static void sort_and_compare_maps(
int proc,
int nbor_proc,
MESH_INFO_PTR mesh,
struct map_list_head *map,
int map_size,
int *sindex
)
{
/*
* Routine to sort a given communication map for a single neighbor processor
* and compare it to the actual communication map for that processor
* (generated by build_elem_comm_maps).
* If the DDirectory were used to build comm maps, this routine could be
* modified to do the assignments to mesh->ecmap_*, rather than do comparisons
* with it.
*/
int i, j;
int cnt = 0;
int indx;
/*
* Sort the given map according to element ids.
* Primary key is determined as in build_elem_comm_maps.
*/
for (i = 0; i < map_size; i++)
sindex[i] = i;
if (proc < nbor_proc)
quicksort_pointer_inc_id_id(sindex, map->glob_id, map->neigh_id,
0, map_size-1);
else
quicksort_pointer_inc_id_id(sindex, map->neigh_id, map->glob_id,
0, map_size-1);
/*
* Compute offset into mesh communication maps for the given nbor proc.
*/
if ((indx = in_list2(nbor_proc, mesh->necmap, mesh->ecmap_id)) == -1) {
printf("%d DDirectory Test: Comm map for nbor proc %d does not exist\n",
proc, nbor_proc);
return;
}
cnt = 0;
for (i = 0; i < indx; i++)
cnt += mesh->ecmap_cnt[i];
/*
* Compare given map to mesh communication map for this nbor proc.
* If the DDirectory code were used to construct maps, assignments
* would be done here (rather than comparisons).
*/
if (map_size != mesh->ecmap_cnt[indx]) {
printf("%d DDirectory Test: Different map size for nbor_proc %d: "
"%d != %d\n", proc, nbor_proc, map_size, mesh->ecmap_cnt[indx]);
return;
}
for (i = 0; i < map_size; i++) {
j = sindex[i];
if (map->elem_id[j] != mesh->ecmap_elemids[i+cnt]) {
printf("%d DDirectory Test: Different element IDs for nbor_proc %d: "
"%d != %d\n", proc, nbor_proc, map->elem_id[j],
mesh->ecmap_elemids[i+cnt]);
}
}
for (i = 0; i < map_size; i++) {
j = sindex[i];
if (map->side_id[j] != mesh->ecmap_sideids[i+cnt]) {
printf("%d DDirectory Test: Different side IDs for nbor_proc %d: "
"%d != %d\n", proc, nbor_proc, map->side_id[j],
mesh->ecmap_sideids[i+cnt]);
}
}
for (i = 0; i < map_size; i++) {
j = sindex[i];
if (map->neigh_id[j] != mesh->ecmap_neighids[i+cnt]) {
printf("%d DDirectory Test: Different neigh IDs for nbor_proc %d: "
ZOLTAN_ID_SPEC " != " ZOLTAN_ID_SPEC "\n", proc, nbor_proc, map->neigh_id[j],
mesh->ecmap_neighids[i+cnt]);
}
}
if (Debug_Driver > 3) {
printf("%d ************* DDirectory Map for %d ***************\n",
proc, nbor_proc);
printf("Local ID\tSide ID\tGlobal ID\tNeigh ID\n");
for (i = 0; i < map_size; i++) {
j = sindex[i];
printf("\t%d\t%d\t" ZOLTAN_ID_SPEC "\t" ZOLTAN_ID_SPEC "\n",
map->elem_id[j], map->side_id[j],
map->glob_id[j], map->neigh_id[j]);
}
}
}
int write_elem_vars(
int Proc,
MESH_INFO_PTR mesh,
PARIO_INFO_PTR pio_info,
int num_exp,
ZOLTAN_ID_PTR exp_gids,
int *exp_procs,
int *exp_to_part
)
{
/* Routine to write processor assignments per element to the nemesis files. */
int iblk;
int i, j;
int tmp;
float ver;
int pexoid, cpu_ws = 0, io_ws = 0;
int max_cnt = 0;
float *vars;
char par_nem_fname[FILENAME_MAX+1];
char tmp_nem_fname[FILENAME_MAX+1];
int Num_Proc;
char cmesg[256];
char *str = "Proc";
/* generate the parallel filename for this processor */
MPI_Comm_size(MPI_COMM_WORLD, &Num_Proc);
gen_par_filename(pio_info->pexo_fname, tmp_nem_fname, pio_info, Proc,
Num_Proc);
/*
* Copy the parallel file to a new file (so we don't write results in the
* cvs version).
