Merge remote-tracking branch 'origin/master' into mpi_skip

Conflicts:
        src/engine.c

README

@@ -26,6 +26,8 @@ Valid options are:
   -f    {int} Overwrite the CPU frequency (Hz) to be used for time measurements
   -g          Run with an external gravitational potential
   -G          Run with self-gravity
+  -n    {int} Execute a fixed number of time steps. Defaults to -1, which means
+              use the time_end parameter to stop. 
   -s          Run with SPH
   -t    {int} The number of threads to use on each MPI rank. Defaults to 1 if not specified.
   -v     [12] Increase the level of verbosity 1: MPI-rank 0 writes 

configure.ac

@@ -72,9 +72,6 @@ if test "$enable_ipo" = "yes"; then
    fi
 fi
 
-# Add libtool support.
-LT_INIT
-
 # Check for MPI. Need to do this before characterising the compiler (C99 mode),
 # as this changes the compiler.
 # We should consider using AX_PROG_CC_MPI to replace AC_PROG_CC when compiling
@@ -151,6 +148,9 @@ AM_CONDITIONAL([HAVEMPI],[test $enable_mpi = "yes"])
 # Indicate that MPIRUN can be modified by an environement variable
 AC_ARG_VAR(MPIRUN, Path to the mpirun command if non-standard)
 
+# Add libtool support (now that CC is defined).
+LT_INIT
+
 # Need C99 and inline support.
 AC_PROG_CC_C99
 AC_C_INLINE
@@ -179,7 +179,7 @@ if test "$enable_opt" = "yes" ; then
    ac_test_CFLAGS="yes"
    CFLAGS="$old_CFLAGS $CFLAGS"
 
-   # Check SSE & AVX support (some overlap with AX_CC_MAXOPT). 
+   # Check SSE & AVX support (some overlap with AX_CC_MAXOPT).
    # Don't use the SIMD_FLAGS result with Intel compilers. The -x<code>
    # value from AX_CC_MAXOPT should be sufficient.
    AX_EXT
@@ -287,12 +287,75 @@ AC_SUBST([METIS_LIBS])
 AC_SUBST([METIS_INCS])
 AM_CONDITIONAL([HAVEMETIS],[test -n "$METIS_LIBS"])
 
-# # Check for zlib.
-# AC_CHECK_LIB([z],[gzopen],[
-#     AC_DEFINE([HAVE_LIBZ],[1],[Set to 1 if zlib is installed.])
-#     LDFLAGS="$LDFLAGS -lz"
-#     ],[])
+#  Check for tcmalloc a fast malloc that is part of the gperftools.
+have_tcmalloc="no"
+AC_ARG_WITH([tcmalloc],
+   [AS_HELP_STRING([--with-tcmalloc],
+      [use tcmalloc library or specify the directory with lib @<:@yes/no@:>@]
+   )],
+   [with_tcmalloc="$withval"],
+   [with_tcmalloc="no"]
+)
+if test "x$with_tcmalloc" != "xno"; then
+   if test "x$with_tcmalloc" != "xyes" && test "x$with_tcmalloc" != "x"; then
+      tclibs="-L$with_tcmalloc -ltcmalloc"
+   else
+      tclibs="-ltcmalloc"
+   fi
+   AC_CHECK_LIB([tcmalloc],[tc_cfree],[have_tcmalloc="yes"],[have_tcmalloc="no"],
+                $tclibs)
+
+   #  Could just have the minimal version.
+   if test "$have_tcmalloc" = "no"; then
+      if test "x$with_tcmalloc" != "xyes" && test "x$with_tcmalloc" != "x"; then
+         tclibs="-L$with_tcmalloc -ltcmalloc_minimal"
+      else
+         tclibs="-ltcmalloc_minimal"
+      fi
+      AC_CHECK_LIB([tcmalloc],[tc_cfree],[have_tcmalloc="yes"],[have_tcmalloc="no"],
+                   $tclibs)
+   fi
 
