static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, const void *sendbuf, void *recvbuf,
MPI_Op op, NBC_Schedule *schedule, NBC_Handle *handle) {
int root, vrank, maxr, vpeer, peer, res;
root = 0; /* this makes the code for ireduce and iallreduce nearly identical - could be changed to improve performance */
RANK2VRANK(rank, vrank, root);
maxr = (int)ceil((log((double)p)/LOG2));
for (int r = 1, firstred = 1 ; r <= maxr ; ++r) {
if ((vrank % (1 << r)) == 0) {
/* we have to receive this round */
vpeer = vrank + (1 << (r - 1));
VRANK2RANK(peer, vpeer, root)
if (peer < p) {
/* we have to wait until we have the data */
res = NBC_Sched_recv (0, true, count, datatype, peer, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
/* this cannot be done until handle->tmpbuf is unused :-( so barrier after the op */
if (firstred && MPI_IN_PLACE != sendbuf) {
/* perform the reduce with the senbuf */
res = NBC_Sched_op (recvbuf, false, sendbuf, false, 0, true, count, datatype, op, schedule, true);
firstred = 0;
} else {
/* perform the reduce in my local buffer */
res = NBC_Sched_op (recvbuf, false, recvbuf, false, 0, true, count, datatype, op, schedule, true);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
} else {
static inline int allred_sched_diss(int rank, int p, int count, MPI_Datatype datatype, void *sendbuf, void *recvbuf, MPI_Op op, NBC_Schedule *schedule, NBC_Handle *handle) {
int root, vrank, r, maxr, firstred, vpeer, peer, res;
root = 0; /* this makes the code for ireduce and iallreduce nearly identical - could be changed to improve performance */
RANK2VRANK(rank, vrank, root);
maxr = (int)ceil((log(p)/LOG2));
firstred = 1;
for(r=1; r<=maxr; r++) {
if((vrank % (1<<r)) == 0) {
/* we have to receive this round */
vpeer = vrank + (1<<(r-1));
VRANK2RANK(peer, vpeer, root)
if(peer<p) {
res = NBC_Sched_recv(0, true, count, datatype, peer, schedule);
if(res != NBC_OK) { 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(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
if(firstred && MPI_IN_PLACE != sendbuf) {
/* perform the reduce with the senbuf */
res = NBC_Sched_op(recvbuf, false, sendbuf, false, 0, true, count, datatype, op, schedule);
firstred = 0;
} else {
/* perform the reduce in my local buffer */
res = NBC_Sched_op(recvbuf, false, recvbuf, false, 0, true, count, datatype, op, schedule);
}
if(res != NBC_OK) { 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(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
}
} else {
static inline int red_sched_binomial (int rank, int p, int root, const void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
MPI_Op op, void *redbuf, NBC_Schedule *schedule, NBC_Handle *handle) {
int vrank, vpeer, peer, res, maxr;
RANK2VRANK(rank, vrank, root);
maxr = (int)ceil((log((double)p)/LOG2));
for (int r = 1, firstred = 1 ; r <= maxr ; ++r) {
if ((vrank % (1 << r)) == 0) {
/* we have to receive this round */
vpeer = vrank + (1 << (r - 1));
VRANK2RANK(peer, vpeer, root)
if (peer < p) {
/* we have to wait until we have the data */
res = NBC_Sched_recv (0, true, count, datatype, peer, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
/* perform the reduce in my local buffer */
/* this cannot be done until handle->tmpbuf is unused :-( so barrier after the op */
if (firstred) {
if (rank == root) {
/* root is the only one who reduces in the receivebuffer
* take data from sendbuf in first round - save copy */
res = NBC_Sched_op (recvbuf, false, sendbuf, false, 0, true, count, datatype, op, schedule, true);
} else {
/* all others may not have a receive buffer
* take data from sendbuf in first round - save copy */
res = NBC_Sched_op ((char *) redbuf - (intptr_t) handle->tmpbuf, true, sendbuf, false, 0, true, count,
datatype, op, schedule, true);
}
firstred = 0;
} else {
if(rank == root) {
/* root is the only one who reduces in the receivebuffer */
res = NBC_Sched_op (recvbuf, false, recvbuf, false, 0, true, count, datatype, op, schedule, true);
} else {
/* all others may not have a receive buffer */
res = NBC_Sched_op ((char *) redbuf - (intptr_t) handle->tmpbuf, true, (char *) redbuf - (intptr_t) handle->tmpbuf,
true, 0, true, count, datatype, op, schedule, true);
}
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
} else {
static inline int red_sched_binomial(int rank, int p, int root, void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, void *redbuf, NBC_Schedule *schedule, NBC_Handle *handle) {
int firstred, vrank, vpeer, peer, res, maxr, r;
RANK2VRANK(rank, vrank, root);
maxr = (int)ceil((log((double)p)/LOG2));
firstred = 1;
for(r=1; r<=maxr; r++) {
if((vrank % (1<<r)) == 0) {
/* we have to receive this round */
vpeer = vrank + (1<<(r-1));
VRANK2RANK(peer, vpeer, root)
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; }
/* perform the reduce in my local buffer */
if(firstred) {
if(rank == root) {
/* root is the only one who reduces in the receivebuffer
* take data from sendbuf in first round - save copy */
res = NBC_Sched_op(recvbuf, false, sendbuf, false, 0, true, count, datatype, op, schedule);
} else {
/* all others may not have a receive buffer
* take data from sendbuf in first round - save copy */
res = NBC_Sched_op((char *)redbuf-(unsigned long)handle->tmpbuf, true, sendbuf, false, 0, true, count, datatype, op, schedule);
}
firstred = 0;
} else {
if(rank == root) {
/* root is the only one who reduces in the receivebuffer */
res = NBC_Sched_op(recvbuf, false, recvbuf, false, 0, true, count, datatype, op, schedule);
} else {
/* all others may not have a receive buffer */
