Current state of compiler differences STOCKFISH

[Fulvio]

I use this to compile the Stockfish-AVX2 with msvc:

Code: Select all

 cl /std:c++17 /DNDEBUG /DUSE_POPCNT /DUSE_AVX2 /DUSE_SSE41 /DUSE_SSSE3 /DUSE_SSE2  /GS- /MT /O2 /Oi /Ot /Oy /GL /EHsc *.cpp nnue/*.cpp nnue/features/*.cpp syzygy/*.cpp /link /LTCG /STACK:reserve=8388608 /OUT:stockfish.exe advapi32.lib

Anyone have something better?

[Karlo Bala]

I use this to compile the Stockfish-AVX2 with msvc:

Code: Select all

 cl /std:c++17 /DNDEBUG /DUSE_POPCNT /DUSE_AVX2 /DUSE_SSE41 /DUSE_SSSE3 /DUSE_SSE2  /GS- /MT /O2 /Oi /Ot /Oy /GL /EHsc *.cpp nnue/*.cpp nnue/features/*.cpp syzygy/*.cpp /link /LTCG /STACK:reserve=8388608 /OUT:stockfish.exe advapi32.lib

Anyone have something better?

Using MSVC compiler with the PGO can be up to 30% faster than without.

[Fulvio]

Using MSVC compiler with the PGO can be up to 30% faster than without.

Cool!
I got ~6% improvement.
I added to the cl command /GENPROFILE
then run "stockfish.exe bench"
and then recompiled replacing /GENPROFILE with /USEPROFILE

[dangi12012]

I use this to compile the Stockfish-AVX2 with msvc:

Code: Select all

 cl /std:c++17 /DNDEBUG /DUSE_POPCNT /DUSE_AVX2 /DUSE_SSE41 /DUSE_SSSE3 /DUSE_SSE2  /GS- /MT /O2 /Oi /Ot /Oy /GL /EHsc *.cpp nnue/*.cpp nnue/features/*.cpp syzygy/*.cpp /link /LTCG /STACK:reserve=8388608 /OUT:stockfish.exe advapi32.lib

Anyone have something better?

Using MSVC compiler with the PGO can be up to 30% faster than without.

But even with +30% its still slower?
Also its like overclocking. You can turn on PGO for clang and gcc as well. But its a lot harder to setup correctly compared to a normal compile.

[Karlo Bala]

Using MSVC compiler with the PGO can be up to 30% faster than without.

Cool!
I got ~6% improvement.
I added to the cl command /GENPROFILE
then run "stockfish.exe bench"
and then recompiled replacing /GENPROFILE with /USEPROFILE

Well, not much

Can you upload somewhere the entire solution, I would like to give it a try (and something to compare with).

[Sopel]

I use this to compile the Stockfish-AVX2 with msvc:

Code: Select all

 cl /std:c++17 /DNDEBUG /DUSE_POPCNT /DUSE_AVX2 /DUSE_SSE41 /DUSE_SSSE3 /DUSE_SSE2  /GS- /MT /O2 /Oi /Ot /Oy /GL /EHsc *.cpp nnue/*.cpp nnue/features/*.cpp syzygy/*.cpp /link /LTCG /STACK:reserve=8388608 /OUT:stockfish.exe advapi32.lib

Anyone have something better?

Using MSVC compiler with the PGO can be up to 30% faster than without.

But even with +30% its still slower?
Also its like overclocking. You can turn on PGO for clang and gcc as well. But its a lot harder to setup correctly compared to a normal compile.

It's literally profile-build instead of build. Just says how terribly useless your measurements are.

[dangi12012]

Using MSVC compiler with the PGO can be up to 30% faster than without.

Cool!
I got ~6% improvement.
I added to the cl command /GENPROFILE
then run "stockfish.exe bench"
and then recompiled replacing /GENPROFILE with /USEPROFILE

Well, not much

Can you upload somewhere the entire solution, I would like to give it a try (and something to compare with).

Yes sure https://1drv.ms/u/s!AoDIyNlDG2QTg8YSjED ... Q?e=wRReGm
You can just copy the sln and the vcxproj file from there. Not more is needed.

BMI2+EVX2 is enabled via settings in the vcxproj.
Share with us if you have some news

[Fulvio]

Can you upload somewhere the entire solution, I would like to give it a try (and something to compare with).

I compiled it directly from the command prompt (x64 Native Tools Command Prompt for VS 2022).
I downloaded Stockfish's source code into a directory "stockfish".
I downloaded the net ( https://tests.stockfishchess.org/api/nn ... 602d6.nnue ) into the directory "stockfish/src".
I opened the VS2022 command prompt and cd into "stockfish/src" and compiled it with:

Code: Select all

cl /std:c++17 /DNDEBUG /DUSE_POPCNT /DUSE_AVX2 /DUSE_SSE41 /DUSE_SSSE3 /DUSE_SSE2  /GS- /MT /O2 /Oi /Ot /Oy /GL /EHsc *.cpp nnue/*.cpp nnue/features/*.cpp syzygy/*.cpp /link /GENPROFILE /LTCG /STACK:reserve=8388608 /OUT:stockfish.exe advapi32.lib

stockfish.exe bench

cl /std:c++17 /DNDEBUG /DUSE_POPCNT /DUSE_AVX2 /DUSE_SSE41 /DUSE_SSSE3 /DUSE_SSE2  /GS- /MT /O2 /Oi /Ot /Oy /GL /EHsc *.cpp nnue/*.cpp nnue/features/*.cpp syzygy/*.cpp /link /USEPROFILE /LTCG /STACK:reserve=8388608 /OUT:stockfish.exe advapi32.lib

[Fulvio]

BMI2+EVX2 is enabled via settings in the vcxproj.
Share with us if you have some news

I get a ~25% slower executable using that vcxproj

[AndrewGrant]

BMI2+EVX2 is enabled via settings in the vcxproj.
Share with us if you have some news

I get a ~25% slower executable using that vcxproj

You happen to have a Ryzen by chance? PEXT could produce that level of slowdown. I notice your original compilation does not include BMI2.