int ompi_coll_libnbc_ibarrier_inter(struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, res, rsize, peer;
NBC_Schedule *schedule;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_remote_size(comm, &rsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_remote_size() (%i)\n", res); return res; }
handle->tmpbuf=(void*)malloc(2*sizeof(char));
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
if (0 == rank) {
for (peer = 1 ; peer < rsize ; ++peer) {
res = NBC_Sched_recv (0, true, 1, MPI_BYTE, peer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
/* synchronize with the remote root */
res = NBC_Sched_recv (0, true, 1, MPI_BYTE, 0, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_send (0, true, 1, MPI_BYTE, 0, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
if (0 == rank) {
/* wait for the remote root */
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
/* inform remote peers that all local peers have entered the barrier */
for (peer = 0 ; peer < rsize ; ++peer) {
res = NBC_Sched_send (0, true, 1, MPI_BYTE, peer, 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; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
/* simple linear Alltoallv */
int ompi_coll_libnbc_ialltoallv_inter (void* sendbuf, int *sendcounts, int *sdispls,
MPI_Datatype sendtype, void* recvbuf, int *recvcounts, int *rdispls,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module)
{
int rank, res, i, rsize;
MPI_Aint sndext, rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
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(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
MPI_Comm_remote_size (comm, &rsize);
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
for (i = 0; i < rsize; i++) {
/* post all sends */
if(sendcounts[i] != 0) {
sbuf = ((char *) sendbuf) + (sdispls[i] * sndext);
res = NBC_Sched_send(sbuf, false, sendcounts[i], sendtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
/* post all receives */
if(recvcounts[i] != 0) {
rbuf = ((char *) recvbuf) + (rdispls[i] * rcvext);
res = NBC_Sched_recv(rbuf, false, recvcounts[i], recvtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int ompi_coll_libnbc_igatherv_inter (void* sendbuf, int sendcount, MPI_Datatype sendtype,
void* recvbuf, int *recvcounts, int *displs, MPI_Datatype recvtype,
int root, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int rank, p, res, i, rsize;
MPI_Aint rcvext;
NBC_Schedule *schedule;
char *rbuf;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Comm_remote_size (comm, &rsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_remote_size() (%i)\n", res); return res; }
if (MPI_ROOT == root) {
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
}
handle->tmpbuf = NULL;
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
/* send to root */
if (MPI_ROOT != root && MPI_PROC_NULL != root) {
/* send msg to root */
res = NBC_Sched_send(sendbuf, false, sendcount, sendtype, root, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
} else if (MPI_ROOT == root) {
for (i = 0 ; i < rsize ; ++i) {
rbuf = ((char *)recvbuf) + (displs[i]*rcvext);
/* root receives message to the right buffer */
res = NBC_Sched_recv(rbuf, false, recvcounts[i], recvtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%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; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int ompi_coll_libnbc_iallgatherv_inter(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int *recvcounts, int *displs,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, res, r, rsize;
MPI_Aint rcvext;
NBC_Schedule *schedule;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_remote_size(comm, &rsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_remote_size() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create, (%i)\n", res); return res; }
/* do rsize rounds */
for (r = 0 ; r < rsize ; ++r) {
char *rbuf = ((char *)recvbuf) + (displs[r]*rcvext);
if (recvcounts[r]) {
res = NBC_Sched_recv(rbuf, false, recvcounts[r], recvtype, r, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
if (sendcount) {
for (r = 0 ; r < rsize ; ++r) {
