I'm testing some autoparallelization on my mactel running g++4.3.3 which I
downloaded from macports.
In every other run, I see my code hang, though I cannot find an error in gdb.
Through a long series of cout-debugging steps, have traced it to the beginning
of an omp loop. The code does not hang on a linux box running g++4.1.0.
The code in question looks like this:
utility::vector1< DOF_Node * > dof_nodes;
//dof_nodes.reserve( min_map.size() );
for ( MinimizerMap::iterator iter = min_map.begin(), iter_e =
min_map.end();
iter != iter_e; ++iter ) {
dof_nodes.push_back( *iter );
}
std::cout << "s" << std::flush;
int const ndofs = dof_nodes.size();
#pragma omp parallel for
for ( int ii = 1; ii <= ndofs; ++ii ) {
std::cout << "*" << std::flush;
DOF_Node & dof_node( *dof_nodes[ ii ] );
std::cout << "(" << ii << "," << dof_nodes[ ii ] << ")" <<
std::flush;
// loop through atoms first moved by this torsion
for ( DOF_Node::AtomIDs::const_iterator
it1=dof_node.atoms().begin(),
it1e = dof_node.atoms().end();
it1 != it1e; ++it1 ) {
id::AtomID const & atom_id( *it1 );
scorefxn.eval_atom_derivative( atom_id, pose,
min_map.domain_map(), dof_node.F1(), dof_node.F2() );
} // atom1
} // tor
//std::cout << "B" << std::flush;
sample output from a hang: notice that it ends with "s"
SEADb0x1SEADf0SEADb0x1SEADf0*(0x2c,0x83933c0)SEADb0x1SEADf0SEADb0x1SEADf0SEADb0x1SEADf0*(0x2d,0x8397aa0)SEADb0x1SEADf0SEADb0x1SEADf0*(0x2e,0x8393360)SEADb0x1SEADf0SEADb0x1SEADf0SEADb0x1SEADf0*(0x2f,0x83c67d0)SEADb0x1SEADf0SEADb0x1SEADf0SEADb0x1SEADf0*(0x30,0x83c6750)SEADb0x1SEADf0*(0x31,0x83c66d0)SEADb0x1SEADf0SEADb0x1SEADf0*(0x32,0x83c6650)SEADb0x1SEADf0SEADb0x1SEADf0SEADb0x1SEADf0*(0x33,0x83c65d0)SEADb0x1SEADf0SEADb0x1SEADf0s
"SEAD" is from the function called within this loop -- Scorefunction Evaluate
Atom Derivative. -- I output SEAD twice, once with a "b" for beginning of the
function and once with an "f" for finished.
(It's unclear to me why the integer ii is getting written out in hex.)
The code hangs in several places besides this loop. I have not found any
similar bug reporting for g++4.3.3 on the mac.
It is certainly possible that there is a bug somewhere else in the code that
results in the hang here. I'm at a loss, though, for how to detect it. If
anyone has even the most mild of suggestions of where to begin, I'd be glad to
hear them.
--
Summary: g++4.3.3 OpenMP (aka omp) for loop hangs
Product: gcc
Version: 4.3.3
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libgomp
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: aleaverfay at gmail dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39217
