int
MPID_Put(const void *origin_addr,
int origin_count,
MPI_Datatype origin_datatype,
int target_rank,
MPI_Aint target_disp,
int target_count,
MPI_Datatype target_datatype,
MPID_Win *win)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_Win_request *req = MPIU_Calloc0(1, MPIDI_Win_request);
req->win = win;
if(win->mpid.request_based != 1)
req->type = MPIDI_WIN_REQUEST_PUT;
else {
req->req_handle = win->mpid.rreq;
req->type = MPIDI_WIN_REQUEST_RPUT;
req->req_handle->mpid.win_req = req;
}
if(win->mpid.sync.origin_epoch_type == win->mpid.sync.target_epoch_type &&
win->mpid.sync.origin_epoch_type == MPID_EPOTYPE_REFENCE){
win->mpid.sync.origin_epoch_type = MPID_EPOTYPE_FENCE;
win->mpid.sync.target_epoch_type = MPID_EPOTYPE_FENCE;
}
if(win->mpid.sync.origin_epoch_type == MPID_EPOTYPE_NONE ||
win->mpid.sync.origin_epoch_type == MPID_EPOTYPE_POST){
MPIU_ERR_SETANDSTMT(mpi_errno, MPI_ERR_RMA_SYNC,
return mpi_errno, "**rmasync");
}
void MPIDI_Coll_comm_create(MPID_Comm *comm)
{
volatile int geom_init = 1;
int i;
MPIDI_Post_geom_create_t geom_post;
TRACE_ERR("MPIDI_Coll_comm_create enter\n");
if (!MPIDI_Process.optimized.collectives)
return;
if(comm->comm_kind != MPID_INTRACOMM) return;
/* Create a geometry */
comm->coll_fns = MPIU_Calloc0(1, MPID_Collops);
MPID_assert(comm->coll_fns != NULL);
if(comm->mpid.geometry != MPIDI_Process.world_geometry)
{
if(unlikely(MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_0 && comm->rank == 0))
fprintf(stderr,"world geom: %p parent geom: %p\n", MPIDI_Process.world_geometry, comm->mpid.parent);
TRACE_ERR("Creating subgeom\n");
/* Change to this at some point */
comm->mpid.tasks = NULL;
for(i=1;i<comm->local_size;i++)
{
/* only if sequential tasks should we use a (single) range.
Multi or reordered ranges are inefficient */
if(MPID_VCR_GET_LPID(comm->vcr, i) != (MPID_VCR_GET_LPID(comm->vcr, i-1) + 1)) {
/* not sequential, use tasklist */
MPID_VCR_GET_LPIDS(comm, comm->mpid.tasks);
break;
}
}
/* Should we use a range? (no task list set) */
if(comm->mpid.tasks == NULL)
{
/* one range, {first rank ... last rank} */
comm->mpid.range.lo = MPID_VCR_GET_LPID(comm->vcr, 0);
comm->mpid.range.hi = MPID_VCR_GET_LPID(comm->vcr, comm->local_size-1);
}
if(unlikely(MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_0 && comm->rank == 0))
fprintf(stderr,"create geometry tasks %p {%u..%u}\n", comm->mpid.tasks, MPID_VCR_GET_LPID(comm->vcr, 0),MPID_VCR_GET_LPID(comm->vcr, comm->local_size-1));
pami_configuration_t config[3];
size_t numconfigs = 0;
#ifdef HAVE_PAMI_GEOMETRY_NONCONTIG
config[0].name = PAMI_GEOMETRY_NONCONTIG;
if(MPIDI_Process.optimized.memory & MPID_OPT_LVL_NONCONTIG)
config[0].value.intval = 0; // Disable non-contig, pamid doesn't use pami for non-contig data collectives
else
config[0].value.intval = 1; // Enable non-contig even though pamid doesn't use pami for non-contig data collectives,
// we still possibly want those collectives for other reasons.