res = NBC_Sched_send(sendbuf, false, sendcount, sendtype, r, 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; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int ompi_coll_libnbc_ibcast_inter(void *buffer, int count, MPI_Datatype datatype, int root,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int rank, p, res, size, peer;
NBC_Schedule *schedule;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Type_size(datatype, &size);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
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; }
if(root != MPI_PROC_NULL) {
/* send to all others */
if(root == MPI_ROOT) {
int remsize;
res = MPI_Comm_remote_size(comm, &remsize);
if(MPI_SUCCESS != res) { printf("MPI_Comm_remote_size() failed\n"); return res; }
for (peer=0;peer<remsize;peer++) {
/* send msg to peer */
res = NBC_Sched_send(buffer, false, count, datatype, peer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
} else {
/* recv msg from root */
res = NBC_Sched_recv(buffer, false, count, datatype, root, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%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; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int ompi_coll_libnbc_iallreduce_inter(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module)
{
int rank, res, size, rsize;
MPI_Aint ext;
NBC_Schedule *schedule;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_remote_size(comm, &rsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_remote_size() (%i)\n", res); return res; }
res = MPI_Type_extent(datatype, &ext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_size(datatype, &size);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
handle->tmpbuf = malloc(ext*count);
if(handle->tmpbuf == NULL) { printf("Error in malloc() (%i)\n", res); return NBC_OOR; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
res = allred_sched_linear(rank, rsize, sendbuf, recvbuf, count, datatype, op, ext, size, schedule, handle);
if (NBC_OK != res) { printf("Error in Schedule creation() (%i)\n", res); return res; }
res = NBC_Sched_commit(schedule);
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
/* tmpbuf is freed with the handle */
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;
}
/* the non-blocking reduce */
int ompi_coll_libnbc_ireduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype,
MPI_Op op, int root, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int rank, p, res, segsize, size;
MPI_Aint ext;
NBC_Schedule *schedule;
char *redbuf=NULL, inplace;
#ifdef NBC_CACHE_SCHEDULE
NBC_Reduce_args *args, *found, search;
#endif
enum { NBC_RED_BINOMIAL, NBC_RED_CHAIN } alg;
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;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(datatype, &ext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_size(datatype, &size);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
/* only one node -> copy data */
if((p == 1) && !inplace) {
res = NBC_Copy(sendbuf, count, datatype, recvbuf, count, datatype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
/* algorithm selection */
if(p > 4 || size*count < 65536) {
alg = NBC_RED_BINOMIAL;
if(rank == root) {
/* root reduces in receivebuffer */
handle->tmpbuf = malloc(ext*count);
} else {
/* recvbuf may not be valid on non-root nodes */
handle->tmpbuf = malloc(ext*count*2);
redbuf = ((char*)handle->tmpbuf)+(ext*count);
}
} else {
handle->tmpbuf = malloc(ext*count);
alg = NBC_RED_CHAIN;
segsize = 16384/2;
}
if (NULL == handle->tmpbuf) { printf("Error in malloc() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf=sendbuf;
search.recvbuf=recvbuf;
search.count=count;
search.datatype=datatype;
search.op=op;
search.root=root;
found = (NBC_Reduce_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_REDUCE], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
switch(alg) {
case NBC_RED_BINOMIAL:
res = red_sched_binomial(rank, p, root, sendbuf, recvbuf, count, datatype, op, redbuf, schedule, handle);
break;
case NBC_RED_CHAIN:
res = red_sched_chain(rank, p, root, sendbuf, recvbuf, count, datatype, op, ext, size, schedule, handle, segsize);
break;
}
