int MPIDO_Scatterv(const void *sendbuf,
const int *sendcounts,
const int *displs,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
int root,
MPID_Comm *comm_ptr,
int *mpierrno)
{
#ifndef HAVE_PAMI_IN_PLACE
if (sendbuf == MPI_IN_PLACE)
{
MPID_Abort (NULL, 0, 1, "'MPI_IN_PLACE' requries support for `PAMI_IN_PLACE`");
return -1;
}
#endif
int tmp, i, pamidt = 1;
int contig ATTRIBUTE((unused));
int ssize ATTRIBUTE((unused));
int rsize ATTRIBUTE((unused));
MPID_Datatype *dt_ptr = NULL;
MPI_Aint send_true_lb=0, recv_true_lb;
char *sbuf, *rbuf;
pami_type_t stype, rtype;
const int rank = comm_ptr->rank;
const int size = comm_ptr->local_size;
#if ASSERT_LEVEL==0
/* We can't afford the tracing in ndebug/performance libraries */
const unsigned verbose = 0;
#else
const unsigned verbose = (MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_ALL) && (rank == 0);
#endif
const struct MPIDI_Comm* const mpid = &(comm_ptr->mpid);
const int selected_type = mpid->user_selected_type[PAMI_XFER_SCATTERV_INT];
if(selected_type == MPID_COLL_USE_MPICH)
{
if(unlikely(verbose))
fprintf(stderr,"Using MPICH scatterv algorithm\n");
MPIDI_Update_last_algorithm(comm_ptr, "SCATTERV_MPICH");
#if CUDA_AWARE_SUPPORT
if(MPIDI_Process.cuda_aware_support_on)
{
MPI_Aint sdt_extent,rdt_extent;
MPID_Datatype_get_extent_macro(sendtype, sdt_extent);
MPID_Datatype_get_extent_macro(recvtype, rdt_extent);
char *scbuf = NULL;
char *rcbuf = NULL;
int is_send_dev_buf = (rank == root) ? MPIDI_cuda_is_device_buf(sendbuf) : 0;
int is_recv_dev_buf = MPIDI_cuda_is_device_buf(recvbuf);
if(is_send_dev_buf)
{
//Since displs can be non-continous, we need to calculate max buffer size
int highest_displs = displs[size - 1];
int highest_sendcount = sendcounts[size - 1];
size_t stotal_buf;
for(i = 0; i < size; i++)
{
if(displs[i]+sendcounts[i] > highest_displs+highest_sendcount)
{
highest_displs = displs[i];
highest_sendcount = sendcounts[i];
}
}
stotal_buf = (highest_displs+highest_sendcount)*sdt_extent;
scbuf = MPIU_Malloc(stotal_buf);
cudaError_t cudaerr = CudaMemcpy(scbuf, sendbuf, stotal_buf, cudaMemcpyDeviceToHost);
if (cudaSuccess != cudaerr)
fprintf(stderr, "cudaMemcpy failed: %s\n", CudaGetErrorString(cudaerr));
}
else
scbuf = sendbuf;
if(is_recv_dev_buf)
{
rcbuf = MPIU_Malloc(recvcount * rdt_extent);
CudaMemcpy(rcbuf, recvbuf, recvcount * rdt_extent, cudaMemcpyDeviceToHost);
}
else
rcbuf = recvbuf;
int cuda_res = MPIR_Scatterv(scbuf, sendcounts, displs, sendtype, rcbuf, recvcount, recvtype, root, comm_ptr, mpierrno);
if(is_send_dev_buf)MPIU_Free(scbuf);
if(is_recv_dev_buf)
{
cudaError_t cudaerr = CudaMemcpy(recvbuf, rcbuf, recvcount * rdt_extent, cudaMemcpyHostToDevice);
if (cudaSuccess != cudaerr)
fprintf(stderr, "cudaMemcpy failed: %s\n", CudaGetErrorString(cudaerr));
MPIU_Free(rcbuf);
}
return cuda_res;
}
else
#endif
return MPIR_Scatterv(sendbuf, sendcounts, displs, sendtype,
recvbuf, recvcount, recvtype,
root, comm_ptr, mpierrno);
}
pami_xfer_t scatterv;
pami_algorithm_t my_scatterv;
const pami_metadata_t *my_md = (pami_metadata_t *)NULL;
volatile unsigned scatterv_active = 1;
int queryreq = 0;
if(selected_type == MPID_COLL_OPTIMIZED)
{
TRACE_ERR("Optimized scatterv %s was selected\n",
mpid->opt_protocol_md[PAMI_XFER_SCATTERV_INT][0].name);
my_scatterv = mpid->opt_protocol[PAMI_XFER_SCATTERV_INT][0];
my_md = &mpid->opt_protocol_md[PAMI_XFER_SCATTERV_INT][0];
queryreq = mpid->must_query[PAMI_XFER_SCATTERV_INT][0];
}
else
{
TRACE_ERR("User selected %s for scatterv\n",
mpid->user_selected[PAMI_XFER_SCATTERV_INT]);
my_scatterv = mpid->user_selected[PAMI_XFER_SCATTERV_INT];
my_md = &mpid->user_metadata[PAMI_XFER_SCATTERV_INT];
queryreq = selected_type;
}
if((recvbuf != MPI_IN_PLACE) && MPIDI_Datatype_to_pami(recvtype, &rtype, -1, NULL, &tmp) != MPI_SUCCESS)
pamidt = 0;
if(MPIDI_Datatype_to_pami(sendtype, &stype, -1, NULL, &tmp) != MPI_SUCCESS)
pamidt = 0;
if(pamidt == 0 || selected_type == MPID_COLL_USE_MPICH)
{
if(unlikely(verbose))
fprintf(stderr,"Using MPICH scatterv algorithm\n");
TRACE_ERR("Scatterv using MPICH\n");
MPIDI_Update_last_algorithm(comm_ptr, "SCATTERV_MPICH");
return MPIR_Scatterv(sendbuf, sendcounts, displs, sendtype,
recvbuf, recvcount, recvtype,
root, comm_ptr, mpierrno);
}
MPIDI_Datatype_get_info(1, sendtype, contig, ssize, dt_ptr, send_true_lb);
sbuf = (char *)sendbuf + send_true_lb;
rbuf = recvbuf;
if(rank == root)
{
if(recvbuf == MPI_IN_PLACE)
{
if(unlikely(verbose))
fprintf(stderr,"scatterv MPI_IN_PLACE buffering\n");
rbuf = PAMI_IN_PLACE;
}
else
{
MPIDI_Datatype_get_info(1, recvtype, contig, rsize, dt_ptr, recv_true_lb);
