fix all lambda lib and cuSTL related errors if compiling with IntelIcc compiler.

[heikoburau]

No description provided.

[psychocoderHPC]

Very nice. 👍
It would by nice if you can add same comments in the changed files. Fix something is the right time to comment files.

[ax3l]

This pull should fix #76 ...

[ax3l]

Way cleaner now 👍
Remember: ICC testing is not included in the compile suite, please test by hand...

[ax3l]

I tested it now by hand mainline:dev vs Heikman:fix/IntelIcc:

• gcc/4.6 + boost/1.49: ☑ | ☑
• icc/12.0 + boost/1.49: ☐ | ☒
• icc/12.0 + boost/1.54: ☐ | ☒
• icc/12.1 + boost/1.49: ☒ | ☒
• icc/12.1 + boost/1.54: ☒ | ☒

Unfortunately: not fixed...

@Heikman What did you test with?

CUDA 5.0: ICC 12.X
CUDA 5.5: ICC 12.1

[psychocoderHPC]

I have looked over the code and can only see a nicer implementation of the old code. Can you please explane what was the problem which trigger the bug #76 and how you solved it.

[ax3l]

The problems seems to be freaking variadic templates in include/boost/proto/detail/preprocessed/expr_variadic.hpp ...
Probably gcc knows them already and one has to compile a separate boost for icc...

Actually, I do not want to rely on proto if it causes that on every system we run they have to compile boost for each compiler independently because of that...

Edit: ICC states it knows them since 12.1+: link

[ax3l]

@Heikman I think the proto includes are outdated, aren't they?

diff --git a/src/libPMacc/include/lambda/placeholder.h b/src/libPMacc/include/lambda/placeholder.h
index 78c18eb..ee381b9 100644
--- a/src/libPMacc/include/lambda/placeholder.h
+++ b/src/libPMacc/include/lambda/placeholder.h
@@ -23,9 +23,6 @@
 #define LAMBDA_PLACEHOLDER_H

 #include <boost/mpl/int.hpp>
-#include <boost/proto/core.hpp>
-
-namespace proto = boost::proto;

 namespace PMacc
 {
@@ -35,7 +32,7 @@ namespace lambda
 template<int I>
 struct placeholder {};

-} // lamnda
+} // lambda
 } // PMacc

[heikoburau]

I didn't mean that I solved all Intel compiler bugs but only the cuSTL and lambdaLib related ones, sorry ;)
At least one error comes from the hdf5 code. I used icc/12.1, boost/1.54.0 and cuda/5.0

The problem I solved were some out-of-range compile-time accesses on a mpl::vector in Expression.hpp. But as they were kind of optimized out by gcc they only appeared along with icc.

@ax3l You are right with proto, they are definitly outdated!

[psychocoderHPC]

@ax3l and @Heikman
I have delete the include of proto and add
#define BOOST_NO_CXX11_VARIADIC_TEMPLATES 1 to libPMacc/include/types.h

Merge #101 to this branch

Than all template bugs are gone but we get new errors (with icc 13.1) and some warnings

/usr/include/fenv.h(86): error: exception specification is incompatible with that of previous function "fegetround" (declared at line 63 of "/opt/pkg/compiler/intel/fc/13.1/composer_xe_2013.3.163/compiler/include/float.h")
  extern "C" int fegetround() throw();
                              ^

UPDATE:
This default linux file /usr/include/fenv.h ships int fegetround() throw() but intel compiler only ship int fegetround()

[ax3l]

I also ended up with fenv.h yesterday evening, caused by "newer than 4.4" gnu headers icc is not compatible with but using. One can by hand set the CPLUS_INCLUDE_PATH to the /usr/include/c++/4.xx/ dirs, but that's a hack.

I also added BOOST_NO_CXX11_VARIADIC_TEMPLATES but in our CMake icc-section since this is an icc bug in the experimental implementation of c++11 variadic templates, I guess link.
There are some more Flags to play with: -U__GXX_EXPERIMENTAL_CXX0X__ and -std=c++0x.

Work around the HDF5 related errors by unloading hdf5 and splash or setting export SPLASH_ROOT=.

@Heikman Can you please remove proto as I stated here ? Afterwards that's ready to get merged.

[ax3l]

Fixed icc 12.1+ with boost now.
They catched that error for icc <1200, probably it is still not working.

diff --git a/src/picongpu/CMakeLists.txt b/src/picongpu/CMakeLists.txt
index ebdb7ee..c9a2b49 100644
--- a/src/picongpu/CMakeLists.txt
+++ b/src/picongpu/CMakeLists.txt
@@ -236,6 +236,9 @@ if(CMAKE_COMPILER_IS_GNUCXX)
 # ICC
 elseif("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Intel")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wall")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DBOOST_NO_VARIADIC_TEMPLATES")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DBOOST_NO_CXX11_VARIADIC_TEMPLATES")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DBOOST_NO_FENV_H")
 # PGI
 elseif("${CMAKE_CXX_COMPILER_ID}" STREQUAL "PGI")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Minform=inform")

[ax3l]

@Heikman I think that pull request is fine.
I will remove proto in a new pull request and I will set the icc flags there, too.