if (NBC_OK != res) { printf("Error in Schedule creation() (%i)\n", res); return res; }
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Reduce_args*)malloc(sizeof(NBC_Alltoall_args));
args->sendbuf=sendbuf;
args->recvbuf=recvbuf;
args->count=count;
args->datatype=datatype;
args->op=op;
args->root=root;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_REDUCE], args, args, 0);
if(res != 0) printf("error in dict_insert() (%i)\n", res);
/* increase number of elements for Reduce */
if(++handle->comminfo->NBC_Dict_size[NBC_REDUCE] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_REDUCE], &handle->comminfo->NBC_Dict_size[NBC_REDUCE]);
}
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
/* tmpbuf is freed with the handle */
return NBC_OK;
}
/* simple linear MPI_Ialltoall the (simple) algorithm just sends to all nodes */
int ompi_coll_libnbc_ialltoall(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, p, res, a2asize, sndsize, datasize;
NBC_Schedule *schedule;
MPI_Aint rcvext, sndext;
#ifdef NBC_CACHE_SCHEDULE
NBC_Alltoall_args *args, *found, search;
#endif
char *rbuf, *sbuf, inplace;
enum {NBC_A2A_LINEAR, NBC_A2A_PAIRWISE, NBC_A2A_DISS} alg;
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;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_size(sendtype, &sndsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
/* algorithm selection */
a2asize = sndsize*sendcount*p;
/* this number is optimized for TCP on odin.cs.indiana.edu */
if((p <= 8) && ((a2asize < 1<<17) || (sndsize*sendcount < 1<<12))) {
/* just send as fast as we can if we have less than 8 peers, if the
* total communicated size is smaller than 1<<17 *and* if we don't
* have eager messages (msgsize < 1<<13) */
alg = NBC_A2A_LINEAR;
} else if(a2asize < (1<<12)*p) {
/*alg = NBC_A2A_DISS;*/
alg = NBC_A2A_LINEAR;
} else
alg = NBC_A2A_LINEAR; /*NBC_A2A_PAIRWISE;*/
if(!inplace) {
/* copy my data to receive buffer */
rbuf = ((char *)recvbuf) + (rank*recvcount*rcvext);
sbuf = ((char *)sendbuf) + (rank*sendcount*sndext);
res = NBC_Copy(sbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
/* allocate temp buffer if we need one */
if(alg == NBC_A2A_DISS) {
/* only A2A_DISS needs buffers */
if(NBC_Type_intrinsic(sendtype)) {
datasize = sndext*sendcount;
} else {
res = MPI_Pack_size(sendcount, sendtype, comm, &datasize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack_size() (%i)\n", res); return res; }
}
/* allocate temporary buffers */
if(p % 2 == 0) {
handle->tmpbuf=malloc(datasize*p*2);
} else {
/* we cannot divide p by two, so alloc more to be safe ... */
handle->tmpbuf=malloc(datasize*(p/2+1)*2*2);
}
/* phase 1 - rotate n data blocks upwards into the tmpbuffer */
#if OPAL_CUDA_SUPPORT
if(NBC_Type_intrinsic(sendtype) && !(opal_cuda_check_bufs((char *)sendbuf, (char *)recvbuf))) {
#else
if(NBC_Type_intrinsic(sendtype)) {
#endif /* OPAL_CUDA_SUPPORT */
/* contiguous - just copy (1st copy) */
memcpy(handle->tmpbuf, (char*)sendbuf+datasize*rank, datasize*(p-rank));
if(rank != 0) memcpy((char*)handle->tmpbuf+datasize*(p-rank), sendbuf, datasize*(rank));
} else {
int pos=0;
/* non-contiguous - pack */
res = MPI_Pack((char*)sendbuf+rank*sendcount*sndext, (p-rank)*sendcount, sendtype, handle->tmpbuf, (p-rank)*datasize, &pos, comm);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack() (%i)\n", res); return res; }
if(rank != 0) {
pos = 0;
MPI_Pack(sendbuf, rank*sendcount, sendtype, (char*)handle->tmpbuf+datasize*(p-rank), rank*datasize, &pos, comm);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack() (%i)\n", res); return res; }
}
}
} else {
handle->tmpbuf=NULL;
}
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf=sendbuf;
search.sendcount=sendcount;
search.sendtype=sendtype;
search.recvbuf=recvbuf;
search.recvcount=recvcount;
search.recvtype=recvtype;
found = (NBC_Alltoall_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLTOALL], &search);
if(found == NULL) {
#endif
/* not found - generate new schedule */
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
switch(alg) {
case NBC_A2A_LINEAR:
res = a2a_sched_linear(rank, p, sndext, rcvext, schedule, sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
