int NBC_Comm_neighbors(MPI_Comm comm, int maxindegree, int sources[], int sourceweights[], int maxoutdegree, int destinations[], int destweights[]) {
int topo, res;
int index = 0;
int indeg, outdeg, wgtd;
res = NBC_Comm_neighbors_count(comm, &indeg, &outdeg, &wgtd);
if(indeg > maxindegree && outdeg > maxoutdegree) return NBC_INVALID_PARAM; /* we want to return *all* neighbors */
res = MPI_Topo_test(comm, &topo);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Topo_test() (%i)\n", res); return res; }
switch(topo) {
case MPI_CART: /* cartesian */
{
int ndims, i, rpeer, speer;
res = MPI_Cartdim_get(comm, &ndims);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Cartdim_get() (%i)\n", res); return res; }
for(i = 0; i<ndims; i++) {
res = MPI_Cart_shift(comm, i, 1, &rpeer, &speer);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Cart_shift() (%i)\n", res); return res; }
sources[index] = destinations[index] = rpeer; index++;
sources[index] = destinations[index] = speer; index++;
}
}
break;
case MPI_GRAPH: /* graph */
{
int rank;
MPI_Comm_rank(comm, &rank);
res = MPI_Graph_neighbors(comm, rank, maxindegree, sources);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Graph_neighbors_count() (%i)\n", res); return res; }
for(int i=0; i<maxindegree; i++) destinations[i] = sources[i];
}
break;
case MPI_DIST_GRAPH: /* dist graph */
{
res = MPI_Dist_graph_neighbors(comm, maxindegree, sources, sourceweights, maxoutdegree, destinations, destweights);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Graph_neighbors_count() (%i)\n", res); return res; }
}
break;
case MPI_UNDEFINED:
return NBC_INVALID_TOPOLOGY_COMM;
break;
default:
return NBC_INVALID_PARAM;
break;
}
return NBC_OK;
}
int ompi_coll_libnbc_ineighbor_allgatherv(void *sbuf, int scount, MPI_Datatype stype, void *rbuf,
int *rcounts, int *displs, MPI_Datatype rtype,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module) {
int rank, size, res, worldsize;
MPI_Aint sndext, rcvext;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
res = NBC_Init_handle(comm, coll_req, libnbc_module);
handle = *coll_req;
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
res = MPI_Comm_size(comm, &size);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Comm_size(MPI_COMM_WORLD, &worldsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Type_extent(stype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(rtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
char inplace;
NBC_Schedule *schedule;
#ifdef NBC_CACHE_SCHEDULE
NBC_Ineighbor_allgatherv_args *args, *found, search;
#endif
NBC_IN_PLACE(sbuf, rbuf, inplace);
handle->tmpbuf=NULL;
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sbuf=sbuf;
search.scount=scount;
search.stype=stype;
search.rbuf=rbuf;
search.rcount=rcount;
search.rtype=rtype;
found = (NBC_Ineighbor_allgatherv_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLGATHERV], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create, res = %i\n", res); return res; }
{
int indegree, outdegree, weighted, *srcs, *dsts, i;
res = NBC_Comm_neighbors_count(comm, &indegree, &outdegree, &weighted);
if(res != NBC_OK) return res;
srcs = (int*)malloc(sizeof(int)*indegree);
dsts = (int*)malloc(sizeof(int)*outdegree);
res = NBC_Comm_neighbors(comm, indegree, srcs, MPI_UNWEIGHTED, outdegree, dsts, MPI_UNWEIGHTED);
if(res != NBC_OK) return res;
if(inplace) { /* we need an extra buffer to be deadlock-free */
int sumrcounts=0;
int offset=0;
for(i=0; i<indegree; ++i) sumrcounts += rcounts[i];
handle->tmpbuf = malloc(rcvext*sumrcounts);
for(i = 0; i < indegree; i++) {
if(srcs[i] != MPI_PROC_NULL) {
