Designing an analysis friendly Stockfish?

[keoki010]

OK, sorry for all the builds. Here's a version which does what I just explained above.

jb

Still doing the same thing. Looks like it is still hanging in TB search.
I'm using Aquarium 4.0.6.

[Jeremy Bernstein]

I had Stockfish crashing in this position:

8/8/3b4/7p/2K4P/3Nk1n1/5p2/8 w - -

How many threads are you using? I can't reproduce using a single core -- it iterates to 100 and then waits for a stop (when go infinite is issued). In this case, it should be possible to break early from the search, but that's a different issue.

Jeremy

[keoki010]

I had Stockfish crashing in this position:

8/8/3b4/7p/2K4P/3Nk1n1/5p2/8 w - -

How many threads are you using? I can't reproduce using a single core -- it iterates to 100 and then waits for a stop (when go infinite is issued). In this case, it should be possible to break early from the search, but that's a different issue.

Jeremy

Sorry that fixed the problem. I was using the wrong version. Using 12 threads.

[keoki010]

I had Stockfish crashing in this position:

8/8/3b4/7p/2K4P/3Nk1n1/5p2/8 w - -

How many threads are you using? I can't reproduce using a single core -- it iterates to 100 and then waits for a stop (when go infinite is issued). In this case, it should be possible to break early from the search, but that's a different issue.

Jeremy

Sorry that fixed the problem. I was using the wrong version. Using 12 threads.

I also tried running Uly's problem and here's what I got.
[Mate in 17] d=100 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:04) 682kN
[< Mate in 17] d=100 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:04) 682kN
[< Mate in 17] d=99 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:04) 681kN
[< Mate in 17] d=98 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:04) 681kN
[Mate in 17] d=97 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:04) 681kN
[< Mate in 17] d=97 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 681kN
[Mate in 17] d=96 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 681kN
[< Mate in 17] d=96 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 681kN
[Mate in 17] d=95 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 681kN
[< Mate in 17] d=95 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 680kN
[Mate in 17] d=94 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 680kN
[< Mate in 17] d=94 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 680kN
[Mate in 17] d=93 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 680kN
[< Mate in 17] d=93 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 680kN
[Mate in 17] d=92 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 680kN
[< Mate in 17] d=92 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 680kN
[Mate in 17] d=91 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 680kN
[< Mate in 17] d=91 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 679kN
[Mate in 17] d=90 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 679kN
[< Mate in 17] d=90 1.Ne5 f1Q+ 2.Kd5 Qb5+ 3.Kxd6 Nf5+ 4.Ke6 Qxe5+ 5.Kxe5 (0:00:03) 679kN

[Jeremy Bernstein]

OK, please try this build. I found (thanks to Jouni Uski) a somewhat serious bug in the TB code, where capture moves from a 6-piece position could have triggered a premature search stop, if a tablebase hit resulted.

Uly's position gives me a mate in 14 for black (after Nxf2) using 2 threads. Any other move is even worse for white.

[Uly]

I'm using 4 threads. The position was reached in a game in where Stockfish played against itself at 5/sec move (so it reached that position move by move, and then crashed). I lost the game after the GUI crash, so it's going to be impossible to try to reproduce. Will report again if I see something similar.

[keoki010]

OK, please try this build. I found (thanks to Jouni Uski) a somewhat serious bug in the TB code, where capture moves from a 6-piece position could have triggered a premature search stop, if a tablebase hit resulted.

Uly's position gives me a mate in 14 for black (after Nxf2) using 2 threads. Any other move is even worse for white.

Is there a big difference between Nalimov and Griviota hits. The last build I tried seemed to start getting less hits for stockfish then for R4. I will try the last build and let you know.

[keoki010]

OK, please try this build. I found (thanks to Jouni Uski) a somewhat serious bug in the TB code, where capture moves from a 6-piece position could have triggered a premature search stop, if a tablebase hit resulted.

Uly's position gives me a mate in 14 for black (after Nxf2) using 2 threads. Any other move is even worse for white.

This build seems to not get as deep as the last build depth wise but seems to get better results on analysis. I'll play with it a little more tomorrow. I have to get off my computer now.

[Jeremy Bernstein]

I re-ran the tournament from the other day, removing PAMC_GTBg and adding Gran2c. Looks like there's no problem that affects play strength. Anyone testing these builds should probably update to Gran2c. A 32-bit version (identical code to the 64-bit) is attached to this message.

Jeremy

Engine                          Score                                                                                                      St                                                                                                   St                                                                                                   St                                                                                                   St    S-B
Stockfish_201_PA_GTB_Gran2c_x64 174,0/300 ···································································································· 0001==0111111101101010100010000001011010001010011110110101010100000001110=11010010=10111110000000011 011=1101000100110100011111101010110=11101101011111010000011110011000=111100101=000111110111101010101 0111000011111011010111111001111011111110100111000111011011011=101101110111011111111011010111010=1101  24191,
Stockfish_201_PA_GTB_Gran2_x64  173,0/300 1110==1000000010010101011101111110100101110101100001001010101011111110001=00101101=01000001111111100 ···································································································· 001100100111001111101111111111111=011011011001000001001111010110011100011011101011011111100000011101 1010101001=000111011111011110110110000010110011100101111111011110011111110011=110101000001110000=111  24402,
Stockfish_201_64_PA_La          127,0/300 100=0010111011001011100000010101001=00010010100000101111100001100111=000011010=111000001000010101010 110011011000110000010000000000000=100100100110111110110000101001100011100100010100100000011111100010 ···································································································· 10000100101010101011001001001110010100000=10111001011001010011=110000111010010111100100110010010=100  19874,
Stockfish-201-64-ja             126,0/300 1000111100000100101000000110000100000001011000111000100100100=010010001000100000000100101000101=0010 0101010110=111000100000100001001001111101001100011010000000100001100000001100=001010111110001111=000 01111011010101010100110110110001101011111=01000110100110101100=001111000101101000011011001101101=011 ····································································································  19322,

600 games played / Tournament is finished
Level: Blitz 0:05/0,1
Hardware: Intel(R) Xeon(R) CPU  W3520 @ 2.67GHz 2659 MHz with 1,0 GB Memory
Operating system: Windows 7 Professional Professional (Build 7600) 64 bit

[Jeremy Bernstein]

Here's an OSX version, too, if you want to complete your collection. Intel-only 32/64-bit fat binary.