+ return group_out ? smpi_comm_new(group_out, NULL) : MPI_COMM_NULL;
+}
+
+void smpi_comm_use(MPI_Comm comm){
+ if (comm == MPI_COMM_UNINITIALIZED)
+ comm = smpi_process_comm_world();
+ comm->refcount++;
+}
+
+void smpi_comm_unuse(MPI_Comm comm){
+ if (comm == MPI_COMM_UNINITIALIZED)
+ comm = smpi_process_comm_world();
+ comm->refcount--;
+ if(comm->refcount==0){
+ if(comm->intra_comm != MPI_COMM_NULL)
+ smpi_comm_unuse(comm->intra_comm);
+ if(comm->leaders_comm != MPI_COMM_NULL)
+ smpi_comm_unuse(comm->leaders_comm);
+ if(comm->non_uniform_map !=NULL)
+ xbt_free(comm->non_uniform_map);
+ if(comm->leaders_map !=NULL)
+ xbt_free(comm->leaders_map);
+ if(comm->attributes !=NULL){
+ xbt_dict_cursor_t cursor = NULL;
+ int* key;
+ void * value;
+ int flag;
+ xbt_dict_foreach(comm->attributes, cursor, key, value){
+ smpi_comm_key_elem elem = xbt_dict_get_or_null(smpi_comm_keyvals, (const char*)key);
+ if(elem && elem->delete_fn)
+ elem->delete_fn(comm, *key, value, &flag);
+ }
+ }
+ xbt_free(comm);
+ }
+}
+
+static int
+compare_ints (const void *a, const void *b)
+{
+ const int *da = (const int *) a;
+ const int *db = (const int *) b;
+
+ return (*da > *db) - (*da < *db);
+}
+
+void smpi_comm_init_smp(MPI_Comm comm){
+ int leader = -1;
+
+ if (comm == MPI_COMM_UNINITIALIZED)
+ comm = smpi_process_comm_world();
+
+ int comm_size =smpi_comm_size(comm);
+
+ // If we are in replay - perform an ugly hack
+ // say to SimGrid that we are not in replay for a while, because we need
+ // the buffers to be copied for the following calls
+ int replaying = 0; //cache data to set it back again after
+ if(smpi_process_get_replaying()){
+ replaying=1;
+ smpi_process_set_replaying(0);
+ }
+
+ if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
+ smpi_switch_data_segment(smpi_process_index());
+ }
+ //identify neighbours in comm
+ //get the indexes of all processes sharing the same simix host
+ xbt_swag_t process_list = simcall_host_get_process_list(SIMIX_host_self());
+ int intra_comm_size = 0;
+ //only one process/node, disable SMP support and return
+// if(intra_comm_size==1){
+// smpi_comm_set_intra_comm(comm, MPI_COMM_SELF);
+// //smpi_comm_set_leaders_comm(comm, comm);
+// smpi_process_set_comm_intra(MPI_COMM_SELF);
+// return;
+// }
+
+
+ int i =0;
+ int min_index=INT_MAX;//the minimum index will be the leader
+ smx_process_t process = NULL;
+ xbt_swag_foreach(process, process_list) {
+ //is_in_comm=0;
+ int index = SIMIX_process_get_PID(process) -1;
+
+ if(smpi_group_rank(smpi_comm_group(comm), index)!=MPI_UNDEFINED){
+ intra_comm_size++;
+ //the process is in the comm
+ if(index < min_index)
+ min_index=index;
+ i++;
+ }
+ }
+ XBT_DEBUG("number of processes deployed on my node : %d", intra_comm_size);
+ MPI_Group group_intra = smpi_group_new(intra_comm_size);
+ i=0;
+ process = NULL;
+ xbt_swag_foreach(process, process_list) {
+ //is_in_comm=0;
+ int index = SIMIX_process_get_PID(process) -1;
+ if(smpi_group_rank(smpi_comm_group(comm), index)!=MPI_UNDEFINED){
+ smpi_group_set_mapping(group_intra, index, i);
+ i++;
+ }
+ }
+
+
+ MPI_Comm comm_intra = smpi_comm_new(group_intra, NULL);
+ //MPI_Comm shmem_comm = smpi_process_comm_intra();
+ //int intra_rank = smpi_comm_rank(shmem_comm);
+
+
+ //if(smpi_process_index()==min_index)
+ leader=min_index;
+
+ int * leaders_map= (int*)xbt_malloc0(sizeof(int)*comm_size);
+ int * leader_list= (int*)xbt_malloc0(sizeof(int)*comm_size);
+ for(i=0; i<comm_size; i++){
+ leader_list[i]=-1;
+ }
+
+ smpi_coll_tuned_allgather_mpich(&leader, 1, MPI_INT , leaders_map, 1, MPI_INT, comm);
+
+ if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
+ smpi_switch_data_segment(smpi_process_index());
+ }
+
+ if(!comm->leaders_map){
+ comm->leaders_map= leaders_map;
+ }else{
+ xbt_free(leaders_map);
+ }
+ int j=0;
+ int leader_group_size = 0;
+ for(i=0; i<comm_size; i++){
+ int already_done=0;
+ for(j=0;j<leader_group_size; j++){
