Stockfish 1.6 - strange behavior in this position

[royb]

I think the problem is in the Makefile. I used the Makefile from the Stockfish-1.5 build to make the 1.6 version and this behavior is now gone (...Rc1 not preferred EVER -- and that's a good thing).

Roy

Hi Roy,

could you pelase verify if the following patch fixes the problem for you ?

Code: Select all

--- a/src/Makefile
+++ b/src/Makefile
@@ -106,6 +106,7 @@ gcc:
 	$(MAKE) \
 	CXX='g++' \
 	CXXFLAGS="$(GCCFLAGS)" \
+	CXXFLAGS+='-fno-strict-aliasing' \	
 	all

Thanks
Marco


P.S: BTW I strongly suggest Intel compiler for Linux, binary are way faster. Use 'make icc-profile' or 'make icc-profile-popcnt' if you have a CPU with POPCNT support and the make will do all the steps to build up a pgo binary in a fully automatic fashion.

Marco -- I made the change you requested and did a 'make clean' and 'make' and the sum for the binary produced is the same as it was before, so I do not expect any change in behavior.

Roy

[mcostalba]

I think the problem is in the Makefile. I used the Makefile from the Stockfish-1.5 build to make the 1.6 version and this behavior is now gone (...Rc1 not preferred EVER -- and that's a good thing).

Roy

Hi Roy,

could you pelase verify if the following patch fixes the problem for you ?

Code: Select all

--- a/src/Makefile
+++ b/src/Makefile
@@ -106,6 +106,7 @@ gcc:
 	$(MAKE) \
 	CXX='g++' \
 	CXXFLAGS="$(GCCFLAGS)" \
+	CXXFLAGS+='-fno-strict-aliasing' \	
 	all

Thanks
Marco


P.S: BTW I strongly suggest Intel compiler for Linux, binary are way faster. Use 'make icc-profile' or 'make icc-profile-popcnt' if you have a CPU with POPCNT support and the make will do all the steps to build up a pgo binary in a fully automatic fashion.

Marco -- I made the change you requested and did a 'make clean' and 'make' and the sum for the binary produced is the same as it was before, so I do not expect any change in behavior.

Roy

I am not able to check the Linux compile until tomorrow, if you have an intel compiler installed you could try to build the Intel binary and see as example with the floowing command line

./stockfish bench 128 1 12 default depth

If node count is the same in both cases, for gcc and Intel. If you post the node count I can verify also under Windows (in this release we have fixed the problem of 1.5.1 of showing different node counts on different platforms).

[royb]

I think the problem is in the Makefile. I used the Makefile from the Stockfish-1.5 build to make the 1.6 version and this behavior is now gone (...Rc1 not preferred EVER -- and that's a good thing).

Roy

Hi Roy,

could you pelase verify if the following patch fixes the problem for you ?

Code: Select all

--- a/src/Makefile
+++ b/src/Makefile
@@ -106,6 +106,7 @@ gcc:
 	$(MAKE) \
 	CXX='g++' \
 	CXXFLAGS="$(GCCFLAGS)" \
+	CXXFLAGS+='-fno-strict-aliasing' \	
 	all

Thanks
Marco


P.S: BTW I strongly suggest Intel compiler for Linux, binary are way faster. Use 'make icc-profile' or 'make icc-profile-popcnt' if you have a CPU with POPCNT support and the make will do all the steps to build up a pgo binary in a fully automatic fashion.

Marco -- I made the change you requested and did a 'make clean' and 'make' and the sum for the binary produced is the same as it was before, so I do not expect any change in behavior.

Roy

I am not able to check the Linux compile until tomorrow, if you have an intel compiler installed you could try to build the Intel binary and see as example with the floowing command line

./stockfish bench 128 1 12 default depth

If node count is the same in both cases, for gcc and Intel. If you post the node count I can verify also under Windows (in this release we have fixed the problem of 1.5.1 of showing different node counts on different platforms).

I do not have an Intel compiler, so we will have to wait till tomorrow. Thanks for the suggestions!

Roy