break;
case NBC_A2A_DISS:
res = a2a_sched_diss(rank, p, sndext, rcvext, schedule, sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, handle);
break;
case NBC_A2A_PAIRWISE:
res = a2a_sched_pairwise(rank, p, sndext, rcvext, schedule, sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
break;
}
if (NBC_OK != 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_Alltoall_args*)malloc(sizeof(NBC_Alltoall_args));
args->sendbuf=sendbuf;
args->sendcount=sendcount;
args->sendtype=sendtype;
args->recvbuf=recvbuf;
args->recvcount=recvcount;
args->recvtype=recvtype;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLTOALL], 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_ALLTOALL] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLTOALL], &handle->comminfo->NBC_Dict_size[NBC_ALLTOALL]);
/*if(!rank) printf("[%i] removing %i elements - new size: %i \n", rank, SCHED_DICT_UPPER-SCHED_DICT_LOWER, handle->comminfo->NBC_Alltoall_size);*/
}
/*if(!rank) printf("[%i] added new schedule to tree - number %i\n", rank, handle->comminfo->NBC_Dict_size[NBC_ALLTOALL]);*/
} 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 ompi_coll_libnbc_ialltoall_inter (void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, res, i, rsize;
MPI_Aint sndext, rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_remote_size (comm, &rsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_remote_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(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
for (i = 0; i < rsize; i++) {
/* post all sends */
sbuf = ((char *) sendbuf) + (i * sendcount * sndext);
res = NBC_Sched_send(sbuf, false, sendcount, sendtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
/* post all receives */
rbuf = ((char *) recvbuf) + (i * recvcount * rcvext);
res = NBC_Sched_recv(rbuf, false, recvcount, recvtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int ompi_coll_libnbc_iallreduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module)
{
int rank, p, res;
OPAL_PTRDIFF_TYPE ext, lb;
NBC_Schedule *schedule;
size_t size;
#ifdef NBC_CACHE_SCHEDULE
NBC_Allreduce_args *args, *found, search;
#endif
enum { NBC_ARED_BINOMIAL, NBC_ARED_RING } alg;
char inplace;
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;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
res = ompi_datatype_get_extent(datatype, &lb, &ext);
if (OMPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = ompi_datatype_type_size (datatype, &size);
if (OMPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
handle->tmpbuf = malloc(ext*count);
if(handle->tmpbuf == NULL) { printf("Error in malloc() (%i)\n", res); return NBC_OOR; }
if((p == 1) && !inplace) {
/* for a single node - copy data to receivebuf */
res = NBC_Copy(sendbuf, count, datatype, recvbuf, count, datatype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
/* algorithm selection */
if(p < 4 || size*count < 65536 || inplace) {
alg = NBC_ARED_BINOMIAL;
} else {
alg = NBC_ARED_RING;
}
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf=sendbuf;
search.recvbuf=recvbuf;
search.count=count;
search.datatype=datatype;
search.op=op;
found = (NBC_Allreduce_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLREDUCE], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
switch(alg) {
case NBC_ARED_BINOMIAL:
res = allred_sched_diss(rank, p, count, datatype, sendbuf, recvbuf, op, schedule, handle);
break;
case NBC_ARED_RING:
res = allred_sched_ring(rank, p, count, datatype, sendbuf, recvbuf, op, size, ext, schedule, handle);
break;
}
if (NBC_OK != res) { printf("Error in Schedule creation() (%i)\n", res); return res; }
res = NBC_Sched_commit(schedule);
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Allreduce_args*)malloc(sizeof(NBC_Allreduce_args));
args->sendbuf=sendbuf;
args->recvbuf=recvbuf;
args->count=count;
args->datatype=datatype;
args->op=op;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLREDUCE], 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_ALLREDUCE] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLREDUCE], &handle->comminfo->NBC_Dict_size[NBC_ALLREDUCE]);
}
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