*/
sprintf(cmesg, "%s.blot", pio_info->pexo_fname);
gen_par_filename(cmesg, par_nem_fname, pio_info, Proc,
Num_Proc);
fcopy(tmp_nem_fname, par_nem_fname);
if ((pexoid = ex_open(par_nem_fname, EX_WRITE, &cpu_ws, &io_ws,
&ver)) < 0) {
sprintf(cmesg,"fatal: could not open parallel Exodus II file %s",
par_nem_fname);
Gen_Error(0, cmesg);
return 0;
}
if (ex_put_var_names(pexoid, "e", 1, &str) < 0) {
Gen_Error(0, "Error returned from ex_put_var_names.");
return 0;
}
/* Get max number of elements in an element block; alloc vars array to size */
for (iblk = 0; iblk < mesh->num_el_blks; iblk++)
max_cnt = (mesh->eb_cnts[iblk] > max_cnt ? mesh->eb_cnts[iblk] : max_cnt);
vars = (float *) malloc(max_cnt * sizeof(float));
/* Must write data by element block; gather the data */
for (iblk = 0; iblk < mesh->num_el_blks; iblk++) {
for (j = 0, i = 0; i < mesh->num_elems; i++) {
if (mesh->elements[i].elem_blk == iblk) {
/* Element is in block; see whether it is to be exported. */
if ((tmp=in_list2((int)mesh->elements[i].globalID, num_exp, (int *) exp_gids)) != -1)
vars[j++] = (Output.Plot_Partition ? (float) (exp_to_part[tmp])
: (float) (exp_procs[tmp]));
else
vars[j++] = (Output.Plot_Partition ? mesh->elements[i].my_part
: (float) (Proc));
}
}
if (ex_put_elem_var(pexoid, 1, 1, mesh->eb_ids[iblk],
mesh->eb_cnts[iblk], vars) < 0) {
Gen_Error(0, "fatal: Error returned from ex_put_elem_var");
return 0;
}
}
safe_free((void **)(void *) &vars);
/* Close the parallel file */
if(ex_close (pexoid) < 0) {
Gen_Error(0, "fatal: Error returned from ex_close");
return 0;
}
return 1;
}
int build_elem_comm_maps(int proc, MESH_INFO_PTR mesh)
{
/*
* Build element communication maps, given a distributed mesh.
* This routine builds initial communication maps for Chaco input
* (for Nemesis, initial communication maps are read from the Nemesis file)
* and rebuilds communication maps after data migration.
*
* One communication map per neighboring processor is built.
* The corresponding maps on neighboring processors
* must be sorted in the same order, so that neighboring processors do not
* have to use ghost elements. For each communication map's pair of
* processors, the lower-numbered processor determines the order of the
* elements in the communication map. The sort key is the elements' global
* IDs on the lower-number processor; the secondary key is the neighboring
* elements global IDs. The secondary key is used when a single element
* must communicate with more than one neighbor.
*/
const char *yo = "build_elem_comm_maps";
int i, j;
ELEM_INFO *elem;
ZOLTAN_ID_TYPE iadj_elem;
int iadj_proc;
int indx;
int num_alloc_maps;
int max_adj = 0;
int max_adj_per_map;
int cnt, offset;
int *sindex = NULL;
int tmp;
struct map_list_head *tmp_maps = NULL, *map = NULL;
DEBUG_TRACE_START(proc, yo);
/*
* Free the old maps, if they exist.
*/
if (mesh->ecmap_id != NULL) {
safe_free((void **) &(mesh->ecmap_id));
safe_free((void **) &(mesh->ecmap_cnt));
safe_free((void **) &(mesh->ecmap_elemids));
safe_free((void **) &(mesh->ecmap_sideids));
safe_free((void **) &(mesh->ecmap_neighids));
mesh->necmap = 0;
}
/*
* Look for off-processor adjacencies.
* Loop over all elements
*/
num_alloc_maps = MAP_ALLOC;
mesh->ecmap_id = (int *) malloc(num_alloc_maps * sizeof(int));
mesh->ecmap_cnt = (int *) malloc(num_alloc_maps * sizeof(int));
tmp_maps = (struct map_list_head*) malloc(num_alloc_maps
* sizeof(struct map_list_head));
if (mesh->ecmap_id == NULL || mesh->ecmap_cnt == NULL || tmp_maps == NULL) {
Gen_Error(0, "Fatal: insufficient memory");
DEBUG_TRACE_END(proc, yo);
return 0;
}
for (i = 0; i < mesh->num_elems; i++) {
elem = &(mesh->elements[i]);
for (j = 0; j < elem->adj_len; j++) {
/* Skip NULL adjacencies (sides that are not adjacent to another elem). */
if (elem->adj[j] == ZOLTAN_ID_INVALID) continue;
iadj_elem = elem->adj[j];
iadj_proc = elem->adj_proc[j];
if (iadj_proc != proc) {
/*
* Adjacent element is off-processor.
* Add this element to the temporary data structure for
* the appropriate neighboring processor.
*/
if ((indx = in_list2(iadj_proc, mesh->necmap, mesh->ecmap_id)) == -1) {
/*
* Start a new communication map.