rbuf = (char *)recvbuf + recv_true_lb;
}
}
scatterv.cb_done = cb_scatterv;
scatterv.cookie = (void *)&scatterv_active;
scatterv.cmd.xfer_scatterv_int.root = MPIDI_Task_to_endpoint(MPID_VCR_GET_LPID(comm_ptr->vcr, root), 0);
scatterv.algorithm = my_scatterv;
scatterv.cmd.xfer_scatterv_int.rcvbuf = rbuf;
scatterv.cmd.xfer_scatterv_int.sndbuf = sbuf;
scatterv.cmd.xfer_scatterv_int.stype = stype;
scatterv.cmd.xfer_scatterv_int.rtype = rtype;
scatterv.cmd.xfer_scatterv_int.stypecounts = (int *) sendcounts;
scatterv.cmd.xfer_scatterv_int.rtypecount = recvcount;
scatterv.cmd.xfer_scatterv_int.sdispls = (int *) displs;
if(unlikely(queryreq == MPID_COLL_ALWAYS_QUERY ||
queryreq == MPID_COLL_CHECK_FN_REQUIRED))
{
metadata_result_t result = {0};
TRACE_ERR("querying scatterv protocol %s, type was %d\n",
my_md->name, queryreq);
if(my_md->check_fn == NULL)
{
/* process metadata bits */
if((!my_md->check_correct.values.inplace) && (recvbuf == MPI_IN_PLACE))
result.check.unspecified = 1;
/* Can't check ranges like this. Non-local. Comment out for now.
if(my_md->check_correct.values.rangeminmax)
{
MPI_Aint data_true_lb;
MPID_Datatype *data_ptr;
int data_size, data_contig;
MPIDI_Datatype_get_info(recvcount, recvtype, data_contig, data_size, data_ptr, data_true_lb);
if((my_md->range_lo <= data_size) &&
(my_md->range_hi >= data_size))
;
else
{
result.check.range = 1;
if(unlikely(verbose))
{
fprintf(stderr,"message size (%u) outside range (%zu<->%zu) for %s.\n",
data_size,
my_md->range_lo,
my_md->range_hi,
my_md->name);
}
}
}
*/
}
else /* calling the check fn is sufficient */
result = my_md->check_fn(&scatterv);
TRACE_ERR("bitmask: %#X\n", result.bitmask);
result.check.nonlocal = 0; /* #warning REMOVE THIS WHEN IMPLEMENTED */
if(result.bitmask)
{
if(unlikely(verbose))
fprintf(stderr,"Query failed for %s. Using MPICH scatterv.\n", my_md->name);
MPIDI_Update_last_algorithm(comm_ptr, "SCATTERV_MPICH");
return MPIR_Scatterv(sendbuf, sendcounts, displs, sendtype,
recvbuf, recvcount, recvtype,
root, comm_ptr, mpierrno);
}
if(my_md->check_correct.values.asyncflowctl && !(--(comm_ptr->mpid.num_requests)))
{
comm_ptr->mpid.num_requests = MPIDI_Process.optimized.num_requests;
int tmpmpierrno;
if(unlikely(verbose))
fprintf(stderr,"Query barrier required for %s\n", my_md->name);
MPIDO_Barrier(comm_ptr, &tmpmpierrno);
}
}
MPIDI_Update_last_algorithm(comm_ptr, my_md->name);
if(unlikely(verbose))
{
unsigned long long int threadID;
MPIU_Thread_id_t tid;
MPIU_Thread_self(&tid);
threadID = (unsigned long long int)tid;
fprintf(stderr,"<%llx> Using protocol %s for scatterv on %u\n",
threadID,
my_md->name,
(unsigned) comm_ptr->context_id);
}
MPIDI_Post_coll_t scatterv_post;
MPIDI_Context_post(MPIDI_Context[0], &scatterv_post.state,
MPIDI_Pami_post_wrapper, (void *)&scatterv);
TRACE_ERR("Waiting on active %d\n", scatterv_active);
MPID_PROGRESS_WAIT_WHILE(scatterv_active);
TRACE_ERR("Leaving MPIDO_Scatterv\n");
return 0;
}
int MPIDO_Barrier(MPID_Comm *comm_ptr, int *mpierrno)
{
TRACE_ERR("Entering MPIDO_Barrier\n");
volatile unsigned active=1;
MPIDI_Post_coll_t barrier_post;
pami_xfer_t barrier;
pami_algorithm_t my_barrier;
pami_metadata_t *my_barrier_md;
int queryreq = 0;
if(comm_ptr->mpid.user_selected_type[PAMI_XFER_BARRIER] == MPID_COLL_USE_MPICH)
{
if(unlikely(MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_ALL && comm_ptr->rank == 0))
fprintf(stderr,"Using MPICH barrier\n");
TRACE_ERR("Using MPICH Barrier\n");
return MPIR_Barrier(comm_ptr, mpierrno);
}
barrier.cb_done = cb_barrier;
barrier.cookie = (void *)&active;
if(comm_ptr->mpid.user_selected_type[PAMI_XFER_BARRIER] == MPID_COLL_OPTIMIZED)
{
TRACE_ERR("Optimized barrier (%s) was pre-selected\n", comm_ptr->mpid.opt_protocol_md[PAMI_XFER_BARRIER][0].name);
my_barrier = comm_ptr->mpid.opt_protocol[PAMI_XFER_BARRIER][0];
my_barrier_md = &comm_ptr->mpid.opt_protocol_md[PAMI_XFER_BARRIER][0];
queryreq = comm_ptr->mpid.must_query[PAMI_XFER_BARRIER][0];
}
else
{
TRACE_ERR("Barrier (%s) was specified by user\n", comm_ptr->mpid.user_metadata[PAMI_XFER_BARRIER].name);
my_barrier = comm_ptr->mpid.user_selected[PAMI_XFER_BARRIER];
my_barrier_md = &comm_ptr->mpid.user_metadata[PAMI_XFER_BARRIER];
queryreq = comm_ptr->mpid.user_selected_type[PAMI_XFER_BARRIER];
}
barrier.algorithm = my_barrier;
/* There is no support for query-required barrier protocols here */
MPID_assert_always(queryreq != MPID_COLL_ALWAYS_QUERY);
MPID_assert_always(queryreq != MPID_COLL_CHECK_FN_REQUIRED);
/* TODO Name needs fixed somehow */
MPIDI_Update_last_algorithm(comm_ptr, my_barrier_md->name);
