/* the non-blocking reduce */
int ompi_coll_libnbc_ireduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype,
MPI_Op op, int root, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int rank, p, res, segsize, size;
MPI_Aint ext;
NBC_Schedule *schedule;
char *redbuf=NULL, inplace;
#ifdef NBC_CACHE_SCHEDULE
NBC_Reduce_args *args, *found, search;
#endif
enum { NBC_RED_BINOMIAL, NBC_RED_CHAIN } alg;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(datatype, &ext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_size(datatype, &size);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
/* only one node -> copy data */
if((p == 1) && !inplace) {
res = NBC_Copy(sendbuf, count, datatype, recvbuf, count, datatype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
/* algorithm selection */
if(p > 4 || size*count < 65536) {
alg = NBC_RED_BINOMIAL;
if(rank == root) {
/* root reduces in receivebuffer */
handle->tmpbuf = malloc(ext*count);
} else {
/* recvbuf may not be valid on non-root nodes */
handle->tmpbuf = malloc(ext*count*2);
redbuf = ((char*)handle->tmpbuf)+(ext*count);
}
} else {
handle->tmpbuf = malloc(ext*count);
alg = NBC_RED_CHAIN;
segsize = 16384/2;
}
if (NULL == handle->tmpbuf) { printf("Error in malloc() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf=sendbuf;
search.recvbuf=recvbuf;
search.count=count;
search.datatype=datatype;
search.op=op;
search.root=root;
found = (NBC_Reduce_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_REDUCE], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
switch(alg) {
case NBC_RED_BINOMIAL:
res = red_sched_binomial(rank, p, root, sendbuf, recvbuf, count, datatype, op, redbuf, schedule, handle);
break;
case NBC_RED_CHAIN:
res = red_sched_chain(rank, p, root, sendbuf, recvbuf, count, datatype, op, ext, size, schedule, handle, segsize);
break;
}
if (NBC_OK != res) { printf("Error in Schedule creation() (%i)\n", res); return res; }
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Reduce_args*)malloc(sizeof(NBC_Alltoall_args));
args->sendbuf=sendbuf;
args->recvbuf=recvbuf;
args->count=count;
args->datatype=datatype;
args->op=op;
args->root=root;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_REDUCE], args, args, 0);
if(res != 0) printf("error in dict_insert() (%i)\n", res);
/* increase number of elements for Reduce */
if(++handle->comminfo->NBC_Dict_size[NBC_REDUCE] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_REDUCE], &handle->comminfo->NBC_Dict_size[NBC_REDUCE]);
}
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
/* tmpbuf is freed with the handle */
return NBC_OK;
}
/* simple linear MPI_Ialltoall the (simple) algorithm just sends to all nodes */
int ompi_coll_libnbc_ialltoall(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, p, res, a2asize, sndsize, datasize;
NBC_Schedule *schedule;
MPI_Aint rcvext, sndext;
#ifdef NBC_CACHE_SCHEDULE
NBC_Alltoall_args *args, *found, search;
#endif
char *rbuf, *sbuf, inplace;
enum {NBC_A2A_LINEAR, NBC_A2A_PAIRWISE, NBC_A2A_DISS} alg;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_size(sendtype, &sndsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
/* algorithm selection */
a2asize = sndsize*sendcount*p;
/* this number is optimized for TCP on odin.cs.indiana.edu */
if((p <= 8) && ((a2asize < 1<<17) || (sndsize*sendcount < 1<<12))) {
/* just send as fast as we can if we have less than 8 peers, if the
* total communicated size is smaller than 1<<17 *and* if we don't
* have eager messages (msgsize < 1<<13) */
alg = NBC_A2A_LINEAR;
} else if(a2asize < (1<<12)*p) {
/*alg = NBC_A2A_DISS;*/
alg = NBC_A2A_LINEAR;
} else
alg = NBC_A2A_LINEAR; /*NBC_A2A_PAIRWISE;*/
if(!inplace) {
/* copy my data to receive buffer */
rbuf = ((char *)recvbuf) + (rank*recvcount*rcvext);
sbuf = ((char *)sendbuf) + (rank*sendcount*sndext);
res = NBC_Copy(sbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
/* allocate temp buffer if we need one */
if(alg == NBC_A2A_DISS) {
/* only A2A_DISS needs buffers */
if(NBC_Type_intrinsic(sendtype)) {
datasize = sndext*sendcount;
} else {
res = MPI_Pack_size(sendcount, sendtype, comm, &datasize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack_size() (%i)\n", res); return res; }
}
/* allocate temporary buffers */
if(p % 2 == 0) {
handle->tmpbuf=malloc(datasize*p*2);
} else {
/* we cannot divide p by two, so alloc more to be safe ... */
handle->tmpbuf=malloc(datasize*(p/2+1)*2*2);
}
/* phase 1 - rotate n data blocks upwards into the tmpbuffer */
#if OPAL_CUDA_SUPPORT
if(NBC_Type_intrinsic(sendtype) && !(opal_cuda_check_bufs((char *)sendbuf, (char *)recvbuf))) {
#else
if(NBC_Type_intrinsic(sendtype)) {
#endif /* OPAL_CUDA_SUPPORT */
/* contiguous - just copy (1st copy) */
memcpy(handle->tmpbuf, (char*)sendbuf+datasize*rank, datasize*(p-rank));
if(rank != 0) memcpy((char*)handle->tmpbuf+datasize*(p-rank), sendbuf, datasize*(rank));
} else {
int pos=0;
/* non-contiguous - pack */
res = MPI_Pack((char*)sendbuf+rank*sendcount*sndext, (p-rank)*sendcount, sendtype, handle->tmpbuf, (p-rank)*datasize, &pos, comm);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack() (%i)\n", res); return res; }