/* tmpbuf is freed with the handle */
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 ompi_coll_libnbc_ireduce_scatter_inter(void* sendbuf, void* recvbuf, int *recvcounts, MPI_Datatype datatype,
MPI_Op op, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int peer, rank, r, res, count, rsize, offset;
MPI_Aint ext;
NBC_Schedule *schedule;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_remote_size(comm, &rsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_remote_size() (%i)\n", res); return res; }
MPI_Type_extent(datatype, &ext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return NBC_OOR; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
count = 0;
for (r = 0 ; r < rsize ; ++r) count += recvcounts[r];
handle->tmpbuf = malloc(2 * ext * count);
if(handle->tmpbuf == NULL) { printf("Error in malloc()\n"); return NBC_OOR; }
/* send my data to the remote root */
res = NBC_Sched_send(sendbuf, false, count, datatype, 0, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
if (0 == rank) {
res = NBC_Sched_recv((void *) 0, true, count, datatype, 0, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
for (peer = 1 ; peer < rsize ; ++peer) {
res = NBC_Sched_recv((void *)(ext * count), true, count, datatype, peer, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
res = NBC_Sched_op((void *) 0, true, (void *)(ext * count), true, (void *) 0, true, count, datatype, op, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_op() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
}
/* exchange data with remote root for scatter phase (we *could* use the local communicator to do the scatter) */
res = NBC_Sched_recv((void *)(ext * count), true, count, datatype, 0, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_send((void *) 0, true, count, datatype, 0, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
/* scatter */
for (peer = 0, offset = ext * count ; peer < rsize ; ++peer) {
res = NBC_Sched_send((void *)(uintptr_t) offset, true, recvcounts[peer], datatype, peer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
offset += recvcounts[peer] * ext;
}
}
/* receive my block */
res = NBC_Sched_recv(recvbuf, false, recvcounts[rank], datatype, 0, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
/* tmpbuf is freed with the handle */
return NBC_OK;
}
/* Dissemination implementation of MPI_Ibarrier */
int ompi_coll_libnbc_ibarrier(struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int round, rank, p, maxround, res, recvpeer, sendpeer;
NBC_Schedule *schedule;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
handle->tmpbuf=(void*)malloc(2*sizeof(char));
#ifdef NBC_CACHE_SCHEDULE
/* there only one argument set per communicator -> hang it directly at
* the tree-position, NBC_Dict_size[...] is 0 for not initialized and
* 1 for initialized. NBC_Dict[...] is a pointer to the schedule in
* this case */
if(handle->comminfo->NBC_Dict_size[NBC_BARRIER] == 0) {
/* we did not init it yet */
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
round = -1;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
maxround = (int)ceil((log((double)p)/LOG2)-1);
do {
round++;
sendpeer = (rank + (1<<round)) % p;
/* add p because modulo does not work with negative values */
recvpeer = ((rank - (1<<round))+p) % p;
/* send msg to sendpeer */
res = NBC_Sched_send((void*)0, true, 1, MPI_BYTE, sendpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
/* recv msg from recvpeer */
res = NBC_Sched_recv((void*)1, true, 1, MPI_BYTE, recvpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
/* end communication round */
if(round < maxround){
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
}
} while (round < maxround);
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* add it */
handle->comminfo->NBC_Dict[NBC_BARRIER] = (hb_tree*)schedule;
handle->comminfo->NBC_Dict_size[NBC_BARRIER] = 1;
} else {
/* we found it */
schedule = (NBC_Schedule*)handle->comminfo->NBC_Dict[NBC_BARRIER];
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
/* simple linear MPI_Iallgatherv
* the algorithm uses p-1 rounds
* first round:
* each node sends to it's left node (rank+1)%p sendcount elements