if(unlikely(MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_ALL && comm_ptr->rank == 0))
{
unsigned long long int threadID;
MPIU_Thread_id_t tid;
MPIU_Thread_self(&tid);
threadID = (unsigned long long int)tid;
fprintf(stderr,"<%llx> Using protocol %s for barrier on %u\n",
threadID,
my_barrier_md->name,
/* comm_ptr->rank,comm_ptr->local_size,comm_ptr->remote_size,*/
(unsigned) comm_ptr->context_id);
}
TRACE_ERR("%s barrier\n", MPIDI_Process.context_post.active>0?"posting":"invoking");
MPIDI_Context_post(MPIDI_Context[0], &barrier_post.state,
MPIDI_Pami_post_wrapper, (void *)&barrier);
TRACE_ERR("barrier %s rc: %d\n", MPIDI_Process.context_post.active>0?"posted":"invoked", rc);
TRACE_ERR("advance spinning\n");
MPID_PROGRESS_WAIT_WHILE(active);
TRACE_ERR("exiting mpido_barrier\n");
return 0;
}
int
MPIDO_Allgather(const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPID_Comm * comm_ptr,
int *mpierrno)
{
/* *********************************
* Check the nature of the buffers
* *********************************
*/
/* MPIDO_Coll_config config = {1,1,1,1,1,1};*/
int config[6], i;
MPID_Datatype * dt_null = NULL;
MPI_Aint send_true_lb = 0;
MPI_Aint recv_true_lb = 0;
int rc, comm_size = comm_ptr->local_size;
size_t send_size = 0;
size_t recv_size = 0;
volatile unsigned allred_active = 1;
volatile unsigned allgather_active = 1;
pami_xfer_t allred;
for (i=0;i<6;i++) config[i] = 1;
pami_metadata_t *my_md;
allred.cb_done = allred_cb_done;
allred.cookie = (void *)&allred_active;
/* Pick an algorithm that is guaranteed to work for the pre-allreduce */
/* TODO: This needs selection for fast(er|est) allreduce protocol */
allred.algorithm = comm_ptr->mpid.coll_algorithm[PAMI_XFER_ALLREDUCE][0][0];
allred.cmd.xfer_allreduce.sndbuf = (void *)config;
allred.cmd.xfer_allreduce.stype = PAMI_TYPE_SIGNED_INT;
allred.cmd.xfer_allreduce.rcvbuf = (void *)config;
allred.cmd.xfer_allreduce.rtype = PAMI_TYPE_SIGNED_INT;
allred.cmd.xfer_allreduce.stypecount = 6;
allred.cmd.xfer_allreduce.rtypecount = 6;
allred.cmd.xfer_allreduce.op = PAMI_DATA_BAND;
char use_tree_reduce, use_alltoall, use_bcast, use_pami, use_opt;
char *rbuf = NULL, *sbuf = NULL;
use_alltoall = comm_ptr->mpid.allgathers[2];
use_tree_reduce = comm_ptr->mpid.allgathers[0];
use_bcast = comm_ptr->mpid.allgathers[1];
use_pami =
(comm_ptr->mpid.user_selected_type[PAMI_XFER_ALLGATHER] == MPID_COLL_USE_MPICH) ? 0 : 1;
/* if(sendbuf == MPI_IN_PLACE) use_pami = 0;*/
use_opt = use_alltoall || use_tree_reduce || use_bcast || use_pami;
TRACE_ERR("flags before: b: %d a: %d t: %d p: %d\n", use_bcast, use_alltoall, use_tree_reduce, use_pami);
if(!use_opt)
{
if(unlikely(MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_ALL && comm_ptr->rank == 0))
fprintf(stderr,"Using MPICH allgather algorithm\n");
TRACE_ERR("No options set/available; using MPICH for allgather\n");
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHER_MPICH");
return MPIR_Allgather(sendbuf, sendcount, sendtype,
recvbuf, recvcount, recvtype,
comm_ptr, mpierrno);
}
if ((sendcount < 1 && sendbuf != MPI_IN_PLACE) || recvcount < 1)
return MPI_SUCCESS;
/* Gather datatype information */
MPIDI_Datatype_get_info(recvcount,
recvtype,
config[MPID_RECV_CONTIG],
recv_size,
dt_null,
recv_true_lb);
send_size = recv_size;
rbuf = (char *)recvbuf+recv_true_lb;
if(sendbuf != MPI_IN_PLACE)
{
if(unlikely(MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_ALL))
fprintf(stderr,"allgather MPI_IN_PLACE buffering\n");
MPIDI_Datatype_get_info(sendcount,
sendtype,
config[MPID_SEND_CONTIG],
send_size,
dt_null,
send_true_lb);
sbuf = (char *)sendbuf+send_true_lb;
}
else
{
sbuf = (char *)recvbuf+recv_size*comm_ptr->rank;
}
/* fprintf(stderr,"sendount: %d, recvcount: %d send_size: %zd
recv_size: %zd\n", sendcount, recvcount, send_size,
recv_size);*/
/* verify everyone's datatype contiguity */
/* Check buffer alignment now, since we're pre-allreducing anyway */
/* Only do this if one of the glue protocols is likely to be used */
if(use_alltoall || use_tree_reduce || use_bcast)
{
config[MPID_ALIGNEDBUFFER] =
!((long)sendbuf & 0x0F) && !((long)recvbuf & 0x0F);
/* #warning need to determine best allreduce for short messages */
if(comm_ptr->mpid.preallreduces[MPID_ALLGATHER_PREALLREDUCE])
{
TRACE_ERR("Preallreducing in allgather\n");
MPIDI_Post_coll_t allred_post;
MPIDI_Context_post(MPIDI_Context[0], &allred_post.state,
MPIDI_Pami_post_wrapper, (void *)&allred);
MPID_PROGRESS_WAIT_WHILE(allred_active);
}
use_alltoall = comm_ptr->mpid.allgathers[2] &&
config[MPID_RECV_CONTIG] && config[MPID_SEND_CONTIG];;
