/* requires a contiguous/continous buffer on root though */
int MPIDO_Scatterv_bcast(void *sendbuf,
int *sendcounts,
int *displs,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
int root,
MPID_Comm *comm_ptr,
int *mpierrno)
{
const int rank = comm_ptr->rank;
const int size = comm_ptr->local_size;
char *tempbuf;
int i, sum = 0, dtsize, rc=0;
for (i = 0; i < size; i++)
if (sendcounts > 0)
sum += sendcounts[i];
MPIDI_Datatype_get_data_size(1,
recvtype,
dtsize);
if (rank != root)
{
tempbuf = MPIU_Malloc(dtsize * sum);
if (!tempbuf)
return MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
__FUNCTION__,
__LINE__,
MPI_ERR_OTHER,
"**nomem", 0);
}
else
tempbuf = sendbuf;
/* Switch to comm->coll_fns->fn() */
rc = MPIDO_Bcast(tempbuf, sum, sendtype, root, comm_ptr, mpierrno);
if(rank == root && recvbuf == MPI_IN_PLACE)
return rc;
memcpy(recvbuf, tempbuf + displs[rank], sendcounts[rank] * dtsize);
if (rank != root)
MPIU_Free(tempbuf);
return rc;
}
/* ****************************************************************** */
int MPIDO_Allgather_bcast(const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
MPI_Datatype recvtype,
MPI_Aint send_true_lb,
MPI_Aint recv_true_lb,
size_t send_size,
size_t recv_size,
MPID_Comm * comm_ptr,
int *mpierrno)
{
int i, np, rc = 0;
MPI_Aint extent;
const int rank = comm_ptr->rank;
np = comm_ptr ->local_size;
MPID_Datatype_get_extent_macro(recvtype, extent);
MPIU_Ensure_Aint_fits_in_pointer ((MPIU_VOID_PTR_CAST_TO_MPI_AINT recvbuf +
np * recvcount * extent));
if (sendbuf != MPI_IN_PLACE)
{
void *destbuf = recvbuf + rank * recvcount * extent;
MPIR_Localcopy(sendbuf,
sendcount,
sendtype,
destbuf,
recvcount,
recvtype);
}
/* this code should either abort on first error or somehow aggregate
* error codes, esp since it calls internal routines */
for (i = 0; i < np; i++)
{
void *destbuf = recvbuf + i * recvcount * extent;
/* Switch to comm->coll_fns->fn() */
rc = MPIDO_Bcast(destbuf,
recvcount,
recvtype,
i,
comm_ptr,
mpierrno);
}
return rc;
}
/* ****************************************************************** */
int MPIDO_Allgatherv_bcast(const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
const int *recvcounts,
int buffer_sum,
const int *displs,
MPI_Datatype recvtype,
MPI_Aint send_true_lb,
MPI_Aint recv_true_lb,
size_t send_size,
size_t recv_size,
MPID_Comm * comm_ptr,
int *mpierrno)
{
const int rank = comm_ptr->rank;
TRACE_ERR("Entering MPIDO_Allgatherv_bcast\n");
int i, rc=MPI_ERR_INTERN;
MPI_Aint extent;
MPID_Datatype_get_extent_macro(recvtype, extent);
if (sendbuf != MPI_IN_PLACE)
{
void *destbuffer = recvbuf + displs[rank] * extent;
MPIR_Localcopy(sendbuf,
sendcount,
sendtype,
destbuffer,
recvcounts[rank],
recvtype);
}
TRACE_ERR("Calling MPIDO_Bcasts in MPIDO_Allgatherv_bcast\n");
for (i = 0; i < comm_ptr->local_size; i++)
{
void *destbuffer = recvbuf + displs[i] * extent;
/* Switch to comm->coll_fns->fn() */
rc = MPIDO_Bcast(destbuffer,
recvcounts[i],
recvtype,
i,
comm_ptr,
mpierrno);
}
TRACE_ERR("Leaving MPIDO_Allgatherv_bcast\n");
return rc;
}
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;
}
