void omp_report_mask(){
int nthrds, thrd; //Thread info
int ncpus, nel_set;
static int ** proc_mask;
int i,j, ierr;
char * dummy;
thrd = omp_get_thread_num();
nthrds = omp_get_num_threads();
ncpus = (int) sysconf(_SC_NPROCESSORS_ONLN);
if(omp_get_num_procs() != ncpus){
printf("ERROR: ncpus_by_omp=%d, ncpus_sched=%d\n",omp_get_num_procs(),ncpus);
exit(1);
}
#pragma omp single
{
proc_mask = malloc(sizeof(int*)*nthrds);
for(i=0;i<nthrds;i++) proc_mask[i] = malloc(sizeof(int)*ncpus );
for(i=0;i<nthrds;i++) for(j=0;j<ncpus;j++) proc_mask[i][j] =0;
}
ierr = boundto(&nel_set,proc_mask[thrd]);
#pragma omp barrier
#pragma omp single
{
print_mask(1, dummy, 0, 0,0, ncpus, nthrds,1, proc_mask[thrd]); //print header
for(thrd=0;thrd<nthrds;thrd++){
print_mask(0, dummy, 0, thrd,0, ncpus, nthrds,1, proc_mask[thrd]);
}
}
}
int hybrid_report_mask(){
// General
int i,j,ierr;
int id, rid,tid;
int in_mpi, in_omp;
int thrd, nthrds;
int ncpus, nel_set;
// Mask storage
static int ** omp_proc_mask;
static int * omp_mask_pac;
char *dummy;
// MPI specific Variables
int rank, nranks;
MPI_Request *request;
MPI_Status *status;
static int multi_node = 0;
static char *all_names;
static int max_name_len;
int name_len;
char proc_name[MPI_MAX_PROCESSOR_NAME];
char l,p;
int tpc; // hwthreads/core
Maskopts opts;
// get print_speed fast or slow (f|c); listing cores or SMT (c|s)
p = opts.get_p();
l = opts.get_l();
tpc=get_threads_per_node();
// In MPI and parallel region ?
MPI_Initialized(&in_mpi);
in_omp = omp_in_parallel();
if(in_mpi == 0){
printf("ERROR: ***** Must call hybrid_report_mask() in MPI program. ***** \n");
exit(1);
}
// Get rank number & no of ranks via MPI
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
if(in_omp == 0){
if(rank == 0){
printf(" ***** When using 1 thread, Intel OpenMP MAY report "
"\"not in a parallel region\" (Uh!)***** \n");
printf(" ***** Each row will only have a rank number (no \"0\" thread_id). \n");
printf("WARNING: ***** Unspecified results if hybrid_report_mask "
"not called in parallel region of MPI code section. ***** \n");
}
}
thrd = omp_get_thread_num(); // thread id
nthrds = omp_get_num_threads(); // Number of Threads
// Get number of cpus (this gives no.
// of cpu_ids in /proc/cpuinfo)
ncpus = (int) sysconf(_SC_NPROCESSORS_ONLN);
// Working only with MPI processes (masters)
#pragma omp master
{
// Get a list of nodes from all ranks.
MPI_Get_processor_name(proc_name,&name_len);
MPI_Allreduce(&name_len, &max_name_len, 1,MPI_INT, MPI_MAX, MPI_COMM_WORLD);
all_names = (char *) malloc(sizeof(int*)*nranks*(max_name_len+1));
MPI_Gather( proc_name, max_name_len+1 , MPI_CHAR,
all_names, max_name_len+1, MPI_CHAR,
0, MPI_COMM_WORLD);
// If multiple nodes, make muti_node non-zero.