+   if test "$have_tcmalloc" = "yes"; then
+      TCMALLOC_LIBS="$tclibs"
+
+      # These are recommended for GCC.
+      if test "$ax_cv_c_compiler_vendor" = "gnu"; then
+         CFLAGS="$CFLAGS -fno-builtin-malloc -fno-builtin-calloc -fno-builtin-realloc -fno-builtin-free"
+      fi
+   else
+      TCMALLOC_LIBS=""
+   fi
+fi
+AC_SUBST([TCMALLOC_LIBS])
+AM_CONDITIONAL([HAVETCMALLOC],[test -n "$TCMALLOC_LIBS"])
+
+#  Check for -lprofiler usually part of the gpreftools along with tcmalloc.
+have_profiler="no"
+AC_ARG_WITH([profiler],
+   [AS_HELP_STRING([--with-profiler],
+      [use cpu profiler library or specify the directory with lib @<:@yes/no@:>@]
+   )],
+   [with_profiler="$withval"],
+   [with_profiler="yes"]
+)
+if test "x$with_profiler" != "xno"; then
+   if test "x$with_profiler" != "xyes" && test "x$with_profiler" != "x"; then
+      proflibs="-L$with_profiler -lprofiler"
+   else
+      proflibs="-lprofiler"
+   fi
+   AC_CHECK_LIB([profiler],[ProfilerFlush],[have_profiler="yes"],[have_profiler="no"],
+                $proflibs)
+
+   if test "$have_profiler" = "yes"; then
+      PROFILER_LIBS="$proflibs"
+   else
+      PROFILER_LIBS=""
+   fi
+fi
+AC_SUBST([PROFILER_LIBS])
+AM_CONDITIONAL([HAVEPROFILER],[test -n "$PROFILER_LIBS"])
 
 # Check for HDF5. This is required.
 AX_LIB_HDF5
@@ -410,6 +473,8 @@ AC_MSG_RESULT([
     - parallel     : $have_parallel_hdf5
    Metis enabled   : $have_metis
    libNUMA enabled : $have_numa
+   Using tcmalloc  : $have_tcmalloc
+   CPU profiler    : $have_profiler
 ])
 
 # Generate output.

examples/Makefile.am

@@ -21,13 +21,17 @@ MYFLAGS = -DTIMER
 # Add the source directory and debug to CFLAGS
 AM_CFLAGS = -I../src $(HDF5_CPPFLAGS)
 
-AM_LDFLAGS =
+AM_LDFLAGS = $(HDF5_LDFLAGS)
 
-MPI_THREAD_LIBS = @MPI_THREAD_LIBS@
+# Extra libraries.
+EXTRA_LIBS = $(HDF5_LIBS) $(PROFILER_LIBS) $(TCMALLOC_LIBS)
+
+# MPI libraries.
 MPI_LIBS = $(METIS_LIBS) $(MPI_THREAD_LIBS)
 MPI_FLAGS = -DWITH_MPI $(METIS_INCS)
 
-# Set-up the library
+
+# Programs.
 bin_PROGRAMS = swift swift_fixdt
 
 # Build MPI versions as well?
@@ -45,20 +49,20 @@ endif
 # Sources for swift
 swift_SOURCES = main.c
 swift_CFLAGS = $(MYFLAGS) $(AM_CFLAGS) -DENGINE_POLICY="engine_policy_keep $(ENGINE_POLICY_SETAFFINITY)"
-swift_LDADD =  ../src/.libs/libswiftsim.a $(HDF5_LDFLAGS) $(HDF5_LIBS)
+swift_LDADD =  ../src/.libs/libswiftsim.a $(EXTRA_LIBS)
 
 swift_fixdt_SOURCES = main.c
 swift_fixdt_CFLAGS = $(MYFLAGS) $(AM_CFLAGS) -DENGINE_POLICY="engine_policy_fixdt | engine_policy_keep $(ENGINE_POLICY_SETAFFINITY)"
-swift_fixdt_LDADD =  ../src/.libs/libswiftsim.a $(HDF5_LDFLAGS) $(HDF5_LIBS)
+swift_fixdt_LDADD =  ../src/.libs/libswiftsim.a $(EXTRA_LIBS)
 