if(rank != 0) {
pos = 0;
MPI_Pack(sendbuf, rank*sendcount, sendtype, (char*)handle->tmpbuf+datasize*(p-rank), rank*datasize, &pos, comm);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Pack() (%i)\n", res); return res; }
}
}
} else {
handle->tmpbuf=NULL;
}
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf=sendbuf;
search.sendcount=sendcount;
search.sendtype=sendtype;
search.recvbuf=recvbuf;
search.recvcount=recvcount;
search.recvtype=recvtype;
found = (NBC_Alltoall_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLTOALL], &search);
if(found == NULL) {
#endif
/* not found - generate new schedule */
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
switch(alg) {
case NBC_A2A_LINEAR:
res = a2a_sched_linear(rank, p, sndext, rcvext, schedule, sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
break;
case NBC_A2A_DISS:
res = a2a_sched_diss(rank, p, sndext, rcvext, schedule, sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm, handle);
break;
case NBC_A2A_PAIRWISE:
res = a2a_sched_pairwise(rank, p, sndext, rcvext, schedule, sendbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
break;
}
if (NBC_OK != res) { return res; }
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Alltoall_args*)malloc(sizeof(NBC_Alltoall_args));
args->sendbuf=sendbuf;
args->sendcount=sendcount;
args->sendtype=sendtype;
args->recvbuf=recvbuf;
args->recvcount=recvcount;
args->recvtype=recvtype;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLTOALL], args, args, 0);
if(res != 0) printf("error in dict_insert() (%i)\n", res);
/* increase number of elements for A2A */
if(++handle->comminfo->NBC_Dict_size[NBC_ALLTOALL] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLTOALL], &handle->comminfo->NBC_Dict_size[NBC_ALLTOALL]);
/*if(!rank) printf("[%i] removing %i elements - new size: %i \n", rank, SCHED_DICT_UPPER-SCHED_DICT_LOWER, handle->comminfo->NBC_Alltoall_size);*/
}
/*if(!rank) printf("[%i] added new schedule to tree - number %i\n", rank, handle->comminfo->NBC_Dict_size[NBC_ALLTOALL]);*/
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int ompi_coll_libnbc_ialltoall_inter (void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, res, i, rsize;
MPI_Aint sndext, rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_remote_size (comm, &rsize);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_remote_size() (%i)\n", res); return res; }
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
for (i = 0; i < rsize; i++) {
/* post all sends */
sbuf = ((char *) sendbuf) + (i * sendcount * sndext);
res = NBC_Sched_send(sbuf, false, sendcount, sendtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
/* post all receives */
rbuf = ((char *) recvbuf) + (i * recvcount * rcvext);
res = NBC_Sched_recv(rbuf, false, recvcount, recvtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
static inline int NBC_Start_round(NBC_Handle *handle) {
int num; /* number of operations */
int i, res, ret=NBC_OK;
char* ptr;
NBC_Fn_type type;
NBC_Args_send sendargs;
NBC_Args_recv recvargs;
NBC_Args_op opargs;
NBC_Args_copy copyargs;
NBC_Args_unpack unpackargs;
NBC_Schedule myschedule;
void *buf1, *buf2, *buf3;
/* get round-schedule address */
myschedule = (NBC_Schedule*)((char*)*handle->schedule + handle->row_offset);
ptr = (char*) myschedule;
NBC_GET_BYTES(ptr,num);
NBC_DEBUG(10, "start_round round at address %p : posting %i operations\n", myschedule, num);
for (i=0; i<num; i++) {
NBC_GET_BYTES(ptr,type);
switch(type) {
case SEND:
NBC_DEBUG(5," SEND (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,sendargs);
NBC_DEBUG(5,"*buf: %p, count: %i, type: %lu, dest: %i, tag: %i)\n", sendargs.buf, sendargs.count, (unsigned long)sendargs.datatype, sendargs.dest, handle->tag);
/* get an additional request */
handle->req_count++;
/* get buffer */
if(sendargs.tmpbuf) {
buf1=(char*)handle->tmpbuf+(long)sendargs.buf;
} else {
buf1=sendargs.buf;
}
#ifdef NBC_TIMING
Isend_time -= MPI_Wtime();
#endif
handle->req_array = (MPI_Request*)realloc((void*)handle->req_array, (handle->req_count)*sizeof(MPI_Request));
NBC_CHECK_NULL(handle->req_array);
res = MCA_PML_CALL(isend(buf1, sendargs.count, sendargs.datatype, sendargs.dest, handle->tag, MCA_PML_BASE_SEND_STANDARD, handle->comm, handle->req_array+handle->req_count-1));
if(OMPI_SUCCESS != res) { printf("Error in MPI_Isend(%lu, %i, %lu, %i, %i, %lu) (%i)\n", (unsigned long)buf1, sendargs.count, (unsigned long)sendargs.datatype, sendargs.dest, handle->tag, (unsigned long)handle->comm, res); ret=res; goto error; }
#ifdef NBC_TIMING
Isend_time += MPI_Wtime();
#endif
break;
case RECV:
NBC_DEBUG(5, " RECV (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,recvargs);
NBC_DEBUG(5, "*buf: %p, count: %i, type: %lu, source: %i, tag: %i)\n", recvargs.buf, recvargs.count, (unsigned long)recvargs.datatype, recvargs.source, handle->tag);
/* get an additional request - TODO: req_count NOT thread safe */
handle->req_count++;
/* get buffer */
if(recvargs.tmpbuf) {
buf1=(char*)handle->tmpbuf+(long)recvargs.buf;
} else {
buf1=recvargs.buf;
