Author: leonid
Date: 19:34:32 12/22/00
Go up one level in this thread
On December 21, 2000 at 11:23:48, Ulrich Tuerke wrote: >On December 20, 2000 at 12:31:51, Thomas Mayer wrote: > >>Hi Dann, >> >>>Phalanx has a really slow motor. Usually 100K NPS or so. Gullydeckel has a >>>barn-burner -- often well over one million NPS on my machine. If we could >>>couple Gullydeckel's move generator with Phalanx's eval, it might rule the >>>world. >>you allways tells us how incredible good Gullydeckel's move generator is... >>can you give us some numbers ? >>Here is my testing position, Vincent Diepeveen has put my nose on it to look how >>things are going: >> >>[D] rnbqkbnr/ppp2ppp/8/3pp3/3PP3/8/PPP2PPP/RNBQKBNR w KQkq d6 0 3 >> >>When I start with Quark, Quark produces in this position about 3.500.000 moves >>per sec. (measured this just with a huge for loop and allways call move gen and >>after that calculate created moves divided by seconds) >>At the moment Quark creates 8.300.000 moves per second here on PII/400 >>Vincent Diepeveen said to me that Diep creates on PIII/450 here 15.500.000 moves Those numbers sound to me as too high after my verification. Even if all those moves are not verified. Strange! Just put counter on my program and counted how many completely legal moves are generated in one second on AMD 400Mhz. Number is very close to what I expected: 1.600.000 moves in one second. My number for this position I expected since average NPS for this position, when search done by brute force is 233.324. Program usually generated 5 times more moves that it used. So it generate even without any special counter somewhere around 233k*5=1165k per sec for this position. Without any evaluation it jump up to 1600k. But never ever 13000k. >>per second and crafty about 7.700.000 moves per second on PIII/450 >>Can you give us some numbers what Gullydeckels move generator does here... >>(Quark and Diep have 0x88 board representation, Crafty we all know... :) >> >>Greets, Thomas >> >>P.S.: Working on move gen is worth a lot, it is incredible, what the improvement >>brings in kNPS... And also it is really funny... At the beginning I thought my >>move gen is really fast... then comes Vincent... some hours coding brings me to >>7M moves... and at this point I had no idea how to get faster... but it seems to >>be never over... I think also for Quark there is still much to improve here... >>when I only think about Vincents number... :) > >Hi Thomas, >I can't agree fully here. At least for Comet, this is not true. I have analyzed >by means of a profiler some time ago that comet consumes its time essentially in >evaluation, and partly in MakeMove (=update of data structures). The time which >is consumed while in pure mive generation is very small compared to these times. > >So I concluded that I can't really win much by optimizing move generation. Even >if I get a factor of 2 there, I will hardly notice in the end (in the final nps, >including everything). > >May be other progs waist their time elsewhere ? >Regards, Uli
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.