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((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(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 bcast_sched_binomial(int rank, int p, int root, NBC_Schedule *schedule, void *buffer, int count, MPI_Datatype datatype) {
int maxr, vrank, peer, r, res;
maxr = (int)ceil((log((double)p)/LOG2));
RANK2VRANK(rank, vrank, root);
/* receive from the right hosts */
if(vrank != 0) {
for(r=0; r<maxr; r++) {
if((vrank >= (1<<r)) && (vrank < (1<<(r+1)))) {
VRANK2RANK(peer, vrank-(1<<r), root);
res = NBC_Sched_recv(buffer, false, count, datatype, peer, schedule);
if (NBC_OK != res) { 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; }
}
/* now send to the right hosts */
for(r=0; r<maxr; r++) {
if(((vrank + (1<<r) < p) && (vrank < (1<<r))) || (vrank == 0)) {
VRANK2RANK(peer, vrank+(1<<r), root);
res = NBC_Sched_send(buffer, false, count, datatype, peer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
}
return NBC_OK;
}
/* simple chained MPI_Ibcast */
static inline int bcast_sched_chain(int rank, int p, int root, NBC_Schedule *schedule, void *buffer, int count, MPI_Datatype datatype, int fragsize, size_t size) {
int res, vrank, rpeer, speer, numfrag, fragcount, thiscount;
MPI_Aint ext;
char *buf;
RANK2VRANK(rank, vrank, root);
VRANK2RANK(rpeer, vrank-1, root);
VRANK2RANK(speer, vrank+1, root);
res = ompi_datatype_type_extent(datatype, &ext);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res);
return res;
}
if (count == 0) {
return OMPI_SUCCESS;
}
numfrag = count * size/fragsize;
if ((count * size) % fragsize != 0) {
numfrag++;
}
fragcount = count/numfrag;
for (int fragnum = 0 ; fragnum < numfrag ; ++fragnum) {
buf = (char *) buffer + fragnum * fragcount * ext;
thiscount = fragcount;
if (fragnum == numfrag-1) {
/* last fragment may not be full */
thiscount = count - fragcount * fragnum;
}
/* root does not receive */
if (vrank != 0) {
res = NBC_Sched_recv (buf, false, thiscount, datatype, rpeer, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
/* last rank does not send */
if (vrank != p-1) {
res = NBC_Sched_send (buf, false, thiscount, datatype, speer, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
/* this barrier here seems awaward but isn't!!!! */
if (vrank == 0) {
res = NBC_Sched_barrier (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
}
}
return OMPI_SUCCESS;
}
static inline int a2a_sched_pairwise(int rank, int p, MPI_Aint sndext, MPI_Aint rcvext, NBC_Schedule* schedule, void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
int res, r, sndpeer, rcvpeer;
char *rbuf, *sbuf;
res = NBC_OK;
if(p < 2) return res;
for(r=1;r<p;r++) {
sndpeer = (rank+r)%p;
rcvpeer = (rank-r+p)%p;
rbuf = ((char *) recvbuf) + (rcvpeer*recvcount*rcvext);
res = NBC_Sched_recv(rbuf, false, recvcount, recvtype, rcvpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
sbuf = ((char *) sendbuf) + (sndpeer*sendcount*sndext);
res = NBC_Sched_send(sbuf, false, sendcount, sendtype, sndpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
if (r < p) {
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barr() (%i)\n", res); return res; }
}
}
return res;
}
int ompi_coll_libnbc_ibarrier_inter(struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, res, rsize, peer;
NBC_Schedule *schedule;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_remote_size(comm, &rsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_remote_size() (%i)\n", res); return res; }
handle->tmpbuf=(void*)malloc(2*sizeof(char));
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
if (0 == rank) {
for (peer = 1 ; peer < rsize ; ++peer) {
res = NBC_Sched_recv (0, true, 1, MPI_BYTE, peer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
/* synchronize with the remote root */
res = NBC_Sched_recv (0, true, 1, MPI_BYTE, 0, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_send (0, true, 1, MPI_BYTE, 0, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
if (0 == rank) {
/* wait for the remote root */
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
/* inform remote peers that all local peers have entered the barrier */
for (peer = 0 ; peer < rsize ; ++peer) {
res = NBC_Sched_send (0, true, 1, MPI_BYTE, peer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
}
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
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 {
/* simple chained MPI_Ibcast */