 # Sources for swift_mpi, do we need an affinity policy for MPI?
 swift_mpi_SOURCES = main.c
 swift_mpi_CFLAGS = $(MYFLAGS) $(AM_CFLAGS) $(MPI_FLAGS) -DENGINE_POLICY="engine_policy_keep $(ENGINE_POLICY_SETAFFINITY)"
-swift_mpi_LDADD =  ../src/.libs/libswiftsim_mpi.a $(HDF5_LDFLAGS) $(HDF5_LIBS) $(MPI_LIBS)
+swift_mpi_LDADD =  ../src/.libs/libswiftsim_mpi.a $(MPI_LIBS) $(EXTRA_LIBS)
 
 swift_fixdt_mpi_SOURCES = main.c
 swift_fixdt_mpi_CFLAGS = $(MYFLAGS) $(AM_CFLAGS) $(MPI_FLAGS) -DENGINE_POLICY="engine_policy_fixdt | engine_policy_keep $(ENGINE_POLICY_SETAFFINITY)"
-swift_fixdt_mpi_LDADD =  ../src/.libs/libswiftsim_mpi.a $(HDF5_LDFLAGS) $(HDF5_LIBS) $(MPI_LIBS)
+swift_fixdt_mpi_LDADD =  ../src/.libs/libswiftsim_mpi.a $(MPI_LIBS) $(EXTRA_LIBS)
 
 # Scripts to generate ICs
 EXTRA_DIST = UniformBox/makeIC.py UniformBox/run.sh UniformBox/uniformBox.yml \

examples/main.c

@@ -74,6 +74,7 @@ void print_help_message() {
   printf("  %2s %8s %s\n", "-g", "",
          "Run with an external gravitational potential");
   printf("  %2s %8s %s\n", "-G", "", "Run with self-gravity");
+  printf("  %2s %8s %s\n", "-n", "{int}", "Execute a fixed number of time steps");
   printf("  %2s %8s %s\n", "-s", "", "Run with SPH");
   printf("  %2s %8s %s\n", "-t", "{int}",
          "The number of threads to use on each MPI rank. Defaults to 1 if not "
@@ -138,6 +139,7 @@ int main(int argc, char *argv[]) {
   int with_aff = 0;
   int dry_run = 0;
   int dump_tasks = 0;
+  int nsteps = -1;
   int with_cosmology = 0;
   int with_external_gravity = 0;
   int with_self_gravity = 0;
@@ -150,7 +152,7 @@ int main(int argc, char *argv[]) {
 
   /* Parse the parameters */
   int c;
-  while ((c = getopt(argc, argv, "acdef:gGhst:v:y:")) != -1) switch (c) {
+  while ((c = getopt(argc, argv, "acdef:gGhn:st:v:y:")) != -1) switch (c) {
       case 'a':
         with_aff = 1;
         break;
@@ -179,6 +181,13 @@ int main(int argc, char *argv[]) {
       case 'h':
         if (myrank == 0) print_help_message();
         return 0;
+      case 'n':
+        if (sscanf(optarg, "%d", &nsteps) != 1) {
+          if (myrank == 0) printf("Error parsing fixed number of steps.\n");
+          if (myrank == 0) print_help_message();
+          return 1;
+        }
+        break;
       case 's':
         with_hydro = 1;
         break;
@@ -322,18 +331,21 @@ int main(int argc, char *argv[]) {
   size_t Ngas = 0, Ngpart = 0;
   double dim[3] = {0., 0., 0.};
   int periodic = 0;
+  int flag_entropy_ICs = 0;
   if (myrank == 0) clocks_gettime(&tic);
 #if defined(WITH_MPI)
 #if defined(HAVE_PARALLEL_HDF5)
   read_ic_parallel(ICfileName, dim, &parts, &gparts, &Ngas, &Ngpart, &periodic,
-                   myrank, nr_nodes, MPI_COMM_WORLD, MPI_INFO_NULL, dry_run);
+                   &flag_entropy_ICs, myrank, nr_nodes, MPI_COMM_WORLD,
+                   MPI_INFO_NULL, dry_run);
 #else
   read_ic_serial(ICfileName, dim, &parts, &gparts, &Ngas, &Ngpart, &periodic,
-                 myrank, nr_nodes, MPI_COMM_WORLD, MPI_INFO_NULL, dry_run);
+                 &flag_entropy_ICs, myrank, nr_nodes, MPI_COMM_WORLD,
+                 MPI_INFO_NULL, dry_run);
 #endif
 #else
   read_ic_single(ICfileName, dim, &parts, &gparts, &Ngas, &Ngpart, &periodic,
-                 dry_run);
+                 &flag_entropy_ICs, dry_run);
 #endif
   if (myrank == 0) {
     clocks_gettime(&toc);
@@ -354,7 +366,7 @@ int main(int argc, char *argv[]) {
     free(parts);
     parts = NULL;
     for (size_t k = 0; k < Ngpart; ++k)
-      if (gparts[k].id > 0) error("Linking problem");
+      if (gparts[k].id_or_neg_offset < 0) error("Linking problem");
     Ngas = 0;
   }
 
