Merge release-5-0 into master

[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 e4128ca049b12e804c88883b5217261f86adbab5..56e616967c3f91d482408d2e4806c5f1f4fb4c37 100644 (file)
--- a/src/gromacs/mdlib/nbnxn_search.c
+++ b/src/gromacs/mdlib/nbnxn_search.c
@@ -33,34 +33,36 @@
  * the research papers on the package. Check out http://www.gromacs.org.
  */
 
-#ifdef HAVE_CONFIG_H
-#include <config.h>
-#endif
+#include "gmxpre.h"
+
+#include "nbnxn_search.h"
+
+#include "config.h"
 
+#include <assert.h>
 #include <math.h>
 #include <string.h>
-#include <assert.h>
 
-#include "types/commrec.h"
-#include "macros.h"
+#include "gromacs/legacyheaders/gmx_omp_nthreads.h"
+#include "gromacs/legacyheaders/macros.h"
+#include "gromacs/legacyheaders/nrnb.h"
+#include "gromacs/legacyheaders/ns.h"
+#include "gromacs/legacyheaders/types/commrec.h"
 #include "gromacs/math/utilities.h"
 #include "gromacs/math/vec.h"
-#include "nbnxn_consts.h"
-/* nbnxn_internal.h included gromacs/simd/macros.h */
-#include "nbnxn_internal.h"
-#ifdef GMX_NBNXN_SIMD
-#include "gromacs/simd/vector_operations.h"
-#endif
-#include "nbnxn_atomdata.h"
-#include "nbnxn_search.h"
-#include "gmx_omp_nthreads.h"
-#include "nrnb.h"
-#include "ns.h"
-
+#include "gromacs/mdlib/nb_verlet.h"
+#include "gromacs/mdlib/nbnxn_atomdata.h"
+#include "gromacs/mdlib/nbnxn_consts.h"
 #include "gromacs/pbcutil/ishift.h"
 #include "gromacs/pbcutil/pbc.h"
 #include "gromacs/utility/smalloc.h"
 
+/* nbnxn_internal.h included gromacs/simd/macros.h */
+#include "gromacs/mdlib/nbnxn_internal.h"
+#ifdef GMX_SIMD
+#include "gromacs/simd/vector_operations.h"
+#endif
+
 #ifdef NBNXN_SEARCH_BB_SIMD4
 /* Always use 4-wide SIMD for bounding box calculations */
 
@@ -418,6 +420,12 @@ 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]);
@@ -438,6 +446,8 @@ 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)
         {
@@ -1695,8 +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)
         {
@@ -1812,7 +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;
         }
@@ -1828,12 +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);
@@ -1911,6 +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];
 
@@ -1937,7 +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;
@@ -2922,10 +2944,10 @@ static void make_cluster_list_simple(const nbnxn_grid_t *gridj,
 }
 
 #ifdef GMX_NBNXN_SIMD_4XN
-#include "nbnxn_search_simd_4xn.h"
+#include "gromacs/mdlib/nbnxn_search_simd_4xn.h"
 #endif
 #ifdef GMX_NBNXN_SIMD_2XNN
-#include "nbnxn_search_simd_2xnn.h"
+#include "gromacs/mdlib/nbnxn_search_simd_2xnn.h"
 #endif
 
 /* Plain C or SIMD4 code for making a pair list of super-cell sci vs scj.
@@ -4453,6 +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)
     {
@@ -4493,7 +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;