res = NBC_Sched_recv((char*)0+offset, true, rcounts[i], rtype, srcs[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
offset += rcounts[i]*rcvext;
}
for(i = 0; i < outdegree; i++) {
if(dsts[i] != MPI_PROC_NULL) {
res = NBC_Sched_send((char*)sbuf, false, scount, stype, dsts[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
}
/* unpack from buffer */
offset=0;
for(i = 0; i < indegree; i++) {
if(srcs[i] != MPI_PROC_NULL) {
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
res = NBC_Sched_copy((char*)0+offset, true, rcounts[i], rtype, (char*)rbuf+displs[i]*rcvext, false, rcounts[i], rtype, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_copy() (%i)\n", res); return res; }
}
offset += rcounts[i]*rcvext;
}
} else { /* non INPLACE case */
/* simply loop over neighbors and post send/recv operations */
for(i = 0; i < indegree; i++) {
if(srcs[i] != MPI_PROC_NULL) {
res = NBC_Sched_recv((char*)rbuf+displs[i]*rcvext, false, rcounts[i], rtype, srcs[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
for(i = 0; i < outdegree; i++) {
if(dsts[i] != MPI_PROC_NULL) {
res = NBC_Sched_send((char*)sbuf, false, scount, stype, dsts[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
}
}
}
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Ineighbor_allgatherv_args*)malloc(sizeof(NBC_Ineighbor_allgatherv_args));
args->sbuf=sbuf;
args->scount=scount;
args->stype=stype;
args->rbuf=rbuf;
args->rcount=rcount;
args->rtype=rtype;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLGATHERV], args, args, 0);
if(res != 0) printf("error in dict_insert() (%i)\n", res);
/* increase number of elements for A2A */
if(++handle->comminfo->NBC_Dict_size[NBC_NEIGHBOR_ALLGATHERV] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLGATHERV], &handle->comminfo->NBC_Dict_size[NBC_NEIGHBOR_ALLGATHERV]);
}
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int NBC_Ineighbor_alltoallw(void *sbuf, int *scounts, MPI_Datatype *stypes,
void *rbuf, int *rcounts, MPI_Datatype *rtypes, MPI_Comm comm, NBC_Handle* handle) {
int rank, size, res, worldsize, i;
MPI_Aint *sndexts, *rcvexts;
double t[10];
t[0] = PMPI_Wtime();
res = NBC_Init_handle(handle, comm);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
res = MPI_Comm_size(comm, &size);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Comm_size(MPI_COMM_WORLD, &worldsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
t[1] = PMPI_Wtime();
char inplace;
NBC_Schedule *schedule;
#ifdef NBC_CACHE_SCHEDULE
NBC_Ineighbor_alltoallw_args *args, *found, search;
#endif
NBC_IN_PLACE(sbuf, rbuf, inplace);
handle->tmpbuf=NULL;
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sbuf=sbuf;
search.scount=scount;
search.stype=stype;
search.rbuf=rbuf;
search.rcount=rcount;
search.rtype=rtype;
found = (NBC_Ineighbor_alltoallw_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLTOALLW], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create, res = %i\n", res); return res; }
{
int indegree, outdegree, weighted, *srcs, *dsts, i;
res = NBC_Comm_neighbors_count(comm, &indegree, &outdegree, &weighted);
if(res != NBC_OK) return res;
srcs = (int*)malloc(sizeof(int)*indegree);
dsts = (int*)malloc(sizeof(int)*outdegree);
sndexts = (MPI_Aint*)malloc(sizeof(MPI_Aint)*outdegree);
for(i=0; i<outdegree; ++i) {
res = MPI_Type_extent(stypes[i], &sndexts[i]);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
}
rcvexts = (MPI_Aint*)malloc(sizeof(MPI_Aint)*indegree);
for(i=0; i<indegree; ++i) {
res = MPI_Type_extent(rtypes[i], &rcvexts[i]);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
}
res = NBC_Comm_neighbors(comm, indegree, srcs, MPI_UNWEIGHTED, outdegree, dsts, MPI_UNWEIGHTED);
if(res != NBC_OK) return res;
if(inplace) { /* we need an extra buffer to be deadlock-free */