static inline int bcast_sched_chain(int rank, int p, int root, NBC_Schedule *schedule, void *buffer, int count, MPI_Datatype datatype, int fragsize, int size) {
int res, vrank, rpeer, speer, numfrag, fragnum, fragcount, thiscount;
MPI_Aint ext;
char *buf;
RANK2VRANK(rank, vrank, root);
VRANK2RANK(rpeer, vrank-1, root);
VRANK2RANK(speer, vrank+1, root);
res = MPI_Type_extent(datatype, &ext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
if(count == 0) return NBC_OK;
numfrag = count*size/fragsize;
if((count*size)%fragsize != 0) numfrag++;
fragcount = count/numfrag;
/*if(!rank) printf("numfrag: %i, count: %i, size: %i, fragcount: %i\n", numfrag, count, size, fragcount);*/
for(fragnum = 0; fragnum < numfrag; fragnum++) {
buf = (char*)buffer+fragnum*fragcount*ext;
thiscount = fragcount;
if(fragnum == numfrag-1) {
/* last fragment may not be full */
thiscount = count-fragcount*fragnum;
}
/* root does not receive */
if(vrank != 0) {
res = NBC_Sched_recv(buf, false, thiscount, datatype, rpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
}
/* last rank does not send */
if(vrank != p-1) {
res = NBC_Sched_send(buf, false, thiscount, datatype, speer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
/* this barrier here seems awaward but isn't!!!! */
if(vrank == 0) res = NBC_Sched_barrier(schedule);
}
}
return NBC_OK;
}
static inline int bcast_sched_binomial(int rank, int p, int root, NBC_Schedule *schedule, void *buffer, int count, MPI_Datatype datatype) {
int maxr, vrank, peer, res;
maxr = (int)ceil((log((double)p)/LOG2));
RANK2VRANK(rank, vrank, root);
/* receive from the right hosts */
if (vrank != 0) {
for (int r = 0 ; r < maxr ; ++r) {
if ((vrank >= (1 << r)) && (vrank < (1 << (r + 1)))) {
VRANK2RANK(peer, vrank - (1 << r), root);
res = NBC_Sched_recv (buffer, false, count, datatype, peer, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
}
res = NBC_Sched_barrier (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
/* now send to the right hosts */
for (int r = 0 ; r < maxr ; ++r) {
if (((vrank + (1 << r) < p) && (vrank < (1 << r))) || (vrank == 0)) {
VRANK2RANK(peer, vrank + (1 << r), root);
res = NBC_Sched_send (buffer, false, count, datatype, peer, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
}
return OMPI_SUCCESS;
}
/* Dissemination implementation of MPI_Ibarrier */
static int nbc_barrier_init(struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_3_0_t *module, bool persistent)
{
int rank, p, maxround, res, recvpeer, sendpeer;
NBC_Schedule *schedule;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
#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)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
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)) {
OBJ_RELEASE(schedule);
return res;
}
/* recv msg from recvpeer */
res = NBC_Sched_recv (NULL, false, 0, MPI_BYTE, recvpeer, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
/* end communication round */
if (round < maxround) {
res = NBC_Sched_barrier (schedule);
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
/* add it */
libnbc_module->NBC_Dict[NBC_BARRIER] = (hb_tree *) schedule;
libnbc_module->NBC_Dict_size[NBC_BARRIER] = 1;
}
OBJ_RETAIN(schedule);
#endif
res = NBC_Schedule_request(schedule, comm, libnbc_module, persistent, request, NULL);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
return OMPI_SUCCESS;
}
int ompi_coll_libnbc_ineighbor_allgatherv(void *sbuf, int scount, MPI_Datatype stype, void *rbuf,
int *rcounts, int *displs, MPI_Datatype rtype,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module) {
int rank, size, res, worldsize;
MPI_Aint sndext, rcvext;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
res = NBC_Init_handle(comm, coll_req, libnbc_module);
handle = *coll_req;
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
res = MPI_Comm_size(comm, &size);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Comm_size(MPI_COMM_WORLD, &worldsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Type_extent(stype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(rtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
char inplace;
NBC_Schedule *schedule;
#ifdef NBC_CACHE_SCHEDULE
NBC_Ineighbor_allgatherv_args *args, *found, search;
#endif
NBC_IN_PLACE(sbuf, rbuf, inplace);
handle->tmpbuf=NULL;
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sbuf=sbuf;
search.scount=scount;
search.stype=stype;
search.rbuf=rbuf;
search.rcount=rcount;