/* Note: some of the glue protocols use recv_size*comm_size rather than
* recv_size so we use that for comparison here, plus we pass that in
* to those protocols. */
use_tree_reduce = comm_ptr->mpid.allgathers[0] &&
config[MPID_RECV_CONTIG] && config[MPID_SEND_CONTIG] &&
config[MPID_RECV_CONTINUOUS] && (recv_size*comm_size % sizeof(int) == 0);
use_bcast = comm_ptr->mpid.allgathers[1];
TRACE_ERR("flags after: b: %d a: %d t: %d p: %d\n", use_bcast, use_alltoall, use_tree_reduce, use_pami);
}
if(use_pami)
{
TRACE_ERR("Using PAMI-level allgather protocol\n");
pami_xfer_t allgather;
allgather.cb_done = allgather_cb_done;
allgather.cookie = (void *)&allgather_active;
allgather.cmd.xfer_allgather.rcvbuf = rbuf;
allgather.cmd.xfer_allgather.sndbuf = sbuf;
allgather.cmd.xfer_allgather.stype = PAMI_TYPE_BYTE;
allgather.cmd.xfer_allgather.rtype = PAMI_TYPE_BYTE;
allgather.cmd.xfer_allgather.stypecount = send_size;
allgather.cmd.xfer_allgather.rtypecount = recv_size;
if(comm_ptr->mpid.user_selected_type[PAMI_XFER_ALLGATHER] == MPID_COLL_OPTIMIZED)
{
allgather.algorithm = comm_ptr->mpid.opt_protocol[PAMI_XFER_ALLGATHER][0];
my_md = &comm_ptr->mpid.opt_protocol_md[PAMI_XFER_ALLGATHER][0];
}
else
{
allgather.algorithm = comm_ptr->mpid.user_selected[PAMI_XFER_ALLGATHER];
my_md = &comm_ptr->mpid.user_metadata[PAMI_XFER_ALLGATHER];
}
if(unlikely( comm_ptr->mpid.user_selected_type[PAMI_XFER_ALLGATHER] == MPID_COLL_ALWAYS_QUERY ||
comm_ptr->mpid.user_selected_type[PAMI_XFER_ALLGATHER] == MPID_COLL_CHECK_FN_REQUIRED))
{
metadata_result_t result = {0};
TRACE_ERR("Querying allgather protocol %s, type was: %d\n",
my_md->name,
comm_ptr->mpid.user_selected_type[PAMI_XFER_ALLGATHER]);
result = my_md->check_fn(&allgather);
TRACE_ERR("bitmask: %#X\n", result.bitmask);
if(!result.bitmask)
{
fprintf(stderr,"Query failed for %s.\n",
my_md->name);
}
}
if(unlikely(MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_ALL && comm_ptr->rank == 0))
{
unsigned long long int threadID;
MPIU_Thread_id_t tid;
MPIU_Thread_self(&tid);
threadID = (unsigned long long int)tid;
fprintf(stderr,"<%llx> Using protocol %s for allgather on %u\n",
threadID,
my_md->name,
(unsigned) comm_ptr->context_id);
}
TRACE_ERR("Calling PAMI_Collective with allgather structure\n");
MPIDI_Post_coll_t allgather_post;
MPIDI_Context_post(MPIDI_Context[0], &allgather_post.state, MPIDI_Pami_post_wrapper, (void *)&allgather);
TRACE_ERR("Allgather %s\n", MPIDI_Process.context_post.active>0?"posted":"invoked");
MPIDI_Update_last_algorithm(comm_ptr, my_md->name);
MPID_PROGRESS_WAIT_WHILE(allgather_active);
TRACE_ERR("Allgather done\n");
return PAMI_SUCCESS;
}
if(use_tree_reduce)
{
TRACE_ERR("Using allgather via allreduce\n");
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHER_OPT_ALLREDUCE");
rc = MPIDO_Allgather_allreduce(sendbuf, sendcount, sendtype,
recvbuf, recvcount, recvtype,
send_true_lb, recv_true_lb, send_size, recv_size*comm_size, comm_ptr, mpierrno);
return rc;
}
if(use_alltoall)
{
TRACE_ERR("Using allgather via alltoall\n");
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHER_OPT_ALLTOALL");
rc = MPIDO_Allgather_alltoall(sendbuf, sendcount, sendtype,
recvbuf, recvcount, recvtype,
send_true_lb, recv_true_lb, send_size, recv_size*comm_size, comm_ptr, mpierrno);
return rc;
}
if(use_bcast)
{
TRACE_ERR("Using allgather via bcast\n");
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHER_OPT_BCAST");
rc = MPIDO_Allgather_bcast(sendbuf, sendcount, sendtype,
recvbuf, recvcount, recvtype,
send_true_lb, recv_true_lb, send_size, recv_size*comm_size, comm_ptr, mpierrno);
return rc;
}
/* Nothing used yet; dump to MPICH */
if(unlikely(MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_ALL && comm_ptr->rank == 0))
fprintf(stderr,"Using MPICH allgather algorithm\n");
TRACE_ERR("Using allgather via mpich\n");
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHER_MPICH");
return MPIR_Allgather(sendbuf, sendcount, sendtype,
recvbuf, recvcount, recvtype,
comm_ptr, mpierrno);
}
int MPIDO_Bcast(void *buffer,
int count,
MPI_Datatype datatype,
int root,
MPID_Comm *comm_ptr,
int *mpierrno)
{
TRACE_ERR("in mpido_bcast\n");
const size_t BCAST_LIMIT = 0x40000000;
int data_contig, rc;
void *data_buffer = NULL,
*noncontig_buff = NULL;
volatile unsigned active = 1;
MPI_Aint data_true_lb = 0;
MPID_Datatype *data_ptr;
MPID_Segment segment;
MPIDI_Post_coll_t bcast_post;
const struct MPIDI_Comm* const mpid = &(comm_ptr->mpid);
const int rank = comm_ptr->rank;
#if ASSERT_LEVEL==0
/* We can't afford the tracing in ndebug/performance libraries */
const unsigned verbose = 0;
