int MPIDI_VCRT_Create(int size, struct MPIDI_VCRT **vcrt_ptr)
{
MPIDI_VCRT_t * vcrt;
int mpi_errno = MPI_SUCCESS;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_VCRT_CREATE);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_VCRT_CREATE);
MPIU_CHKPMEM_MALLOC(vcrt, MPIDI_VCRT_t *, sizeof(MPIDI_VCRT_t) + (size - 1) * sizeof(MPIDI_VC_t *), mpi_errno, "**nomem");
vcrt->handle = HANDLE_SET_KIND(0, HANDLE_KIND_INVALID);
MPIU_Object_set_ref(vcrt, 1);
vcrt->size = size;
*vcrt_ptr = vcrt;
fn_exit:
MPIU_CHKPMEM_COMMIT();
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_VCRT_CREATE);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPID_nem_ptl_poll_init(void)
{
int mpi_errno = MPI_SUCCESS;
int i;
MPIU_CHKPMEM_DECL(NUM_OVERFLOW_ME);
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_PTL_POLL_INIT);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_PTL_POLL_INIT);
/* create overflow buffers */
for (i = 0; i < NUM_OVERFLOW_ME; ++i) {
MPIU_CHKPMEM_MALLOC(overflow_buf[i], void *, OVERFLOW_LENGTH, mpi_errno, "overflow buffer");
mpi_errno = append_overflow(i);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
fn_exit:
MPIU_CHKPMEM_COMMIT();
fn_exit2:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_PTL_POLL_INIT);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit2;
}
int MPIR_Comm_map_dup(MPID_Comm * newcomm, MPID_Comm * src_comm, MPIR_Comm_map_dir_t dir)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Comm_map_t *mapper;
MPIU_CHKPMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_MAP_DUP);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_MAP_DUP);
MPIU_CHKPMEM_MALLOC(mapper, MPIR_Comm_map_t *, sizeof(MPIR_Comm_map_t), mpi_errno, "mapper");
mapper->type = MPIR_COMM_MAP_DUP;
mapper->src_comm = src_comm;
mapper->dir = dir;
mapper->next = NULL;
MPL_LL_APPEND(newcomm->mapper_head, newcomm->mapper_tail, mapper);
fn_exit:
MPIU_CHKPMEM_COMMIT();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_MAP_DUP);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
int MPID_nem_scif_init(MPIDI_PG_t * pg_p, int pg_rank, char **bc_val_p, int *val_max_sz_p)
{
int mpi_errno = MPI_SUCCESS;
int ret;
int i;
MPIU_CHKPMEM_DECL(2);
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_SCIF_INIT);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_SCIF_INIT);
/* first make sure that our private fields in the vc fit into the
* area provided */
MPIU_Assert(sizeof(MPID_nem_scif_vc_area) <= MPID_NEM_VC_NETMOD_AREA_LEN);
MPID_nem_scif_nranks = pg_p->size;
MPID_nem_scif_myrank = pg_rank;
/* set up listener socket */
if (MPID_nem_scif_myrank < MPID_nem_scif_nranks - 1) {
listen_fd = scif_open();
MPIU_ERR_CHKANDJUMP1(listen_fd == -1, mpi_errno, MPI_ERR_OTHER,
"**scif_open", "**scif_open %s", MPIU_Strerror(errno));
listen_port = scif_bind(listen_fd, 0);
MPIU_ERR_CHKANDJUMP1(listen_port == -1, mpi_errno, MPI_ERR_OTHER,
"**scif_bind", "**scif_bind %s", MPIU_Strerror(errno));
ret = scif_listen(listen_fd, MPID_nem_scif_nranks);
MPIU_ERR_CHKANDJUMP1(ret == -1, mpi_errno, MPI_ERR_OTHER,
"**scif_listen", "**scif_listen %s", MPIU_Strerror(errno));
}
/* create business card */
mpi_errno = MPID_nem_scif_get_business_card(pg_rank, bc_val_p, val_max_sz_p);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
MPIU_CHKPMEM_MALLOC(MPID_nem_scif_conns, scifconn_t *,
MPID_nem_scif_nranks * sizeof(scifconn_t), mpi_errno,
"connection table");
memset(MPID_nem_scif_conns, 0, MPID_nem_scif_nranks * sizeof(scifconn_t));
for (i = 0; i < MPID_nem_scif_nranks; ++i)
MPID_nem_scif_conns[i].fd = -1;
MPIU_CHKPMEM_MALLOC(MPID_nem_scif_recv_buf, char *,