}
#ifdef NBC_TIMING
Irecv_time -= MPI_Wtime();
#endif
handle->req_array = (MPI_Request*)realloc((void*)handle->req_array, (handle->req_count)*sizeof(MPI_Request));
NBC_CHECK_NULL(handle->req_array);
res = MCA_PML_CALL(irecv(buf1, recvargs.count, recvargs.datatype, recvargs.source, handle->tag, handle->comm, handle->req_array+handle->req_count-1));
if(OMPI_SUCCESS != res) { printf("Error in MPI_Irecv(%lu, %i, %lu, %i, %i, %lu) (%i)\n", (unsigned long)buf1, recvargs.count, (unsigned long)recvargs.datatype, recvargs.source, handle->tag, (unsigned long)handle->comm, res); ret=res; goto error; }
#ifdef NBC_TIMING
Irecv_time += MPI_Wtime();
#endif
break;
case OP:
NBC_DEBUG(5, " OP (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,opargs);
NBC_DEBUG(5, "*buf1: %p, buf2: %p, buf3: %p, count: %i, type: %lu)\n", opargs.buf1, opargs.buf2, opargs.buf3, opargs.count, (unsigned long)opargs.datatype);
/* get buffers */
if(opargs.tmpbuf1) {
buf1=(char*)handle->tmpbuf+(long)opargs.buf1;
} else {
buf1=opargs.buf1;
}
if(opargs.tmpbuf2) {
buf2=(char*)handle->tmpbuf+(long)opargs.buf2;
} else {
buf2=opargs.buf2;
}
if(opargs.tmpbuf3) {
buf3=(char*)handle->tmpbuf+(long)opargs.buf3;
} else {
buf3=opargs.buf3;
}
ompi_3buff_op_reduce(opargs.op, buf1, buf2, buf3, opargs.count, opargs.datatype);
break;
case COPY:
NBC_DEBUG(5, " COPY (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,copyargs);
NBC_DEBUG(5, "*src: %lu, srccount: %i, srctype: %lu, *tgt: %lu, tgtcount: %i, tgttype: %lu)\n", (unsigned long)copyargs.src, copyargs.srccount, (unsigned long)copyargs.srctype, (unsigned long)copyargs.tgt, copyargs.tgtcount, (unsigned long)copyargs.tgttype);
/* get buffers */
if(copyargs.tmpsrc) {
buf1=(char*)handle->tmpbuf+(long)copyargs.src;
} else {
buf1=copyargs.src;
}
if(copyargs.tmptgt) {
buf2=(char*)handle->tmpbuf+(long)copyargs.tgt;
} else {
buf2=copyargs.tgt;
}
res = NBC_Copy(buf1, copyargs.srccount, copyargs.srctype, buf2, copyargs.tgtcount, copyargs.tgttype, handle->comm);
if(res != NBC_OK) { printf("NBC_Copy() failed (code: %i)\n", res); ret=res; goto error; }
break;
case UNPACK:
NBC_DEBUG(5, " UNPACK (offset %li) ", (long)ptr-(long)myschedule);
NBC_GET_BYTES(ptr,unpackargs);
NBC_DEBUG(5, "*src: %lu, srccount: %i, srctype: %lu, *tgt: %lu\n", (unsigned long)unpackargs.inbuf, unpackargs.count, (unsigned long)unpackargs.datatype, (unsigned long)unpackargs.outbuf);
/* get buffers */
if(unpackargs.tmpinbuf) {
buf1=(char*)handle->tmpbuf+(long)unpackargs.inbuf;
} else {
buf1=unpackargs.outbuf;
}
if(unpackargs.tmpoutbuf) {
buf2=(char*)handle->tmpbuf+(long)unpackargs.outbuf;
} else {
buf2=unpackargs.outbuf;
}
res = NBC_Unpack(buf1, unpackargs.count, unpackargs.datatype, buf2, handle->comm);
if(res != NBC_OK) { printf("NBC_Unpack() failed (code: %i)\n", res); ret=res; goto error; }
break;
default:
printf("NBC_Start_round: bad type %li at offset %li\n", (long)type, (long)ptr-(long)myschedule);
ret=NBC_BAD_SCHED;
goto error;
}
}
/* check if we can make progress - not in the first round, this allows us to leave the
* initialization faster and to reach more overlap
*
* threaded case: calling progress in the first round can lead to a
* deadlock if NBC_Free is called in this round :-( */
if(handle->row_offset != sizeof(int)) {
res = NBC_Progress(handle);
if((NBC_OK != res) && (NBC_CONTINUE != res)) { printf("Error in NBC_Progress() (%i)\n", res); ret=res; goto error; }
}
error:
return ret;
}
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;
}
static int nbc_allgather_init(const void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_3_0_t *module, bool persistent)
{
int rank, p, res;
MPI_Aint rcvext;
NBC_Schedule *schedule;
char *rbuf, inplace;
#ifdef NBC_CACHE_SCHEDULE
NBC_Allgather_args *args, *found, search;
#endif
enum { NBC_ALLGATHER_LINEAR, NBC_ALLGATHER_RDBL} alg;
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);
int is_commsize_pow2 = !(p & (p - 1));
if (libnbc_iallgather_algorithm == 0) {
alg = NBC_ALLGATHER_LINEAR;
} else {
/* user forced dynamic decision */
if (libnbc_iallgather_algorithm == 1) {
alg = NBC_ALLGATHER_LINEAR;
} else if (libnbc_iallgather_algorithm == 2 && is_commsize_pow2) {
alg = NBC_ALLGATHER_RDBL;
} else {
alg = NBC_ALLGATHER_LINEAR;
}
}
res = ompi_datatype_type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) {
return res;
}
if (inplace) {
sendtype = recvtype;
sendcount = recvcount;
} else if (!persistent) { /* for persistent, the copy must be scheduled */
/* copy my data to receive buffer */
rbuf = (char *) recvbuf + rank * recvcount * rcvext;
res = NBC_Copy (sendbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
if (1 == p && (!persistent || inplace)) {
return nbc_get_noop_request(persistent, request);
}
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf = sendbuf;
search.sendcount = sendcount;
search.sendtype = sendtype;
search.recvbuf = recvbuf;
search.recvcount = recvcount;
search.recvtype = recvtype;
found = (NBC_Allgather_args *) hb_tree_search ((hb_tree*)libnbc_module->NBC_Dict[NBC_ALLGATHER], &search);
if (NULL == found) {
#endif
