void
sc_flops_count (sc_flopinfo_t * fi)
{
double seconds;
float rtime, ptime;
long long flpops;
seconds = sc_MPI_Wtime ();
sc_flops_papi (&rtime, &ptime, &flpops, &fi->mflops);
fi->iwtime = seconds - fi->seconds;
fi->cwtime += fi->iwtime;
fi->iptime = ptime - fi->cptime;
fi->cptime = ptime;
fi->iflpops = flpops - fi->cflpops;
fi->cflpops = flpops;
#ifdef SC_PAPI
fi->irtime = rtime - fi->crtime;
fi->crtime = rtime;
#else
fi->irtime = (float) fi->iwtime;
fi->crtime = (float) fi->cwtime;
#endif
fi->seconds = seconds;
}
void
sc_flops_start (sc_flopinfo_t * fi)
{
float rtime, ptime, mflops;
long long flpops;
fi->seconds = sc_MPI_Wtime ();
sc_flops_papi (&rtime, &ptime, &flpops, &mflops); /* ignore results */
fi->cwtime = 0.;
fi->crtime = fi->cptime = 0.;
fi->cflpops = 0;
fi->iwtime = 0.;
fi->irtime = fi->iptime = fi->mflops = 0.;
fi->iflpops = 0;
}
static void
run_load (sc_MPI_Comm mpicomm, p4est_connectivity_t * conn, int level)
{
int mpiret;
double elapsed_create, elapsed_partition, elapsed_balance;
#ifdef LOADCONN_VTK
char filename[BUFSIZ];
#endif
p4est_t *p4est;
P4EST_GLOBAL_PRODUCTIONF ("Run load on level %d\n", level);
/* create and refine the forest */
mpiret = sc_MPI_Barrier (mpicomm);
SC_CHECK_MPI (mpiret);
elapsed_create = -sc_MPI_Wtime ();
p4est = p4est_new_ext (mpicomm, conn, 0, level, 1, 0, NULL, NULL);
level_shift = 4;
refine_level = level + level_shift;
p4est_refine (p4est, 1, refine_fractal, NULL);
elapsed_create += sc_MPI_Wtime ();
#ifdef LOADCONN_VTK
snprintf (filename, BUFSIZ, "loadconn%d_%02d_C", P4EST_DIM, level);
p4est_vtk_write_file (p4est, NULL, filename);
#endif
/* partition the forest */
mpiret = sc_MPI_Barrier (mpicomm);
SC_CHECK_MPI (mpiret);
elapsed_partition = -sc_MPI_Wtime ();
p4est_partition (p4est, 0, NULL);
elapsed_partition += sc_MPI_Wtime ();
/* balance the forest */
mpiret = sc_MPI_Barrier (mpicomm);
SC_CHECK_MPI (mpiret);
elapsed_balance = -sc_MPI_Wtime ();
p4est_balance (p4est, P4EST_CONNECT_FULL, NULL);
elapsed_balance += sc_MPI_Wtime ();
#ifdef LOADCONN_VTK
snprintf (filename, BUFSIZ, "loadconn%d_%02d_B", P4EST_DIM, level);
p4est_vtk_write_file (p4est, NULL, filename);
#endif
/* report timings */
P4EST_GLOBAL_PRODUCTIONF ("Timings %d: %g %g %g\n", level, elapsed_create,
elapsed_partition, elapsed_balance);
p4est_destroy (p4est);
}
static void
test_loadsave (p4est_connectivity_t * connectivity, const char *prefix,
sc_MPI_Comm mpicomm, int mpirank)
{
int mpiret, retval;
unsigned csum, csum2;
double elapsed, wtime;
p4est_connectivity_t *conn2;
p4est_t *p4est, *p4est2;
sc_statinfo_t stats[STATS_COUNT];
char conn_name[BUFSIZ];
char p4est_name[BUFSIZ];
snprintf (conn_name, BUFSIZ, "%s.%s", prefix, P4EST_CONN_SUFFIX);
snprintf (p4est_name, BUFSIZ, "%s.%s", prefix, P4EST_FOREST_SUFFIX);
P4EST_GLOBAL_INFOF ("Using file names %s and %s\n", conn_name, p4est_name);
p4est = p4est_new_ext (mpicomm, connectivity, 0, 0, 0,