MPID_NEM_SCIF_RECV_MAX_PKT_LEN, mpi_errno, "SCIF temporary buffer");
MPIU_CHKPMEM_COMMIT();
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_SCIF_INIT);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
int MPIDI_CH3U_Win_allocate_shared(MPI_Aint size, int disp_unit, MPID_Info *info, MPID_Comm *comm_ptr,
void **base_ptr, MPID_Win **win_ptr)
{
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_self_ptr = NULL;
MPID_Group *group_comm, *group_self;
int result;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_WIN_ALLOCATE_SHARED);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_WIN_ALLOCATE_SHARED);
#ifdef HAVE_ERROR_CHECKING
/* The baseline CH3 implementation only works with MPI_COMM_SELF */
MPID_Comm_get_ptr( MPI_COMM_SELF, comm_self_ptr );
mpi_errno = MPIR_Comm_group_impl(comm_ptr, &group_comm);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Comm_group_impl(comm_self_ptr, &group_self);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Group_compare_impl(group_comm, group_self, &result);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Group_free_impl(group_comm);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Group_free_impl(group_self);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
if (result != MPI_IDENT) {
MPIU_ERR_SETANDJUMP(mpi_errno, MPI_ERR_RMA_SHARED, "**ch3|win_shared_comm");
}
#endif
mpi_errno = MPIDI_CH3U_Win_allocate(size, disp_unit, info, comm_ptr,
base_ptr, win_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_CHKPMEM_MALLOC((*win_ptr)->shm_base_addrs, void **,
1 /* comm_size */ * sizeof(void *),
mpi_errno, "(*win_ptr)->shm_base_addrs");
(*win_ptr)->shm_base_addrs[0] = *base_ptr;
/* Register the shared memory window free function, which will free the
memory allocated here. */
(*win_ptr)->RMAFns.Win_free = MPIDI_SHM_Win_free;
fn_exit:
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_WIN_ALLOCATE_SHARED);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
static int handle_mprobe(const ptl_event_t *e)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *const req = e->user_ptr;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_HANDLE_PROBE);
MPIDI_FUNC_ENTER(MPID_STATE_HANDLE_PROBE);
if (e->ni_fail_type == PTL_NI_NO_MATCH) {
REQ_PTL(req)->found = FALSE;
goto finish_mprobe;
}
REQ_PTL(req)->found = TRUE;
req->status.MPI_SOURCE = NPTL_MATCH_GET_RANK(e->match_bits);
req->status.MPI_TAG = NPTL_MATCH_GET_TAG(e->match_bits);
MPIR_STATUS_SET_COUNT(req->status, NPTL_HEADER_GET_LENGTH(e->hdr_data));
MPIDI_Request_set_sync_send_flag(req, e->hdr_data & NPTL_SSEND);
MPIU_CHKPMEM_MALLOC(req->dev.tmpbuf, void *, e->mlength, mpi_errno, "tmpbuf");
MPIU_Memcpy((char *)req->dev.tmpbuf, e->start, e->mlength);
req->dev.recv_data_sz = e->mlength;
if (!(e->hdr_data & NPTL_LARGE)) {
MPIDI_Request_set_msg_type(req, MPIDI_REQUEST_EAGER_MSG);
}
else {
MPIU_Assert (e->mlength == PTL_LARGE_THRESHOLD);
req->dev.match.parts.tag = req->status.MPI_TAG;
req->dev.match.parts.context_id = NPTL_MATCH_GET_CTX(e->match_bits);
req->dev.match.parts.rank = req->status.MPI_SOURCE;
MPIDI_Request_set_msg_type(req, MPIDI_REQUEST_RNDV_MSG);
}
/* At this point we know the ME is unlinked. Invalidate the handle to
prevent further accesses, e.g. an attempted cancel. */
REQ_PTL(req)->put_me = PTL_INVALID_HANDLE;
req->dev.recv_pending_count = 1;
finish_mprobe:
mpi_errno = MPID_Request_complete(req);
if (mpi_errno != MPI_SUCCESS) {
MPIR_ERR_POP(mpi_errno);
}
fn_exit:
MPIU_CHKPMEM_COMMIT();
MPIDI_FUNC_EXIT(MPID_STATE_HANDLE_PROBE);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