@@ -466,7 +478,7 @@ int main(int argc, char *argv[]) {
 #endif
 
   /* Initialise the particles */
-  engine_init_particles(&e);
+  engine_init_particles(&e, flag_entropy_ICs);
 
   /* Legend */
   if (myrank == 0)
@@ -474,7 +486,7 @@ int main(int argc, char *argv[]) {
            "Updates", "g-Updates", "Wall-clock time", clocks_getunit());
 
   /* Main simulation loop */
-  for (int j = 0; !engine_is_done(&e); j++) {
+  for (int j = 0; !engine_is_done(&e) && e.step != nsteps; j++) {
 
 /* Repartition the space amongst the nodes? */
 #ifdef WITH_MPI

src/Makefile.am

@@ -24,6 +24,13 @@ AM_LDFLAGS = $(HDF5_LDFLAGS) -version-info 0:0:0
 # The git command, if available.
 GIT_CMD = @GIT_CMD@
 
+# Additional dependencies for shared libraries.
+EXTRA_LIBS = $(HDF5_LIBS) $(PROFILER_LIBS) $(TCMALLOC_LIBS)
+
+# MPI libraries.
+MPI_LIBS = $(METIS_LIBS) $(MPI_THREAD_LIBS)
+MPI_FLAGS = -DWITH_MPI $(METIS_INCS)
+
 # Build the libswiftsim library
 lib_LTLIBRARIES = libswiftsim.la
 # Build a MPI-enabled version too?
@@ -65,11 +72,13 @@ nobase_noinst_HEADERS = approx_math.h atomic.h cycle.h error.h inline.h kernel_h
 
 # Sources and flags for regular library
 libswiftsim_la_SOURCES = $(AM_SOURCES)
+libswiftsim_la_CFLAGS = $(AM_CFLAGS)
+libswiftsim_la_LDFLAGS = $(AM_LDFLAGS) $(EXTRA_LIBS)
 
 # Sources and flags for MPI library
 libswiftsim_mpi_la_SOURCES = $(AM_SOURCES)
-libswiftsim_mpi_la_CFLAGS = $(AM_CFLAGS) -DWITH_MPI $(METIS_INCS)
-libswiftsim_mpi_la_LDFLAGS = $(AM_LDFLAGS) -DWITH_MPI $(METIS_LIBS)
+libswiftsim_mpi_la_CFLAGS = $(AM_CFLAGS) $(MPI_FLAGS)
+libswiftsim_mpi_la_LDFLAGS = $(AM_LDFLAGS) $(MPI_LIBS) $(EXTRA_LIBS)
 libswiftsim_mpi_la_SHORTNAME = mpi
 
 

src/cell.c

@@ -222,7 +222,7 @@ int cell_pack_ti_ends(struct cell *c, int *ti_ends) {
 
   /* Pack this cell's data. */
   ti_ends[0] = c->ti_end_min;
-  
+
   /* Fill in the progeny, depth-first recursion. */
   int count = 1;
   for (int k = 0; k < 8; k++)
@@ -238,7 +238,7 @@ int cell_unpack_ti_ends(struct cell *c, int *ti_ends) {
 
   /* Unpack this cell's data. */
   c->ti_end_min = ti_ends[0];
-  
+
   /* Fill in the progeny, depth-first recursion. */
   int count = 1;
   for (int k = 0; k < 8; k++)
@@ -410,9 +410,11 @@ void cell_gunlocktree(struct cell *c) {
  * @brief Sort the parts into eight bins along the given pivots.
  *
  * @param c The #cell array to be sorted.
+ * @param parts_offset Offset of the cell parts array relative to the
+ *        space's parts array, i.e. c->parts - s->parts.
  */
 
-void cell_split(struct cell *c) {
+void cell_split(struct cell *c, ptrdiff_t parts_offset) {
 
   int i, j;
   const int count = c->count, gcount = c->gcount;
@@ -530,8 +532,7 @@ void cell_split(struct cell *c) {
   }
 
