gladius wrote: I'm seeing it crash with stockfish-231-64-ja.exe, however, with my local compile it doesn't crash. I wonder if the pgo is hitting a bug?
Yes I confirm I can reproduce the crash with Jim's 64bit binaries, both popcount and not popcount, on my Windows 7. The 32bit one seems to work instead. Moreover my local 64 bit (MSVC) compile doesn't shows this behaviour.
$ ./stockfish
Stockfish 2.3.1 by Tord Romstad, Marco Costalba and Joona Kiiski
setoption name uci_chess960 value true
position fen qnnbbrkr/pppppppp/8/8/8/8/PPPPPPPP/QNNBBRKR w HFhf - 0 0
go infinite
Yes it crashed for me too. I found the GCC optimizations that were causing the crash. It was the Graphite loop transformation switches '-floop-interchange' , '-floop-strip-mine' & '-floop-block'. A pity really as these gave a small speedup when used.
I will recompile the 64 bit versions and post them here when done.
Jim Ablett wrote: Yes it crashed for me too. I found the GCC optimizations that were causing the crash. It was the Graphite loop transformation switches '-floop-interchange' , '-floop-strip-mine' & '-floop-block'. A pity really as these gave a small speedup when used.
Does it crashes for the Window build or also for Linux one? In the latter case could you compile me a crashing version for Linux with debug info embedded so that I could try to understand in wich part of code it crashes?
Also doe sit crashes with any of the 3 or when you enable all of them? In this case you may be could enable only the non crashing ones.
Jim Ablett wrote: Yes it crashed for me too. I found the GCC optimizations that were causing the crash. It was the Graphite loop transformation switches '-floop-interchange' , '-floop-strip-mine' & '-floop-block'. A pity really as these gave a small speedup when used.
Does it crashes for the Window build or also for Linux one? In the latter case could you compile me a crashing version for Linux with debug info embedded so that I could try to understand in wich part of code it crashes?
Also doe sit crashes with any of the 3 or when you enable all of them? In this case you may be could enable only the non crashing ones.
No problem with the Linux versions which use identical switches.
Makes me think this is a compiler bug in Mingw compiler although my Windows builds still run faster than Linux ones on same machine without graphite loop optimizations.
Anyway I have recompiled the 64 bit windows versions. Very very slightly slower than before on my machine but your mileage may vary.
Jim Ablett wrote: Yes it crashed for me too. I found the GCC optimizations that were causing the crash. It was the Graphite loop transformation switches '-floop-interchange' , '-floop-strip-mine' & '-floop-block'. A pity really as these gave a small speedup when used.
Does it crashes for the Window build or also for Linux one? In the latter case could you compile me a crashing version for Linux with debug info embedded so that I could try to understand in wich part of code it crashes?
Also doe sit crashes with any of the 3 or when you enable all of them? In this case you may be could enable only the non crashing ones.
No problem with the Linux versions which use identical switches.
Makes me think this is a compiler bug in Mingw compiler although my Windows builds still run faster than Linux ones on same machine without graphite loop optimizations.
Anyway I have recompiled the 64 bit windows versions. Very very slightly slower than before on my machine but your mileage may vary.
Quocvuong82 wrote:Stockfish 2.3.1 x64 JA (SSE4.2 and non SSE4.2) is bug (gen moves)
Fen: rn2k2r/pp1b1p1p/4pp2/qB6/1b1NP3/2N5/PP3PPP/R2QK2R w KQkq - 0 0
perft 5
Nodes searched : 76978078
2.3.1 w32
perft 5
Nodes searched : 73001686
I have just downloaded the package of the URL of your post I quote from, and I can only check Windows versions, 32-bit, legacy and not legacy versions, and both of them give perft(5) = 73001686 (number confirmed by JetChess 1.0.0.0). I hope that this bug in 64-bit versions is fixed with your new compiles.
Thanks to everybody that help SF to be better every day.
Ajedrecista wrote: I hope that this bug in 64-bit versions is fixed with your new compiles.
I confirm new JIm's compiles are correct in all versions: 32bits, 64bits and 64+popcount
I guess we are at the end of this troubled releasing. Thanks to Jim to have worked hard at late Saturday night to fix the miscompile, thanks to Ray and all other guys to have reported bugs and verified new versions are correct.
Stockfish bench signature of this 2.3.1 version is: 5423738
That is different from 2.3, it means above the chess960 bug fix there is also a small added evaluation tweak that should increase ELO a bit (especially in the 64bit popcount version), but anyhow I do not suggest testers to redo the ongoing tests because the increase should be well within usual public lists error bars.
Ajedrecista wrote: I hope that this bug in 64-bit versions is fixed with your new compiles.
I confirm new JIm's compiles are correct in all versions: 32bits, 64bits and 64+popcount
I guess we are at the end of this troubled releasing. Thanks to Jim to have worked hard at late Saturday night to fix the miscompile, thanks to Ray and all other guys to have reported bugs and verified new versions are correct.
Stockfish bench signature of this 2.3.1 version is: 5423738
That is different from 2.3, it means above the chess960 bug fix there is also a small added evaluation tweak that should increase ELO a bit (especially in the 64bit popcount version), but anyhow I do not suggest testers to redo the ongoing tests because the increase should be well within usual public lists error bars.
At this level it is very difficult to improve the strength of stockfish. It was a great effort from the team of stockfish, Jim, Gary, etc., and has achieved a great improvement thanks to the collaboration of all.
Thank you very much to all
