Dear András,
I have made some progress trying to compile the framework with Intel's compiler.
The reason why I try is that I have an account at the university's leo3 HPC with Intel
processors, and I hope for increased performance.
In debug mode, the testsuitePhysics is passed. In release mode, testsuitePhysics is passed
except tests 4 and 64, where I get SIGBUS or SIGSEGV depending on the intel compiler version.
This happens in cpputils::details::concatenateHelper upon return std::copy(t.begin(),t.end(),iter);
Unfortunately my own script also crashes with SIGBUS.
Preliminary tests show that the scripts which succeed are indeed faster, though I have to investigate
this more systematically.
I will also try to find the exact compiler flags which cause the crashes.
Best regards
Raimar
P.S. let me know if you want to experiment with this on teazer, I have a version of gsl compiled
with the intel compiler and I can send you the user-config.jam which points to this library.
The changes that were necessary to compile the code are in the branch cppqed/raimar/intel.
Here is a summary:
1. namespace problems which icpc has for some reasons
========================================================================
2. Undefined references to some static const members, fixed by e.g.
=== modified file 'utils/include/impl/BlitzArraySliceIterator.tcc'
--- utils/include/impl/BlitzArraySliceIterator.tcc 2011-09-19 11:29:15 +0000
+++ utils/include/impl/BlitzArraySliceIterator.tcc 2012-04-18 11:49:57 +0000
@@ -205,6 +205,9 @@
template<int RANK, typename V>
typename IndexerBase<RANK,V>::Idx IndexerBase<RANK,V>::cache_;
+template<int RANK, typename V>
+const typename IndexerBase<RANK,V>::Helper IndexerBase<RANK,V>::helper_ = IndexerBase<RANK,V>::Helper();
+
========================================================================
3. Here I am not sure why gcc compiles the code, as the constructor which is called is missing:
=== modified file 'elements/frees/MultiLevel.h'
--- elements/frees/MultiLevel.h 2011-08-11 13:21:25 +0000
+++ elements/frees/MultiLevel.h 2012-04-18 11:49:57 +0000
@@ -316,7 +316,7 @@
// here, the compiler cannot deduce NL
// anymore
const VP& etas, const VL& gammas,
- const Averaged& averaged=multilevel::DiagonalDO(NL))
+ const Averaged& averaged=multilevel::DiagonalDO("",NL))
{
return RETURN_type(new PumpedLossyMultiLevelSch<NL,VP,VL,Averaged>(deltas,etas,gammas,averaged));
}
@@ -325,7 +325,7 @@
template<int NL, typename VP, typename VL, typename Averaged>
inline
RETURN_type
-makePumpedLossyMultiLevelSch(const multilevel::ParsPumpedLossy<NL,VP,VL>& p, const Averaged& averaged=multilevel::DiagonalDO(NL))
+makePumpedLossyMultiLevelSch(const multilevel::ParsPumpedLossy<NL,VP,VL>& p, const Averaged& averaged=multilevel::DiagonalDO("",NL))
{
return RETURN_type(new PumpedLossyMultiLevelSch<NL,VP,VL,Averaged>(p.deltas,p.etas,p.gammas,averaged));
}
=========================================================================
4. Here I have no clue what is going on, adding the public keyword to a struct should not make any difference.
But icpc complains with "... is inaccessible" without it.
=== modified file 'elements/interactions/impl/MLJC.tcc'
--- elements/interactions/impl/MLJC.tcc 2011-10-07 23:10:35 +0000
+++ elements/interactions/impl/MLJC.tcc 2012-04-18 11:49:57 +0000
@@ -23,6 +23,7 @@
template<int NL, typename VC> template<int N1, int N2>
struct Base<NL,VC>::ModeDynamics : public tmptools::pair_c<N1,N2>
{
+public:
ModeDynamics(const MultiLevelBase<NL>& ml, const ModeBase& mode, const dcomp& g)
: a (-g*mode::aop(&mode)),
adagger((g*mode::aop(&mode)).dagger())
|