static int sched_get_cid_nonblock(MPID_Comm * comm_ptr, MPID_Comm * newcomm,
MPIU_Context_id_t * ctx0, MPIU_Context_id_t * ctx1,
MPID_Sched_t s, MPID_Comm_kind_t gcn_cid_kind)
{
int mpi_errno = MPI_SUCCESS;
struct gcn_state *st = NULL;
MPIU_CHKPMEM_DECL(1);
if (initialize_context_mask) {
context_id_init();
}
MPIU_CHKPMEM_MALLOC(st, struct gcn_state *, sizeof(struct gcn_state), mpi_errno, "gcn_state");
st->ctx0 = ctx0;
st->ctx1 = ctx1;
if (gcn_cid_kind == MPID_INTRACOMM) {
st->comm_ptr = comm_ptr;
st->comm_ptr_inter = NULL;
}
else {
st->comm_ptr = comm_ptr->local_comm;
st->comm_ptr_inter = comm_ptr;
}
st->s = s;
st->gcn_cid_kind = gcn_cid_kind;
*(st->ctx0) = 0;
st->own_eager_mask = 0;
st->first_iter = 1;
st->new_comm = newcomm;
st->own_mask = 0;
if (eager_nelem < 0) {
/* Ensure that at least one word of deadlock-free context IDs is
* always set aside for the base protocol */
MPIU_Assert(MPIR_CVAR_CTXID_EAGER_SIZE >= 0 &&
MPIR_CVAR_CTXID_EAGER_SIZE < MPIR_MAX_CONTEXT_MASK - 1);
eager_nelem = MPIR_CVAR_CTXID_EAGER_SIZE;
}
mpi_errno = MPID_Sched_cb(&sched_cb_gcn_copy_mask, st, s);
if (mpi_errno)
MPIR_ERR_POP(mpi_errno);
MPID_SCHED_BARRIER(s);
MPIU_CHKPMEM_COMMIT();
fn_exit:
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
static int handler_recv_dequeue_unpack_large(const ptl_event_t *e)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *const rreq = e->user_ptr;
MPIDI_VC_t *vc;
MPI_Aint last;
void *buf;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_HANDLER_RECV_DEQUEUE_UNPACK_LARGE);
MPIDI_FUNC_ENTER(MPID_STATE_HANDLER_RECV_DEQUEUE_UNPACK_LARGE);
MPIU_Assert(e->type == PTL_EVENT_PUT || e->type == PTL_EVENT_PUT_OVERFLOW);
MPIDI_Comm_get_vc(rreq->comm, NPTL_MATCH_GET_RANK(e->match_bits), &vc);
dequeue_req(e);
if (!(e->hdr_data & NPTL_LARGE)) {
/* all data has already been received; we're done */
mpi_errno = handler_recv_unpack_complete(e);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
goto fn_exit;
}
if (e->type == PTL_EVENT_PUT_OVERFLOW)
buf = e->start;
else
buf = REQ_PTL(rreq)->chunk_buffer[0];
MPIU_Assert(e->mlength == PTL_LARGE_THRESHOLD);
last = PTL_LARGE_THRESHOLD;
MPID_Segment_unpack(rreq->dev.segment_ptr, rreq->dev.segment_first, &last, buf);
MPIU_Assert(last == PTL_LARGE_THRESHOLD);
rreq->dev.segment_first += PTL_LARGE_THRESHOLD;
MPIU_Free(REQ_PTL(rreq)->chunk_buffer[0]);
MPIU_CHKPMEM_MALLOC(REQ_PTL(rreq)->chunk_buffer[0], void *, rreq->dev.segment_size - rreq->dev.segment_first,
mpi_errno, "chunk_buffer");
big_get(REQ_PTL(rreq)->chunk_buffer[0], rreq->dev.segment_size - rreq->dev.segment_first, vc, e->match_bits, rreq);
fn_exit:
MPIU_CHKPMEM_COMMIT();
fn_exit2:
MPIDI_FUNC_EXIT(MPID_STATE_HANDLER_RECV_DEQUEUE_UNPACK_LARGE);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit2;
}
int MPIR_Comm_map_irregular(MPID_Comm * newcomm, MPID_Comm * src_comm,
int *src_mapping, int src_mapping_size,
MPIR_Comm_map_dir_t dir, MPIR_Comm_map_t ** map)
{
int mpi_errno = MPI_SUCCESS;
MPIR_Comm_map_t *mapper;
MPIU_CHKPMEM_DECL(3);
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_MAP_IRREGULAR);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_MAP_IRREGULAR);
MPIU_CHKPMEM_MALLOC(mapper, MPIR_Comm_map_t *, sizeof(MPIR_Comm_map_t), mpi_errno, "mapper");
mapper->type = MPIR_COMM_MAP_IRREGULAR;
mapper->src_comm = src_comm;
mapper->dir = dir;
mapper->src_mapping_size = src_mapping_size;
if (src_mapping) {
mapper->src_mapping = src_mapping;
mapper->free_mapping = 0;
}
else {
MPIU_CHKPMEM_MALLOC(mapper->src_mapping, int *,
src_mapping_size * sizeof(int), mpi_errno, "mapper mapping");
