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;
NBC_Schedule *schedule;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
rank = ompi_comm_rank (comm);
rsize = ompi_comm_remote_size (comm);
res = NBC_Init_handle(comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* ensure the schedule is released with the handle on error */
handle->schedule = schedule;
if (0 == rank) {
for (int peer = 1 ; peer < rsize ; ++peer) {
res = NBC_Sched_recv (NULL, false, 0, MPI_BYTE, peer, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
}
/* synchronize with the remote root */
res = NBC_Sched_recv (NULL, false, 0, MPI_BYTE, 0, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
res = NBC_Sched_send (NULL, false, 0, MPI_BYTE, 0, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (0 == rank) {
/* wait for the remote root */
res = NBC_Sched_barrier (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* inform remote peers that all local peers have entered the barrier */
for (int peer = 1; peer < rsize ; ++peer) {
res = NBC_Sched_send (NULL, false, 0, MPI_BYTE, peer, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
/* 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 rank, p, maxround, res, recvpeer, sendpeer;
NBC_Schedule *schedule;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
res = NBC_Init_handle(comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
#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 (libnbc_module->NBC_Dict_size[NBC_BARRIER] == 0) {
/* we did not init it yet */
#endif
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* ensure the schedule is released with the handle on error */
handle->schedule = schedule;
maxround = (int)ceil((log((double)p)/LOG2)-1);
for (int round = 0 ; round <= maxround ; ++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 (NULL, false, 0, MPI_BYTE, sendpeer, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* recv msg from recvpeer */
res = NBC_Sched_recv (NULL, false, 0, MPI_BYTE, recvpeer, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* end communication round */
if (round < maxround) {
res = NBC_Sched_barrier (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
#ifdef NBC_CACHE_SCHEDULE
/* add it */
libnbc_module->NBC_Dict[NBC_BARRIER] = (hb_tree *) schedule;
libnbc_module->NBC_Dict_size[NBC_BARRIER] = 1;
} else {
/* we found it */
handle->schedule = schedule = (NBC_Schedule *) libnbc_module->NBC_Dict[NBC_BARRIER];
}
OBJ_RETAIN(schedule);
#endif
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
/* simple linear Alltoallw */
int ompi_coll_libnbc_ialltoallw_inter (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_1_0_t *module)
{
int res, rsize;
NBC_Schedule *schedule;
char *rbuf, *sbuf;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
rsize = ompi_comm_remote_size (comm);
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
for (int i = 0 ; i < rsize ; ++i) {
/* post all sends */
if (sendcounts[i] != 0) {
sbuf = (char *) sendbuf + sdispls[i];
res = NBC_Sched_send (sbuf, false, sendcounts[i], sendtypes[i], i, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
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, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
/* simple linear Alltoallv */
int ompi_coll_libnbc_ialltoallv(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, p, res, i;
MPI_Aint sndext, 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(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;
}
int ompi_coll_libnbc_ineighbor_alltoallv(void *sbuf, int *scounts, int *sdispls, MPI_Datatype stype,
void *rbuf, int *rcounts, int *rdispls, MPI_Datatype rtype,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module) {
int res, indegree, outdegree, *srcs, *dsts;
MPI_Aint sndext, rcvext;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_Schedule *schedule;
res = MPI_Type_extent (stype, &sndext);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in MPI_Type_extent() (%i)", res);
return res;
}
res = MPI_Type_extent (rtype, &rcvext);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in MPI_Type_extent() (%i)", res);
return res;
}
#ifdef NBC_CACHE_SCHEDULE
NBC_Ineighbor_alltoallv_args *args, *found, search;
/* 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_alltoallv_args *) hb_tree_search ((hb_tree *) libnbc_module->NBC_Dict[NBC_NEIGHBOR_ALLTOALLV],
&search);
if (NULL == found) {
#endif
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
res = NBC_Comm_neighbors (comm, &srcs, &indegree, &dsts, &outdegree);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
/* simply loop over neighbors and post send/recv operations */
for (int i = 0 ; i < indegree ; ++i) {
if (srcs[i] != MPI_PROC_NULL) {
res = NBC_Sched_recv ((char *) rbuf + rdispls[i] * rcvext, false, rcounts[i], rtype, srcs[i], schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
break;
}
}
}
free (srcs);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
free (dsts);
return res;
}
for (int i = 0 ; i < outdegree ; ++i) {
if (dsts[i] != MPI_PROC_NULL) {
res = NBC_Sched_send ((char *) sbuf + sdispls[i] * sndext, false, scounts[i], stype, dsts[i], schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
break;
}
}
}
free (dsts);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Ineighbor_alltoallv_args *) malloc (sizeof (args));
if (NULL != 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 *) libnbc_module->NBC_Dict[NBC_NEIGHBOR_ALLTOALLV], args, args, 0);
if (0 == res) {
OBJ_RETAIN(schedule);
/* increase number of elements for A2A */
if (++libnbc_module->NBC_Dict_size[NBC_NEIGHBOR_ALLTOALLV] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_NEIGHBOR_ALLTOALLV],
&libnbc_module->NBC_Dict_size[NBC_NEIGHBOR_ALLTOALLV]);
}
} else {
NBC_Error("error in dict_insert() (%i)", res);
free (args);
}
}
} else {
/* found schedule */
schedule = found->schedule;
OBJ_RETAIN(schedule);
}
#endif
res = NBC_Init_handle(comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
/* 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;
}
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;
