Merge pull request #536 from akohlmey/fix-nvcc-openmp-conflicts

Implement workaround for NVCC incompatibilities with OpenMP directives

src/USER-OMP/pair_comb_omp.cpp

@@ -484,7 +484,7 @@ double PairCombOMP::yasu_char(double *qf_fix, int &igroup)
 
         qfo_field(&params[iparam_ij],rsq1,iq,jq,fqji,fqjj);
         fqi   += jq * fqij + fqji;
-#if defined(_OPENMP)
+#if defined(_OPENMP) && !defined(__NVCC__)
 #pragma omp atomic
 #endif
         qf[j] += (iq * fqij + fqjj);
@@ -511,13 +511,13 @@ double PairCombOMP::yasu_char(double *qf_fix, int &igroup)
 
         qfo_short(&params[iparam_ij],i,nj,rsq1,iq,jq,fqij,fqjj);
         fqi += fqij;
-#if defined(_OPENMP)
+#if defined(_OPENMP) && !defined(__NVCC__)
 #pragma omp atomic
 #endif
         qf[j] += fqjj;
       }
 
-#if defined(_OPENMP)
+#if defined(_OPENMP) && !defined(__NVCC__)
 #pragma omp atomic
 #endif
       qf[i] += fqi;

src/USER-OMP/reaxc_torsion_angles_omp.cpp

@@ -69,7 +69,9 @@ void Torsion_AnglesOMP( reax_system *system, control_params *control,
   double total_Econ = 0;
   int  nthreads = control->nthreads;
 
+#if defined(_OPENMP)
 #pragma omp parallel default(shared) reduction(+: total_Etor, total_Econ)
+#endif
   {
   int i, j, k, l, pi, pj, pk, pl, pij, plk;
   int type_i, type_j, type_k, type_l;
@@ -125,7 +127,9 @@ void Torsion_AnglesOMP( reax_system *system, control_params *control,
                                     system->N, system->pair_ptr->eatom,
                                     system->pair_ptr->vatom, thr);
 
+#if defined(_OPENMP)
 #pragma omp for schedule(static)
+#endif
   for (j = 0; j < system->N; ++j) {
     start_j = Start_Index(j, bonds);
     end_j = End_Index(j, bonds);
@@ -137,7 +141,9 @@ void Torsion_AnglesOMP( reax_system *system, control_params *control,
     }
   }
 
+#if defined(_OPENMP)
 #pragma omp for schedule(dynamic,50)
+#endif
   for (j = 0; j < natoms; ++j) {
     type_j = system->my_atoms[j].type;
     Delta_j = workspace->Delta_boc[j];

src/USER-OMP/reaxc_valence_angles_omp.cpp

@@ -124,8 +124,9 @@ void Valence_AnglesOMP( reax_system *system, control_params *control,
   int  nthreads = control->nthreads;
   int  num_thb_intrs = 0;
   int  TWICE = 2;
-
+#if defined(_OPENMP)
 #pragma omp parallel default(shared) reduction(+:total_Eang, total_Epen, total_Ecoa, num_thb_intrs)
+#endif
   {
     int i, j, pi, k, pk, t;
     int type_i, type_j, type_k;
@@ -180,7 +181,9 @@ void Valence_AnglesOMP( reax_system *system, control_params *control,
 
     const int per_thread = thb_intrs->num_intrs / nthreads;
 
+#if defined(_OPENMP)
 #pragma omp for schedule(dynamic,50)
+#endif
     for (j = 0; j < system->N; ++j) {
       type_j = system->my_atoms[j].type;
       _my_offset[j] = 0;
@@ -251,11 +254,14 @@ void Valence_AnglesOMP( reax_system *system, control_params *control,
     } // for(j)
 
     // Wait for all threads to finish counting angles
+#if defined(_OPENMP) && !defined(__NVCC__)
 #pragma omp barrier
-
+#endif
     // Master thread uses angle counts to compute offsets
     // This can be threaded
+#if defined(_OPENMP) && !defined(__NVCC__)
 #pragma omp master
+#endif
     {
       int current_count = 0;
       int m = _my_offset[0];
@@ -269,12 +275,15 @@ void Valence_AnglesOMP( reax_system *system, control_params *control,
     }
 
     // All threads wait till master thread finished computing offsets
+#if defined(_OPENMP) && !defined(__NVCC__)
 #pragma omp barrier
-
+#endif
     // Original loop, but now using precomputed offsets
     // Safe to use all threads available, regardless of threads tasked above
     // We also now skip over atoms that have no angles assigned
+#if defined(_OPENMP)
 #pragma omp for schedule(dynamic,50)//(dynamic,chunksize)//(guided)
+#endif
     for (j = 0; j < system->N; ++j) {         // Ray: the first one with system->N
       type_j = system->my_atoms[j].type;
       if(type_j < 0) continue;