Stockfish v2.1 now available

[mcostalba]

So far I see:
[ul][*]Re-merging of most remaining duplicated code (search() & root_search(), and code common to pawn & material-imbalance hashes)
[*]Removal of UCI knobs controlling search extensions
[*]Briefer opening book implementation
[*]Removal of "banner" comments, tighter text-wrap of comments, micro-reorganizations
[*]Header file reorganization[/ul]
[quote="muxecoid"The reported NPS is much lower than 2.0.1. Is it counted differently now?

Thanks for the summary, actually is quite accurate

Also works fine with the 32-bit Linux compile. Strange.

Yes, I will try to investigate when I have a bit of time, probably this week-end.

[Gian-Carlo Pascutto]

People who want to squeeze the extra speed out and compile with GCC 4.6 can manually add the extra compiler flag in Makefile.

If you actually read my post, you will see that's what I did and it seems to expose a bug, even in GCC 4.6.

[Gian-Carlo Pascutto]

Well, I think most Linux distros (at least debian stable) ship with GCC 4.4.

FYI, the latest Ubuntu ships 4.5.2.

[Laskos]

Wow really impressive so far. Stockfish 2.1 - Rybka 4.1 24,5 - 15,5
and Stockfish 2.1 - Critter 1.01 22,5 - 17,5 (2.0 scored 16 - 24 and 20 - 20 in same matches)! Surprisingly engine size is only 1/3 of 2.0 size.

Jouni

It is also a very good example of how little one should rely on such short matches for statistical evidence. Since there is no question R4 is stronger than Critter, the fact that it scored worse than Critter in parallel matches is helpful reinforcement.

6,600 games 1s + 0.1s in LittleBlitzer, SF 2.1 vs SF 2.01, 64 bit engines

Code: Select all

Games Completed = 6600 of 6600 (Avg game length = 12.820 sec)
Settings = RR/16MB/1000ms+100ms/M 10000cp for 1000 moves, D 200 moves/PGN:D:\LittleBlitzer-2.72\swcr.pgn(3395)
Time = 88804 sec elapsed, 0 sec remaining
 1.  Stockfish 2.1 JA 64bit   	3544.5/6600	1924-1435-3241  	(L: m=1435 t=0 i=0 a=0) 	(tpm=97.1 d=12.1 nps=959707)
 2.  Stockfish 2.0.1 JA 64bit 	3055.5/6600	1435-1924-3241  	(L: m=1924 t=0 i=0 a=0)	 (tpm=95.7 d=11.8 nps=935155)

Elostat statistics

Code: Select all

    Program                            Score       %      Elo    +   -   Draws

  1 Stockfish 2.1 JA 64bit         : 3544.5/6600  53.7    3213   6   6   49.1 %
  2 Stockfish 2.0.1 JA 64bit       : 3055.5/6600  46.3    3187   6   6   49.1%

+26 +/- 6 Elo points against SF 2.01 95% confidence intervals. Not the best way to see the real difference (would need a gauntlet with several different engines), but the improvement is clear. SF 2.1 seems ~2% faster and is going to ~0.3 plies deeper.

Kai

[zamar]

+26 +/- 6 Elo points against SF 2.01 95% confidence intervals. Not the best way to see the real difference (would need a gauntlet with several different engines), but the improvement is clear. SF 2.1 seems ~2% faster and is going to ~0.3 plies deeper.

Kai

Thanks for the test Kai, it's well in line with our own tests. However ist should be noted that self-play often exaggarates the improvement, so the real improvement should be somewhat less.

[Jamal Bubker]

Tord, Marco, Joona, Jim , thanks very much guys !

[Amstaff]

Hello thanks for the great update. I just loaded Stockfish v2.1 into arena along with the optional opening book.bin, but I can't seem to get the program to use it's opening book. The configuration looks okay and the book.bin is there.
Any thoughts would be helpful.
Thanks,
Gerald

[Jim Ablett]

Here is an alternative Windows 32 bit executable compiled with Intel compiler (other was Msvc).
Opening book function works with this build.
One problem though, on my machine Intel build produces slightly less nps.

http://dl.dropbox.com/u/5047625/stockfi ... tel-ja.zip

Jim.

[bob]

I'm using GCC 4.6. The speed gain from LTO is bigger than the last few years of GCC development combined, so it's worth verifying whether Stockfish triggers any bugs there or not.

Is this somehow supposed to be better than "one big compile?" I've always had a bunch of modules, and then for max speed, I have a crafty.c file that just #includes every source file, in what I consider the most cache-friendly order...

[ernest]

One problem though, on my machine Intel build produces slightly less nps.

Hi Jim,
Difficult to compare on my Intel Core2 Duo, with default options and 1 thread (for reproducibility).
setoption name threads value 1
go infinite

I get completely different analysis:
Intel compiler
info depth 21 seldepth 34 multipv 1 score cp 36 nodes 19983802 nps 796357 time 2
5094 pv e2e4 e7e5 g1f3 g8f6 f3e5 d7d6 e5f3 f6e4 d1e2 d8e7 b1c3 e4c3 d2c3 e7e2 f1
e2 b8c6 c1e3 c8f5 e1c1 f8e7 e2d3 f5d3 d1d3 e8g8 d3d5
info depth 22

Msvc compiler
info depth 21 seldepth 32 multipv 1 score cp 20 nodes 32817584 nps 775884 time 4
2297 pv e2e4 e7e6 g1f3 d7d5 b1c3 g8f6 e4d5 e6d5 d2d4 f8d6 f1d3 e8g8 e1g1 b8c6 c1
g5 c6b4 c3b5 c8g4 f1e1 h7h6 b5d6 d8d6
info depth 22