   /* Re-link the gparts. */
-  for (int k = 0; k < count; k++)
-    if (parts[k].gpart != NULL) parts[k].gpart->part = &parts[k];
+  part_relink_gparts(parts, count, parts_offset);
 
 #ifdef SWIFT_DEBUG_CHECKS
   /* Verify that _all_ the parts have been assigned to a cell. */
@@ -626,8 +627,7 @@ void cell_split(struct cell *c) {
   }
 
   /* Re-link the parts. */
-  for (int k = 0; k < gcount; k++)
-    if (gparts[k].id > 0) gparts[k].part->gpart = &gparts[k];
+  part_relink_parts(gparts, gcount, parts - parts_offset);
 }
 
 /**

src/cell.h

@@ -24,6 +24,9 @@
 #define SWIFT_CELL_H
 
 /* Includes. */
+#include <stddef.h>
+
+/* Local includes. */
 #include "lock.h"
 #include "multipole.h"
 #include "part.h"
@@ -178,7 +181,7 @@ struct cell {
   ((int)(k) + (cdim)[2] * ((int)(j) + (cdim)[1] * (int)(i)))
 
 /* Function prototypes. */
-void cell_split(struct cell *c);
+void cell_split(struct cell *c, ptrdiff_t parts_offset);
 int cell_locktree(struct cell *c);
 void cell_unlocktree(struct cell *c);
 int cell_glocktree(struct cell *c);

src/common_io.c

@@ -516,12 +516,10 @@ void prepare_dm_gparts(struct gpart* const gparts, size_t Ndm) {
 
   /* Let's give all these gparts a negative id */
   for (size_t i = 0; i < Ndm; ++i) {
-
     /* 0 or negative ids are not allowed */
-    if (gparts[i].id <= 0)
-      error("0 or negative ID for DM particle %zd: ID=%lld", i, gparts[i].id);
-
-    gparts[i].id = -gparts[i].id;
+    if (gparts[i].id_or_neg_offset <= 0)
+      error("0 or negative ID for DM particle %zd: ID=%lld", i,
+            gparts[i].id_or_neg_offset);
   }
 }
 
@@ -555,7 +553,7 @@ void duplicate_hydro_gparts(struct part* const parts,
     gparts[i + Ndm].mass = parts[i].mass;
 
     /* Link the particles */
-    gparts[i + Ndm].part = &parts[i];
+    gparts[i + Ndm].id_or_neg_offset = -i;
     parts[i].gpart = &gparts[i + Ndm];
   }
 }
@@ -580,9 +578,8 @@ void collect_dm_gparts(const struct gpart* const gparts, size_t Ntot,
      * gparts[i].part); */
 
     /* And collect the DM ones */
-    if (gparts[i].id < 0LL) {
-      memcpy(&dmparts[count], &gparts[i], sizeof(struct gpart));
-      dmparts[count].id = -dmparts[count].id;
+    if (gparts[i].id_or_neg_offset > 0) {
+      dmparts[count] = gparts[i];
       count++;
     }
   }

src/debug.c

@@ -59,8 +59,8 @@
  *
  * (Should be used for debugging only as it runs in O(N).)
  */
-void printParticle(struct part *parts, struct xpart *xparts, long long int id,
-                   size_t N) {
+void printParticle(const struct part *parts, struct xpart *xparts,
+                   long long int id, size_t N) {
 
   int found = 0;
 
@@ -82,24 +82,27 @@ void printParticle(struct part *parts, struct xpart *xparts, long long int id,
  * the standard output.
  *
  * @param gparts The array of g-particles.
+ * @param parts The array of particles.
  * @param id The id too look for.
  * @param N The size of the array of g-particles.
  *
  * (Should be used for debugging only as it runs in O(N).)
  */
-void printgParticle(struct gpart *gparts, long long int id, size_t N) {
+void printgParticle(const struct gpart *gparts, const struct part *parts,
+                    long long int id, size_t N) {
 
   int found = 0;
 