#else
const unsigned verbose = (MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_ALL) && (rank == 0);
#endif
const int selected_type = mpid->user_selected_type[PAMI_XFER_BROADCAST];
/* Must calculate data_size based on count=1 in case it's total size is > integer */
int data_size_one;
MPIDI_Datatype_get_info(1, datatype,
data_contig, data_size_one, data_ptr, data_true_lb);
/* do this calculation once and use twice */
const size_t data_size_sz = (size_t)data_size_one*(size_t)count;
if(unlikely(verbose))
fprintf(stderr,"bcast count %d, size %d (%#zX), root %d, buffer %p\n",
count,data_size_one, (size_t)data_size_one*(size_t)count, root,buffer);
if(unlikely( data_size_sz > BCAST_LIMIT) )
{
void *new_buffer=buffer;
int c, new_count = (int)BCAST_LIMIT/data_size_one;
MPID_assert(new_count > 0);
for(c=1; ((size_t)c*(size_t)new_count) <= (size_t)count; ++c)
{
if ((rc = MPIDO_Bcast(new_buffer,
new_count,
datatype,
root,
comm_ptr,
mpierrno)) != MPI_SUCCESS)
return rc;
new_buffer = (char*)new_buffer + (size_t)data_size_one*(size_t)new_count;
}
new_count = count % new_count; /* 0 is ok, just returns no-op */
return MPIDO_Bcast(new_buffer,
new_count,
datatype,
root,
comm_ptr,
mpierrno);
}
/* Must use data_size based on count for byte bcast processing.
Previously calculated as a size_t but large data_sizes were
handled above so this cast to int should be fine here.
*/
const int data_size = (int)data_size_sz;
if(selected_type == MPID_COLL_USE_MPICH || data_size == 0)
{
if(unlikely(verbose))
fprintf(stderr,"Using MPICH bcast algorithm\n");
MPIDI_Update_last_algorithm(comm_ptr,"BCAST_MPICH");
return MPIR_Bcast_intra(buffer, count, datatype, root, comm_ptr, mpierrno);
}
data_buffer = (char *)buffer + data_true_lb;
if(!data_contig)
{
noncontig_buff = MPIU_Malloc(data_size);
data_buffer = noncontig_buff;
if(noncontig_buff == NULL)
{
MPID_Abort(NULL, MPI_ERR_NO_SPACE, 1,
"Fatal: Cannot allocate pack buffer");
}
if(rank == root)
{
DLOOP_Offset last = data_size;
MPID_Segment_init(buffer, count, datatype, &segment, 0);
MPID_Segment_pack(&segment, 0, &last, noncontig_buff);
}
}
pami_xfer_t bcast;
pami_algorithm_t my_bcast;
const pami_metadata_t *my_md = (pami_metadata_t *)NULL;
int queryreq = 0;
bcast.cb_done = cb_bcast;
bcast.cookie = (void *)&active;
bcast.cmd.xfer_broadcast.root = MPIDI_Task_to_endpoint(MPID_VCR_GET_LPID(comm_ptr->vcr, root), 0);
bcast.algorithm = mpid->user_selected[PAMI_XFER_BROADCAST];
bcast.cmd.xfer_broadcast.buf = data_buffer;
bcast.cmd.xfer_broadcast.type = PAMI_TYPE_BYTE;
/* Needs to be sizeof(type)*count since we are using bytes as * the generic type */
bcast.cmd.xfer_broadcast.typecount = data_size;
if(selected_type == MPID_COLL_OPTIMIZED)
{
TRACE_ERR("Optimized bcast (%s) and (%s) were pre-selected\n",
mpid->opt_protocol_md[PAMI_XFER_BROADCAST][0].name,
mpid->opt_protocol_md[PAMI_XFER_BROADCAST][1].name);
if(mpid->cutoff_size[PAMI_XFER_BROADCAST][1] != 0)/* SSS: There is FCA cutoff (FCA only sets cutoff for [PAMI_XFER_BROADCAST][1]) */
{
if(data_size <= mpid->cutoff_size[PAMI_XFER_BROADCAST][1])
{
my_bcast = mpid->opt_protocol[PAMI_XFER_BROADCAST][1];
my_md = &mpid->opt_protocol_md[PAMI_XFER_BROADCAST][1];
queryreq = mpid->must_query[PAMI_XFER_BROADCAST][1];
}
else
{
return MPIR_Bcast_intra(buffer, count, datatype, root, comm_ptr, mpierrno);
}
}
if(data_size > mpid->cutoff_size[PAMI_XFER_BROADCAST][0])
{
my_bcast = mpid->opt_protocol[PAMI_XFER_BROADCAST][1];
my_md = &mpid->opt_protocol_md[PAMI_XFER_BROADCAST][1];
queryreq = mpid->must_query[PAMI_XFER_BROADCAST][1];
}
else
{
my_bcast = mpid->opt_protocol[PAMI_XFER_BROADCAST][0];
my_md = &mpid->opt_protocol_md[PAMI_XFER_BROADCAST][0];
queryreq = mpid->must_query[PAMI_XFER_BROADCAST][0];
}
}
else
{
TRACE_ERR("Bcast (%s) was specified by user\n",
mpid->user_metadata[PAMI_XFER_BROADCAST].name);
my_bcast = mpid->user_selected[PAMI_XFER_BROADCAST];
my_md = &mpid->user_metadata[PAMI_XFER_BROADCAST];
queryreq = selected_type;
}
bcast.algorithm = my_bcast;
if(unlikely(queryreq == MPID_COLL_ALWAYS_QUERY ||
queryreq == MPID_COLL_CHECK_FN_REQUIRED))
{
metadata_result_t result = {0};
TRACE_ERR("querying bcast protocol %s, type was: %d\n",
my_md->name, queryreq);
if(my_md->check_fn != NULL) /* calling the check fn is sufficient */
{
metadata_result_t result = {0};
result = my_md->check_fn(&bcast);
result.check.nonlocal = 0; /* #warning REMOVE THIS WHEN IMPLEMENTED */
}
else /* no check_fn, manually look at the metadata fields */
{
TRACE_ERR("Optimzed selection line %d\n",__LINE__);
/* Check if the message range if restricted */