*/
if (mesh->necmap >= num_alloc_maps) {
num_alloc_maps += MAP_ALLOC;
mesh->ecmap_id = (int *) realloc(mesh->ecmap_id,
num_alloc_maps * sizeof(int));
mesh->ecmap_cnt = (int *) realloc(mesh->ecmap_cnt,
num_alloc_maps * sizeof(int));
tmp_maps = (struct map_list_head *) realloc(tmp_maps,
num_alloc_maps * sizeof(struct map_list_head));
if (mesh->ecmap_id == NULL || mesh->ecmap_cnt == NULL ||
tmp_maps == NULL) {
Gen_Error(0, "Fatal: insufficient memory");
DEBUG_TRACE_END(proc, yo);
return 0;
}
}
mesh->ecmap_id[mesh->necmap] = iadj_proc;
mesh->ecmap_cnt[mesh->necmap] = 0;
map = &(tmp_maps[mesh->necmap]);
map->glob_id = (ZOLTAN_ID_TYPE *) malloc(MAP_ALLOC * sizeof(ZOLTAN_ID_TYPE));
map->elem_id = (int *) malloc(MAP_ALLOC * sizeof(int));
map->side_id = (int *) malloc(MAP_ALLOC * sizeof(int));
map->neigh_id = (ZOLTAN_ID_TYPE *) malloc(MAP_ALLOC * sizeof(ZOLTAN_ID_TYPE));
if (map->glob_id == NULL || map->elem_id == NULL ||
map->side_id == NULL || map->neigh_id == NULL) {
Gen_Error(0, "Fatal: insufficient memory");
DEBUG_TRACE_END(proc, yo);
return 0;
}
map->map_alloc_size = MAP_ALLOC;
indx = mesh->necmap;
mesh->necmap++;
}
/* Add to map for indx. */
map = &(tmp_maps[indx]);
if (mesh->ecmap_cnt[indx] >= map->map_alloc_size) {
map->map_alloc_size += MAP_ALLOC;
map->glob_id = (ZOLTAN_ID_TYPE *) realloc(map->glob_id, map->map_alloc_size * sizeof(ZOLTAN_ID_TYPE));
map->elem_id = (int *) realloc(map->elem_id,
map->map_alloc_size * sizeof(int));
map->side_id = (int *) realloc(map->side_id,
map->map_alloc_size * sizeof(int));
map->neigh_id = (ZOLTAN_ID_TYPE *) realloc(map->neigh_id, map->map_alloc_size * sizeof(ZOLTAN_ID_TYPE));
if (map->glob_id == NULL || map->elem_id == NULL ||
map->side_id == NULL || map->neigh_id == NULL) {
Gen_Error(0, "Fatal: insufficient memory");
DEBUG_TRACE_END(proc, yo);
return 0;
}
}
tmp = mesh->ecmap_cnt[indx];
map->glob_id[tmp] = elem->globalID;
map->elem_id[tmp] = i;
map->side_id[tmp] = j+1; /* side is determined by position in
adj array (+1 since not 0-based). */
map->neigh_id[tmp] = iadj_elem;
mesh->ecmap_cnt[indx]++;
max_adj++;
}
}
}
/*
* If no communication maps, don't need to do anything else.
*/
if (mesh->necmap > 0) {
/*
* Allocate data structure for element communication map arrays.
*/
mesh->ecmap_elemids = (int *) malloc(max_adj * sizeof(int));
mesh->ecmap_sideids = (int *) malloc(max_adj * sizeof(int));
mesh->ecmap_neighids = (ZOLTAN_ID_TYPE *) malloc(max_adj * sizeof(ZOLTAN_ID_TYPE));
/*
* Allocate temporary memory for sort index.
*/
max_adj_per_map = 0;
for (i = 0; i < mesh->necmap; i++)
if (mesh->ecmap_cnt[i] > max_adj_per_map)
max_adj_per_map = mesh->ecmap_cnt[i];
sindex = (int *) malloc(max_adj_per_map * sizeof(int));
cnt = 0;
for (i = 0; i < mesh->necmap; i++) {
map = &(tmp_maps[i]);
for (j = 0; j < mesh->ecmap_cnt[i]; j++)
sindex[j] = j;
/*
* Sort the map so that adjacent processors have the same ordering
* for the communication.
* Assume the ordering of the lower-numbered processor in the pair
* of communicating processors.
*/
if (proc < mesh->ecmap_id[i])
quicksort_pointer_inc_id_id(sindex, map->glob_id, map->neigh_id,
0, mesh->ecmap_cnt[i]-1);
else
quicksort_pointer_inc_id_id(sindex, map->neigh_id, map->glob_id,
0, mesh->ecmap_cnt[i]-1);
/*
* Copy sorted data into elem map arrays.
*/
offset = cnt;
for (j = 0; j < mesh->ecmap_cnt[i]; j++) {
mesh->ecmap_elemids[offset] = map->elem_id[sindex[j]];
mesh->ecmap_sideids[offset] = map->side_id[sindex[j]];
mesh->ecmap_neighids[offset] = map->neigh_id[sindex[j]];
offset++;
}
cnt += mesh->ecmap_cnt[i];
}
}
/* Free temporary data structure. */
for (i = 0; i < mesh->necmap; i++) {
safe_free((void **) &(tmp_maps[i].glob_id));
safe_free((void **) &(tmp_maps[i].elem_id));
safe_free((void **) &(tmp_maps[i].side_id));
safe_free((void **) &(tmp_maps[i].neigh_id));
}
safe_free((void **) &tmp_maps);
safe_free((void **) &sindex);
if (Test.DDirectory)
compare_maps_with_ddirectory_results(proc, mesh);
DEBUG_TRACE_END(proc, yo);
return 1;
}