   /* Look for the particle. */
   for (size_t i = 0; i < N; i++)
-    if (gparts[i].id == -id) {
-      printf("## gParticle[%zd] (DM) :\n id=%lld ", i, -gparts[i].id);
+    if (gparts[i].id_or_neg_offset == id) {
+      printf("## gParticle[%zd] (DM) :\n id=%lld", i, id);
       gravity_debug_particle(&gparts[i]);
       found = 1;
       break;
-    } else if (gparts[i].id > 0 && gparts[i].part->id == id) {
-      printf("## gParticle[%zd] (hydro) :\n id=%lld ", i, gparts[i].id);
+    } else if (gparts[i].id_or_neg_offset < 0 &&
+               parts[-gparts[i].id_or_neg_offset].id == id) {
+      printf("## gParticle[%zd] (hydro) :\n id=%lld", i, id);
       gravity_debug_particle(&gparts[i]);
       found = 1;
       break;
@@ -114,9 +117,9 @@ void printgParticle(struct gpart *gparts, long long int id, size_t N) {
  * @param p The particle to print
  * @param xp The extended data ot the particle to print
  */
-void printParticle_single(struct part *p, struct xpart *xp) {
+void printParticle_single(const struct part *p, const struct xpart *xp) {
 
-  printf("## Particle: id=%lld ", p->id);
+  printf("## Particle: id=%lld", p->id);
   hydro_debug_particle(p, xp);
   printf("\n");
 }
@@ -128,7 +131,7 @@ void printParticle_single(struct part *p, struct xpart *xp) {
  */
 void printgParticle_single(struct gpart *gp) {
 
-  printf("## g-Particle: id=%lld ", gp->id);
+  printf("## g-Particle: id=%lld ", gp->id_or_neg_offset);
   gravity_debug_particle(gp);
   printf("\n");
 }

src/debug.h

@@ -19,14 +19,15 @@
 #ifndef SWIFT_DEBUG_H
 #define SWIFT_DEBUG_H
 
-struct part;
-struct gpart;
-struct xpart;
+/* Includes. */
+#include "cell.h"
+#include "part.h"
 
-void printParticle(struct part *parts, struct xpart *xparts, long long int id,
-                   size_t N);
-void printgParticle(struct gpart *parts, long long int id, size_t N);
-void printParticle_single(struct part *p, struct xpart *xp);
+void printParticle(const struct part *parts, struct xpart *xparts,
+                   long long int id, size_t N);
+void printgParticle(const struct gpart *gparts, const struct part *parts,
+                    long long int id, size_t N);
+void printParticle_single(const struct part *p, const struct xpart *xp);
 void printgParticle_single(struct gpart *gp);
 
 #ifdef HAVE_METIS

src/engine.c

@@ -102,8 +102,8 @@ struct link *engine_addlink(struct engine *e, struct link *l, struct task *t) {
 }
 
 /**
- * @brief Generate the hierarchical tasks for a hierarchy of cells - i.e. all
- * the O(Npart) tasks.
+ * @brief Generate the gravity hierarchical tasks for a hierarchy of cells -
+ * i.e. all the O(Npart) tasks.
  *
  * Tasks are only created here. The dependencies will be added later on.
  *
@@ -111,8 +111,8 @@ struct link *engine_addlink(struct engine *e, struct link *l, struct task *t) {
  * @param c The #cell.
  * @param super The super #cell.
  */
-void engine_make_hierarchical_tasks(struct engine *e, struct cell *c,
-                                    struct cell *super) {
+void engine_make_gravity_hierarchical_tasks(struct engine *e, struct cell *c,
+                                            struct cell *super) {
 
   struct scheduler *s = &e->sched;
   const int is_with_external_gravity =
@@ -132,36 +132,91 @@ void engine_make_hierarchical_tasks(struct engine *e, struct cell *c,
     if (c->nodeID == e->nodeID) {
 
       /* Add the init task. */
-      c->init = scheduler_addtask(s, task_type_init, task_subtype_none, 0, 0, c,
-                                  NULL, 0);
+      if (c->init == NULL)
+        c->init = scheduler_addtask(s, task_type_init, task_subtype_none, 0, 0,
+                                    c, NULL, 0);
 
       /* Add the drift task. */
-      c->drift = scheduler_addtask(s, task_type_drift, task_subtype_none, 0, 0,
-                                   c, NULL, 0);
+      if (c->drift == NULL)
+        c->drift = scheduler_addtask(s, task_type_drift, task_subtype_none, 0,
+                                     0, c, NULL, 0);
 
