Merge release-5-0 into master

[alexxy/gromacs.git] / src / programs / mdrun / pme_loadbal.c

diff --git a/src/programs/mdrun/pme_loadbal.c b/src/programs/mdrun/pme_loadbal.c
index d6fd68f37269d10692d625172706519c053c2cf3..dd2cd1a2e9f049b01ff6de90078d249f925c0cf7 100644 (file)
--- a/src/programs/mdrun/pme_loadbal.c
+++ b/src/programs/mdrun/pme_loadbal.c
@@ -32,23 +32,27 @@
  * To help us fund GROMACS development, we humbly ask that you cite
  * the research papers on the package. Check out http://www.gromacs.org.
  */
-#ifdef HAVE_CONFIG_H
-#include <config.h>
-#endif
+#include "gmxpre.h"
 
-#include "gromacs/utility/smalloc.h"
-#include "types/commrec.h"
-#include "network.h"
-#include "calcgrid.h"
-#include "pme.h"
-#include "vec.h"
-#include "domdec.h"
-#include "nbnxn_cuda_data_mgmt.h"
-#include "force.h"
-#include "macros.h"
-#include "md_logging.h"
 #include "pme_loadbal.h"
 
+#include "config.h"
+
+#include "gromacs/legacyheaders/calcgrid.h"
+#include "gromacs/legacyheaders/domdec.h"
+#include "gromacs/legacyheaders/force.h"
+#include "gromacs/legacyheaders/macros.h"
+#include "gromacs/legacyheaders/md_logging.h"
+#include "gromacs/legacyheaders/network.h"
+#include "gromacs/legacyheaders/pme.h"
+#include "gromacs/legacyheaders/sim_util.h"
+#include "gromacs/legacyheaders/types/commrec.h"
+#include "gromacs/math/vec.h"
+#include "gromacs/mdlib/nbnxn_cuda/nbnxn_cuda_data_mgmt.h"
+#include "gromacs/pbcutil/pbc.h"
+#include "gromacs/utility/cstringutil.h"
+#include "gromacs/utility/smalloc.h"
+
 /* Parameters and setting for one PP-PME setup */
 typedef struct {
     real      rcut_coulomb;    /* Coulomb cut-off                              */
@@ -436,17 +440,17 @@ static void switch_to_stage1(pme_load_balancing_t pme_lb)
     pme_lb->cur = pme_lb->start - 1;
 }
 
-gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
-                          t_commrec           *cr,
-                          FILE                *fp_err,
-                          FILE                *fp_log,
-                          t_inputrec          *ir,
-                          t_state             *state,
-                          double               cycles,
-                          interaction_const_t *ic,
-                          nonbonded_verlet_t  *nbv,
-                          gmx_pme_t           *pmedata,
-                          gmx_int64_t          step)
+gmx_bool pme_load_balance(pme_load_balancing_t        pme_lb,
+                          t_commrec                  *cr,
+                          FILE                       *fp_err,
+                          FILE                       *fp_log,
+                          t_inputrec                 *ir,
+                          t_state                    *state,
+                          double                      cycles,
+                          interaction_const_t        *ic,
+                          struct nonbonded_verlet_t  *nbv,
+                          gmx_pme_t                  *pmedata,
+                          gmx_int64_t                 step)
 {
     gmx_bool     OK;
     pme_setup_t *set;
@@ -698,30 +702,26 @@ gmx_bool pme_load_balance(pme_load_balancing_t pme_lb,
     }
 
     bUsesSimpleTables = uses_simple_tables(ir->cutoff_scheme, nbv, 0);
-    if (pme_lb->cutoff_scheme == ecutsVERLET &&
-        nbv->grp[0].kernel_type == nbnxnk8x8x8_CUDA)
-    {
-        nbnxn_cuda_pme_loadbal_update_param(nbv->cu_nbv, ic);
-
-        /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
-         * also sharing texture references. To keep the code simple, we don't
-         * treat texture references as shared resources, but this means that
-         * the coulomb_tab texture ref will get updated by multiple threads.
-         * Hence, to ensure that the non-bonded kernels don't start before all
-         * texture binding operations are finished, we need to wait for all ranks
-         * to arrive here before continuing.
-         *
-         * Note that we could omit this barrier if GPUs are not shared (or
-         * texture objects are used), but as this is initialization code, there
-         * is not point in complicating things.
-         */
+    nbnxn_cuda_pme_loadbal_update_param(nbv, ic);
+
+    /* With tMPI + GPUs some ranks may be sharing GPU(s) and therefore
+     * also sharing texture references. To keep the code simple, we don't
+     * treat texture references as shared resources, but this means that
+     * the coulomb_tab texture ref will get updated by multiple threads.
+     * Hence, to ensure that the non-bonded kernels don't start before all
+     * texture binding operations are finished, we need to wait for all ranks
+     * to arrive here before continuing.
+     *
+     * Note that we could omit this barrier if GPUs are not shared (or
+     * texture objects are used), but as this is initialization code, there
+     * is not point in complicating things.
+     */
 #ifdef GMX_THREAD_MPI
-        if (PAR(cr))
-        {
-            gmx_barrier(cr);
-        }
-#endif  /* GMX_THREAD_MPI */
+    if (PAR(cr) && use_GPU(nbv))
+    {
+        gmx_barrier(cr);
     }
+#endif  /* GMX_THREAD_MPI */
 
     /* Usually we won't need the simple tables with GPUs.
      * But we do with hybrid acceleration and with free energy.