search.rtype=rtype;
found = (NBC_Ineighbor_allgatherv_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLGATHERV], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create, res = %i\n", res); return res; }
{
int indegree, outdegree, weighted, *srcs, *dsts, i;
res = NBC_Comm_neighbors_count(comm, &indegree, &outdegree, &weighted);
if(res != NBC_OK) return res;
srcs = (int*)malloc(sizeof(int)*indegree);
dsts = (int*)malloc(sizeof(int)*outdegree);
res = NBC_Comm_neighbors(comm, indegree, srcs, MPI_UNWEIGHTED, outdegree, dsts, MPI_UNWEIGHTED);
if(res != NBC_OK) return res;
if(inplace) { /* we need an extra buffer to be deadlock-free */
int sumrcounts=0;
int offset=0;
for(i=0; i<indegree; ++i) sumrcounts += rcounts[i];
handle->tmpbuf = malloc(rcvext*sumrcounts);
for(i = 0; i < indegree; i++) {
if(srcs[i] != MPI_PROC_NULL) {
res = NBC_Sched_recv((char*)0+offset, true, rcounts[i], rtype, srcs[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
offset += rcounts[i]*rcvext;
}
for(i = 0; i < outdegree; i++) {
if(dsts[i] != MPI_PROC_NULL) {
res = NBC_Sched_send((char*)sbuf, false, scount, stype, dsts[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
}
/* unpack from buffer */
offset=0;
for(i = 0; i < indegree; i++) {
if(srcs[i] != MPI_PROC_NULL) {
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
res = NBC_Sched_copy((char*)0+offset, true, rcounts[i], rtype, (char*)rbuf+displs[i]*rcvext, false, rcounts[i], rtype, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_copy() (%i)\n", res); return res; }
}
offset += rcounts[i]*rcvext;
}
} else { /* non INPLACE case */
/* simply loop over neighbors and post send/recv operations */
for(i = 0; i < indegree; i++) {
if(srcs[i] != MPI_PROC_NULL) {
res = NBC_Sched_recv((char*)rbuf+displs[i]*rcvext, false, rcounts[i], rtype, srcs[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
for(i = 0; i < outdegree; i++) {
if(dsts[i] != MPI_PROC_NULL) {
res = NBC_Sched_send((char*)sbuf, false, scount, stype, dsts[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
}
}
}
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Ineighbor_allgatherv_args*)malloc(sizeof(NBC_Ineighbor_allgatherv_args));
args->sbuf=sbuf;
args->scount=scount;
args->stype=stype;
args->rbuf=rbuf;
args->rcount=rcount;
args->rtype=rtype;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLGATHERV], args, args, 0);
if(res != 0) printf("error in dict_insert() (%i)\n", res);
/* increase number of elements for A2A */
if(++handle->comminfo->NBC_Dict_size[NBC_NEIGHBOR_ALLGATHERV] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLGATHERV], &handle->comminfo->NBC_Dict_size[NBC_NEIGHBOR_ALLGATHERV]);
}
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
/* 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;
}
static inline int a2a_sched_diss(int rank, int p, MPI_Aint sndext, MPI_Aint rcvext, NBC_Schedule* schedule, void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm, NBC_Handle *handle) {
int res, i, r, speer, rpeer, datasize, offset, virtp;
char *rbuf, *rtmpbuf, *stmpbuf;
res = NBC_OK;
if(p < 2) return res;
if(NBC_Type_intrinsic(sendtype)) {
datasize = sndext*sendcount;
} else {
res = MPI_Pack_size(sendcount, sendtype, comm, &datasize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack_size() (%i)\n", res); return res; }
}
/* allocate temporary buffers */
if(p % 2 == 0) {
rtmpbuf = (char*)handle->tmpbuf+datasize*p;
stmpbuf = (char*)handle->tmpbuf+datasize*(p+p/2);
} else {
/* we cannot divide p by two, so alloc more to be safe ... */
virtp = (p/2+1)*2;
rtmpbuf = (char*)handle->tmpbuf+datasize*p;
stmpbuf = (char*)handle->tmpbuf+datasize*(p+virtp/2);
}
/* phase 2 - communicate */
/*printf("[%i] temp buffer is at %lu of size %i, maxround: %i\n", rank, (unsigned long)handle->tmpbuf, (int)datasize*p*(1<<maxround), maxround);*/
for(r = 1; r < p; r<<=1) {
offset = 0;
for(i=1; i<p; i++) {
/* test if bit r is set in rank number i */
if((i&r) == r) {
/* copy data to sendbuffer (2nd copy) - could be avoided using iovecs */
/*printf("[%i] round %i: copying element %i to buffer %lu\n", rank, r, i, (unsigned long)(stmpbuf+offset));*/
NBC_Sched_copy((void*)(long)(i*datasize), true, datasize, MPI_BYTE, stmpbuf+offset-(unsigned long)handle->tmpbuf, true, datasize, MPI_BYTE, schedule);
offset += datasize;
}
}
speer = ( rank + r) % p;
/* add p because modulo does not work with negative values */
rpeer = ((rank - r)+p) % p;