if(my_md->check_correct.values.rangeminmax)
{
if((my_md->range_lo <= data_size) &&
(my_md->range_hi >= data_size))
; /* ok, algorithm selected */
else
{
result.check.range = 1;
if(unlikely(verbose))
{
fprintf(stderr,"message size (%u) outside range (%zu<->%zu) for %s.\n",
data_size,
my_md->range_lo,
my_md->range_hi,
my_md->name);
}
}
}
/* \todo check the rest of the metadata */
}
TRACE_ERR("bitmask: %#X\n", result.bitmask);
if(result.bitmask)
{
if(unlikely(verbose))
fprintf(stderr,"Using MPICH bcast algorithm - query fn failed\n");
MPIDI_Update_last_algorithm(comm_ptr,"BCAST_MPICH");
return MPIR_Bcast_intra(buffer, count, datatype, root, comm_ptr, mpierrno);
}
if(my_md->check_correct.values.asyncflowctl && !(--(comm_ptr->mpid.num_requests)))
{
comm_ptr->mpid.num_requests = MPIDI_Process.optimized.num_requests;
int tmpmpierrno;
if(unlikely(verbose))
fprintf(stderr,"Query barrier required for %s\n", my_md->name);
MPIDO_Barrier(comm_ptr, &tmpmpierrno);
}
}
if(unlikely(verbose))
{
unsigned long long int threadID;
MPIU_Thread_id_t tid;
MPIU_Thread_self(&tid);
threadID = (unsigned long long int)tid;
fprintf(stderr,"<%llx> Using protocol %s for bcast on %u\n",
threadID,
my_md->name,
(unsigned) comm_ptr->context_id);
}
MPIDI_Context_post(MPIDI_Context[0], &bcast_post.state, MPIDI_Pami_post_wrapper, (void *)&bcast);
MPIDI_Update_last_algorithm(comm_ptr, my_md->name);
MPID_PROGRESS_WAIT_WHILE(active);
TRACE_ERR("bcast done\n");
if(!data_contig)
{
if(rank != root)
MPIR_Localcopy(noncontig_buff, data_size, MPI_CHAR,
buffer, count, datatype);
MPIU_Free(noncontig_buff);
}
TRACE_ERR("leaving bcast\n");
return 0;
}
int
MPIDO_Allgatherv(const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
const int *recvcounts,
const int *displs,
MPI_Datatype recvtype,
MPID_Comm * comm_ptr,
int *mpierrno)
{
#ifndef HAVE_PAMI_IN_PLACE
if (sendbuf == MPI_IN_PLACE)
{
MPID_Abort (NULL, 0, 1, "'MPI_IN_PLACE' requries support for `PAMI_IN_PLACE`");
return -1;
}
#endif
TRACE_ERR("Entering MPIDO_Allgatherv\n");
/* function pointer to be used to point to approperiate algorithm */
/* Check the nature of the buffers */
MPID_Datatype *dt_null = NULL;
MPI_Aint send_true_lb = 0;
MPI_Aint recv_true_lb = 0;
size_t send_size = 0;
size_t recv_size = 0;
int config[6];
int scount=sendcount;
int i, rc, buffer_sum = 0;
const int size = comm_ptr->local_size;
char use_tree_reduce, use_alltoall, use_bcast, use_pami, use_opt;
char *sbuf, *rbuf;
const int rank = comm_ptr->rank;
const struct MPIDI_Comm* const mpid = &(comm_ptr->mpid);
int queryreq = 0;
#if ASSERT_LEVEL==0
/* We can't afford the tracing in ndebug/performance libraries */
const unsigned verbose = 0;
#else
const unsigned verbose = (MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_ALL) && (rank == 0);
#endif
const int selected_type = mpid->user_selected_type[PAMI_XFER_ALLGATHERV_INT];
pami_xfer_t allred;
volatile unsigned allred_active = 1;
volatile unsigned allgatherv_active = 1;
pami_type_t stype, rtype;
int tmp;
const pami_metadata_t *my_md = (pami_metadata_t *)NULL;
for(i=0;i<6;i++) config[i] = 1;
allred.cb_done = allred_cb_done;
allred.cookie = (void *)&allred_active;
allred.algorithm = mpid->coll_algorithm[PAMI_XFER_ALLREDUCE][0][0];
allred.cmd.xfer_allreduce.sndbuf = (void *)config;
allred.cmd.xfer_allreduce.stype = PAMI_TYPE_SIGNED_INT;
allred.cmd.xfer_allreduce.rcvbuf = (void *)config;
allred.cmd.xfer_allreduce.rtype = PAMI_TYPE_SIGNED_INT;
allred.cmd.xfer_allreduce.stypecount = 6;
allred.cmd.xfer_allreduce.rtypecount = 6;
allred.cmd.xfer_allreduce.op = PAMI_DATA_BAND;
use_alltoall = mpid->allgathervs[2];
use_tree_reduce = mpid->allgathervs[0];
use_bcast = mpid->allgathervs[1];
use_pami = selected_type != MPID_COLL_USE_MPICH;
if((sendbuf != MPI_IN_PLACE) && (MPIDI_Datatype_to_pami(sendtype, &stype, -1, NULL, &tmp) != MPI_SUCCESS))
use_pami = 0;
if(MPIDI_Datatype_to_pami(recvtype, &rtype, -1, NULL, &tmp) != MPI_SUCCESS)
use_pami = 0;
use_opt = use_alltoall || use_tree_reduce || use_bcast || use_pami;
if(!use_opt) /* back to MPICH */
{
if(unlikely(verbose))
fprintf(stderr,"Using MPICH allgatherv type %u.\n",
selected_type);
TRACE_ERR("Using MPICH Allgatherv\n");
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHERV_MPICH");
#if CUDA_AWARE_SUPPORT
if(MPIDI_Process.cuda_aware_support_on)
{
MPI_Aint sdt_extent,rdt_extent;
MPID_Datatype_get_extent_macro(sendtype, sdt_extent);
MPID_Datatype_get_extent_macro(recvtype, rdt_extent);
char *scbuf = NULL;
char *rcbuf = NULL;
int is_send_dev_buf = MPIDI_cuda_is_device_buf(sendbuf);
int is_recv_dev_buf = MPIDI_cuda_is_device_buf(recvbuf);