++numconfigs;
#endif
if(MPIDI_Process.optimized.subcomms)
{
config[numconfigs].name = PAMI_GEOMETRY_OPTIMIZE;
config[numconfigs].value.intval = 1;
++numconfigs;
}
#ifdef HAVE_PAMI_GEOMETRY_MEMORY_OPTIMIZE
if(MPIDI_Process.optimized.memory)
{
config[numconfigs].name = PAMI_GEOMETRY_MEMORY_OPTIMIZE;
config[numconfigs].value.intval = MPIDI_Process.optimized.memory; /* level of optimization */
++numconfigs;
}
#endif
if((MPIDI_Process.optimized.memory & MPID_OPT_LVL_IRREG) && (comm->local_size & (comm->local_size-1)))
{
/* Don't create irregular geometries. Fallback to MPICH only collectives */
geom_init = 0;
comm->mpid.geometry = PAMI_GEOMETRY_NULL;
}
else if(comm->mpid.tasks == NULL)
{
geom_post.client = MPIDI_Client;
geom_post.configs = config;
geom_post.context_offset = 0; /* TODO BES investigate */
geom_post.num_configs = numconfigs;
geom_post.newgeom = &comm->mpid.geometry,
geom_post.parent = PAMI_GEOMETRY_NULL;
geom_post.id = comm->context_id;
geom_post.ranges = &comm->mpid.range;
geom_post.tasks = NULL;;
geom_post.count = (size_t)1;
geom_post.fn = geom_create_cb_done;
geom_post.cookie = (void*)&geom_init;
TRACE_ERR("%s geom_rangelist_create\n", MPIDI_Process.context_post>0?"Posting":"Invoking");
MPIDI_Context_post(MPIDI_Context[0], &geom_post.state,
geom_rangelist_create_wrapper, (void *)&geom_post);
}
else
{
geom_post.client = MPIDI_Client;
geom_post.configs = config;
geom_post.context_offset = 0; /* TODO BES investigate */
geom_post.num_configs = numconfigs;
geom_post.newgeom = &comm->mpid.geometry,
geom_post.parent = PAMI_GEOMETRY_NULL;
geom_post.id = comm->context_id;
geom_post.ranges = NULL;
geom_post.tasks = comm->mpid.tasks;
geom_post.count = (size_t)comm->local_size;
geom_post.fn = geom_create_cb_done;
geom_post.cookie = (void*)&geom_init;
TRACE_ERR("%s geom_tasklist_create\n", MPIDI_Process.context_post>0?"Posting":"Invoking");
MPIDI_Context_post(MPIDI_Context[0], &geom_post.state,
geom_tasklist_create_wrapper, (void *)&geom_post);
}
TRACE_ERR("Waiting for geom create to finish\n");
MPID_PROGRESS_WAIT_WHILE(geom_init);
if(comm->mpid.geometry == PAMI_GEOMETRY_NULL)
{
if(unlikely(MPIDI_Process.verbose >= MPIDI_VERBOSE_DETAILS_0 && comm->rank == 0))
fprintf(stderr,"Created unoptimized communicator id=%u, size=%u\n", (unsigned) comm->context_id,comm->local_size);
MPIU_TestFree(&comm->coll_fns);
return;
}
}
/* Initialize the async flow control in case it will be used. */
comm->mpid.num_requests = MPIDI_Process.optimized.num_requests;
TRACE_ERR("Querying protocols\n");
/* Determine what protocols are available for this comm/geom */
/* These two functions moved to mpid_collselect.c */
MPIDI_Comm_coll_query(comm);
MPIDI_Comm_coll_envvars(comm);
if(MPIDI_Process.optimized.select_colls)
MPIDI_Comm_coll_select(comm);
TRACE_ERR("mpir barrier\n");
int mpierrno = FALSE;
/* Switch to comm->coll_fns->fn() */
MPIDO_Barrier(comm, &mpierrno);
TRACE_ERR("MPIDI_Coll_comm_create exit\n");
}