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (persistent && !inplace) {
/* for nonblocking, data has been copied already */
/* copy my data to receive buffer (= send buffer of NBC_Sched_send) */
rbuf = (char *)recvbuf + rank * recvcount * rcvext;
res = NBC_Sched_copy((void *)sendbuf, false, sendcount, sendtype,
rbuf, false, recvcount, recvtype, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
}
switch (alg) {
case NBC_ALLGATHER_LINEAR:
res = allgather_sched_linear(rank, p, schedule, sendbuf, sendcount, sendtype,
recvbuf, recvcount, recvtype);
break;
case NBC_ALLGATHER_RDBL:
res = allgather_sched_recursivedoubling(rank, p, schedule, sendbuf, sendcount,
sendtype, recvbuf, recvcount, recvtype);
break;
}
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_Allgather_args *) malloc (sizeof (args));
args->sendbuf = sendbuf;
args->sendcount = sendcount;
args->sendtype = sendtype;
args->recvbuf = recvbuf;
args->recvcount = recvcount;
args->recvtype = recvtype;
args->schedule = schedule;
res = hb_tree_insert ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLGATHER], args, args, 0);
if (res != 0) {
free (args);
} else {
OBJ_RETAIN(schedule);
}
/* increase number of elements for A2A */
if (++libnbc_module->NBC_Dict_size[NBC_ALLGATHER] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLGATHER], &libnbc_module->NBC_Dict_size[NBC_ALLGATHER]);
}
} else {
/* found schedule */
schedule = found->schedule;
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_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;
}
/* 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;
}
/* simple linear MPI_Iallgather
* the algorithm uses p-1 rounds
* each node sends the packet it received last round (or has in round 0) to it's right neighbor (modulo p)
* each node receives from it's left (modulo p) neighbor */
int ompi_coll_libnbc_iallgather(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_1_0_t *module)
{
int rank, p, res;
MPI_Aint rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
#ifdef NBC_CACHE_SCHEDULE
NBC_Allgather_args *args, *found, search;
#endif
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 = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) {
return res;
}
if (inplace) {
sendtype = recvtype;
sendcount = recvcount;
} else {
/* copy my data to receive buffer */
rbuf = (char *) recvbuf + rank * recvcount * rcvext;
res = NBC_Copy (sendbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != 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;
found = (NBC_Allgather_args *) hb_tree_search ((hb_tree*)libnbc_module->NBC_Dict[NBC_ALLGATHER], &search);
if (NULL == found) {
#endif
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
sbuf = (char *)recvbuf + rank * recvcount * rcvext;
/* do p-1 rounds */
for(int r = 0 ; r < p ; ++r) {
if(r != rank) {
/* recv from rank r */
rbuf = (char *)recvbuf + r * recvcount * rcvext;
res = NBC_Sched_recv (rbuf, false, recvcount, recvtype, r, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
return res;
}
/* send to rank r - not from the sendbuf to optimize MPI_IN_PLACE */
res = NBC_Sched_send (sbuf, 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;
}
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Allgather_args *) malloc (sizeof (args));
args->sendbuf = sendbuf;
args->sendcount = sendcount;
args->sendtype = sendtype;
args->recvbuf = recvbuf;
args->recvcount = recvcount;
args->recvtype = recvtype;
args->schedule = schedule;
res = hb_tree_insert ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLGATHER], args, args, 0);
if (res != 0) {
free (args);
} else {
OBJ_RETAIN(schedule);
}
/* increase number of elements for A2A */
if (++libnbc_module->NBC_Dict_size[NBC_ALLGATHER] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_ALLGATHER], &libnbc_module->NBC_Dict_size[NBC_ALLGATHER]);
}
} 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)) {
OMPI_COLL_LIBNBC_REQUEST_RETURN(handle);
return res;
}
*request = (ompi_request_t *) handle;
return OMPI_SUCCESS;
}
/* simple linear Alltoallv */
int ompi_coll_libnbc_ialltoallv(const void* sendbuf, const int *sendcounts, const int *sdispls,
MPI_Datatype sendtype, void* recvbuf, const int *recvcounts, const int *rdispls,
MPI_Datatype recvtype, struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_2_0_t *module)
{
int rank, p, res;
MPI_Aint sndext, rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
ptrdiff_t gap, span;
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 (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;
}
/* copy data to receivbuffer */
if (inplace) {
int count = 0;
for (int i = 0; i < p; i++) {
if (recvcounts[i] > count) {
count = recvcounts[i];
}
}
span = opal_datatype_span(&recvtype->super, count, &gap);
if (OPAL_UNLIKELY(0 == span)) {
*request = &ompi_request_empty;
NBC_Return_handle (handle);
return MPI_SUCCESS;
}
handle->tmpbuf = malloc(span);
if (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
sendcounts = recvcounts;
sdispls = rdispls;
} else {
res = ompi_datatype_type_extent (sendtype, &sndext);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res);