mapper->free_mapping = 1;
}
mapper->next = NULL;
MPL_LL_APPEND(newcomm->mapper_head, newcomm->mapper_tail, mapper);
if (map)
*map = mapper;
fn_exit:
MPIU_CHKPMEM_COMMIT();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_MAP_IRREGULAR);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
int MPIR_T_pvar_session_create_impl(MPI_T_pvar_session *session)
{
int mpi_errno = MPI_SUCCESS;
MPIU_CHKPMEM_DECL(1);
*session = MPI_T_PVAR_SESSION_NULL;
MPIU_CHKPMEM_MALLOC(*session, MPI_T_pvar_session, sizeof(**session), mpi_errno, "performance var session");
/* essential for utlist to work */
(*session)->hlist = NULL;
MPIU_CHKPMEM_COMMIT();
fn_exit:
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
int MPIDI_CH3_Connect_to_root (const char *port_name, MPIDI_VC_t **new_vc)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_VC_t * vc;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3_CONNECT_TO_ROOT);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3_CONNECT_TO_ROOT);
*new_vc = NULL; /* so that the err handling knows to cleanup */
MPIU_CHKPMEM_MALLOC (vc, MPIDI_VC_t *, sizeof(MPIDI_VC_t), mpi_errno, "vc");
/* FIXME - where does this vc get freed?
ANSWER (goodell@) - ch3u_port.c FreeNewVC
(but the VC_Destroy is in this file) */
/* init ch3 portion of vc */
MPIDI_VC_Init (vc, NULL, 0);
/* init channel portion of vc */
MPIR_ERR_CHKINTERNAL(!nemesis_initialized, mpi_errno, "Nemesis not initialized");
vc->ch.recv_active = NULL;
MPIDI_CHANGE_VC_STATE(vc, ACTIVE);
*new_vc = vc; /* we now have a valid, disconnected, temp VC */
mpi_errno = MPID_nem_connect_to_root (port_name, vc);
if (mpi_errno) MPIR_ERR_POP (mpi_errno);
MPIU_CHKPMEM_COMMIT();
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3_CONNECT_TO_ROOT);
return mpi_errno;
fn_fail:
/* freeing without giving the lower layer a chance to cleanup can lead to
leaks on error */
if (*new_vc)
MPIDI_CH3_VC_Destroy(*new_vc);
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
int MPID_nem_lmt_dma_initiate_lmt(MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt, MPID_Request *sreq)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_pkt_lmt_rts_t * const rts_pkt = (MPID_nem_pkt_lmt_rts_t *)pkt;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_LMT_DMA_INITIATE_LMT);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_LMT_DMA_INITIATE_LMT);
MPIU_CHKPMEM_MALLOC(sreq->ch.s_cookie, knem_cookie_t *, sizeof(knem_cookie_t), mpi_errno, "s_cookie");
mpi_errno = send_sreq_data(vc, sreq, sreq->ch.s_cookie);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPID_nem_lmt_send_RTS(vc, rts_pkt, sreq->ch.s_cookie, sizeof(knem_cookie_t));
fn_exit:
MPIU_CHKPMEM_COMMIT();
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_LMT_DMA_INITIATE_LMT);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
int MPIDI_CH3U_Comm_register_destroy_hook(int (*hook_fn)(struct MPID_Comm *, void *), void *param)
{
int mpi_errno = MPI_SUCCESS;
hook_elt *elt;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_COMM_REGISTER_DESTROY_HOOK);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_COMM_REGISTER_DESTROY_HOOK);
MPIU_CHKPMEM_MALLOC(elt, hook_elt *, sizeof(hook_elt), mpi_errno, "hook_elt");
elt->hook_fn = hook_fn;
elt->param = param;
MPL_LL_PREPEND(destroy_hooks_head, destroy_hooks_tail, elt);
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_COMM_REGISTER_DESTROY_HOOK);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
/*@
MPI_Dist_graph_create - MPI_DIST_GRAPH_CREATE returns a handle to a new
communicator to which the distributed graph topology information is
attached.