int sumrbytes=0;
int offset=0;
for(i=0; i<indegree; ++i) sumrbytes += rcounts[i]*rcvexts[i];
handle->tmpbuf = malloc(sumrbytes);
for(i = 0; i < indegree; i++) {
res = NBC_Sched_recv((char*)0+offset, true, rcounts[i], rtypes[i], srcs[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
offset += rcounts[i]*rcvexts[i];
}
offset=0;
for(i = 0; i < outdegree; i++) {
res = NBC_Sched_send((char*)sbuf+offset, false, scounts[i], stypes[i], dsts[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
offset += scounts[i]*sndexts[i];
}
/* unpack from buffer */
offset=0;
for(i = 0; i < indegree; i++) {
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
res = NBC_Sched_copy((char*)0+offset, true, rcounts[i], rtypes[i], (char*)rbuf+offset, false, rcounts[i], rtypes[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_copy() (%i)\n", res); return res; }
offset += rcounts[i]*rcvexts[i];
}
} else { /* non INPLACE case */
/* simply loop over neighbors and post send/recv operations */
int offset=0;
for(i = 0; i < indegree; i++) {
res = NBC_Sched_recv((char*)rbuf+offset, false, rcounts[i], rtypes[i], srcs[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
offset += rcounts[i]*rcvexts[i];
}
offset = 0;
for(i = 0; i < outdegree; i++) {
res = NBC_Sched_send((char*)sbuf+offset, false, scounts[i], stypes[i], dsts[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
offset += scounts[i]*sndexts[i];
}
}
}
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Ineighbor_alltoallw_args*)malloc(sizeof(NBC_Ineighbor_alltoallw_args));
args->sbuf=sbuf;
args->scount=scount;
args->stype=stype;
args->rbuf=rbuf;
args->rcount=rcount;
args->rtype=rtype;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLTOALLW], args, args, 0);
if(res != 0) printf("error in dict_insert() (%i)\n", res);
/* increase number of elements for A2A */
if(++handle->comminfo->NBC_Dict_size[NBC_NEIGHBOR_ALLTOALLW] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLTOALLW], &handle->comminfo->NBC_Dict_size[NBC_NEIGHBOR_ALLTOALLW]);
}
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int NBC_Comm_neighbors (ompi_communicator_t *comm, int **sources, int *source_count, int **destinations, int *dest_count) {
int res, indeg, outdeg;
*sources = *destinations = NULL;
res = NBC_Comm_neighbors_count(comm, &indeg, &outdeg);
if (OMPI_SUCCESS != res) {
return res;
}
*source_count = indeg;
*dest_count = outdeg;
if (indeg) {
*sources = malloc (sizeof (int) * indeg);
if (OPAL_UNLIKELY(NULL == *sources)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
} else {
*sources = NULL;
}
if (outdeg) {
*destinations = malloc (sizeof (int) * outdeg);
if (OPAL_UNLIKELY(NULL == *destinations)) {
free (*sources);
*sources = NULL;
return OMPI_ERR_OUT_OF_RESOURCE;
}
} else {
*destinations = NULL;
}
/* silence clang static analyzer warning about NULL-dereference */
if (0 == indeg && 0 == outdeg) {
return OMPI_SUCCESS;
}
if (OMPI_COMM_IS_CART(comm)) {
/* cartesian */
int rpeer, speer;
/* silence clang static analyzer warning */
assert (indeg == outdeg);
for (int dim = 0, i = 0 ; dim < comm->c_topo->mtc.cart->ndims ; ++dim) {
mca_topo_base_cart_shift (comm, dim, 1, &rpeer, &speer);
sources[0][i] = destinations[0][i] = rpeer; i++;
sources[0][i] = destinations[0][i] = speer; i++;
}
} else if (OMPI_COMM_IS_GRAPH(comm)) {
/* graph */
mca_topo_base_graph_neighbors (comm, ompi_comm_rank (comm), indeg, sources[0]);
memcpy (destinations[0], sources[0], indeg * sizeof (int));
} else if (OMPI_COMM_IS_DIST_GRAPH(comm)) {
/* dist graph */
mca_topo_base_dist_graph_neighbors (comm, indeg, sources[0], MPI_UNWEIGHTED, outdeg, destinations[0],
MPI_UNWEIGHTED);
}
return OMPI_SUCCESS;
}