}
/* 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(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int *recvcounts, const 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, speer, rpeer;
MPI_Aint rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
res = ompi_datatype_type_extent (recvtype, &rcvext);
if (OPAL_UNLIKELY(MPI_SUCCESS != res)) {
NBC_Error ("MPI Error in ompi_datatype_type_extent() (%i)", 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 (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
schedule = OBJ_NEW(NBC_Schedule);
if (NULL == schedule) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
sbuf = (char *) recvbuf + displs[rank] * rcvext;
/* do p-1 rounds */
for (int 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, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Sched_send (sbuf, false, sendcount, sendtype, speer, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
/* 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;
}
/* linear iexscan
* working principle:
* 1. each node (but node 0) receives from left neigbor
* 2. performs op
* 3. all but rank p-1 do sends to it's right neigbor and exits
*
*/
int ompi_coll_libnbc_iexscan(const 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_2_0_t *module) {
int rank, p, res;
ptrdiff_t gap, span;
NBC_Schedule *schedule;
#ifdef NBC_CACHE_SCHEDULE
NBC_Scan_args *args, *found, search;
#endif
char inplace;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
res = NBC_Init_handle(comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
span = opal_datatype_span(&datatype->super, count, &gap);
if (0 < rank) {
handle->tmpbuf = malloc(span);
if (handle->tmpbuf == NULL) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (inplace) {
res = NBC_Copy(recvbuf, count, datatype, (char *)handle->tmpbuf-gap, count, datatype, comm);
} else {
res = NBC_Copy(sendbuf, count, datatype, (char *)handle->tmpbuf-gap, count, datatype, comm);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
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;
found = (NBC_Scan_args *) hb_tree_search ((hb_tree *) libnbc_module->NBC_Dict[NBC_EXSCAN], &search);
if (NULL == found) {
#endif
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* make sure the schedule is released with the handle on error */
handle->schedule = schedule;
if (rank != 0) {
res = NBC_Sched_recv (recvbuf, false, count, datatype, rank-1, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
if (rank < p - 1) {
/* we have to wait until we have the data */
res = NBC_Sched_barrier(schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
res = NBC_Sched_op (recvbuf, false, (void *)(-gap), true, count,
datatype, op, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
/* send reduced data onward */
res = NBC_Sched_send ((void *)(-gap), true, count, datatype, rank + 1, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
}
} else if (p > 1) {
if (inplace) {
res = NBC_Sched_send (recvbuf, false, count, datatype, 1, schedule, false);
} else {
res = NBC_Sched_send (sendbuf, false, count, datatype, 1, schedule, false);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
}
res = NBC_Sched_commit(schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Scan_args *) malloc (sizeof (args));
if (NULL != 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 *) libnbc_module->NBC_Dict[NBC_EXSCAN], args, args, 0);
if (0 == res) {
OBJ_RETAIN(schedule);
/* increase number of elements for A2A */
if (++libnbc_module->NBC_Dict_size[NBC_EXSCAN] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_EXSCAN],
&libnbc_module->NBC_Dict_size[NBC_EXSCAN]);
}
} else {
NBC_Error("error in dict_insert() (%i)", res);
free (args);
}
}
} else {
/* found schedule */
schedule = found->schedule;
OBJ_RETAIN(schedule);
}
#endif
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
/* tmpbuf is freed with the handle */
return OMPI_SUCCESS;
}
int ompi_coll_libnbc_igatherv(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_1_0_t *module) {
int rank, p, res;
MPI_Aint rcvext = 0;
NBC_Schedule *schedule;
char *rbuf, inplace;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
if (rank == root) {
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in MPI_Type_extent() (%i)", res);
return res;
}
}
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* send to root */
if (rank != root) {
/* send msg to root */
res = NBC_Sched_send (sendbuf, false, sendcount, sendtype, root, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
} else {
for (int i = 0 ; i < p ; ++i) {
rbuf = (char *) recvbuf + displs[i] * rcvext;
if (i == root) {
if (!inplace) {
/* if I am the root - just copy the message */
res = NBC_Copy (sendbuf, sendcount, sendtype, rbuf, recvcounts[i], recvtype,
comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
} else {
/* root receives message to the right buffer */
res = NBC_Sched_recv (rbuf, false, recvcounts[i], recvtype, i, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
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_1_0_t *module) {
int res, rsize;
MPI_Aint rcvext;
NBC_Schedule *schedule;
char *rbuf;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
rsize = ompi_comm_size (comm);
if (MPI_ROOT == root) {
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in MPI_Type_extent() (%i)", res);
return res;
}
}
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* 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, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
} else if (MPI_ROOT == root) {
for (int 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, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
/* 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_1_0_t *module)
{
int rank, p, res;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
/* 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 (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
for (int 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, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