Input Parameters:
+ comm_old - input communicator (handle)
. n - number of source nodes for which this process specifies edges
(non-negative integer)
. sources - array containing the n source nodes for which this process
specifies edges (array of non-negative integers)
. degrees - array specifying the number of destinations for each source node
in the source node array (array of non-negative integers)
. destinations - destination nodes for the source nodes in the source node
array (array of non-negative integers)
. weights - weights for source to destination edges (array of non-negative
integers or MPI_UNWEIGHTED)
. info - hints on optimization and interpretation of weights (handle)
- reorder - the process may be reordered (true) or not (false) (logical)
Output Parameters:
. comm_dist_graph - communicator with distributed graph topology added (handle)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_ARG
.N MPI_ERR_OTHER
@*/
int MPI_Dist_graph_create(MPI_Comm comm_old, int n, const int sources[],
const int degrees[], const int destinations[],
const int weights[],
MPI_Info info, int reorder, MPI_Comm *comm_dist_graph)
{
int mpi_errno = MPI_SUCCESS;
MPID_Comm *comm_ptr = NULL;
MPID_Comm *comm_dist_graph_ptr = NULL;
MPI_Request *reqs = NULL;
MPIR_Topology *topo_ptr = NULL;
MPIR_Dist_graph_topology *dist_graph_ptr = NULL;
int i;
int j;
int idx;
int comm_size = 0;
int in_capacity;
int out_capacity;
int **rout = NULL;
int **rin = NULL;
int *rin_sizes;
int *rout_sizes;
int *rin_idx;
int *rout_idx;
int *rs;
int in_out_peers[2] = {-1, -1};
int errflag = FALSE;
MPIU_CHKLMEM_DECL(9);
MPIU_CHKPMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPI_DIST_GRAPH_CREATE);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIU_THREAD_CS_ENTER(ALLFUNC,);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_DIST_GRAPH_CREATE);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COMM(comm_old, mpi_errno);
MPIR_ERRTEST_INFO_OR_NULL(info, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Convert MPI object handles to object pointers */
MPID_Comm_get_ptr(comm_old, comm_ptr);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
/* Validate comm_ptr */
MPID_Comm_valid_ptr(comm_ptr, mpi_errno);
/* If comm_ptr is not valid, it will be reset to null */
if (comm_ptr) {
MPIR_ERRTEST_COMM_INTRA(comm_ptr, mpi_errno);
}
MPIR_ERRTEST_ARGNEG(n, "n", mpi_errno);
if (n > 0) {
int have_degrees = 0;
MPIR_ERRTEST_ARGNULL(sources, "sources", mpi_errno);
MPIR_ERRTEST_ARGNULL(degrees, "degrees", mpi_errno);
for (i = 0; i < n; ++i) {
if (degrees[i]) {
have_degrees = 1;
break;
}
}
if (have_degrees) {
MPIR_ERRTEST_ARGNULL(destinations, "destinations", mpi_errno);
if (weights != MPI_UNWEIGHTED)
MPIR_ERRTEST_ARGNULL(weights, "weights", mpi_errno);
}
}
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
/* Implementation based on Torsten Hoefler's reference implementation
* attached to MPI-2.2 ticket #33. */
*comm_dist_graph = MPI_COMM_NULL;
comm_size = comm_ptr->local_size;
/* following the spirit of the old topo interface, attributes do not
* propagate to the new communicator (see MPI-2.1 pp. 243 line 11) */
mpi_errno = MPIR_Comm_copy(comm_ptr, comm_size, &comm_dist_graph_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_Assert(comm_dist_graph_ptr != NULL);
/* rin is an array of size comm_size containing pointers to arrays of
* rin_sizes[x]. rin[x] is locally known number of edges into this process
* from rank x.
