Page 7 of 10

Re: "Roc" - it is an improvement !!

Posted: Tue Feb 14, 2017 10:18 pm
by kranium
Gusev wrote:Overnight, Roc 0.3 crashed 3 times under Arena, while Gull 3.1 ran stable, so after 543 games I see Gull 3.1 ahead, +149-96=298.
Hi Dimitri,

Just an FYI in case you weren't aware: Velasco's "Gull 3.1" is nothing new.
Although it may appear so, (it retains the name 'Gull' and uses the next logical version #)...it's not an official release.

It's a copy of Thinkingalots' original syzygy implementation from a couple years ago, available for download here:

http://www.sdchess.ru/engines/Gull_3s.zip

In fact it's a step backward IMO. The 4 (previously) separate large syzygy TB files are now copied into the middle of the already huge and confusing Gull 3 source file,
making it even more difficult to read. In addition, parts of the Mersenne Twister (GPL) random number code has also been copied in.
Other than that, there's no new code as far as I can see.

At least Tom H. is a programmer...and has done some nice work with the codebase.

Norm

Re: "Roc" - it is an improvement !!

Posted: Tue Feb 14, 2017 11:33 pm
by velmarin
kranium wrote:
Gusev wrote:Overnight, Roc 0.3 crashed 3 times under Arena, while Gull 3.1 ran stable, so after 543 games I see Gull 3.1 ahead, +149-96=298.
Hi Dimitri,

Just an FYI in case you weren't aware: Velasco's "Gull 3.1" is nothing new.
Although it may appear so, (it retains the name 'Gull' and uses the next logical version #)...it's not an official release.

It's a copy of Thinkingalots' original syzygy implementation from a couple years ago, available for download here:

http://www.sdchess.ru/engines/Gull_3s.zip

In fact it's a step backward IMO. The 4 (previously) separate large syzygy TB files are now copied into the middle of the already huge and confusing Gull 3 source file,
making it even more difficult to read. In addition, parts of the Mersenne Twister (GPL) random number code has also been copied in.
Other than that, there's no new code as far as I can see.

At least Tom H. is a programmer...and has done some nice work with the codebase.

Norm
Readme by Gull 3.1

Code: Select all

Chess uci engine with support Syzygy tablebassses

This repository is the original Gull 3.0 code and code probe Syzygy embedded in the file. True to the philosophy of ThinkingALot, all in a file. There are several modifications to compile without errors in Visual Studio in a Windows environment.

An attempt to give functionality to this project at the moment abandoned, correct some bad behaviors and an attempt to improve it.

Modifications and code written by José M. Velasco declared how PUBLIC DOMAIN. First give their deserved recognition to the principal authors of the project Gull, Syzygy end tables and Mersenne twister:

Gull Author Vadim Demichev "ThinkingALot", gull is PUBLIC DOMAIN https://chessprogramming.wikispaces.com/Vadim+Demichev Vadim Demichev on public domain: ... Moreover, I believe that when someone releases an engine as a public domain project it's a kind of invitation "feel free to improve it and make it commercial"! Since nothing prevented the original authors to place a restriction on commercial use of their code.

Author code SYZYGY: Copyright (c) 2011-2013 Ronald de Man https://chessprogramming.wikispaces.com/Ronald+de+Man This file may be redistributed and/or modified without restrictions. Official code Syzygy it can be found here https://github.com/syzygy1/tb

RANDOM MERSENNE TWISTER A C-program for MT19937, with initialization improved 2002/1/26. Coded by Takuji Nishimura and Makoto Matsumoto. Copyright (C) 1997 - 2002, Makoto Matsumoto and Takuji Nishimura, all rights reserved.
Of course, it's still Gull and its author, and PUBLIC DOMAIN.
To steal and defraud others , You know a lot about this, friend Norman.

Re: "Roc" - it is an improvement !!

Posted: Wed Feb 15, 2017 12:31 am
by kranium
velmarin wrote: Of course, it's still Gull and its author, and PUBLIC DOMAIN.
ok...just to be clear
you downloaded Thinkingalot's Gull 3 syzygy,
copied 100's of lines of GPL code into it,
compiled it,
changed the name to Gull 3.1 JV,
released it on Sourceforge and Github via your JVelasco accounts,
and you're proclaiming it 'public domain'...