NBC_Return_handle (handle);
return res;
}
if (sendcounts[rank] != 0) {
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 (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;
}
if (inplace) {
res = a2av_sched_inplace(rank, p, schedule, recvbuf, recvcounts,
rdispls, rcvext, recvtype, gap);
} else {
res = a2av_sched_linear(rank, p, schedule,
sendbuf, sendcounts, sdispls, sndext, sendtype,
recvbuf, recvcounts, rdispls, rcvext, recvtype);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Sched_commit (schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
OBJ_RELEASE(schedule);
return res;
}
res = NBC_Start(handle, schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (handle);
OBJ_RELEASE(schedule);
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(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;
}
/* simple linear MPI_Iscatter */
int ompi_coll_libnbc_iscatter(void* sendbuf, int sendcount, MPI_Datatype sendtype,
void* recvbuf, int recvcount, MPI_Datatype recvtype, int root,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int rank, p, res, i;
MPI_Aint sndext = 0;
NBC_Schedule *schedule;
char *sbuf, inplace;
#ifdef NBC_CACHE_SCHEDULE
NBC_Scatter_args *args, *found, search;
#endif
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
if (rank == root) {
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
}
handle->tmpbuf=NULL;
if((rank == root) && (!inplace)) {
sbuf = ((char *)sendbuf) + (rank*sendcount*sndext);
/* if I am the root - just copy the message (not for MPI_IN_PLACE) */
res = NBC_Copy(sbuf, sendcount, sendtype, recvbuf, recvcount, recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf=sendbuf;
search.sendcount=sendcount;
search.sendtype=sendtype;
search.recvbuf=recvbuf;
search.recvcount=recvcount;
search.recvtype=recvtype;
search.root=root;
found = (NBC_Scatter_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_SCATTER], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
/* receive from root */
if(rank != root) {
/* recv msg from root */
res = NBC_Sched_recv(recvbuf, false, recvcount, recvtype, root, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
} else {
for(i=0;i<p;i++) {
sbuf = ((char *)sendbuf) + (i*sendcount*sndext);
if(i != root) {
/* root sends the right buffer to the right receiver */
res = NBC_Sched_send(sbuf, false, sendcount, sendtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
}
}
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Scatter_args*)malloc(sizeof(NBC_Scatter_args));
args->sendbuf=sendbuf;
args->sendcount=sendcount;
args->sendtype=sendtype;
args->recvbuf=recvbuf;
args->recvcount=recvcount;
args->recvtype=recvtype;
args->root=root;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_SCATTER], args, args, 0);
if(res != 0) printf("error in dict_insert() (%i)\n", res);
/* increase number of elements for A2A */
if(++handle->comminfo->NBC_Dict_size[NBC_SCATTER] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_SCATTER], &handle->comminfo->NBC_Dict_size[NBC_SCATTER]);
}
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
/* simple linear MPI_Iscatterv */
int ompi_coll_libnbc_iscatterv(void* sendbuf, int *sendcounts, int *displs, MPI_Datatype sendtype,
void* recvbuf, int recvcount, MPI_Datatype recvtype, int root,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module) {
int rank, p, res, i;
MPI_Aint sndext;
NBC_Schedule *schedule;
char *sbuf, inplace;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res = MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
/* receive from root */
if(rank != root) {
/* recv msg from root */
res = NBC_Sched_recv(recvbuf, false, recvcount, recvtype, root, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
} else {
for(i=0;i<p;i++) {
sbuf = ((char *)sendbuf) + (displs[i]*sndext);
if(i == root) {
if(!inplace) {
/* if I am the root - just copy the message */
res = NBC_Copy(sbuf, sendcounts[i], sendtype, recvbuf, recvcount, recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
} else {
/* root sends the right buffer to the right receiver */
res = NBC_Sched_send(sbuf, false, sendcounts[i], sendtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
}
}
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
/* 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;
}
/* simple linear Alltoallw */
int ompi_coll_libnbc_ialltoallw(void* sendbuf, int *sendcounts, int *sdispls,
MPI_Datatype sendtypes[], void* recvbuf, int *recvcounts, int *rdispls,
MPI_Datatype recvtypes[], struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module)
{
int rank, p, res, i;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res= MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
/* copy data to receivbuffer */
if((sendcounts[rank] != 0) && !inplace) {
rbuf = ((char *) recvbuf) + rdispls[rank];
sbuf = ((char *) sendbuf) + sdispls[rank];
res = NBC_Copy(sbuf, sendcounts[rank], sendtypes[rank], rbuf, recvcounts[rank], recvtypes[rank], comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
for (i = 0; i < p; i++) {
if (i == rank) { continue; }
/* post all sends */
if(sendcounts[i] != 0) {
