mmx-stuff is rather fast. If i eval a leaf-node with pinned pieces, that may be get attacked and conquered by a pawn-push, i safe probably one or two plies. > >>>I agree, there are other reasons to do it. It must be good for extensions, as >>>Uri says a pinned queen might be reason to extend. >>> >>>One of the best things though, must be if you generate all moves all the time, >>>that would be real nice for extensions and threat detections. >>> >>>But I didn't think any bitboarders generated all moves since bitboards are so >>>good at incremental generation, is this what you do? >>> >>>-S. >> >>I guess yes, but incremental generation is possibly the wrong term. I don't look >>to previously generated moves two plies before. It's a finite state machine move >>generator, trying to generate as less as possible, hoping for a cutoff. > >Yes, that's what I figured :) > >> But my >>final approach with hammer in mind is quite unclear so far. > >The question is, how much do you need those extensions? Yes, i consider pinns in interior nodes for fractional extensions too. I even consider whether the piece was pinned one ply before. >I also like the full knowledge idea, but on the other hand if you slow down >the program 100% just to produce a good extension here and there, it might be >better just to run brute force faster and get the extension simply because you >go that extra ply. When going an extra ply you also see a lot of other stuff of >course. > >-S. OK, i also have some experience with other approaches, and i do not claim that legal move generation is the "best" solution. Gerd
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.