Stockfish 7 beta 1

[beram]

Great work, Bram, can you post a pgn file of the games?

hereby https://drive.google.com/open?id=0BzGb6 ... HdsLVdGUzA

[MikeB]

Stockfish 7 beta 1 is out.

We would kindly ask rating list authors and other interested parties to do test runs with this version and report any encountered issues in this thread.

If you have a Mac, the two exe's enclosed in the zip folder below will most likely be the fastest. "Hand tuned"* for the bmi2 processor, and the other for popcnt. You may have to change permissions "chmod 755" to enable it for your machine. Over 2M nps on the Mac Pro and 2.7M nps on the iMac

Stockfish for the Mac

* "Hand tuned" here means that I modified the makefile options to produce the absolute fastest exe's on my 2010 Mac Pro and and 2015 iMac Could be anywhere 10-15% faster than and standard created exe for the Mac. (ymmv).

[zamar]

Thanks for a detailed report, Peter!

It's a pity that patching source code is required... But your change looks fine. Clearly this a problem with NDK, so there is not much we can do on our side.

Thanks!

[fenchel]

The "artificial" position is from a CCRL 40/4 game. (GrayMatter SVN1604 vs POS 1.20, date 2012-07-16).

POS 1.2 is rated 230 elo weaker than RAM 2.0, which is a RANDOM MOVER! So you're looking at a quality of chess worse than random moves!

The fact that these patzers cannot even mate the opponent with 4 queens should already give you a hint, don't you think ?

Clearly an artificial position.

Interesting; as Lucas is no doubt aware (but I'll say it for people less involved with software development), tools that feed random data to a program are called "fuzzers", and they are used to uncover bugs.

Maybe running depth 1 SF vs RAM could reveal some other issues?

[Jouni]

A proposition: Can we get out of "64" in engine name now? Because it's standard nowadays so only "32" is needed!

[beram]

First testrun on my i5 4200M, 2 cores, 512mB hash, privat testbook 25 lines

Code: Select all

SF 7beta - - Komodo 9.3  Blitz 3m2s     i5 4200M @2500Mhz, 2 cpu (Fritzmark 5,1)
1   Stockfish 7Beta1 64 BMI2   +24  +22/=63/-15 53.50%   53.5/100
2   Komodo 9.3 64-bit          -24  +15/=63/-22 46.50%   46.5/100

53,5 % out of 100games in line with earlier results of SF devs from dec 2015

Code: Select all

SF 051215 - Komodo 9.3 Blitz 3m2s     i5 4200M @2500Mhz, 2 cpu (Fritzmark 5,1)                                    
1   Stockfish 051215 64 BMI2   +21  +27/=52/-21 53.00%   53.0/100
2   Komodo 9.3 64-bit          -21  +21/=52/-27 47.00%   47.0/100

SF 141215 - Komodo 9.3  Blitz 3m2s     i5 4200M @2500Mhz, 2 cpu (Fritzmark 5,1)                                   
1   Stockfish 141215 64 BMI2    +7  +18/=66/-16 51.00%   51.0/100
2   Komodo 9.3 64-bit           -7  +16/=66/-18 49.00%   49.0/100

SF 201215 - Komodo 9.3  Blitz 3m2s     i5 4200M @2500Mhz, 2 cpu (Fritzmark 5,1)                               
1   Stockfish 201215 64 BMI2   +31  +25/=59/-16 54.50%   54.5/100
2   Komodo 9.3 64-bit          -31  +16/=59/-25 45.50%   45.5/100

a second testrun on my AMD 1090T @3200 Mhz 4cpu, (Fritzmark 16,1)
SF7 beta 59,5%

Code: Select all

SF 7beta - Komodo 9.3, Blitz 3m+2s
                               
1   Stockfish 7Beta1 64   +67  +29/=61/-10 59.50%   59.5/100
2   Komodo 9.3 64-bit     -67  +10/=61/-29 40.50%   40.5/100

[Max]

If you have a Mac, the two exe's enclosed in the zip folder below will most likely be the fastest. "Hand tuned"* for the bmi2 processor, and the other for popcnt. You may have to change permissions "chmod 755" to enable it for your machine. Over 2M nps on the Mac Pro and 2.7M nps on the iMac

Stockfish for the Mac

* "Hand tuned" here means that I modified the makefile options to produce the absolute fastest exe's on my 2010 Mac Pro and and 2015 iMac Could be anywhere 10-15% faster than and standard created exe for the Mac. (ymmv).

