Clang vs GCC [stockfish 4]

[Rein Halbersma]

There are plenty of links out there on how to install the latest Clang compiler and C/C++ headers (or the LLVM new libc++ headers). You might want to check clang --version on your system to see if you are up to date.

I think Bob has mentioned MacPorts. He can get everything he wants there.

Well, without specific version info on the compiler and standard headers it' all speculation what went wrong . For C++ at least, Clang 3.3 is fully conforming at pretty much the same level as gcc 4.8.

[JuLieN]

There are plenty of links out there on how to install the latest Clang compiler and C/C++ headers (or the LLVM new libc++ headers). You might want to check clang --version on your system to see if you are up to date.

I think Bob has mentioned MacPorts. He can get everything he wants there.

Well, without specific version info on the compiler and standard headers it' all speculation what went wrong . For C++ at least, Clang 3.3 is fully conforming at pretty much the same level as gcc 4.8.

For reference, my version is:

Code: Select all

Apple LLVM version 4.2 (clang-425.0.28) (based on LLVM 3.2svn)
Target: x86_64-apple-darwin12.4.0
Thread model: posix

I think it's just the one that ships with the current XCode (not the beta for the upcoming one).

[Rein Halbersma]

Clang 3.2 should be good enough, but you should also check your c/c++ headers. If they are provided by gcc 4.2 then you want to upgrade to at least 4.7

[bob]

I compiled the official Stockfish 4's sources with clang, and got an exe significantly faster than the official gcc build for mac included in the archive. Here are the compared results of the bench command and a search to depth 30 in the starting position:

Code: Select all

Official:
===========================
Total time (ms) : 2836
Nodes searched  : 4132374
Nodes/second    : 1457113

info depth 30 seldepth 46 score cp 30 nodes 1699401570 nps 1207059 time 1407886 multipv 1 pv d2d4 g8f6 g1f3 e7e6 c2c4 f8b4 c1d2 b4d2 d1d2 b7b6 e2e3 e8g8 f1e2 f6e4 d2d3 c8b7 e1g1 d7d6 f3d2 e4d2 d3d2 d8g5 g2g3 b8d7 b1c3 d7f6 a1d1
info nodes 1699401570 time 1407886
bestmove d2d4 ponder g8f6

clang:
===========================
Total time (ms) : 2691
Nodes searched  : 4132374
Nodes/second    : 1535627

info depth 30 seldepth 46 score cp 30 nodes 1699401570 nps 1293199 time 1314106 multipv 1 pv d2d4 g8f6 g1f3 e7e6 c2c4 f8b4 c1d2 b4d2 d1d2 b7b6 e2e3 e8g8 f1e2 f6e4 d2d3 c8b7 e1g1 d7d6 f3d2 e4d2 d3d2 d8g5 g2g3 b8d7 b1c3 d7f6 a1d1
info nodes 1699401570 time 1314106
bestmove d2d4 ponder g8f6

the GCC build takes 23mn and 27 seconds, while the clang one takes 21mn and 54 seconds on my corei7 CPU.

I don't say that Clang produces faster code than GCC, but in some occurrences, like this one, it clearly does. So if you have access to both compilers you might want to try them both for your engine.

clang has bugs. The version that came with xcode on my 8-month-old macbook will NOT compile a functional crafty. gcc 4.7.2 works perfectly, however. I don't trust clang, period. Did you PGO and such?

clang is an old gcc version...

This is a known issue for some versions of XCode. It includes an old (v2.9) Clang compiler and an even older C/C++ Standard Library (based on gcc 4.2). There are plenty of links out there on how to install the latest Clang compiler and C/C++ headers (or the LLVM new libc++ headers). You might want to check clang --version on your system to see if you are up to date.

I got this new macbook in December of last year. I had to install xcode myself, and then do the various updates. At that point in time, december or january, clang broke crafty badly. I used macports to download 4.7.2 and build it, and it worked perfectly as it generally always has (at least over last 10+ years, I was an EARLY "long long" user and it was certainly "dicey" to start with...

