Merge branch 'blueGene'

Conflicts:
	examples/test.c

examples/test.c

@@ -554,6 +554,7 @@ int main(int argc, char *argv[]) {
   int c, icount, j, k, N, periodic = 1;
   long long N_total = -1;
   int nr_threads = 1, nr_queues = -1, runs = INT_MAX;
+  int dump_tasks = 0;
   int data[2];
   double dim[3] = {1.0, 1.0, 1.0}, shift[3] = {0.0, 0.0, 0.0};
   double h_max = -1.0, scaling = 1.0;
@@ -605,7 +606,7 @@ int main(int argc, char *argv[]) {
   bzero(&s, sizeof(struct space));
 
   /* Parse the options */
-  while ((c = getopt(argc, argv, "a:c:d:f:g:m:oq:r:s:t:w:z:")) != -1)
+  while ((c = getopt(argc, argv, "a:c:d:f:g:m:q:r:s:t:w:yz:")) != -1)
     switch (c) {
       case 'a':
         if (sscanf(optarg, "%lf", &scaling) != 1)
@@ -663,6 +664,9 @@ int main(int argc, char *argv[]) {
           error("Error parsing sub size.");
         if (myrank == 0) message("sub size set to %i.", space_subsize);
         break;
+      case 'y':
+        dump_tasks = 1;
+        break;
       case 'z':
         if (sscanf(optarg, "%d", &space_splitsize) != 1)
           error("Error parsing split size.");
@@ -756,7 +760,7 @@ int main(int argc, char *argv[]) {
 
   /* Initialize the space with this data. */
   tic = getticks();
-  space_init(&s, dim, parts, N, periodic, h_max);
+  space_init(&s, dim, parts, N, periodic, h_max, myrank == 0);
   if (myrank == 0)
     message("space_init took %.3f ms.",
             ((double)(getticks() - tic)) / CPU_TPS * 1000);
@@ -796,7 +800,7 @@ int main(int argc, char *argv[]) {
 
   /* Initialize the engine with this space. */
   tic = getticks();
-  message("nr_nodes is %i.", nr_nodes);
+  if (myrank == 0) message("nr_nodes is %i.", nr_nodes);
   engine_init(&e, &s, dt_max, nr_threads, nr_queues, nr_nodes, myrank,
               ENGINE_POLICY | engine_policy_steal);
   if (myrank == 0)
@@ -825,8 +829,9 @@ int main(int argc, char *argv[]) {
 #else
     write_output_single(&e, &us);
 #endif
-    message("writing particle properties took %.3f ms.",
-            ((double)(getticks() - tic)) / CPU_TPS * 1000);
+    if (myrank == 0)
+      message("writing particle properties took %.3f ms.",
+              ((double)(getticks() - tic)) / CPU_TPS * 1000);
     fflush(stdout);
   }
 
@@ -916,40 +921,43 @@ int main(int argc, char *argv[]) {
            (double)runner_hist_bins[k]);
 #endif
 