if(is_send_dev_buf)
{
scbuf = MPIU_Malloc(sdt_extent * sendcount);
cudaError_t cudaerr = CudaMemcpy(scbuf, sendbuf, sdt_extent * sendcount, cudaMemcpyDeviceToHost);
if (cudaSuccess != cudaerr)
fprintf(stderr, "cudaMemcpy failed: %s\n", CudaGetErrorString(cudaerr));
}
else
scbuf = sendbuf;
size_t rtotal_buf;
if(is_recv_dev_buf)
{
//Since displs can be non-continous, we need to calculate max buffer size
int highest_displs = displs[size - 1];
int highest_recvcount = recvcounts[size - 1];
for(i = 0; i < size; i++)
{
if(displs[i]+recvcounts[i] > highest_displs+highest_recvcount)
{
highest_displs = displs[i];
highest_recvcount = recvcounts[i];
}
}
rtotal_buf = (highest_displs+highest_recvcount)*rdt_extent;
rcbuf = MPIU_Malloc(rtotal_buf);
if(sendbuf == MPI_IN_PLACE)
{
cudaError_t cudaerr = CudaMemcpy(rcbuf, recvbuf, rtotal_buf, cudaMemcpyDeviceToHost);
if (cudaSuccess != cudaerr)
fprintf(stderr, "cudaMemcpy failed: %s\n", CudaGetErrorString(cudaerr));
}
else
memset(rcbuf, 0, rtotal_buf);
}
else
rcbuf = recvbuf;
int cuda_res = MPIR_Allgatherv(scbuf, sendcount, sendtype, rcbuf, recvcounts, displs, recvtype, comm_ptr, mpierrno);
if(is_send_dev_buf)MPIU_Free(scbuf);
if(is_recv_dev_buf)
{
cudaError_t cudaerr = CudaMemcpy(recvbuf, rcbuf, rtotal_buf, cudaMemcpyHostToDevice);
if (cudaSuccess != cudaerr)
fprintf(stderr, "cudaMemcpy failed: %s\n", CudaGetErrorString(cudaerr));
MPIU_Free(rcbuf);
}
return cuda_res;
}
else
#endif
return MPIR_Allgatherv(sendbuf, sendcount, sendtype,
recvbuf, recvcounts, displs, recvtype,
comm_ptr, mpierrno);
}
MPIDI_Datatype_get_info(1,
recvtype,
config[MPID_RECV_CONTIG],
recv_size,
dt_null,
recv_true_lb);
if(sendbuf == MPI_IN_PLACE)
{
sbuf = PAMI_IN_PLACE;
if(unlikely(verbose))
fprintf(stderr,"allgatherv MPI_IN_PLACE buffering\n");
stype = rtype;
scount = recvcounts[rank];
send_size = recv_size * scount;
}
else
{
MPIDI_Datatype_get_info(sendcount,
sendtype,
config[MPID_SEND_CONTIG],
send_size,
dt_null,
send_true_lb);
sbuf = (char *)sendbuf+send_true_lb;
}
rbuf = (char *)recvbuf+recv_true_lb;
if(use_alltoall || use_bcast || use_tree_reduce)
{
if (displs[0])
config[MPID_RECV_CONTINUOUS] = 0;
for (i = 1; i < size; i++)
{
buffer_sum += recvcounts[i - 1];
if (buffer_sum != displs[i])
{
config[MPID_RECV_CONTINUOUS] = 0;
break;
}
}
buffer_sum += recvcounts[size - 1];
buffer_sum *= recv_size;
/* disable with "safe allgatherv" env var */
if(mpid->preallreduces[MPID_ALLGATHERV_PREALLREDUCE])
{
MPIDI_Post_coll_t allred_post;
MPIDI_Context_post(MPIDI_Context[0], &allred_post.state,
MPIDI_Pami_post_wrapper, (void *)&allred);
MPID_PROGRESS_WAIT_WHILE(allred_active);
}
use_tree_reduce = mpid->allgathervs[0] &&
config[MPID_RECV_CONTIG] && config[MPID_SEND_CONTIG] &&
config[MPID_RECV_CONTINUOUS] && buffer_sum % sizeof(unsigned) == 0;
use_alltoall = mpid->allgathervs[2] &&
config[MPID_RECV_CONTIG] && config[MPID_SEND_CONTIG];
use_bcast = mpid->allgathervs[1];
}
if(use_pami)
{
pami_xfer_t allgatherv;
allgatherv.cb_done = allgatherv_cb_done;
allgatherv.cookie = (void *)&allgatherv_active;
if(selected_type == MPID_COLL_OPTIMIZED)
{
if((mpid->cutoff_size[PAMI_XFER_ALLGATHERV_INT][0] == 0) ||
(mpid->cutoff_size[PAMI_XFER_ALLGATHERV_INT][0] > 0 && mpid->cutoff_size[PAMI_XFER_ALLGATHERV_INT][0] >= send_size))
{
allgatherv.algorithm = mpid->opt_protocol[PAMI_XFER_ALLGATHERV_INT][0];
my_md = &mpid->opt_protocol_md[PAMI_XFER_ALLGATHERV_INT][0];
queryreq = mpid->must_query[PAMI_XFER_ALLGATHERV_INT][0];
}
else
return MPIR_Allgatherv(sendbuf, sendcount, sendtype,
recvbuf, recvcounts, displs, recvtype,
comm_ptr, mpierrno);
}
else
{
allgatherv.algorithm = mpid->user_selected[PAMI_XFER_ALLGATHERV_INT];
my_md = &mpid->user_metadata[PAMI_XFER_ALLGATHERV_INT];
queryreq = selected_type;
}
allgatherv.cmd.xfer_allgatherv_int.sndbuf = sbuf;
allgatherv.cmd.xfer_allgatherv_int.rcvbuf = rbuf;
allgatherv.cmd.xfer_allgatherv_int.stype = stype;
allgatherv.cmd.xfer_allgatherv_int.rtype = rtype;
allgatherv.cmd.xfer_allgatherv_int.stypecount = scount;
allgatherv.cmd.xfer_allgatherv_int.rtypecounts = (int *) recvcounts;
allgatherv.cmd.xfer_allgatherv_int.rdispls = (int *) displs;
if(unlikely (queryreq == MPID_COLL_ALWAYS_QUERY ||
queryreq == MPID_COLL_CHECK_FN_REQUIRED))
{
metadata_result_t result = {0};
TRACE_ERR("Querying allgatherv_int protocol %s, type was %d\n", my_md->name,
selected_type);
if(my_md->check_fn == NULL)
{
/* process metadata bits */
if((!my_md->check_correct.values.inplace) && (sendbuf == MPI_IN_PLACE))
result.check.unspecified = 1;
/* Can't check ranges like this. Non-local. Comment out for now.