[bob]

I compiled the official Stockfish 4's sources with clang, and got an exe significantly faster than the official gcc build for mac included in the archive. Here are the compared results of the bench command and a search to depth 30 in the starting position:

Code: Select all

Official:
===========================
Total time (ms) : 2836
Nodes searched  : 4132374
Nodes/second    : 1457113

info depth 30 seldepth 46 score cp 30 nodes 1699401570 nps 1207059 time 1407886 multipv 1 pv d2d4 g8f6 g1f3 e7e6 c2c4 f8b4 c1d2 b4d2 d1d2 b7b6 e2e3 e8g8 f1e2 f6e4 d2d3 c8b7 e1g1 d7d6 f3d2 e4d2 d3d2 d8g5 g2g3 b8d7 b1c3 d7f6 a1d1
info nodes 1699401570 time 1407886
bestmove d2d4 ponder g8f6

clang:
===========================
Total time (ms) : 2691
Nodes searched  : 4132374
Nodes/second    : 1535627

info depth 30 seldepth 46 score cp 30 nodes 1699401570 nps 1293199 time 1314106 multipv 1 pv d2d4 g8f6 g1f3 e7e6 c2c4 f8b4 c1d2 b4d2 d1d2 b7b6 e2e3 e8g8 f1e2 f6e4 d2d3 c8b7 e1g1 d7d6 f3d2 e4d2 d3d2 d8g5 g2g3 b8d7 b1c3 d7f6 a1d1
info nodes 1699401570 time 1314106
bestmove d2d4 ponder g8f6

According to the bench, your clang binary is about 5% faster than the official one. I've tried all recent versions of both gcc and clang (llvm). My best is a binary that is about 11% faster than the official one. This was produced using gcc-4.6 from MacPorts. popcount was used, PGO was used, but not LTO. Add -fno-tree-pre to the optimization flags.

clang is improving, but for me it has never produced a faster binary than some version of gcc. Don't have icc, since it's not free!

my gcc-4.6 binary:

Code: Select all

Total time (ms) : 3198
Nodes searched  : 4132374
Nodes/second    : 1292174

official binary:

Code: Select all

Total time (ms) : 3552
Nodes searched  : 4132374
Nodes/second    : 1163393

This is on an old Nehalem Xeon running at 2.26 GHz. One thread.

Note that if you run linux, icc is free. If you use a mac, it is not free, however. Seems odd, but oh well...

[zullil]

For reference, my version is:

Code: Select all

Apple LLVM version 4.2 (clang-425.0.28) (based on LLVM 3.2svn)
Target: x86_64-apple-darwin12.4.0
Thread model: posix

I think it's just the one that ships with the current XCode (not the beta for the upcoming one).

Yes, that's the version that comes with Xcode 4.6. It's Apple-specific, and is not quite clang-3.2. It will sometime fail to compile code that causes clang-3.2 no trouble.

[jdart]

I don't think Intel's C/C++ compiler even supports Mac OS; see http://software.intel.com/en-us/intel-compilers. They have a Fortran compiler on that platform, though .

--Jon

[zullil]

I don't think Intel's C/C++ compiler even supports Mac OS

--Jon

It always has, and I think it still does.

http://software.intel.com/en-us/intel-s ... -mac-os-x/

Somebody at Intel should fix that other webpage.

[bob]

I don't think Intel's C/C++ compiler even supports Mac OS; see http://software.intel.com/en-us/intel-compilers. They have a Fortran compiler on that platform, though .

--Jon

I think it is a typo. icc comes with the fortran compiler. I used that when we found the old version of Cray Blitz source. To make a few changes as needed to make it compile on a normal (non-cray) compiler. We used something called "task common" that Cray supported but normal FORTRAN compilers did not. needed to fix that to make things compile and run.

[zullil]

I think it is a typo. icc comes with the fortran compiler.

Apparently, for OS X, the C++ and Fortran compilers are now disjoint products.


http://software.intel.com/en-us/intel-s ... -mac-os-x/