-/* Dump the task data. */
+  /* Dump the task data. */
+  if (dump_tasks) {
 #ifdef WITH_MPI
-  file_thread = fopen("thread_info_MPI.dat", "w");
-  for (j = 0; j < nr_nodes; j++) {
-    MPI_Barrier(MPI_COMM_WORLD);
-    if (j == myrank) {
-      fprintf(file_thread, " %03i 0 0 0 0 %lli 0 0 0 0\n", myrank, e.tic_step);
-      for (k = 0; k < e.sched.nr_tasks; k++)
-        if (!e.sched.tasks[k].skip && !e.sched.tasks[k].implicit)
-          fprintf(
-              file_thread, " %03i %i %i %i %i %lli %lli %i %i %i\n", myrank,
-              e.sched.tasks[k].rid, e.sched.tasks[k].type,
-              e.sched.tasks[k].subtype, (e.sched.tasks[k].cj == NULL),
-              e.sched.tasks[k].tic, e.sched.tasks[k].toc,
-              e.sched.tasks[k].ci->count,
-              (e.sched.tasks[k].cj != NULL) ? e.sched.tasks[k].cj->count : 0,
-              e.sched.tasks[k].flags);
-      fflush(stdout);
-      sleep(1);
+    file_thread = fopen("thread_info_MPI.dat", "w");
+    for (j = 0; j < nr_nodes; j++) {
+      MPI_Barrier(MPI_COMM_WORLD);
+      if (j == myrank) {
+        fprintf(file_thread, " %03i 0 0 0 0 %lli 0 0 0 0\n", myrank,
+                e.tic_step);
+        for (k = 0; k < e.sched.nr_tasks; k++)
+          if (!e.sched.tasks[k].skip && !e.sched.tasks[k].implicit)
+            fprintf(
+                file_thread, " %03i %i %i %i %i %lli %lli %i %i %i\n", myrank,
+                e.sched.tasks[k].rid, e.sched.tasks[k].type,
+                e.sched.tasks[k].subtype, (e.sched.tasks[k].cj == NULL),
+                e.sched.tasks[k].tic, e.sched.tasks[k].toc,
+                e.sched.tasks[k].ci->count,
+                (e.sched.tasks[k].cj != NULL) ? e.sched.tasks[k].cj->count : 0,
+                e.sched.tasks[k].flags);
+        fflush(stdout);
+        sleep(1);
+      }
     }
-  }
-  fclose(file_thread);
+    fclose(file_thread);
 #else
-  file_thread = fopen("thread_info.dat", "w");
-  for (k = 0; k < e.sched.nr_tasks; k++)
-    if (!e.sched.tasks[k].skip && !e.sched.tasks[k].implicit)
-      fprintf(file_thread, " %i %i %i %i %lli %lli %i %i\n",
-              e.sched.tasks[k].rid, e.sched.tasks[k].type,
-              e.sched.tasks[k].subtype, (e.sched.tasks[k].cj == NULL),
-              e.sched.tasks[k].tic, e.sched.tasks[k].toc,
-              e.sched.tasks[k].ci->count,
-              (e.sched.tasks[k].cj == NULL) ? 0 : e.sched.tasks[k].cj->count);
-  fclose(file_thread);
+    file_thread = fopen("thread_info.dat", "w");
+    for (k = 0; k < e.sched.nr_tasks; k++)
+      if (!e.sched.tasks[k].skip && !e.sched.tasks[k].implicit)
+        fprintf(file_thread, " %i %i %i %i %lli %lli %i %i\n",
+                e.sched.tasks[k].rid, e.sched.tasks[k].type,
+                e.sched.tasks[k].subtype, (e.sched.tasks[k].cj == NULL),
+                e.sched.tasks[k].tic, e.sched.tasks[k].toc,
+                e.sched.tasks[k].ci->count,
+                (e.sched.tasks[k].cj == NULL) ? 0 : e.sched.tasks[k].cj->count);
+    fclose(file_thread);
 #endif
+  }
 
   if (with_outputs) {
 /* Write final output. */
@@ -972,7 +980,7 @@ int main(int argc, char *argv[]) {
 #endif
 
   /* Say goodbye. */
-  message("done.");
+  if (myrank == 0) message("done.");
 
   /* All is calm, all is bright. */
   return 0;

src/engine.c

@@ -1370,7 +1370,7 @@ void engine_rebuild(struct engine *e) {
 
   /* Re-build the space. */
   // tic = getticks();
-  space_rebuild(e->s, 0.0);
+  space_rebuild(e->s, 0.0, e->nodeID == 0);
 // message( "space_rebuild took %.3f ms." , (double)(getticks() -
 // tic)/CPU_TPS*1000 );
 
@@ -2041,8 +2041,9 @@ void engine_split(struct engine *e, int *grid) {
   engine_makeproxies(e);
 
   /* Re-allocate the local parts. */
-  message("Re-allocating parts array from %i to %i.", s->size_parts,
-          (int)(s->nr_parts * 1.2));
+  if (e->nodeID == 0)
+    message("Re-allocating parts array from %i to %i.", s->size_parts,
+            (int)(s->nr_parts * 1.2));
   s->size_parts = s->nr_parts * 1.2;
   struct part *parts_new;
   struct xpart *xparts_new;
@@ -2094,13 +2095,16 @@ void engine_init(struct engine *e, struct space *s, float dt, int nr_threads,
     for (i = 1; i < maxint; i *= 2)
       for (j = maxint / i / 2; j < maxint; j += maxint / i)
         if (j < nr_cores && j != 0) cpuid[k++] = j;
-#ifdef WITHMPI
-    printf("engine_init: cpu map is [ ");
+
+    if (nodeID == 0) {
+#ifdef WITH_MPI
+      message("engine_init: cpu map is [ ");
 #else
-    printf("[%03i] engine_init: cpu map is [ ", nodeID);
+      printf("[%03i] engine_init: cpu map is [ ", nodeID);
 #endif
-    for (i = 0; i < nr_cores; i++) printf("%i ", cpuid[i]);
-    printf("].\n");
+      for (i = 0; i < nr_cores; i++) printf("%i ", cpuid[i]);
+      printf("].\n");
+    }
   }
 #endif
 

src/space.c

@@ -149,9 +149,10 @@ void space_rebuild_recycle(struct space *s, struct cell *c) {
  *
  * @param s The #space.
  * @param cell_max Maximum cell edge length.
+ * @param verbose Print messages to stdout or not.
  */
 