res = NBC_Sched_op((char *)redbuf-(unsigned long)handle->tmpbuf, true, (char *)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 {
/* linear iscan
* working principle:
* 1. each node (but node 0) receives from left neighbor
* 2. performs op
* 3. all but rank p-1 do sends to it's right neighbor and exits
*
*/
int ompi_coll_libnbc_iscan(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;
ptrdiff_t gap, span;
NBC_Schedule *schedule;
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);
if (!inplace) {
/* copy data to receivebuf */
res = NBC_Copy (sendbuf, count, datatype, recvbuf, count, datatype, comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
res = NBC_Init_handle(comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
#ifdef NBC_CACHE_SCHEDULE
NBC_Scan_args *args, *found, search;
/* 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_SCAN], &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;
}
/* ensure the schedule is released with the handle */
handle->schedule = schedule;
if(rank != 0) {
span = opal_datatype_span(&datatype->super, count, &gap);
handle->tmpbuf = malloc (span);
if (NULL == handle->tmpbuf) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* we have to wait until we have the data */
res = NBC_Sched_recv ((void *)(-gap), true, count, datatype, rank-1, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
/* perform the reduce in my local buffer */
/* this cannot be done until handle->tmpbuf is unused :-( so barrier after the op */
res = NBC_Sched_op ((void *)(-gap), true, recvbuf, false, count, datatype, op, schedule,
true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
}
if (rank != p-1) {
res = NBC_Sched_send (recvbuf, false, count, datatype, rank+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_SCAN], args, args, 0);
if (0 == res) {
OBJ_RETAIN(schedule);
/* increase number of elements for A2A */
if (++libnbc_module->NBC_Dict_size[NBC_SCAN] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_SCAN],
&libnbc_module->NBC_Dict_size[NBC_SCAN]);
}
} 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_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;
}
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;
}
/* 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_1_0_t *module) {
int rank, p, res;
MPI_Aint ext;
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 = ompi_datatype_type_extent(datatype, &ext);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res);
return res;
}
res = NBC_Init_handle(comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
if (inplace && rank < p - 1) {
/* need more buffer space for the inplace case */
handle->tmpbuf = malloc(ext * count * 2);
} else {
handle->tmpbuf = malloc(ext * count);
}
if (handle->tmpbuf == NULL) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
#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) {
if (inplace && rank < p - 1) {
/* if sendbuf == recvbuf do not clobber the send buffer until it has been combined
* with the incoming data. */
res = NBC_Sched_recv ((void *) (ext * count), true, count, datatype, rank-1, schedule, false);
} else {
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;
}
/* perform the reduce in my temporary buffer */
/* this cannot be done until handle->tmpbuf is unused :-( so barrier after */
if (inplace) {
res = NBC_Sched_op (0, true, sendbuf, false, (void *)(ext * count), true, count,
datatype, op, schedule, true);
} else {
res = NBC_Sched_op (0, true, sendbuf, false, recvbuf, false, 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 (0, true, count, datatype, rank + 1, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
if (inplace) {
/* copy the received data into the receive buffer */
res = NBC_Sched_copy ((void *)(ext * count), true, count, datatype, recvbuf,
false, count, datatype, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
return res;
}
}
}
} else if (p > 1) {
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;
}
/* 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;
void *tmpbuf = NULL;
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);
span = opal_datatype_span(&datatype->super, count, &gap);
if (0 < rank) {
tmpbuf = malloc(span);
if (NULL == tmpbuf) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (inplace) {
res = NBC_Copy(recvbuf, count, datatype, (char *)tmpbuf-gap, count, datatype, comm);
} else {
res = NBC_Copy(sendbuf, count, datatype, (char *)tmpbuf-gap, count, datatype, comm);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
free(tmpbuf);
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)) {
free(tmpbuf);
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (rank != 0) {
res = NBC_Sched_recv (recvbuf, false, count, datatype, rank-1, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
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)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
return res;
}
res = NBC_Sched_op (recvbuf, false, (void *)(-gap), true, count,
datatype, op, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
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)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
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)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
return res;
}
}
res = NBC_Sched_commit(schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
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_Schedule_request(schedule, comm, libnbc_module, request, tmpbuf);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
return res;
}
return OMPI_SUCCESS;
}