if(my_md->check_correct.values.rangeminmax)
{
MPI_Aint data_true_lb;
MPID_Datatype *data_ptr;
int data_size, data_contig;
MPIDI_Datatype_get_info(sendcount, sendtype, data_contig, data_size, data_ptr, data_true_lb);
if((my_md->range_lo <= data_size) &&
(my_md->range_hi >= data_size))
;
else
{
result.check.range = 1;
if(unlikely(verbose))
{
fprintf(stderr,"message size (%u) outside range (%zu<->%zu) for %s.\n",
data_size,
my_md->range_lo,
my_md->range_hi,
my_md->name);
}
}
}
*/
}
else /* calling the check fn is sufficient */
result = my_md->check_fn(&allgatherv);
TRACE_ERR("Allgatherv bitmask: %#X\n", result.bitmask);
result.check.nonlocal = 0; /* #warning REMOVE THIS WHEN IMPLEMENTED */
if(result.bitmask)
{
if(unlikely(verbose))
fprintf(stderr,"Query failed for %s. Using MPICH allgatherv.\n", my_md->name);
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHERV_MPICH");
return MPIR_Allgatherv(sendbuf, sendcount, sendtype,
recvbuf, recvcounts, displs, recvtype,
comm_ptr, mpierrno);
}
if(my_md->check_correct.values.asyncflowctl && !(--(comm_ptr->mpid.num_requests)))
{
comm_ptr->mpid.num_requests = MPIDI_Process.optimized.num_requests;
int tmpmpierrno;
if(unlikely(verbose))
fprintf(stderr,"Query barrier required for %s\n", my_md->name);
MPIDO_Barrier(comm_ptr, &tmpmpierrno);
}
}
if(unlikely(verbose))
{
unsigned long long int threadID;
MPIU_Thread_id_t tid;
MPIU_Thread_self(&tid);
threadID = (unsigned long long int)tid;
fprintf(stderr,"<%llx> Using protocol %s for allgatherv on %u\n",
threadID,
my_md->name,
(unsigned) comm_ptr->context_id);
}
MPIDI_Post_coll_t allgatherv_post;
MPIDI_Context_post(MPIDI_Context[0], &allgatherv_post.state,
MPIDI_Pami_post_wrapper, (void *)&allgatherv);
MPIDI_Update_last_algorithm(comm_ptr, my_md->name);
TRACE_ERR("Rank %d waiting on active %d\n", rank, allgatherv_active);
MPID_PROGRESS_WAIT_WHILE(allgatherv_active);
return PAMI_SUCCESS;
}
/* TODO These need ordered in speed-order */
if(use_tree_reduce)
{
if(unlikely(verbose))
fprintf(stderr,"Using tree reduce allgatherv type %u.\n",
selected_type);
rc = MPIDO_Allgatherv_allreduce(sendbuf, sendcount, sendtype,
recvbuf, recvcounts, buffer_sum, displs, recvtype,
send_true_lb, recv_true_lb, send_size, recv_size,
comm_ptr, mpierrno);
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHERV_OPT_ALLREDUCE");
return rc;
}
if(use_bcast)
{
if(unlikely(verbose))
fprintf(stderr,"Using bcast allgatherv type %u.\n",
selected_type);
rc = MPIDO_Allgatherv_bcast(sendbuf, sendcount, sendtype,
recvbuf, recvcounts, buffer_sum, displs, recvtype,
send_true_lb, recv_true_lb, send_size, recv_size,
comm_ptr, mpierrno);
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHERV_OPT_BCAST");
return rc;
}
if(use_alltoall)
{
if(unlikely(verbose))
fprintf(stderr,"Using alltoall allgatherv type %u.\n",
selected_type);
rc = MPIDO_Allgatherv_alltoall(sendbuf, sendcount, sendtype,
recvbuf, (int *)recvcounts, buffer_sum, displs, recvtype,
send_true_lb, recv_true_lb, send_size, recv_size,
comm_ptr, mpierrno);
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHERV_OPT_ALLTOALL");
return rc;
}
if(unlikely(verbose))
fprintf(stderr,"Using MPICH allgatherv type %u.\n",
selected_type);
TRACE_ERR("Using MPICH for Allgatherv\n");
MPIDI_Update_last_algorithm(comm_ptr, "ALLGATHERV_MPICH");
return MPIR_Allgatherv(sendbuf, sendcount, sendtype,
recvbuf, recvcounts, displs, recvtype,
comm_ptr, mpierrno);
}