-void space_regrid(struct space *s, double cell_max) {
+void space_regrid(struct space *s, double cell_max, int verbose) {
 
   float h_max = s->cell_min / kernel_gamma / space_stretch, dmin;
   int i, j, k, cdim[3], nr_parts = s->nr_parts;
@@ -182,7 +183,7 @@ void space_regrid(struct space *s, double cell_max) {
     h_max = buff;
   }
 #endif
-  message("h_max is %.3e (cell_max=%.3e).", h_max, cell_max);
+  if (verbose) message("h_max is %.3e (cell_max=%.3e).", h_max, cell_max);
 
   /* Get the new putative cell dimensions. */
   for (k = 0; k < 3; k++)
@@ -253,7 +254,9 @@ void space_regrid(struct space *s, double cell_max) {
         }
 
     /* Be verbose about the change. */
-    message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1], cdim[2]);
+    if (verbose)
+      message("set cell dimensions to [ %i %i %i ].", cdim[0], cdim[1],
+              cdim[2]);
     fflush(stdout);
 
   } /* re-build upper-level cells? */
@@ -289,10 +292,11 @@ void space_regrid(struct space *s, double cell_max) {
  *
  * @param s The #space in which to update the cells.
  * @param cell_max Maximal cell size.
+ * @param verbose Print messages to stdout or not
  *
  */
 
-void space_rebuild(struct space *s, double cell_max) {
+void space_rebuild(struct space *s, double cell_max, int verbose) {
 
   int j, k, cdim[3], nr_parts = s->nr_parts, nr_gparts = s->nr_gparts;
   struct cell *restrict c, *restrict cells;
@@ -305,7 +309,7 @@ void space_rebuild(struct space *s, double cell_max) {
   // message( "re)building space..." ); fflush(stdout);
 
   /* Re-grid if necessary, or just re-set the cell data. */
-  space_regrid(s, cell_max);
+  space_regrid(s, cell_max, verbose);
   cells = s->cells;
 
   /* Run through the particles and get their cell index. */
@@ -1073,6 +1077,7 @@ struct cell *space_getcell(struct space *s) {
  * @param N The number of parts in the space.
  * @param periodic flag whether the domain is periodic or not.
  * @param h_max The maximal interaction radius.
+ * @param verbose Print messages to stdout or not
  *
  * Makes a grid of edge length > r_max and fills the particles
  * into the respective cells. Cells containing more than #space_splitsize
@@ -1081,7 +1086,7 @@ struct cell *space_getcell(struct space *s) {
  */
 
 void space_init(struct space *s, double dim[3], struct part *parts, int N,
-                int periodic, double h_max) {
+                int periodic, double h_max, int verbose) {
 
   /* Store eveything in the space. */
   s->dim[0] = dim[0];
@@ -1151,5 +1156,5 @@ void space_init(struct space *s, double dim[3], struct part *parts, int N,
   if (lock_init(&s->lock) != 0) error("Failed to create space spin-lock.");
 
   /* Build the cells and the tasks. */
-  space_regrid(s, h_max);
+  space_regrid(s, h_max, verbose);
 }

src/space.h

@@ -119,7 +119,7 @@ struct cell *space_getcell(struct space *s);
 int space_getsid(struct space *s, struct cell **ci, struct cell **cj,
                  double *shift);
 void space_init(struct space *s, double dim[3], struct part *parts, int N,
-                int periodic, double h_max);
+                int periodic, double h_max, int verbose);
 void space_map_cells_pre(struct space *s, int full,
                          void (*fun)(struct cell *c, void *data), void *data);
 void space_map_parts(struct space *s,
@@ -127,7 +127,7 @@ void space_map_parts(struct space *s,
                      void *data);
 void space_map_cells_post(struct space *s, int full,
                           void (*fun)(struct cell *c, void *data), void *data);
-void space_rebuild(struct space *s, double h_max);
+void space_rebuild(struct space *s, double h_max, int verbose);
 void space_recycle(struct space *s, struct cell *c);
 void space_split(struct space *s, struct cell *c);