sbuf = ((char *) sendbuf) + sdispls[i];
res = NBC_Sched_send(sbuf, false, sendcounts[i], sendtypes[i], i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
/* post all receives */
if(recvcounts[i] != 0) {
rbuf = ((char *) recvbuf) + rdispls[i];
res = NBC_Sched_recv(rbuf, false, recvcounts[i], recvtypes[i], i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
/* the non-blocking reduce */
int ompi_coll_libnbc_ireduce(const 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_t size;
MPI_Aint ext;
NBC_Schedule *schedule;
char *redbuf=NULL, inplace;
enum { NBC_RED_BINOMIAL, NBC_RED_CHAIN } alg;
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 = ompi_datatype_type_size(datatype, &size);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in ompi_datatype_type_size() (%i)", res);
return res;
}
/* only one node -> copy data */
if ((p == 1) && !inplace) {
res = NBC_Copy (sendbuf, count, datatype, recvbuf, count, datatype, comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
*request = &ompi_request_empty;
return OMPI_SUCCESS;
}
res = NBC_Init_handle (comm, &handle, libnbc_module);
if (OPAL_UNLIKELY(OMPI_SUCCESS != 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 (OPAL_UNLIKELY(NULL == handle->tmpbuf)) {
NBC_Return_handle (handle);
return OMPI_ERR_OUT_OF_RESOURCE;
}
#ifdef NBC_CACHE_SCHEDULE
NBC_Reduce_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;
search.root = root;
found = (NBC_Reduce_args *) hb_tree_search ((hb_tree *) libnbc_module->NBC_Dict[NBC_REDUCE], &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;
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 (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_Reduce_args *) malloc (sizeof (args));
if (NULL != 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 *) libnbc_module->NBC_Dict[NBC_REDUCE], args, args, 0);
if (0 == res) {
OBJ_RETAIN(schedule);
/* increase number of elements for Reduce */
if (++libnbc_module->NBC_Dict_size[NBC_REDUCE] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_REDUCE],
&libnbc_module->NBC_Dict_size[NBC_REDUCE]);
}
} 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;
}
/* 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_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_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;
}
/* simple linear Alltoallv */
int NBC_Ialltoallv(void* sendbuf, int *sendcounts, int *sdispls,
MPI_Datatype sendtype, void* recvbuf, int *recvcounts, int *rdispls,
MPI_Datatype recvtype, MPI_Comm comm, NBC_Handle* handle) {
int rank, p, res, i;
MPI_Aint sndext, rcvext;
NBC_Schedule *schedule;
char *rbuf, *sbuf, inplace;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(handle, comm);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
res = MPI_Comm_rank(comm, &rank);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_rank() (%i)\n", res); return res; }
res= MPI_Comm_size(comm, &p);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Comm_size() (%i)\n", res); return res; }
res = MPI_Type_extent(sendtype, &sndext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = MPI_Type_extent(recvtype, &rcvext);
if (MPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc() (%i)\n", res); return res; }
handle->tmpbuf=NULL;
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
/* copy data to receivbuffer */
if((sendcounts[rank] != 0) && !inplace) {
rbuf = ((char *) recvbuf) + (rdispls[rank] * rcvext);
sbuf = ((char *) sendbuf) + (sdispls[rank] * sndext);
res = NBC_Copy(sbuf, sendcounts[rank], sendtype, rbuf, recvcounts[rank], recvtype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
for (i = 0; i < p; i++) {
if (i == rank) { continue; }
/* post all sends */
if(sendcounts[i] != 0) {
sbuf = ((char *) sendbuf) + (sdispls[i] * sndext);
res = NBC_Sched_send(sbuf, false, sendcounts[i], sendtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_send() (%i)\n", res); return res; }
}
/* post all receives */
if(recvcounts[i] != 0) {
rbuf = ((char *) recvbuf) + (rdispls[i] * rcvext);
res = NBC_Sched_recv(rbuf, false, recvcounts[i], recvtype, i, schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_recv() (%i)\n", res); return res; }
}
}
/*NBC_PRINT_SCHED(*schedule);*/
res = NBC_Sched_commit(schedule);
if (NBC_OK != res) { printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
res = NBC_Start(handle, schedule);
if (NBC_OK != res) { printf("Error in NBC_Start() (%i)\n", res); return res; }
return NBC_OK;
}
int ompi_coll_libnbc_igather(const 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_1_0_t *module) {
int rank, p, res;
MPI_Aint rcvext = 0;
NBC_Schedule *schedule;
char *rbuf, inplace = 0;
NBC_Handle *handle;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
rank = ompi_comm_rank (comm);
if (root == rank) {
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
}
p = ompi_comm_size (comm);
if (rank == root) {
res = ompi_datatype_type_extent (recvtype, &rcvext);
if (MPI_SUCCESS != res) {
NBC_Error("MPI Error in ompi_datatype_type_extent() (%i)", res);
return res;
}
}
if (inplace) {
sendcount = recvcount;
sendtype = recvtype;
} else if (rank == root) {