if(rank == 0){
for(id=0;id<nranks;id++){
if( strcmp(&all_names[id*(max_name_len+1)],&all_names[0]) ) multi_node++; }
}
// Create shared storage for masks (only master allocates)
omp_proc_mask = (int **) malloc(sizeof(int*)*nthrds);
for(i=0;i<nthrds;i++) omp_proc_mask[i] = (int * ) malloc(sizeof(int )*ncpus );
for(i=0;i<nthrds;i++) for(j=0;j<ncpus;j++) omp_proc_mask[i][j] =0;
}
#pragma omp barrier
#pragma omp critical // (boundto -- may not be thread safe)
ierr = boundto(&nel_set,omp_proc_mask[thrd]);
#pragma omp barrier
#pragma omp master
{
omp_mask_pac = (int *) malloc(sizeof(int)*nranks*nthrds*ncpus); // need packing space for mpi send/recv
if(rank == 0){
request = (MPI_Request *) malloc(sizeof(MPI_Request)*nranks);
status = (MPI_Status *) malloc(sizeof(MPI_Status )*nranks);
print_mask(1, dummy, multi_node, 0, 0, ncpus, nranks,nthrds, omp_proc_mask[0],tpc,l); //print header
fflush(stdout);
for(tid=0;tid<nthrds;tid++){
print_mask(0, &all_names[tid*(max_name_len+1)], multi_node, 0,tid, ncpus, nranks,nthrds, omp_proc_mask[tid],tpc,l);
}
fflush(stdout);
for(rid=1;rid<nranks;rid++){ // Receive other rank's packed mask arrays
MPI_Irecv(&omp_mask_pac[rid*nthrds*ncpus], nthrds*ncpus, MPI_INT, rid, 99, MPI_COMM_WORLD, &request[rid-1]);
}
MPI_Waitall(nranks-1,&request[0],&status[0]);
for(rid=1;rid<nranks;rid++){ // Print for each rank
for(tid=0;tid<nthrds;tid++){
print_mask(0, &all_names[tid*(max_name_len+1)], multi_node, rid,tid, ncpus, nranks,nthrds, &omp_mask_pac[rid*nthrds*ncpus + tid*ncpus],tpc,l);
if(p == 's') ierr=usleep(300000);
}
}
if(nranks*nthrds > 50)
print_mask(2, dummy, multi_node, 0, 0, ncpus, nranks,nthrds, omp_proc_mask[0],tpc,l); //print header
fflush(stdout);
} // end root printing
else{ //all non-root ranks
// Pack up the ranks' mask arrays (Uh, should have made one array from beginning!)
for( tid=0;tid<nthrds;tid++){
for( id=0; id<ncpus; id++) omp_mask_pac[(tid*ncpus)+id] = omp_proc_mask[tid][id];
if(p == 's') ierr=usleep(300000);
}
// Send to root
MPI_Send(omp_mask_pac, nthrds*ncpus, MPI_INT, 0, 99, MPI_COMM_WORLD);
} // end non-root printing
// Return allocated space
for(i=0;i<nthrds;i++) free(omp_proc_mask[i]);
free(omp_proc_mask);
free(omp_mask_pac);
if(rank == 0 ){ free(request); free(status);}
free(all_names);
} // end of Master
#pragma omp barrier // JIC, what all threads leaving at the same time.
}
int hybrid_report_mask(void){
int thrd, nthrds;
int rank, nranks;
static int multi_node = 0;
int ncpus, nel_set;
static int ** omp_proc_mask;
static int * omp_mask_pac;
char *dummy;
char proc_name[MPI_MAX_PROCESSOR_NAME];
static char * all_names;
int name_len;
static int max_name_len;
// General
int i,j,ierr;
int id, rid,tid;
int in_mpi, in_omp;
// Mask storage
int ** proc_mask;
static int * all_masks=0;
MPI_Initialized(&in_mpi);
in_omp = omp_in_parallel();
if(in_mpi != 0 && in_omp == 0){
// Get number of cpus (this gives no. of cpu_ids in /proc/cpuinfo)
// Get rank number & no of ranks via MPI
ncpus = (int) sysconf(_SC_NPROCESSORS_ONLN);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
// Create a 2-D array for mask
// proc_mask[rank][ncpus] -- for simplicity, size is [ncpus][ncpus]
// Thinking ahead for hybrid code.
// zero out proc_mask[ncpus][ncpus]
// I could have made proc_mask a single array (proc_mask[ncpus]); but didn't
// This is a hold-over from the openmp version that holds everything for all threads.