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, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
/* 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;
}
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_1_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; }
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;
}
int ompi_coll_libnbc_iallgatherv_inter(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, const int *recvcounts, const 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 res, rsize;
MPI_Aint rcvext;
NBC_Schedule *schedule;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
rsize = ompi_comm_remote_size (comm);
res = ompi_datatype_type_extent(recvtype, &rcvext);
if (OPAL_UNLIKELY(MPI_SUCCESS != res)) {
NBC_Error ("MPI Error in ompi_datatype_type_extent() (%i)", res);
return res;
}
schedule = OBJ_NEW(NBC_Schedule);
if (NULL == schedule) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* do rsize rounds */
for (int 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, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
}
if (sendcount) {
for (int r = 0 ; r < rsize ; ++r) {
res = NBC_Sched_send (sendbuf, false, sendcount, sendtype, r, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
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_1_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; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
if(p==1) {
if(!inplace) {
/* single node not in_place: copy data to recvbuf */
res = NBC_Copy(sendbuf, recvcounts[0], datatype, recvbuf, recvcounts[0], datatype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
/* manually complete the request */
(*request)->req_status.MPI_ERROR = OMPI_SUCCESS;
OPAL_THREAD_LOCK(&ompi_request_lock);
ompi_request_complete(*request, true);
OPAL_THREAD_UNLOCK(&ompi_request_lock);
return NBC_OK;
}
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(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((double)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);
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;
}
/* 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_1_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;
}
int ompi_coll_libnbc_iallreduce_inter(const 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_1_0_t *module)
{
int rank, res, size, rsize;
MPI_Aint ext;
NBC_Schedule *schedule;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
rank = ompi_comm_rank (comm);
rsize = ompi_comm_remote_size (comm);
res = MPI_Type_extent(datatype, &ext);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in MPI_Type_extent() (%i)", res);
return res;
}
res = MPI_Type_size(datatype, &size);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in MPI_Type_size() (%i)", res);
return res;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
handle->tmpbuf = malloc (ext * count);
if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* ensure the schedule is released with the handle on error */
handle->schedule = schedule;
res = allred_sched_linear (rank, rsize, sendbuf, recvbuf, count, datatype, op,
ext, size, schedule, handle);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
res = NBC_Sched_commit(schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
res = NBC_Start(handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
/* tmpbuf is freed with the handle */
return OMPI_SUCCESS;
}
/* 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;
}
int ompi_coll_libnbc_iallreduce(const 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_1_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_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
res = ompi_datatype_get_extent(datatype, &lb, &ext);
if (OMPI_SUCCESS != res) {
NBC_Error ("MPI Error in MPI_Type_extent() (%i)", res);
return res;
}
res = ompi_datatype_type_size (datatype, &size);
if (OMPI_SUCCESS != res) {
NBC_Error ("MPI Error in MPI_Type_size() (%i)", res);
return res;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
handle->tmpbuf = malloc (ext * count);
if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
if ((p == 1) && !inplace) {
/* for a single node - copy data to receivebuf */
res = NBC_Copy(sendbuf, count, datatype, recvbuf, count, datatype, comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
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 *) libnbc_module->NBC_Dict[NBC_ALLREDUCE], &search);
if (NULL == found) {
#endif
schedule = OBJ_NEW(NBC_Schedule);
if (NULL == schedule) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* ensure the schedule is released with the handle on error */
handle->schedule = schedule;
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 (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
res = NBC_Sched_commit(schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Allreduce_args *) malloc (sizeof(args));
if (NULL != 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 *) libnbc_module->NBC_Dict[NBC_ALLREDUCE], args, args, 0);
if (0 == res) {
OBJ_RETAIN(schedule);
/* increase number of elements for A2A */
if (++libnbc_module->NBC_Dict_size[NBC_ALLREDUCE] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLREDUCE],
&libnbc_module->NBC_Dict_size[NBC_ALLREDUCE]);
}
} else {
NBC_Error("error in dict_insert() (%i)", res);
free (args);
}
}
} else {
/* found schedule */
schedule = found->schedule;
OBJ_RETAIN(schedule);
}
#endif
res = NBC_Start (handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
*request = (ompi_request_t *) handle;
/* tmpbuf is freed with the handle */
return OMPI_SUCCESS;
}
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_1_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_allgatherv(void *sbuf, int scount, MPI_Datatype stype,
void *rbuf, int *rcounts, int *displs, MPI_Datatype rtype, MPI_Comm comm, NBC_Handle* handle) {
int rank, size, res, worldsize, i;
MPI_Aint sndext, rcvext;
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; }
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; }
t[1] = PMPI_Wtime();
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;
}