rbuf = ((char *)recvbuf) + (rank*recvcount*rcvext);
/* if I am the root - just copy the message (only without MPI_IN_PLACE) */
res = NBC_Copy(sendbuf, sendcount, sendtype, rbuf, recvcount, recvtype, comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
}
#ifdef NBC_CACHE_SCHEDULE
NBC_Gather_args *args, *found, search;
/* 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_Gather_args *) hb_tree_search ((hb_tree *) libnbc_module->NBC_Dict[NBC_GATHER],
&search);
if (NULL == found) {
#endif
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 + i * recvcount * rcvext;
if (i != root) {
/* root receives message to the right buffer */
res = NBC_Sched_recv (rbuf, false, recvcount, 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;
}
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Gather_args *) malloc (sizeof (args));
if (NULL != 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 *) libnbc_module->NBC_Dict[NBC_GATHER], args, args, 0);
if (0 == res) {
OBJ_RETAIN(schedule);
/* increase number of elements for A2A */
if (++libnbc_module->NBC_Dict_size[NBC_GATHER] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_GATHER],
&libnbc_module->NBC_Dict_size[NBC_GATHER]);
}
} 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;
}
int ompi_coll_libnbc_iallreduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
struct ompi_communicator_t *comm, ompi_request_t ** request,
struct mca_coll_base_module_2_0_0_t *module)
{
int rank, p, res;
OPAL_PTRDIFF_TYPE ext, lb;
NBC_Schedule *schedule;
size_t size;
#ifdef NBC_CACHE_SCHEDULE
NBC_Allreduce_args *args, *found, search;
#endif
enum { NBC_ARED_BINOMIAL, NBC_ARED_RING } alg;
char inplace;
NBC_Handle *handle;
ompi_coll_libnbc_request_t **coll_req = (ompi_coll_libnbc_request_t**) request;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
res = NBC_Init_handle(comm, coll_req, libnbc_module);
if(res != NBC_OK) { printf("Error in NBC_Init_handle(%i)\n", res); return res; }
handle = (*coll_req);
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
res = ompi_datatype_get_extent(datatype, &lb, &ext);
if (OMPI_SUCCESS != res) { printf("MPI Error in MPI_Type_extent() (%i)\n", res); return res; }
res = ompi_datatype_type_size (datatype, &size);
if (OMPI_SUCCESS != res) { printf("MPI Error in MPI_Type_size() (%i)\n", res); return res; }
handle->tmpbuf = malloc(ext*count);
if(handle->tmpbuf == NULL) { printf("Error in malloc() (%i)\n", res); return NBC_OOR; }
if((p == 1) && !inplace) {
/* for a single node - copy data to receivebuf */
res = NBC_Copy(sendbuf, count, datatype, recvbuf, count, datatype, comm);
if (NBC_OK != res) { printf("Error in NBC_Copy() (%i)\n", res); return res; }
}
/* algorithm selection */
if(p < 4 || size*count < 65536 || inplace) {
alg = NBC_ARED_BINOMIAL;
} else {
alg = NBC_ARED_RING;
}
#ifdef NBC_CACHE_SCHEDULE
/* search schedule in communicator specific tree */
search.sendbuf=sendbuf;
search.recvbuf=recvbuf;
search.count=count;
search.datatype=datatype;
search.op=op;
found = (NBC_Allreduce_args*)hb_tree_search((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLREDUCE], &search);
if(found == NULL) {
#endif
schedule = (NBC_Schedule*)malloc(sizeof(NBC_Schedule));
if (NULL == schedule) { printf("Error in malloc()\n"); return res; }
res = NBC_Sched_create(schedule);
if(res != NBC_OK) { printf("Error in NBC_Sched_create (%i)\n", res); return res; }
switch(alg) {
case NBC_ARED_BINOMIAL:
res = allred_sched_diss(rank, p, count, datatype, sendbuf, recvbuf, op, schedule, handle);
break;
case NBC_ARED_RING:
res = allred_sched_ring(rank, p, count, datatype, sendbuf, recvbuf, op, size, ext, schedule, handle);
break;
}
if (NBC_OK != res) { printf("Error in Schedule creation() (%i)\n", res); return res; }
res = NBC_Sched_commit(schedule);
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Sched_commit() (%i)\n", res); return res; }
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Allreduce_args*)malloc(sizeof(NBC_Allreduce_args));
args->sendbuf=sendbuf;
args->recvbuf=recvbuf;
args->count=count;
args->datatype=datatype;
args->op=op;
args->schedule=schedule;
res = hb_tree_insert ((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLREDUCE], args, args, 0);
if(res != 0) printf("error in dict_insert() (%i)\n", res);
/* increase number of elements for A2A */
if(++handle->comminfo->NBC_Dict_size[NBC_ALLREDUCE] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe((hb_tree*)handle->comminfo->NBC_Dict[NBC_ALLREDUCE], &handle->comminfo->NBC_Dict_size[NBC_ALLREDUCE]);
}
} else {
/* found schedule */
schedule=found->schedule;
}
#endif
res = NBC_Start(handle, schedule);
if(res != NBC_OK) { free(handle->tmpbuf); printf("Error in NBC_Start() (%i)\n", res); return res; }
/* tmpbuf is freed with the handle */
return NBC_OK;
}
static inline int NBC_Start_round(NBC_Handle *handle) {
int num; /* number of operations */
int res;
char* ptr;
MPI_Request *tmp;
NBC_Fn_type type;
NBC_Args_send sendargs;
NBC_Args_recv recvargs;
NBC_Args_op opargs;
NBC_Args_copy copyargs;
NBC_Args_unpack unpackargs;
void *buf1, *buf2;
/* get round-schedule address */