// For MPI I made a continguous collection array (all_masks).
proc_mask = malloc(sizeof(int*)*ncpus);
for(i=0;i<ncpus;i++) proc_mask[i] = malloc(sizeof(int)*ncpus);
for(i=0;i<ncpus;i++) for(j=0;j<ncpus;j++) proc_mask[i][j] =0;
all_masks = (int *) malloc(sizeof(int)*ncpus*ncpus);
// get map for this processor
ierr=boundto(&nel_set,proc_mask[rank]);
// Gather information to rank 0
MPI_Gather( proc_mask[rank], ncpus, MPI_INT,
all_masks, ncpus, MPI_INT,
0, MPI_COMM_WORLD);
// Get a list of nodes from all ranks.
MPI_Get_processor_name(proc_name,&name_len);
MPI_Allreduce(&name_len, &max_name_len, 1,MPI_INT, MPI_MAX, MPI_COMM_WORLD);
all_names = malloc(sizeof(int*)*nranks*(max_name_len+1));
MPI_Gather( proc_name, max_name_len+1 , MPI_CHAR,
all_names, max_name_len+1, MPI_CHAR,
0, MPI_COMM_WORLD);
// If multiple nodes, make muti_node not equal to 0.
if(rank == 0)
for(id=0;id<nranks;id++){
if( strcmp(&all_names[id*(max_name_len+1)],&all_names[0]) ) multi_node++;
}
} // End of Pure MPI part
if(in_mpi != 0 && in_omp != 0){
if(all_masks == 0) {
printf("ERROR: ***** You must call hybrid_report_mask() in a Pure MPI region first. ***** \n");
exit(1);
}
thrd = omp_get_thread_num();
nthrds = omp_get_num_threads();
ncpus = (int) sysconf(_SC_NPROCESSORS_ONLN);
#pragma omp single
{
omp_proc_mask = malloc(sizeof(int*)*nthrds);
for(i=0;i<nthrds;i++) omp_proc_mask[i] = malloc(sizeof(int)*ncpus );
for(i=0;i<nthrds;i++) for(j=0;j<ncpus;j++) omp_proc_mask[i][j] =0;
}
#pragma omp critical
ierr = boundto(&nel_set,omp_proc_mask[thrd]);
#pragma omp barrier
MPI_Comm_size(MPI_COMM_WORLD, &nranks);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
#pragma omp master
{
omp_mask_pac = (int *) malloc(sizeof(int)*nranks*ncpus); // need packing space for mpi send/recv
if(rank == 0){
print_mask(1, dummy, multi_node, 0, 0, ncpus, nthrds,nranks, omp_proc_mask[0]); //print header
fflush(stdout);
for(tid=0;tid<nthrds;tid++){
print_mask(0, &all_names[tid*(max_name_len+1)], multi_node, 0,tid, ncpus, nthrds,nranks, omp_proc_mask[tid]);
}
fflush(stdout);
for(rid=1;rid<nranks;rid++){
// Receive other rank's packed mask arrays
MPI_Recv(omp_mask_pac, nthrds*ncpus, MPI_INT, rid, 99, MPI_COMM_WORLD, MPI_STATUS_IGNORE);
for(tid=0;tid<nthrds;tid++){
print_mask(0, &all_names[rid*(max_name_len+1)], multi_node, rid,tid, ncpus, nthrds,nranks, &omp_mask_pac[tid*ncpus]);
}
fflush(stdout);
} // rank loop
} // end root printing
else{ //all other ranks
// All non-root ranks send to root.
for(rid=1;rid<nranks;rid++){
// Pack up the ranks' mask arrays (Uh, should have made one array from beginning!)
for( tid=0;tid<nthrds;tid++){
for( id=0; id<ncpus; id++) omp_mask_pac[(tid*ncpus)+id] = omp_proc_mask[tid][id];
}
// Send to root
MPI_Send(omp_mask_pac, nthrds*ncpus, MPI_INT, 0, 99, MPI_COMM_WORLD);
} //all other ranks
} // end non-root printing
MPI_Barrier(MPI_COMM_WORLD);
} // end of Master
#pragma omp barrier
} // end of OpenMP part
}