/*printf("[%i] receiving %i bytes from host %i into rbuf %lu\n", rank, offset, rpeer, (unsigned long)rtmpbuf);*/
res = NBC_Sched_recv(rtmpbuf-(unsigned long)handle->tmpbuf, true, offset, MPI_BYTE, rpeer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
/*printf("[%i] sending %i bytes to host %i from sbuf %lu\n", rank, offset, speer, (unsigned long)stmpbuf);*/
res = NBC_Sched_send(stmpbuf-(unsigned long)handle->tmpbuf, true, offset, MPI_BYTE, speer, 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; }
/* unpack from buffer */
offset = 0;
for(i=1; i<p; i++) {
/* test if bit r is set in rank number i */
if((i&r) == r) {
/* copy data to tmpbuffer (3rd copy) - could be avoided using iovecs */
NBC_Sched_copy(rtmpbuf+offset-(unsigned long)handle->tmpbuf, true, datasize, MPI_BYTE, (void*)(long)(i*datasize), true, datasize, MPI_BYTE, schedule);
offset += datasize;
}
}
}
/* phase 3 - reorder - data is now in wrong order in handle->tmpbuf -
* reorder it into recvbuf */
for(i=0; i<p; i++) {
rbuf = (char*)recvbuf+((rank-i+p)%p)*recvcount*rcvext;
res = NBC_Sched_unpack((void*)(long)(i*datasize), true, recvcount, recvtype, rbuf, false, schedule);
if (NBC_OK != res) { printf("MPI Error in NBC_Sched_pack() (%i)\n", res); return res; }
}
return NBC_OK;
}
int ompi_coll_libnbc_ireduce_scatter_inter(void* sendbuf, void* recvbuf, int *recvcounts, MPI_Datatype datatype,
MPI_Op op, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int peer, rank, r, res, count, rsize, offset;
MPI_Aint ext;
NBC_Schedule *schedule;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_remote_size(comm, &rsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_remote_size() (%i)\n", res); return res; }
MPI_Type_extent(datatype, &ext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return NBC_OOR; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
count = 0;
for (r = 0 ; r < rsize ; ++r) count += recvcounts[r];
handle->tmpbuf = malloc(2 * ext * count);
if(handle->tmpbuf == NULL) { printf("Error in malloc()\n"); return NBC_OOR; }
/* send my data to the remote root */
res = NBC_Sched_send(sendbuf, false, count, datatype, 0, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
if (0 == rank) {
res = NBC_Sched_recv((void *) 0, true, count, datatype, 0, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
for (peer = 1 ; peer < rsize ; ++peer) {
res = NBC_Sched_recv((void *)(ext * count), true, count, datatype, peer, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
res = NBC_Sched_op((void *) 0, true, (void *)(ext * count), true, (void *) 0, true, count, datatype, op, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_op() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
}
/* exchange data with remote root for scatter phase (we *could* use the local communicator to do the scatter) */
res = NBC_Sched_recv((void *)(ext * count), true, count, datatype, 0, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
res = NBC_Sched_send((void *) 0, true, count, datatype, 0, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
/* scatter */
for (peer = 0, offset = ext * count ; peer < rsize ; ++peer) {
res = NBC_Sched_send((void *)(uintptr_t) offset, true, recvcounts[peer], datatype, peer, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
offset += recvcounts[peer] * ext;
}
}
/* receive my block */
res = NBC_Sched_recv(recvbuf, false, recvcounts[rank], datatype, 0, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
/* tmpbuf is freed with the handle */
return NBC_OK;
}
static int nbc_barrier_inter_init(struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_3_0_t *module, bool persistent)
{
int rank, res, rsize;
NBC_Schedule *schedule;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
rank = ompi_comm_rank (comm);
rsize = ompi_comm_remote_size (comm);
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
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)) {
OBJ_RELEASE(schedule);
return res;
}
}
}
/* synchronize with the remote root */
res = NBC_Sched_recv (NULL, false, 0, MPI_BYTE, 0, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Sched_send (NULL, false, 0, MPI_BYTE, 0, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
if (0 == rank) {
/* wait for the remote root */
res = NBC_Sched_barrier (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
/* 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)) {
OBJ_RELEASE(schedule);
return res;
}
}