+ if(comm->leaders_map[i]==leader_list[j]){
+ already_done=1;
+ }
+ }
+ if(!already_done){
+ leader_list[leader_group_size]=comm->leaders_map[i];
+ leader_group_size++;
+ }
+ }
+ qsort(leader_list, leader_group_size, sizeof(int),compare_ints);
+
+ MPI_Group leaders_group = smpi_group_new(leader_group_size);
+
+
+ MPI_Comm leader_comm = MPI_COMM_NULL;
+ if(MPI_COMM_WORLD!=MPI_COMM_UNINITIALIZED && comm!=MPI_COMM_WORLD){
+ //create leader_communicator
+ for (i=0; i< leader_group_size;i++)
+ smpi_group_set_mapping(leaders_group, leader_list[i], i);
+ leader_comm = smpi_comm_new(leaders_group, NULL);
+ smpi_comm_set_leaders_comm(comm, leader_comm);
+ smpi_comm_set_intra_comm(comm, comm_intra);
+
+ //create intracommunicator
+ // smpi_comm_set_intra_comm(comm, smpi_comm_split(comm, *(int*)SIMIX_host_self(), comm_rank));
+ }else{
+ for (i=0; i< leader_group_size;i++)
+ smpi_group_set_mapping(leaders_group, leader_list[i], i);
+
+ leader_comm = smpi_comm_new(leaders_group, NULL);
+ if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL)
+ smpi_comm_set_leaders_comm(comm, leader_comm);
+ smpi_process_set_comm_intra(comm_intra);
+ }
+
+ int is_uniform = 1;
+
+ // Are the nodes uniform ? = same number of process/node
+ int my_local_size=smpi_comm_size(comm_intra);
+ if(smpi_comm_rank(comm_intra)==0) {
+ int* non_uniform_map = xbt_malloc0(sizeof(int)*leader_group_size);
+ smpi_coll_tuned_allgather_mpich(&my_local_size, 1, MPI_INT,
+ non_uniform_map, 1, MPI_INT, leader_comm);
+ for(i=0; i < leader_group_size; i++) {
+ if(non_uniform_map[0] != non_uniform_map[i]) {
+ is_uniform = 0;
+ break;
+ }
+ }
+ if(!is_uniform && smpi_comm_is_uniform(comm)){
+ comm->non_uniform_map= non_uniform_map;
+ }else{
+ xbt_free(non_uniform_map);
+ }
+ comm->is_uniform=is_uniform;
+ }
+ smpi_coll_tuned_bcast_mpich(&(comm->is_uniform),1, MPI_INT, 0, comm_intra );
+
+ if(smpi_privatize_global_variables){ //we need to switch here, as the called function may silently touch global variables
+ smpi_switch_data_segment(smpi_process_index());
+ }
+ // Are the ranks blocked ? = allocated contiguously on the SMP nodes
+ int is_blocked=1;
+ int prev=smpi_group_rank(smpi_comm_group(comm), smpi_group_index(smpi_comm_group(comm_intra), 0));
+ for (i=1; i<my_local_size; i++){
+ int this=smpi_group_rank(smpi_comm_group(comm),smpi_group_index(smpi_comm_group(comm_intra), i));
+ if(this!=prev+1){
+ is_blocked=0;
+ break;
+ }
+ prev = this;
+ }
+
+ int global_blocked;
+ smpi_mpi_allreduce(&is_blocked, &(global_blocked), 1,
+ MPI_INT, MPI_LAND, comm);
+
+ if(MPI_COMM_WORLD==SMPI_UNINITIALIZED || comm==MPI_COMM_WORLD){
+ if(smpi_comm_rank(comm)==0){
+ comm->is_blocked=global_blocked;
+ }
+ }else{
+ comm->is_blocked=global_blocked;
+ }
+ xbt_free(leader_list);
+
+ if(replaying==1)
+ smpi_process_set_replaying(1);
+}
+
+int smpi_comm_attr_delete(MPI_Comm comm, int keyval){
+ smpi_comm_key_elem elem = xbt_dict_get_or_null_ext(smpi_comm_keyvals, (const char*)&keyval, sizeof(int));
+ if(!elem)
+ return MPI_ERR_ARG;
+ if(elem->delete_fn!=MPI_NULL_DELETE_FN){
+ void * value;
+ int flag;
+ if(smpi_comm_attr_get(comm, keyval, &value, &flag)==MPI_SUCCESS){
+ int ret = elem->delete_fn(comm, keyval, value, &flag);
+ if(ret!=MPI_SUCCESS) return ret;
+ }
+ }
+ if(comm->attributes==NULL)
+ return MPI_ERR_ARG;
+
+ xbt_dict_remove_ext(comm->attributes, (const char*)&keyval, sizeof(int));
+ return MPI_SUCCESS;
+}
+
+int smpi_comm_attr_get(MPI_Comm comm, int keyval, void* attr_value, int* flag){
+ smpi_comm_key_elem elem = xbt_dict_get_or_null_ext(smpi_comm_keyvals, (const char*)&keyval, sizeof(int));
+ if(!elem)
+ return MPI_ERR_ARG;
+ xbt_ex_t ex;
+ if(comm->attributes==NULL){
+ *flag=0;
+ return MPI_SUCCESS;
+ }
+ TRY {
+ *(void**)attr_value = xbt_dict_get_ext(comm->attributes, (const char*)&keyval, sizeof(int));
+ *flag=1;
+ }
+ CATCH(ex) {
+ *flag=0;
+ xbt_ex_free(ex);
+ }
+ return MPI_SUCCESS;