Hi Michael,

thanks a lot for your compiles, I appreciate them very much!
On my Macbook did just "make profile-build ARCH=x86-64-modern COMP=gcc“ and get this bench result:

Code: Select all

$ ./stockfish bench > /dev/null 
===========================
Total time (ms) : 4035
Nodes searched  : 8355485
Nodes/second    : 2070752

Your binary is only about half the size , but not faster on my system (core i7 3615qm):

Code: Select all

$ ./stockfish-7-beta-popcnt bench > /dev/null 
===========================
Total time (ms) : 4078
Nodes searched  : 8355485
Nodes/second    : 2048917

May I ask, which compile options did you change in the Makefile?

-Max

[zamar]

Stockfish 7 beta 2 is out:

It corrects the following problems:
- Infinite loop when searching in position: 7k/QQQQR3/2B5/4KN1Q/3QQ3/8/8/4R3 b - - 0 1
- Correction for Trace scores for pawns

[beram]

First testrun on my i5 4200M, 2 cores, 512mB hash, privat testbook 25 lines

Code: Select all

SF 7beta - - Komodo 9.3  Blitz 3m2s     i5 4200M @2500Mhz, 2 cpu (Fritzmark 5,1)
1   Stockfish 7Beta1 64 BMI2   +24  +22/=63/-15 53.50%   53.5/100
2   Komodo 9.3 64-bit          -24  +15/=63/-22 46.50%   46.5/100

53,5 % out of 100games in line with earlier results of SF devs from dec 2015

Code: Select all

SF 051215 - Komodo 9.3 Blitz 3m2s     i5 4200M @2500Mhz, 2 cpu (Fritzmark 5,1)                                    
1   Stockfish 051215 64 BMI2   +21  +27/=52/-21 53.00%   53.0/100
2   Komodo 9.3 64-bit          -21  +21/=52/-27 47.00%   47.0/100

SF 141215 - Komodo 9.3  Blitz 3m2s     i5 4200M @2500Mhz, 2 cpu (Fritzmark 5,1)                                   
1   Stockfish 141215 64 BMI2    +7  +18/=66/-16 51.00%   51.0/100
2   Komodo 9.3 64-bit           -7  +16/=66/-18 49.00%   49.0/100

SF 201215 - Komodo 9.3  Blitz 3m2s     i5 4200M @2500Mhz, 2 cpu (Fritzmark 5,1)                               
1   Stockfish 201215 64 BMI2   +31  +25/=59/-16 54.50%   54.5/100
2   Komodo 9.3 64-bit          -31  +16/=59/-25 45.50%   45.5/100

a second testrun on my AMD 1090T @3200 Mhz 4cpu, (Fritzmark 16,1)
SF7 beta 59,5%

Code: Select all

SF 7beta - Komodo 9.3, Blitz 3m+2s
                               
1   Stockfish 7Beta1 64   +67  +29/=61/-10 59.50%   59.5/100
2   Komodo 9.3 64-bit     -67  +10/=61/-29 40.50%   40.5/100

A third run on the AMD 1090T @3200 Mhz 4cpu, (Fritzmark 16,1) with TCEC7 finalbook, SF7beta 54,69%

Code: Select all

SF 7beta - Komodo 9.3, Blitz 3m+2s
                               
1   Stockfish 7Beta1 64   +33  +14/=42/-8 54.69%   35.0/64
2   Komodo 9.3 64-bit     -33  +8/=42/-14 45.31%   29.0/64

[MikeB]

Hi Michael,

thanks a lot for your compiles, I appreciate them very much!
On my Macbook did just "make profile-build ARCH=x86-64-modern COMP=gcc“ and get this bench result:

Code: Select all

$ ./stockfish bench > /dev/null 
===========================
Total time (ms) : 4035
Nodes searched  : 8355485
Nodes/second    : 2070752

Your binary is only about half the size , but not faster on my system (core i7 3615qm):

Code: Select all

$ ./stockfish-7-beta-popcnt bench > /dev/null 
===========================
Total time (ms) : 4078
Nodes searched  : 8355485
Nodes/second    : 2048917

May I ask, which compile options did you change in the Makefile?

-Max

Hi Max,

For gcc , try this on your machine

Code: Select all

Section 3.9: CXXFLAGS +=-flto -O2 -mtune=native -mssse3 -mpopcnt -DUSE_POPCNT

and comment out -flto in section 3.11 below. When you profile, -flto should be in front of the profile options so as you don't overrule pgo with flto. Generally speaking, when you profile on your own machine, they will be the fastest exe's for obvious reasons.

With respect to the exe size, I'm using the compressor upx
I then added an alias to the bash_profile:

Code: Select all

alias upxl="/applications/upx/./upx2 --lzma"  

(renamed upx to upx2)
so from anywhere, when i compile stockfish or crafty, I can compress the exe by typing:

Code: Select all

upxl <filename> -o <output-filename>

There is no loss in speed and it saves bandwidth/hardrive space