* each node begins with it's right node (rank-11)%p and receives from it recvcounts[(rank+1)%p] elements
* second round:
* each node sends to node (rank+2)%p sendcount elements
* each node receives from node (rank-2)%p recvcounts[(rank+2)%p] elements */
int ompi_coll_libnbc_iallgatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int *recvcounts, int *displs,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, p, res, r, speer, rpeer;
MPI_Aint rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
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;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create, (%i)\n", res); return res; }
if (inplace) {
sendtype = recvtype;
sendcount = recvcounts[rank];
} else {
/* copy my data to receive buffer */
rbuf = ((char *)recvbuf) + (displs[rank]*rcvext);
res = NBC_Copy(sendbuf, sendcount, sendtype, rbuf, recvcounts[rank], recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
sbuf = ((char*) recvbuf) + (displs[rank]*rcvext);
/* do p-1 rounds */
for(r=1;r<p;r++) {
speer = (rank+r)%p;
rpeer = (rank-r+p)%p;
rbuf = ((char *)recvbuf) + (displs[rpeer]*rcvext);
res = NBC_Sched_recv(rbuf, false, recvcounts[rpeer], recvtype, rpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_send(sbuf, false, sendcount, sendtype, speer, 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; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
/* simple linear MPI_Iscatter */
int ompi_coll_libnbc_iscatter(void* sendbuf, int sendcount, MPI_Datatype sendtype,
void* recvbuf, int recvcount, MPI_Datatype recvtype, int root,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int rank, p, res, i;
MPI_Aint sndext = 0;
NBC_Schedule *schedule;
char *sbuf, inplace;
#ifdef NBC_CACHE_SCHEDULE
NBC_Scatter_args *args, *found, search;
#endif
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;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
if (rank == root) {
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
}
handle->tmpbuf=NULL;
if((rank == root) && (!inplace)) {
sbuf = ((char *)sendbuf) + (rank*sendcount*sndext);
/* if I am the root - just copy the message (not for MPI_IN_PLACE) */
res = NBC_Copy(sbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf=sendbuf;
search.sendcount=sendcount;
search.sendtype=sendtype;
search.recvbuf=recvbuf;
search.recvcount=recvcount;
search.recvtype=recvtype;
search.root=root;
found = (NBC_Scatter_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_SCATTER], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
/* receive from root */
if(rank != root) {
/* recv msg from root */
res = NBC_Sched_recv(recvbuf, false, recvcount, recvtype, root, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
} else {
for(i=0;i<p;i++) {
sbuf = ((char *)sendbuf) + (i*sendcount*sndext);
if(i != root) {
/* root sends the right buffer to the right receiver */
res = NBC_Sched_send(sbuf, false, sendcount, sendtype, 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_Scatter_args*)malloc(sizeof(NBC_Scatter_args));
args->sendbuf=sendbuf;
args->sendcount=sendcount;
args->sendtype=sendtype;
args->recvbuf=recvbuf;
args->recvcount=recvcount;
args->recvtype=recvtype;
args->root=root;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_SCATTER], 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_SCATTER] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_SCATTER], &handle->comminfo->NBC_Dict_size[NBC_SCATTER]);
}
} 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;
}
/* simple linear MPI_Iscatterv */
int ompi_coll_libnbc_iscatterv(void* sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype,
void* recvbuf, int recvcount, MPI_Datatype recvtype, int root,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int rank, p, res, i;
MPI_Aint sndext;
NBC_Schedule *schedule;
char *sbuf, inplace;
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;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
/* receive from root */
if(rank != root) {
/* recv msg from root */
res = NBC_Sched_recv(recvbuf, false, recvcount, recvtype, root, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
} else {
for(i=0;i<p;i++) {