Re: New engine "Roc"

Posted: Wed Feb 15, 2017 10:11 am
by Adam Hair
THyer wrote:Roc 0.3 has been released. There are now three binaries:

x64 (without HNI)
x64_HNI
x86 (32-bit; has no tablebase support)
The results of my CCRL 40/4 testing using the CCRL 40/4 Single CPU Pure database (computed with Ordo):

Code: Select all

   # PLAYER                              :  RATING  ERROR  POINTS  PLAYED   (%)
   7 Roc 0.3 64-bit                      :    3241     33   377.5     700  53.9
   8 Gull 3 64-bit                       :    3223   ----   666.0    1287  51.7

Re: New engine "Roc"

Posted: Wed Feb 15, 2017 2:52 pm
by THyer
Thanks, Adam. There seems to be now a range of measurements of Roc's improvement, with yours near the middle.

Were you using the HNI build? I'm trying to get a handle on the two crash reports, and one hypothesis is that my crude Windows port of Fathom is incompatible with the HNI build.

Re: New engine "Roc"

Posted: Wed Feb 15, 2017 4:55 pm
by kranium
THyer wrote:Thanks, Adam. There seems to be now a range of measurements of Roc's improvement, with yours near the middle.

Were you using the HNI build? I'm trying to get a handle on the two crash reports, and one hypothesis is that my crude Windows port of Fathom is incompatible with the HNI build.
Hi Tom...

If you're having issues with TB stability or implementation, Vadim's original (pre-Fathom) source works well.
I have a stable version here (VS2015) that I've worked on (played with) from time to time in the last year.

-source code cleaned up, simplified, and broken up into source and header files
-a couple of VS code analysis fixes (ex: gen_kpk() was causing stack exceed error)
so it's now implemented using uint8 Kpk_gen[2][64][64][64]
-compiler warnings resolved up to level 4
-benchmark and perft utilities added
-small change to phase/piece value calculation adding 3-5 elo:

Settings = Gauntlet/128MB/1000ms+100ms/M 500cp for 6 moves, D 120 moves/PGN:10000.pgn(10000)
(Avg game length = 13.252 sec)

Code: Select all

1 Seagull 138 : 3002 9795/38207 (+9795,=19150,-9262) 50.7 %			
2 Gull 3      : 2998 9262/38207 (+9262,=19150,-9795) 49.3 %


games = 38207
win% = 0.506975
elo = +4.84715
los = 0.999926

Since you're pursuing this project I'd be happy to share it with you...
you're welcome to use any/all of it as you see fit.

Norm

Re: New engine "Roc"

Posted: Wed Feb 15, 2017 5:19 pm
by THyer
I would be grateful for that. Dmitry is using the HNI version and experiencing crashes. So I suspect it is unsound.

Re: New engine "Roc"

Posted: Wed Feb 15, 2017 5:36 pm
by kranium
check you mailbox...
good luck with your project!

Re: New engine "Roc"

Posted: Wed Feb 15, 2017 8:44 pm
by Gusev
Just now I witnessed the error message pop up again! It says, "Roc_0_3_x64_HNI.exe has stopped working
A problem caused the program to stop working correctly. Windows will close the program and notify you if a solution is available
Debug / Close program".
I will be very grateful if a stable version of Roc/Seagull supporting both AVX2/BMI2 and Syzygy is developed. Then I'll have it run in Purdue FOSCEC Season 4.
THyer wrote:I would be grateful for that. Dmitry is using the HNI version and experiencing crashes. So I suspect it is unsound.

Re: New engine "Roc"

Posted: Wed Feb 15, 2017 10:08 pm
by Adam Hair
THyer wrote:Thanks, Adam. There seems to be now a range of measurements of Roc's improvement, with yours near the middle.

Were you using the HNI build? I'm trying to get a handle on the two crash reports, and one hypothesis is that my crude Windows port of Fathom is incompatible with the HNI build.
No. My processor is too old to run the HNI build.