Avoid using function calls in OpenMP directives

[alexxy/gromacs.git] / src / gromacs / mdlib / nbnxn_search.c

diff --git a/src/gromacs/mdlib/nbnxn_search.c b/src/gromacs/mdlib/nbnxn_search.c
index c06a5d10f1701d3d9981617c23622f525a44a1d2..3114bebc22b048559369dede69c9774cf4d8f21a 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_search.c
+++ b/src/gromacs/mdlib/nbnxn_search.c
@@ -419,6 +419,12 @@ static real grid_atom_density(int n, rvec corner0, rvec corner1)
 {
     rvec size;
 
+    if (n == 0)
+    {
+        /* To avoid zero density we use a minimum of 1 atom */
+        n = 1;
+    }
+
     rvec_sub(corner1, corner0, size);
 
     return n/(size[XX]*size[YY]*size[ZZ]);
@@ -439,6 +445,8 @@ static int set_grid_size_xy(const nbnxn_search_t nbs,
 
     if (n > grid->na_sc)
     {
+        assert(atom_density > 0);
+
         /* target cell length */
         if (grid->bSimple)
         {
@@ -1697,8 +1705,8 @@ static void calc_cell_indices(const nbnxn_search_t nbs,
     }
 
     /* Sort the super-cell columns along z into the sub-cells. */
-#pragma omp parallel for num_threads(nbs->nthread_max) schedule(static)
-    for (thread = 0; thread < nbs->nthread_max; thread++)
+#pragma omp parallel for num_threads(nthread) schedule(static)
+    for (thread = 0; thread < nthread; thread++)
     {
         if (grid->bSimple)
         {
@@ -1814,7 +1822,8 @@ void nbnxn_put_on_grid(nbnxn_search_t nbs,
         nbs->ePBC = ePBC;
         copy_mat(box, nbs->box);
 
-        if (atom_density >= 0)
+        /* Avoid zero density */
+        if (atom_density > 0)
         {
             grid->atom_density = atom_density;
         }
@@ -1830,12 +1839,21 @@ void nbnxn_put_on_grid(nbnxn_search_t nbs,
          * for the local atoms (dd_zone=0).
          */
         nbs->natoms_nonlocal = a1 - nmoved;
+
+        if (debug)
+        {
+            fprintf(debug, "natoms_local = %5d atom_density = %5.1f\n",
+                    nbs->natoms_local, grid->atom_density);
+        }
     }
     else
     {
         nbs->natoms_nonlocal = max(nbs->natoms_nonlocal, a1);
     }
 
+    /* We always use the home zone (grid[0]) for setting the cell size,
+     * since determining densities for non-local zones is difficult.
+     */
     nc_max_grid = set_grid_size_xy(nbs, grid,
                                    dd_zone, n-nmoved, corner0, corner1,
                                    nbs->grid[0].atom_density);
@@ -1913,6 +1931,7 @@ void nbnxn_grid_add_simple(nbnxn_search_t    nbs,
     float        *bbcz;
     nbnxn_bb_t   *bb;
     int           ncd, sc;
+    int           nthreads gmx_unused;
 
     grid = &nbs->grid[0];
 
@@ -1939,7 +1958,8 @@ void nbnxn_grid_add_simple(nbnxn_search_t    nbs,
     bbcz = grid->bbcz_simple;
     bb   = grid->bb_simple;
 
-#pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
+    nthreads = gmx_omp_nthreads_get(emntPairsearch);
+#pragma omp parallel for num_threads(nthreads) schedule(static)
     for (sc = 0; sc < grid->nc; sc++)
     {
         int c, tx, na;
@@ -3440,7 +3460,6 @@ static void make_fep_list(const nbnxn_search_t    nbs,
                              * Note that the charge has been set to zero,
                              * but we need to avoid 0/0, as perturbed atoms
                              * can be on top of each other.
-                             * (and the LJ parameters have not been zeroed)
                              */
                             nbl->cj[cj_ind].excl &= ~(1U << (i*nbl->na_cj + j));
                         }
@@ -4456,6 +4475,7 @@ static void combine_nblists(int nnbl, nbnxn_pairlist_t **nbl,
 {
     int nsci, ncj4, nexcl;
     int n, i;
+    int nthreads gmx_unused;
 
     if (nblc->bSimple)
     {
@@ -4496,7 +4516,8 @@ static void combine_nblists(int nnbl, nbnxn_pairlist_t **nbl,
     /* Each thread should copy its own data to the combined arrays,
      * as otherwise data will go back and forth between different caches.
      */
-#pragma omp parallel for num_threads(gmx_omp_nthreads_get(emntPairsearch)) schedule(static)
+    nthreads = gmx_omp_nthreads_get(emntPairsearch);
+#pragma omp parallel for num_threads(nthreads) schedule(static)
     for (n = 0; n < nnbl; n++)
     {
         int                     sci_offset;