sbuf = ((char *)sendbuf) + (displs[i]*sndext);
if(i == root) {
if(!inplace) {
/* if I am the root - just copy the message */
res = NBC_Copy(sbuf, sendcounts[i], sendtype, recvbuf, recvcount, recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
} else {
/* root sends the right buffer to the right receiver */
res = NBC_Sched_send(sbuf, false, sendcounts[i], sendtype, 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; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
/* simple linear Alltoallw */
int ompi_coll_libnbc_ialltoallw(void* sendbuf, int *sendcounts, int *sdispls,
MPI_Datatype sendtypes[], void* recvbuf, int *recvcounts, int *rdispls,
MPI_Datatype recvtypes[], struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module)
{
int rank, p, res, i;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
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;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res= MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
/* copy data to receivbuffer */
if((sendcounts[rank] != 0) && !inplace) {
rbuf = ((char *) recvbuf) + rdispls[rank];
sbuf = ((char *) sendbuf) + sdispls[rank];
res = NBC_Copy(sbuf, sendcounts[rank], sendtypes[rank], rbuf, recvcounts[rank], recvtypes[rank], comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
for (i = 0; i < p; i++) {
if (i == rank) { continue; }
/* post all sends */
if(sendcounts[i] != 0) {
sbuf = ((char *) sendbuf) + sdispls[i];
res = NBC_Sched_send(sbuf, false, sendcounts[i], sendtypes[i], i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
/* post all receives */
if(recvcounts[i] != 0) {
rbuf = ((char *) recvbuf) + rdispls[i];
res = NBC_Sched_recv(rbuf, false, recvcounts[i], recvtypes[i], i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int ompi_coll_libnbc_ibcast(void *buffer, int count, MPI_Datatype datatype, int root,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, p, res, size, segsize;
NBC_Schedule *schedule;
#ifdef NBC_CACHE_SCHEDULE
NBC_Bcast_args *args, *found, search;
#endif
enum { NBC_BCAST_LINEAR, NBC_BCAST_BINOMIAL, NBC_BCAST_CHAIN } alg;
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);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Type_size(datatype, &size);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
segsize = 16384;
/* algorithm selection */
if(p <= 4) {
alg = NBC_BCAST_LINEAR;
} else if(size*count < 65536) {
alg = NBC_BCAST_BINOMIAL;
} else if(size*count < 524288) {
alg = NBC_BCAST_CHAIN;
segsize = 16384/2;
} else {
alg = NBC_BCAST_CHAIN;
segsize = 65536/2;
}
handle->tmpbuf=NULL;
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.buffer=buffer;
search.count=count;
search.datatype=datatype;
search.root=root;
found = (NBC_Bcast_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_BCAST], &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; }
switch(alg) {
case NBC_BCAST_LINEAR:
res = bcast_sched_linear(rank, p, root, schedule, buffer, count, datatype);
break;
case NBC_BCAST_BINOMIAL:
res = bcast_sched_binomial(rank, p, root, schedule, buffer, count, datatype);
break;
case NBC_BCAST_CHAIN:
res = bcast_sched_chain(rank, p, root, schedule, buffer, count, datatype, segsize, size);
break;
}
if (NBC_OK != res) { printf("Error in Schedule creation() (%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_Bcast_args*)malloc(sizeof(NBC_Bcast_args));
args->buffer=buffer;
args->count=count;
args->datatype=datatype;
args->root=root;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_BCAST], 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_BCAST] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_BCAST], &handle->comminfo->NBC_Dict_size[NBC_BCAST]);
}
} 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 ompi_coll_libnbc_ireduce_scatter(void* sendbuf, void* recvbuf, int *recvcounts, MPI_Datatype datatype,
MPI_Op op, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int peer, rank, maxr, p, r, res, count, offset, firstred;
MPI_Aint ext;
char *redbuf, *sbuf, inplace;
NBC_Schedule *schedule;
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;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