sizeof (int), init_fn, NULL);
p4est_refine (p4est, 1, refine_fn, init_fn);
test_deflate (p4est);
/* save, synchronize, load connectivity and compare */
if (mpirank == 0) {
retval = p4est_connectivity_save (conn_name, connectivity);
SC_CHECK_ABORT (retval == 0, "connectivity_save failed");
}
mpiret = sc_MPI_Barrier (mpicomm);
SC_CHECK_MPI (mpiret);
wtime = sc_MPI_Wtime ();
conn2 = p4est_connectivity_load (conn_name, NULL);
elapsed = sc_MPI_Wtime () - wtime;
sc_stats_set1 (stats + STATS_CONN_LOAD, elapsed, "conn load");
SC_CHECK_ABORT (p4est_connectivity_is_equal (connectivity, conn2),
"load/save connectivity mismatch A");
p4est_connectivity_destroy (conn2);
/* save, synchronize, load p4est and compare */
wtime = sc_MPI_Wtime ();
p4est_save (p4est_name, p4est, 1);
elapsed = sc_MPI_Wtime () - wtime;
sc_stats_set1 (stats + STATS_P4EST_SAVE1, elapsed, "p4est save 1");
wtime = sc_MPI_Wtime ();
p4est2 = p4est_load (p4est_name, mpicomm, sizeof (int), 1, NULL, &conn2);
elapsed = sc_MPI_Wtime () - wtime;
sc_stats_set1 (stats + STATS_P4EST_LOAD1a, elapsed, "p4est load 1a");
SC_CHECK_ABORT (p4est_connectivity_is_equal (connectivity, conn2),
"load/save connectivity mismatch Ba");
SC_CHECK_ABORT (p4est_is_equal (p4est, p4est2, 1),
"load/save p4est mismatch Ba");
p4est_destroy (p4est2);
p4est_connectivity_destroy (conn2);
wtime = sc_MPI_Wtime ();
p4est2 = p4est_load (p4est_name, mpicomm, 0, 0, NULL, &conn2);
elapsed = sc_MPI_Wtime () - wtime;
sc_stats_set1 (stats + STATS_P4EST_LOAD1b, elapsed, "p4est load 1b");
SC_CHECK_ABORT (p4est_connectivity_is_equal (connectivity, conn2),
"load/save connectivity mismatch Bb");
SC_CHECK_ABORT (p4est_is_equal (p4est, p4est2, 0),
"load/save p4est mismatch Bb");
test_deflate (p4est2);
p4est_destroy (p4est2);
p4est_connectivity_destroy (conn2);
/* partition and balance */
p4est_partition (p4est, 0, NULL);
p4est_balance (p4est, P4EST_CONNECT_FULL, init_fn);
csum = p4est_checksum (p4est);
sc_stats_set1 (stats + STATS_P4EST_ELEMS,
(double) p4est->local_num_quadrants, "p4est elements");
/* save, synchronize, load p4est and compare */
wtime = sc_MPI_Wtime ();
p4est_save (p4est_name, p4est, 0);
elapsed = sc_MPI_Wtime () - wtime;
sc_stats_set1 (stats + STATS_P4EST_SAVE2, elapsed, "p4est save 2");
wtime = sc_MPI_Wtime ();
p4est2 = p4est_load (p4est_name, mpicomm, sizeof (int), 0, NULL, &conn2);
elapsed = sc_MPI_Wtime () - wtime;
sc_stats_set1 (stats + STATS_P4EST_LOAD2, elapsed, "p4est load 2");
SC_CHECK_ABORT (p4est_connectivity_is_equal (connectivity, conn2),
"load/save connectivity mismatch C");
SC_CHECK_ABORT (p4est_is_equal (p4est, p4est2, 0),
"load/save p4est mismatch C");
p4est_destroy (p4est2);
p4est_connectivity_destroy (conn2);
/* save, synchronize, load p4est and compare */
wtime = sc_MPI_Wtime ();
p4est_save (p4est_name, p4est, 1);
elapsed = sc_MPI_Wtime () - wtime;
sc_stats_set1 (stats + STATS_P4EST_SAVE3, elapsed, "p4est save 3");
