To Uri, about perft speeds

[Albert Bertilsson]

Geschrieben von: / Posted by: Albert Bertilsson at 14 February 2003 14:29:59:

Hello Uri!
You've previously stated that Movei is one of the fastest perft testers around for free, how fast is it?
You also stated that you had some special perft tester that was even faster, how fast is it?
Since this performance is very dependent on machine specs please relate the results to some other engine, like yace or crafty.
My current best results with perft 6 calculation from the initial position is:
Sharper 0.00
8492 ms (doing perft only, avoiding some DoMove/UndoMove)
20669 ms (generating all positions)
Yace 0.99.50
24.944 s
Crafty 18.12
28.70 s
XP 1800+, 512MB SDRAM, WinXP.
Note that this Sharper version doesn't generate Zobrist keys, attacktables or anything useful, nor does use any hashtables for speed up.
When testing with Zobrist key generation on it's about the speed of yace (24846 ms).
I'm very interested to hear your results.
/Regards Albert Bertilsson

[Sune Fischer]

Geschrieben von: / Posted by: Sune Fischer at 14 February 2003 14:54:07:
Als Antwort auf: / In reply to: To Uri, about perft speeds geschrieben von: / posted by: Albert Bertilsson at 14 February 2003 14:29:59:

I'm very interested to hear your results.
/Regards Albert Bertilsson

Some day someone is going to implement hashing in perft, just to spite Uri.
IT'S A DEBUG TOOL!!!!


-S.

[Albert Bertilsson]

Geschrieben von: / Posted by: Albert Bertilsson at 14 February 2003 15:09:04:
Als Antwort auf: / In reply to: Re: To Uri, about perft speeds geschrieben von: / posted by: Sune Fischer at 14 February 2003 14:54:07:

IT'S A DEBUG TOOL!!!!

Yes of course it is, but it's also a part of the engine, if you want to make something really great, all the parts must also be great!
And even if I don't finish the engine it would be nice to have made the fastest debug tool around =).
/Regards Albert

[Uri Blass]

Geschrieben von: / Posted by: Uri Blass at 14 February 2003 15:46:15:
Als Antwort auf: / In reply to: To Uri, about perft speeds geschrieben von: / posted by: Albert Bertilsson at 14 February 2003 14:29:59:

Hello Uri!
You've previously stated that Movei is one of the fastest perft testers around for free, how fast is it?
You also stated that you had some special perft tester that was even faster, how fast is it?
Since this performance is very dependent on machine specs please relate the results to some other engine, like yace or crafty.
My current best results with perft 6 calculation from the initial position is:
Sharper 0.00
8492 ms (doing perft only, avoiding some DoMove/UndoMove)
20669 ms (generating all positions)
Yace 0.99.50
24.944 s
Crafty 18.12
28.70 s
XP 1800+, 512MB SDRAM, WinXP.
Note that this Sharper version doesn't generate Zobrist keys, attacktables or anything useful, nor does use any hashtables for speed up.
When testing with Zobrist key generation on it's about the speed of yace (24846 ms).
I'm very interested to hear your results.
/Regards Albert Bertilsson

I am not sure
I have not the exact programs that you use and relatively to slightly older Crafty I get better result but relatively to Yace0.99.56 I get worse results
I suggest that you download public Movei and run perft and multiply your result by 12.308/13.129
I get the following results in seconds
WCrafty18.10 50.29
Yace 0.99.56 30.406(I guess that it is the same as Yace0.99.50 but I am not sure about it)
Public Perft 13.129
latest perft 12.308
Public Movei(0.0799) 15.423
latest Movei(0.079992) 17.555


Note that It was not so important for me to be the fastest in calculating perft and developing the chess engine was more important for me.
I also do not use Zobrist key to speed things in generating perft but I use attack tables for my move generator(not only in the perft function).
Regards,
Uri

[Albert Bertilsson]

Geschrieben von: / Posted by: Albert Bertilsson at 14 February 2003 16:19:37:
Als Antwort auf: / In reply to: Re: To Uri, about perft speeds geschrieben von: / posted by: Uri Blass at 14 February 2003 15:46:15:

Hello Uri!
I've tested Movei and Perft.exe on my machine and your result's are very good!
Movei: 8763 ms, damn that's fast, specially considering that you have attackboards too!
Perft tool: 7641 * 0,937 = 7160 ms, really fast compared to my tool (8492 ms).
I've found no other engines even close to these values, and that of course renders some questions...
Board representation?
I use standard 0x88, with PieceLists for both sides. Is there any thing faster?
Zobrist?
You said that you didn't use it when doind perft. My DoMove does the Zobrist key updating, how do you exclude that when doing perft? Do you have seperate DoMove for perft or is there some trick to avoid updating the Zobrist key in the DoMove function?
Congrats to the really good results, perhaps some day Sharper will beat that, I'll dive into the compiler and post results if I can improve something...
/Regards Albert Bertilsson

[Sune Fischer]

Geschrieben von: / Posted by: Sune Fischer at 14 February 2003 16:20:08:
Als Antwort auf: / In reply to: Re: To Uri, about perft speeds geschrieben von: / posted by: Albert Bertilsson at 14 February 2003 15:09:04:

IT'S A DEBUG TOOL!!!!

Yes of course it is, but it's also a part of the engine, if you want to make something really great, all the parts must also be great!
And even if I don't finish the engine it would be nice to have made the fastest debug tool around =).
/Regards Albert

Sure, but what has this got to do with perft?
Alright, lets see how fast hashing will be then..
-S.

[Uri Blass]

Geschrieben von: / Posted by: Uri Blass at 14 February 2003 16:43:35:
Als Antwort auf: / In reply to: Results with Movei and some more questions... geschrieben von: / posted by: Albert Bertilsson at 14 February 2003 16:19:37:

Hello Uri!
I've tested Movei and Perft.exe on my machine and your result's are very good!
Movei: 8763 ms, damn that's fast, specially considering that you have attackboards too!
Perft tool: 7641 * 0,937 = 7160 ms, really fast compared to my tool (8492 ms).
I've found no other engines even close to these values, and that of course renders some questions...
Board representation?
I use standard 0x88, with PieceLists for both sides. Is there any thing faster?
Zobrist?
You said that you didn't use it when doind perft. My DoMove does the Zobrist key updating, how do you exclude that when doing perft? Do you have seperate DoMove for perft or is there some trick to avoid updating the Zobrist key in the DoMove function?
Congrats to the really good results, perhaps some day Sharper will beat that, I'll dive into the compiler and post results if I can improve something...
/Regards Albert Bertilsson

info[64]
info give me in one number both the color of the square and the piece on the board.
I use special array to give me knight moves or bishop directions.
for example leftdown[x] is usually x-9 but it gets -1 in some cases when bishop cannot go from x to x-9.
I do not use bitboard and not 0x88.
I do not know
Maybe 0x88 is faster than what I do but I did not try it.
The trick is simply
#if MODE != MODE_PERFT
For example when I undo moves I do
#if MODE != MODE_PERFT
zob=zob^zobrist[0][0][0];
I see now that I can do my perft even faster because there are places when I forgot to write
#if MODE != MODE_PERFT
I did not spend a lot of time in trying to do it faster unless it was in a way that can also do movei faster.
Thanks
Uri