MPI_Type_extent(datatype, &ext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return NBC_OOR; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
maxr = (int)ceil((log(p)/LOG2));
count = 0;
for(r=0;r<p;r++) count += recvcounts[r];
handle->tmpbuf = malloc(ext*count*2);
if(handle->tmpbuf == NULL) { printf("Error in malloc()\n"); return NBC_OOR; }
redbuf = ((char*)handle->tmpbuf)+(ext*count);
/* copy data to redbuf if we only have a single node */
if((p==1) && !inplace) {
res = NBC_Copy(sendbuf, count, datatype, redbuf, count, datatype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
firstred = 1;
for(r=1; r<=maxr; r++) {
if((rank % (1<<r)) == 0) {
/* we have to receive this round */
peer = rank + (1<<(r-1));
if(peer<p) {
res = NBC_Sched_recv(0, true, count, datatype, peer, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
/* we have to wait until we have the data */
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
if(firstred) {
/* take reduce data from the sendbuf in the first round -> save copy */
res = NBC_Sched_op(redbuf-(unsigned long)handle->tmpbuf, true, sendbuf, false, 0, true, count, datatype, op, schedule);
firstred = 0;
} else {
/* perform the reduce in my local buffer */
res = NBC_Sched_op(redbuf-(unsigned long)handle->tmpbuf, true, redbuf-(unsigned long)handle->tmpbuf, true, 0, true, count, datatype, op, schedule);
}
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_op() (%i)\n", res); return res; }
/* this cannot be done until handle->tmpbuf is unused :-( */
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
}
} else {
/* we have to send this round */
peer = rank - (1<<(r-1));
if(firstred) {
/* we have to send the senbuf */
res = NBC_Sched_send(sendbuf, false, count, datatype, peer, schedule);
} else {
/* we send an already reduced value from redbuf */
res = NBC_Sched_send(redbuf-(unsigned long)handle->tmpbuf, true, count, datatype, peer, schedule);
}
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
/* leave the game */
break;
}
}
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
/* rank 0 is root and sends - all others receive */
if(rank != 0) {
res = NBC_Sched_recv(recvbuf, false, recvcounts[rank], datatype, 0, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
if(rank == 0) {
offset = 0;
for(r=1;r<p;r++) {
offset += recvcounts[r-1];
sbuf = ((char *)redbuf) + (offset*ext);
/* root sends the right buffer to the right receiver */
res = NBC_Sched_send(sbuf-(unsigned long)handle->tmpbuf, true, recvcounts[r], datatype, r, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
res = NBC_Sched_copy(redbuf-(unsigned long)handle->tmpbuf, true, recvcounts[0], datatype, recvbuf, false, recvcounts[0], datatype, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_copy() (%i)\n", res); return res; }
}
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
/* tmpbuf is freed with the handle */
return NBC_OK;
}
/* simple linear Alltoallv */
int NBC_Ialltoallv(void* sendbuf, int *sendcounts, int *sdispls,
MPI_Datatype sendtype, void* recvbuf, int *recvcounts, int *rdispls,
MPI_Datatype recvtype, MPI_Comm comm, NBC_Handle* handle) {
int rank, p, res, i;
MPI_Aint sndext, rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(handle, comm);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%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_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
/* copy data to receivbuffer */
if((sendcounts[rank] != 0) && !inplace) {
rbuf = ((char *) recvbuf) + (rdispls[rank] * rcvext);
sbuf = ((char *) sendbuf) + (sdispls[rank] * sndext);
res = NBC_Copy(sbuf, sendcounts[rank], sendtype, rbuf, recvcounts[rank], recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
for (i = 0; i < p; i++) {
if (i == rank) { continue; }
/* post all sends */
if(sendcounts[i] != 0) {
sbuf = ((char *) sendbuf) + (sdispls[i] * sndext);
res = NBC_Sched_send(sbuf, false, sendcounts[i], sendtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
/* post all receives */
if(recvcounts[i] != 0) {
rbuf = ((char *) recvbuf) + (rdispls[i] * rcvext);
res = NBC_Sched_recv(rbuf, false, recvcounts[i], recvtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