       /* Add the kick task that matches the policy. */
       if (is_fixdt) {
-        c->kick = scheduler_addtask(s, task_type_kick_fixdt, task_subtype_none,
-                                    0, 0, c, NULL, 0);
+        if (c->kick == NULL)
+          c->kick = scheduler_addtask(s, task_type_kick_fixdt,
+                                      task_subtype_none, 0, 0, c, NULL, 0);
       } else {
-        c->kick = scheduler_addtask(s, task_type_kick, task_subtype_none, 0, 0,
-                                    c, NULL, 0);
+        if (c->kick == NULL)
+          c->kick = scheduler_addtask(s, task_type_kick, task_subtype_none, 0,
+                                      0, c, NULL, 0);
       }
 
-      if (c->count > 0) {
+      if (is_with_external_gravity)
+        c->grav_external = scheduler_addtask(
+            s, task_type_grav_external, task_subtype_none, 0, 0, c, NULL, 0);
+    }
+  }
 
-        /* Generate the ghost task. */
-        c->ghost = scheduler_addtask(s, task_type_ghost, task_subtype_none, 0,
-                                     0, c, NULL, 0);
-      }
+  /* Set the super-cell. */
+  c->super = super;
+
+  /* Recurse. */
+  if (c->split)
+    for (int k = 0; k < 8; k++)
+      if (c->progeny[k] != NULL)
+        engine_make_gravity_hierarchical_tasks(e, c->progeny[k], super);
+}
 
-      if (c->gcount > 0) {
+/**
+ * @brief Generate the hydro hierarchical tasks for a hierarchy of cells -
+ * i.e. all the O(Npart) tasks.
+ *
+ * Tasks are only created here. The dependencies will be added later on.
+ *
+ * @param e The #engine.
+ * @param c The #cell.
+ * @param super The super #cell.
+ */
+void engine_make_hydro_hierarchical_tasks(struct engine *e, struct cell *c,
+                                          struct cell *super) {
+
+  struct scheduler *s = &e->sched;
+  const int is_fixdt = (e->policy & engine_policy_fixdt) == engine_policy_fixdt;
 
-        /* Add the external gravity tasks */
-        if (is_with_external_gravity)
-          c->grav_external = scheduler_addtask(
-              s, task_type_grav_external, task_subtype_none, 0, 0, c, NULL, 0);
+  /* Is this the super-cell? */
+  if (super == NULL && (c->density != NULL || (c->count > 0 && !c->split))) {
+
+    /* This is the super cell, i.e. the first with density tasks attached. */
+    super = c;
+
+    /* Local tasks only... */
+    if (c->nodeID == e->nodeID) {
+
+      /* Add the init task. */
+      if (c->init == NULL)
+        c->init = scheduler_addtask(s, task_type_init, task_subtype_none, 0, 0,
+                                    c, NULL, 0);
+
+      /* Add the drift task. */
+      if (c->drift == NULL)
+        c->drift = scheduler_addtask(s, task_type_drift, task_subtype_none, 0,
+                                     0, c, NULL, 0);
+
+      /* Add the kick task that matches the policy. */
+      if (is_fixdt) {
+        if (c->kick == NULL)
+          c->kick = scheduler_addtask(s, task_type_kick_fixdt,
+                                      task_subtype_none, 0, 0, c, NULL, 0);
+      } else {
+        if (c->kick == NULL)
+          c->kick = scheduler_addtask(s, task_type_kick, task_subtype_none, 0,
+                                      0, c, NULL, 0);
       }
+
+      /* Generate the ghost task. */
+      c->ghost = scheduler_addtask(s, task_type_ghost, task_subtype_none, 0, 0,
+                                   c, NULL, 0);
     }
   }
 
@@ -172,7 +227,7 @@ void engine_make_hierarchical_tasks(struct engine *e, struct cell *c,
   if (c->split)
     for (int k = 0; k < 8; k++)
       if (c->progeny[k] != NULL)
-        engine_make_hierarchical_tasks(e, c->progeny[k], super);
+        engine_make_hydro_hierarchical_tasks(e, c->progeny[k], super);
 }
 
 /**
@@ -270,11 +325,11 @@ void engine_redistribute(struct engine *e) {
         }
 
 #ifdef SWIFT_DEBUG_CHECKS
-        if (s->parts[k].gpart->id < 0)