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Schedule_request(schedule, comm, libnbc_module, persistent, request, NULL);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
return OMPI_SUCCESS;
}
int NBC_Ineighbor_alltoallw(void *sbuf, int *scounts, MPI_Datatype *stypes,
void *rbuf, int *rcounts, MPI_Datatype *rtypes, MPI_Comm comm, NBC_Handle* handle) {
int rank, size, res, worldsize, i;
MPI_Aint *sndexts, *rcvexts;
double t[10];
t[0] = PMPI_Wtime();
res = NBC_Init_handle(handle, comm);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
res = MPI_Comm_size(comm, &size);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Comm_size(MPI_COMM_WORLD, &worldsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
t[1] = PMPI_Wtime();
char inplace;
NBC_Schedule *schedule;
#ifdef NBC_CACHE_SCHEDULE
NBC_Ineighbor_alltoallw_args *args, *found, search;
#endif
NBC_IN_PLACE(sbuf, rbuf, inplace);
handle->tmpbuf=NULL;
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sbuf=sbuf;
search.scount=scount;
search.stype=stype;
search.rbuf=rbuf;
search.rcount=rcount;
search.rtype=rtype;
found = (NBC_Ineighbor_alltoallw_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLTOALLW], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create, res = %i\n", res); return res; }
{
int indegree, outdegree, weighted, *srcs, *dsts, i;
res = NBC_Comm_neighbors_count(comm, &indegree, &outdegree, &weighted);
if(res != NBC_OK) return res;
srcs = (int*)malloc(sizeof(int)*indegree);
dsts = (int*)malloc(sizeof(int)*outdegree);
sndexts = (MPI_Aint*)malloc(sizeof(MPI_Aint)*outdegree);
for(i=0; i<outdegree; ++i) {
res = MPI_Type_extent(stypes[i], &sndexts[i]);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
}
rcvexts = (MPI_Aint*)malloc(sizeof(MPI_Aint)*indegree);
for(i=0; i<indegree; ++i) {
res = MPI_Type_extent(rtypes[i], &rcvexts[i]);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
}
res = NBC_Comm_neighbors(comm, indegree, srcs, MPI_UNWEIGHTED, outdegree, dsts, MPI_UNWEIGHTED);
if(res != NBC_OK) return res;
if(inplace) { /* we need an extra buffer to be deadlock-free */
int sumrbytes=0;
int offset=0;
for(i=0; i<indegree; ++i) sumrbytes += rcounts[i]*rcvexts[i];
handle->tmpbuf = malloc(sumrbytes);
for(i = 0; i < indegree; i++) {
res = NBC_Sched_recv((char*)0+offset, true, rcounts[i], rtypes[i], srcs[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
offset += rcounts[i]*rcvexts[i];
}
offset=0;
for(i = 0; i < outdegree; i++) {
res = NBC_Sched_send((char*)sbuf+offset, false, scounts[i], stypes[i], dsts[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
offset += scounts[i]*sndexts[i];
}
/* unpack from buffer */
offset=0;
for(i = 0; i < indegree; i++) {
res = NBC_Sched_barrier(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_barrier() (%i)\n", res); return res; }
res = NBC_Sched_copy((char*)0+offset, true, rcounts[i], rtypes[i], (char*)rbuf+offset, false, rcounts[i], rtypes[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_copy() (%i)\n", res); return res; }
offset += rcounts[i]*rcvexts[i];
}
} else { /* non INPLACE case */
/* simply loop over neighbors and post send/recv operations */
int offset=0;
for(i = 0; i < indegree; i++) {
res = NBC_Sched_recv((char*)rbuf+offset, false, rcounts[i], rtypes[i], srcs[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
offset += rcounts[i]*rcvexts[i];
}
offset = 0;
for(i = 0; i < outdegree; i++) {
res = NBC_Sched_send((char*)sbuf+offset, false, scounts[i], stypes[i], dsts[i], schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
offset += scounts[i]*sndexts[i];
}
}
}
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Ineighbor_alltoallw_args*)malloc(sizeof(NBC_Ineighbor_alltoallw_args));
args->sbuf=sbuf;
args->scount=scount;
args->stype=stype;
args->rbuf=rbuf;
args->rcount=rcount;
args->rtype=rtype;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLTOALLW], args, args, 0);
if(res != 0) printf("error in dict_insert() (%i)\n", res);
/* increase number of elements for A2A */
if(++handle->comminfo->NBC_Dict_size[NBC_NEIGHBOR_ALLTOALLW] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_NEIGHBOR_ALLTOALLW], &handle->comminfo->NBC_Dict_size[NBC_NEIGHBOR_ALLTOALLW]);
}
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int ompi_coll_libnbc_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;
}
/* 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;
}
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;
}
/* 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;
}