ptr = handle->schedule->data + handle->row_offset;
NBC_GET_BYTES(ptr,num);
NBC_DEBUG(10, "start_round round at offset %d : posting %i operations\n", handle->row_offset, num);
for (int i = 0 ; i < num ; ++i) {
int offset = (intptr_t)(ptr - handle->schedule->data);
memcpy (&type, ptr, sizeof (type));
switch(type) {
case SEND:
NBC_DEBUG(5," SEND (offset %li) ", offset);
NBC_GET_BYTES(ptr,sendargs);
NBC_DEBUG(5,"*buf: %p, count: %i, type: %p, dest: %i, tag: %i)\n", sendargs.buf,
sendargs.count, sendargs.datatype, sendargs.dest, handle->tag);
/* get an additional request */
handle->req_count++;
/* get buffer */
if(sendargs.tmpbuf) {
buf1=(char*)handle->tmpbuf+(long)sendargs.buf;
} else {
buf1=(void *)sendargs.buf;
}
#ifdef NBC_TIMING
Isend_time -= MPI_Wtime();
#endif
tmp = (MPI_Request *) realloc ((void *) handle->req_array, handle->req_count * sizeof (MPI_Request));
if (NULL == tmp) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
handle->req_array = tmp;
res = MCA_PML_CALL(isend(buf1, sendargs.count, sendargs.datatype, sendargs.dest, handle->tag,
MCA_PML_BASE_SEND_STANDARD, sendargs.local?handle->comm->c_local_comm:handle->comm,
handle->req_array+handle->req_count - 1));
if (OMPI_SUCCESS != res) {
NBC_Error ("Error in MPI_Isend(%lu, %i, %p, %i, %i, %lu) (%i)", (unsigned long)buf1, sendargs.count,
sendargs.datatype, sendargs.dest, handle->tag, (unsigned long)handle->comm, res);
return res;
}
#ifdef NBC_TIMING
Isend_time += MPI_Wtime();
#endif
break;
case RECV:
NBC_DEBUG(5, " RECV (offset %li) ", offset);
NBC_GET_BYTES(ptr,recvargs);
NBC_DEBUG(5, "*buf: %p, count: %i, type: %p, source: %i, tag: %i)\n", recvargs.buf, recvargs.count,
recvargs.datatype, recvargs.source, handle->tag);
/* get an additional request - TODO: req_count NOT thread safe */
handle->req_count++;
/* get buffer */
if(recvargs.tmpbuf) {
buf1=(char*)handle->tmpbuf+(long)recvargs.buf;
} else {
buf1=recvargs.buf;
}
#ifdef NBC_TIMING
Irecv_time -= MPI_Wtime();
#endif
tmp = (MPI_Request *) realloc ((void *) handle->req_array, handle->req_count * sizeof (MPI_Request));
if (NULL == tmp) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
handle->req_array = tmp;
res = MCA_PML_CALL(irecv(buf1, recvargs.count, recvargs.datatype, recvargs.source, handle->tag, recvargs.local?handle->comm->c_local_comm:handle->comm,
handle->req_array+handle->req_count-1));
if (OMPI_SUCCESS != res) {
NBC_Error("Error in MPI_Irecv(%lu, %i, %p, %i, %i, %lu) (%i)", (unsigned long)buf1, recvargs.count,
recvargs.datatype, recvargs.source, handle->tag, (unsigned long)handle->comm, res);
return res;
}
#ifdef NBC_TIMING
Irecv_time += MPI_Wtime();
#endif
break;
case OP:
NBC_DEBUG(5, " OP2 (offset %li) ", offset);
NBC_GET_BYTES(ptr,opargs);
NBC_DEBUG(5, "*buf1: %p, buf2: %p, count: %i, type: %p)\n", opargs.buf1, opargs.buf2,
opargs.count, opargs.datatype);
/* get buffers */
if(opargs.tmpbuf1) {
buf1=(char*)handle->tmpbuf+(long)opargs.buf1;
} else {
buf1=(void *)opargs.buf1;
}
if(opargs.tmpbuf2) {
buf2=(char*)handle->tmpbuf+(long)opargs.buf2;
} else {
buf2=opargs.buf2;
}
ompi_op_reduce(opargs.op, buf1, buf2, opargs.count, opargs.datatype);
break;
case COPY:
NBC_DEBUG(5, " COPY (offset %li) ", offset);
NBC_GET_BYTES(ptr,copyargs);
NBC_DEBUG(5, "*src: %lu, srccount: %i, srctype: %p, *tgt: %lu, tgtcount: %i, tgttype: %p)\n",
(unsigned long) copyargs.src, copyargs.srccount, copyargs.srctype,
(unsigned long) copyargs.tgt, copyargs.tgtcount, copyargs.tgttype);
/* get buffers */
if(copyargs.tmpsrc) {
buf1=(char*)handle->tmpbuf+(long)copyargs.src;
} else {
buf1=copyargs.src;
}
if(copyargs.tmptgt) {
buf2=(char*)handle->tmpbuf+(long)copyargs.tgt;
} else {
buf2=copyargs.tgt;
}
res = NBC_Copy (buf1, copyargs.srccount, copyargs.srctype, buf2, copyargs.tgtcount, copyargs.tgttype,
handle->comm);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
break;
case UNPACK:
NBC_DEBUG(5, " UNPACK (offset %li) ", offset);
NBC_GET_BYTES(ptr,unpackargs);
NBC_DEBUG(5, "*src: %lu, srccount: %i, srctype: %p, *tgt: %lu\n", (unsigned long) unpackargs.inbuf,
unpackargs.count, unpackargs.datatype, (unsigned long) unpackargs.outbuf);
/* get buffers */
if(unpackargs.tmpinbuf) {
buf1=(char*)handle->tmpbuf+(long)unpackargs.inbuf;
} else {
buf1=unpackargs.inbuf;
}
if(unpackargs.tmpoutbuf) {
buf2=(char*)handle->tmpbuf+(long)unpackargs.outbuf;
} else {
buf2=unpackargs.outbuf;
}
res = NBC_Unpack (buf1, unpackargs.count, unpackargs.datatype, buf2, handle->comm);
if (OMPI_SUCCESS != res) {
NBC_Error ("NBC_Unpack() failed (code: %i)", res);
return res;
}
break;
default:
NBC_Error ("NBC_Start_round: bad type %li at offset %li", (long)type, offset);
return OMPI_ERROR;
}
}
/* check if we can make progress - not in the first round, this allows us to leave the
* initialization faster and to reach more overlap
*
* threaded case: calling progress in the first round can lead to a
* deadlock if NBC_Free is called in this round :-( */
if (handle->row_offset) {
res = NBC_Progress(handle);
if ((NBC_OK != res) && (NBC_CONTINUE != res)) {
return OMPI_ERROR;
}
}
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;
}