wtime = sc_MPI_Wtime ();
p4est2 = p4est_load (p4est_name, mpicomm, sizeof (int), 0, NULL, &conn2);
elapsed = sc_MPI_Wtime () - wtime;
sc_stats_set1 (stats + STATS_P4EST_LOAD3, elapsed, "p4est load 3");
SC_CHECK_ABORT (p4est_connectivity_is_equal (connectivity, conn2),
"load/save connectivity mismatch D");
SC_CHECK_ABORT (p4est_is_equal (p4est, p4est2, 0),
"load/save p4est mismatch D");
p4est_destroy (p4est2);
p4est_connectivity_destroy (conn2);
/* Test autopartition load feature */
wtime = sc_MPI_Wtime ();
p4est2 = p4est_load_ext (p4est_name, mpicomm, sizeof (int), 0,
1, 0, NULL, &conn2);
elapsed = sc_MPI_Wtime () - wtime;
csum2 = p4est_checksum (p4est2);
sc_stats_set1 (stats + STATS_P4EST_LOAD4, elapsed, "p4est load 4");
SC_CHECK_ABORT (p4est_connectivity_is_equal (connectivity, conn2),
"load/save connectivity mismatch E");
SC_CHECK_ABORT (mpirank != 0 || csum == csum2,
"load/save p4est mismatch E");
p4est_destroy (p4est2);
p4est_connectivity_destroy (conn2);
/* destroy data structures */
p4est_destroy (p4est);
/* compute and print timings */
sc_stats_compute (mpicomm, STATS_COUNT, stats);
sc_stats_print (p4est_package_id, SC_LP_STATISTICS,
STATS_COUNT, stats, 0, 1);
}
int
main (int argc, char **argv)
{
int i, i1, i2, i3, i3last, i4, i4last, temp, count;
size_t s, swaps1, swaps2, swaps3, total1, total2, total3;
ssize_t searched;
int *pi;
sc_array_t *a1, *a2, *a3, *a4;
int mpiret;
double start, elapsed_pqueue, elapsed_qsort;
mpiret = sc_MPI_Init (&argc, &argv);
SC_CHECK_MPI (mpiret);
sc_init (sc_MPI_COMM_WORLD, 1, 1, NULL, SC_LP_DEFAULT);
a1 = sc_array_new (sizeof (int));
a2 = sc_array_new (sizeof (int));
a3 = sc_array_new (sizeof (int));
a4 = sc_array_new (sizeof (int));
#ifdef THEBIGTEST
count = 325323;
#else
count = 3251;
#endif
SC_INFOF ("Test pqueue with count %d\n", count);
start = -sc_MPI_Wtime ();
swaps1 = swaps2 = swaps3 = 0;
total1 = total2 = total3 = 0;
for (i = 0; i < count; ++i) {
*(int *) sc_array_push (a1) = i;
s = sc_array_pqueue_add (a1, &temp, compar);
swaps1 += ((s > 0) ? 1 : 0);
total1 += s;
*(int *) sc_array_push (a2) = count - i - 1;
s = sc_array_pqueue_add (a2, &temp, compar);
swaps2 += ((s > 0) ? 1 : 0);
total2 += s;
*(int *) sc_array_push (a3) = (15 * i) % 172;
s = sc_array_pqueue_add (a3, &temp, compar);
swaps3 += ((s > 0) ? 1 : 0);
total3 += s;
}
SC_CHECK_ABORT (swaps1 == 0 && total1 == 0, "pqueue_add");
SC_VERBOSEF (" Swaps %lld %lld %lld Total %lld %lld %lld\n",
(long long) swaps1, (long long) swaps2, (long long) swaps3,
(long long) total1, (long long) total2, (long long) total3);
temp = 52;
searched = sc_array_bsearch (a1, &temp, compar);
SC_CHECK_ABORT (searched != -1, "array_bsearch_index");
pi = (int *) sc_array_index_ssize_t (a1, searched);
SC_CHECK_ABORT (*pi == temp, "array_bsearch");
i3last = -1;
swaps1 = swaps2 = swaps3 = 0;
total1 = total2 = total3 = 0;
for (i = 0; i < count; ++i) {
