Merge pull request #679 from stanmoore1/kk_reax_comm

ReaxC optimizations

src/KOKKOS/fix_qeq_reax_kokkos.cpp

@@ -63,6 +63,7 @@ FixQEqReaxKokkos(LAMMPS *lmp, int narg, char **arg) :
 
   nmax = nmax = m_cap = 0;
   allocated_flag = 0;
+  nprev = 4;
 }
 
 /* ---------------------------------------------------------------------- */
@@ -158,15 +159,15 @@ void FixQEqReaxKokkos<DeviceType>::init_hist()
 {
   int i,j;
 
-  k_s_hist = DAT::tdual_ffloat_2d("qeq/kk:s_hist",atom->nmax,5);
+  k_s_hist = DAT::tdual_ffloat_2d("qeq/kk:s_hist",atom->nmax,nprev);
   d_s_hist = k_s_hist.template view<DeviceType>();
   h_s_hist = k_s_hist.h_view;
-  k_t_hist = DAT::tdual_ffloat_2d("qeq/kk:t_hist",atom->nmax,5);
+  k_t_hist = DAT::tdual_ffloat_2d("qeq/kk:t_hist",atom->nmax,nprev);
   d_t_hist = k_t_hist.template view<DeviceType>();
   h_t_hist = k_t_hist.h_view;
 
   for( i = 0; i < atom->nmax; i++ )
-    for( j = 0; j < 5; j++ )
+    for( j = 0; j < nprev; j++ )
       k_s_hist.h_view(i,j) = k_t_hist.h_view(i,j) = 0.0;
 
   k_s_hist.template modify<LMPHostType>();
@@ -334,11 +335,11 @@ void FixQEqReaxKokkos<DeviceType>::allocate_array()
     d_d = k_d.template view<DeviceType>();
     h_d = k_d.h_view;
 
-    k_s_hist = DAT::tdual_ffloat_2d("qeq/kk:s_hist",nmax,5);
+    k_s_hist = DAT::tdual_ffloat_2d("qeq/kk:s_hist",nmax,nprev);
     d_s_hist = k_s_hist.template view<DeviceType>();
     h_s_hist = k_s_hist.h_view;
 
-    k_t_hist = DAT::tdual_ffloat_2d("qeq/kk:t_hist",nmax,5);
+    k_t_hist = DAT::tdual_ffloat_2d("qeq/kk:t_hist",nmax,nprev);
     d_t_hist = k_t_hist.template view<DeviceType>();
     h_t_hist = k_t_hist.h_view;
   }
@@ -368,7 +369,7 @@ void FixQEqReaxKokkos<DeviceType>::zero_item(int ii) const
     d_o[i] = 0.0;
     d_r[i] = 0.0;
     d_d[i] = 0.0;
-    //for( int j = 0; j < 5; j++ )
+    //for( int j = 0; j < nprev; j++ )
       //d_s_hist(i,j) = d_t_hist(i,j) = 0.0;
   }
 
@@ -1087,7 +1088,7 @@ void FixQEqReaxKokkos<DeviceType>::calculate_q_item(int ii) const
   if (mask[i] & groupbit) {
     q(i) = d_s[i] - delta * d_t[i];
 
-    for (int k = 4; k > 0; --k) {
+    for (int k = nprev-1; k > 0; --k) {
       d_s_hist(i,k) = d_s_hist(i,k-1);
       d_t_hist(i,k) = d_t_hist(i,k-1);
     }
@@ -1173,7 +1174,7 @@ double FixQEqReaxKokkos<DeviceType>::memory_usage()
 {
   double bytes;
 
-  bytes = atom->nmax*5*2 * sizeof(F_FLOAT); // s_hist & t_hist
+  bytes = atom->nmax*nprev*2 * sizeof(F_FLOAT); // s_hist & t_hist
   bytes += atom->nmax*8 * sizeof(F_FLOAT); // storage
   bytes += n_cap*2 * sizeof(int); // matrix...
   bytes += m_cap * sizeof(int);

src/KOKKOS/pair_reaxc_kokkos.cpp

@@ -131,6 +131,8 @@ template<class DeviceType>
 void PairReaxCKokkos<DeviceType>::init_style()
 {
   PairReaxC::init_style();
+  if (fix_reax) modify->delete_fix("REAXC"); // not needed in the Kokkos version
+  fix_reax = NULL;
 
   // irequest = neigh request made by parent class
 
@@ -555,8 +557,8 @@ void PairReaxCKokkos<DeviceType>::Deallocate_Lookup_Tables()
 
   ntypes = atom->ntypes;
 
-  for( i = 0; i < ntypes; ++i ) {
-    for( j = i; j < ntypes; ++j )
+  for( i = 0; i <= ntypes; ++i ) {
+    for( j = i; j <= ntypes; ++j )
       if( LR[i][j].n ) {
         sfree( LR[i][j].y, "LR[i,j].y" );
         sfree( LR[i][j].H, "LR[i,j].H" );

src/USER-OMP/fix_qeq_reax_omp.cpp

@@ -703,7 +703,7 @@ void FixQEqReaxOMP::calculate_Q()
       q[i] = s[i] - u * t[i];
 
       // backup s & t
-      for (int k = 4; k > 0; --k) {
+      for (int k = nprev-1; k > 0; --k) {
         s_hist[i][k] = s_hist[i][k-1];
         t_hist[i][k] = t_hist[i][k-1];
       }

src/USER-REAXC/fix_qeq_reax.cpp

@@ -95,7 +95,7 @@ FixQEqReax::FixQEqReax(LAMMPS *lmp, int narg, char **arg) :
   pack_flag = 0;
   s = NULL;
   t = NULL;
-  nprev = 5;
+  nprev = 4;
 
   Hdia_inv = NULL;
   b_s = NULL;
@@ -817,7 +817,7 @@ void FixQEqReax::calculate_Q()
       q[i] = s[i] - u * t[i];
 
       /* backup s & t */
-      for (k = 4; k > 0; --k) {
+      for (k = nprev-1; k > 0; --k) {
         s_hist[i][k] = s_hist[i][k-1];
         t_hist[i][k] = t_hist[i][k-1];
       }