/*
* reduce_log_inter
*
* Function: - reduction using O(N) algorithm
* Accepts: - same as MPI_Reduce()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_cuda_reduce(const void *sbuf, void *rbuf, int count,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
int root, struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_cuda_module_t *s = (mca_coll_cuda_module_t*) module;
ptrdiff_t true_lb, true_extent, lb, extent;
char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2 = NULL;
const char *sbuf2;
size_t bufsize;
int rc;
ompi_datatype_get_extent(dtype, &lb, &extent);
ompi_datatype_get_true_extent(dtype, &true_lb, &true_extent);
bufsize = true_extent + (ptrdiff_t)(count - 1) * extent;
if ((MPI_IN_PLACE != sbuf) && (opal_cuda_check_bufs((char *)sbuf, NULL))) {
sbuf1 = (char*)malloc(bufsize);
if (NULL == sbuf1) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
opal_cuda_memcpy_sync(sbuf1, sbuf, bufsize);
sbuf2 = sbuf; /* save away original buffer */
sbuf = sbuf1 - lb;
}
if (opal_cuda_check_bufs(rbuf, NULL)) {
rbuf1 = (char*)malloc(bufsize);
if (NULL == rbuf1) {
if (NULL != sbuf1) free(sbuf1);
return OMPI_ERR_OUT_OF_RESOURCE;
}
opal_cuda_memcpy_sync(rbuf1, rbuf, bufsize);
rbuf2 = rbuf; /* save away original buffer */
rbuf = rbuf1 - lb;
}
rc = s->c_coll.coll_reduce((void *) sbuf, rbuf, count,
dtype, op, root, comm,
s->c_coll.coll_reduce_module);
if (NULL != sbuf1) {
free(sbuf1);
}
if (NULL != rbuf1) {
rbuf = rbuf2;
opal_cuda_memcpy_sync(rbuf, rbuf1, bufsize);
free(rbuf1);
}
return rc;
}
/*
* reduce_scatter_block
*
* Function: - reduce then scatter
* Accepts: - same as MPI_Reduce_scatter_block()
* Returns: - MPI_SUCCESS or error code
*
* Algorithm:
* reduce and scatter (needs to be cleaned
* up at some point)
*/
int
mca_coll_cuda_reduce_scatter_block(const void *sbuf, void *rbuf, int rcount,
struct ompi_datatype_t *dtype,
struct ompi_op_t *op,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_cuda_module_t *s = (mca_coll_cuda_module_t*) module;
ptrdiff_t gap;
char *rbuf1 = NULL, *sbuf1 = NULL, *rbuf2 = NULL;
const char *sbuf2;
size_t sbufsize, rbufsize;
int rc;
rbufsize = opal_datatype_span(&dtype->super, rcount, &gap);
sbufsize = rbufsize * ompi_comm_size(comm);
if ((MPI_IN_PLACE != sbuf) && (opal_cuda_check_bufs((char *)sbuf, NULL))) {
sbuf1 = (char*)malloc(sbufsize);
if (NULL == sbuf1) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
opal_cuda_memcpy_sync(sbuf1, sbuf, sbufsize);
sbuf2 = sbuf; /* save away original buffer */
sbuf = sbuf1 - gap;
}
if (opal_cuda_check_bufs(rbuf, NULL)) {
rbuf1 = (char*)malloc(rbufsize);
if (NULL == rbuf1) {
if (NULL != sbuf1) free(sbuf1);
return OMPI_ERR_OUT_OF_RESOURCE;
}
opal_cuda_memcpy_sync(rbuf1, rbuf, rbufsize);
rbuf2 = rbuf; /* save away original buffer */
rbuf = rbuf1 - gap;
}
rc = s->c_coll.coll_reduce_scatter_block(sbuf, rbuf, rcount, dtype, op, comm,
s->c_coll.coll_reduce_scatter_block_module);
if (NULL != sbuf1) {
free(sbuf1);
}
if (NULL != rbuf1) {
rbuf = rbuf2;
opal_cuda_memcpy_sync(rbuf, rbuf1, rbufsize);
free(rbuf1);
}
return rc;
}
int32_t opal_datatype_copy_content_same_ddt( const opal_datatype_t* datatype, int32_t count,
char* destination_base, char* source_base )
{
OPAL_PTRDIFF_TYPE extent;
size_t iov_len_local;
int32_t (*fct)( const opal_datatype_t*, int32_t, char*, char*);
#if OPAL_CUDA_SUPPORT
bool cuda_device_bufs = opal_cuda_check_bufs(destination_base, source_base);
#endif
DO_DEBUG( opal_output( 0, "opal_datatype_copy_content_same_ddt( %p, %d, dst %p, src %p )\n",
(void*)datatype, count, destination_base, source_base ); );
/* 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;
}