s = sc_array_pqueue_pop (a1, &i1, compar);
swaps1 += ((s > 0) ? 1 : 0);
total1 += s;
s = sc_array_pqueue_pop (a2, &i2, compar);
swaps2 += ((s > 0) ? 1 : 0);
total2 += s;
s = sc_array_pqueue_pop (a3, &i3, compar);
swaps3 += ((s > 0) ? 1 : 0);
total3 += s;
SC_CHECK_ABORT (i == i1 && i == i2, "pqueue_pop");
SC_CHECK_ABORT (i3 >= i3last, "pqueue_pop");
i3last = i3;
}
SC_VERBOSEF (" Swaps %lld %lld %lld Total %lld %lld %lld\n",
(long long) swaps1, (long long) swaps2, (long long) swaps3,
(long long) total1, (long long) total2, (long long) total3);
elapsed_pqueue = start + sc_MPI_Wtime ();
sc_array_destroy (a1);
sc_array_destroy (a2);
sc_array_destroy (a3);
SC_INFOF ("Test array sort with count %d\n", count);
start = -sc_MPI_Wtime ();
/* the resize is done to be comparable with the above procedure */
for (i = 0; i < count; ++i) {
*(int *) sc_array_push (a4) = (15 * i) % 172;
}
sc_array_sort (a4, compar);
i4last = -1;
for (i = 0; i < count; ++i) {
i4 = *(int *) sc_array_index_int (a4, i);
SC_CHECK_ABORT (i4 >= i4last, "array_sort");
i4last = i4;
}
sc_array_resize (a4, 0);
elapsed_qsort = start + sc_MPI_Wtime ();
SC_STATISTICSF ("Test timings pqueue %g qsort %g\n",
elapsed_pqueue, 3. * elapsed_qsort);
sc_array_destroy (a4);
sc_finalize ();
mpiret = sc_MPI_Finalize ();
SC_CHECK_MPI (mpiret);
return 0;
}
static void
run_bricks (sc_MPI_Comm mpicomm, int per, int l, int rlevel)
{
int mpiret;
int tcount;
double elapsed_create, elapsed_partition, elapsed_balance;
#ifdef BRICKS_VTK
char filename[BUFSIZ];
#endif
p4est_connectivity_t *conn;
p4est_t *p4est;
P4EST_GLOBAL_PRODUCTIONF ("Run bricks on level %d/%d\n", l, rlevel);
P4EST_ASSERT (l <= rlevel);
/* create and refine the forest */
mpiret = sc_MPI_Barrier (mpicomm);
SC_CHECK_MPI (mpiret);
elapsed_create = -sc_MPI_Wtime ();
tcount = 1 << l;
#ifndef P4_TO_P8
conn = p4est_connectivity_new_brick (tcount, tcount, per, per);
#else
conn = p8est_connectivity_new_brick (tcount, tcount, tcount, per, per, per);
#endif
p4est = p4est_new_ext (mpicomm, conn, 0, rlevel - l, 1, 0, NULL, NULL);
level_shift = 4;
refine_level = rlevel - l + level_shift;
p4est_refine (p4est, 1, refine_fractal, NULL);
elapsed_create += sc_MPI_Wtime ();
/* partition the forest */
mpiret = sc_MPI_Barrier (mpicomm);
SC_CHECK_MPI (mpiret);
elapsed_partition = -sc_MPI_Wtime ();
p4est_partition (p4est, 0, NULL);
elapsed_partition += sc_MPI_Wtime ();
/* balance the forest */
mpiret = sc_MPI_Barrier (mpicomm);
SC_CHECK_MPI (mpiret);
elapsed_balance = -sc_MPI_Wtime ();
p4est_balance (p4est, P4EST_CONNECT_FULL, NULL);
elapsed_balance += sc_MPI_Wtime ();
/* postprocessing */
P4EST_GLOBAL_PRODUCTIONF ("Timings %g %g %g\n", elapsed_create,
elapsed_partition, elapsed_balance);
#ifdef BRICKS_VTK
snprintf (filename, BUFSIZ, "brick_%02d_%02d_B", rlevel, l);
p4est_vtk_write_file (p4est, NULL, filename);
#endif
p4est_destroy (p4est);
p4est_connectivity_destroy (conn);
}