*
* rout is an array of comm_size containing pointers to arrays of
* rout_sizes[x]. rout[x] is the locally known number of edges out of this
* process to rank x. */
MPIU_CHKLMEM_MALLOC(rout, int **, comm_size*sizeof(int*), mpi_errno, "rout");
MPIU_CHKLMEM_MALLOC(rin, int **, comm_size*sizeof(int*), mpi_errno, "rin");
MPIU_CHKLMEM_MALLOC(rin_sizes, int *, comm_size*sizeof(int), mpi_errno, "rin_sizes");
MPIU_CHKLMEM_MALLOC(rout_sizes, int *, comm_size*sizeof(int), mpi_errno, "rout_sizes");
MPIU_CHKLMEM_MALLOC(rin_idx, int *, comm_size*sizeof(int), mpi_errno, "rin_idx");
MPIU_CHKLMEM_MALLOC(rout_idx, int *, comm_size*sizeof(int), mpi_errno, "rout_idx");
memset(rout, 0, comm_size*sizeof(int*));
memset(rin, 0, comm_size*sizeof(int*));
memset(rin_sizes, 0, comm_size*sizeof(int));
memset(rout_sizes, 0, comm_size*sizeof(int));
memset(rin_idx, 0, comm_size*sizeof(int));
memset(rout_idx, 0, comm_size*sizeof(int));
/* compute array sizes */
idx = 0;
for (i = 0; i < n; ++i) {
MPIU_Assert(sources[i] < comm_size);
for (j = 0; j < degrees[i]; ++j) {
MPIU_Assert(destinations[idx] < comm_size);
/* rout_sizes[i] is twice as long as the number of edges to be
* sent to rank i by this process */
rout_sizes[sources[i]] += 2;
rin_sizes[destinations[idx]] += 2;
++idx;
}
}
/* allocate arrays */
for (i = 0; i < comm_size; ++i) {
/* can't use CHKLMEM macros b/c we are in a loop */
if (rin_sizes[i]) {
rin[i] = MPIU_Malloc(rin_sizes[i] * sizeof(int));
}
if (rout_sizes[i]) {
rout[i] = MPIU_Malloc(rout_sizes[i] * sizeof(int));
}
}
/* populate arrays */
idx = 0;
for (i = 0; i < n; ++i) {
/* TODO add this assert as proper error checking above */
int s_rank = sources[i];
MPIU_Assert(s_rank < comm_size);
MPIU_Assert(s_rank >= 0);
for (j = 0; j < degrees[i]; ++j) {
int d_rank = destinations[idx];
int weight = (weights == MPI_UNWEIGHTED ? 0 : weights[idx]);
/* TODO add this assert as proper error checking above */
MPIU_Assert(d_rank < comm_size);
MPIU_Assert(d_rank >= 0);
/* XXX DJG what about self-edges? do we need to drop one of these
* cases when there is a self-edge to avoid double-counting? */
/* rout[s][2*x] is the value of d for the j'th edge between (s,d)
* with weight rout[s][2*x+1], where x is the current end of the
* outgoing edge list for s. x==(rout_idx[s]/2) */
rout[s_rank][rout_idx[s_rank]++] = d_rank;
rout[s_rank][rout_idx[s_rank]++] = weight;
/* rin[d][2*x] is the value of s for the j'th edge between (s,d)
* with weight rout[d][2*x+1], where x is the current end of the
* incoming edge list for d. x==(rin_idx[d]/2) */
rin[d_rank][rin_idx[d_rank]++] = s_rank;
rin[d_rank][rin_idx[d_rank]++] = weight;
++idx;
}
}
for (i = 0; i < comm_size; ++i) {
/* sanity check that all arrays are fully populated*/
MPIU_Assert(rin_idx[i] == rin_sizes[i]);
MPIU_Assert(rout_idx[i] == rout_sizes[i]);
}
MPIU_CHKLMEM_MALLOC(rs, int *, 2*comm_size*sizeof(int), mpi_errno, "red-scat source buffer");
for (i = 0; i < comm_size; ++i) {
rs[2*i] = (rin_sizes[i] ? 1 : 0);
rs[2*i+1] = (rout_sizes[i] ? 1 : 0);
}
/* compute the number of peers I will recv from */
mpi_errno = MPIR_Reduce_scatter_block_impl(rs, in_out_peers, 2, MPI_INT, MPI_SUM, comm_ptr, &errflag);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
MPIU_Assert(in_out_peers[0] <= comm_size && in_out_peers[0] >= 0);
MPIU_Assert(in_out_peers[1] <= comm_size && in_out_peers[1] >= 0);
idx = 0;
/* must be 2*comm_size requests because we will possibly send inbound and
* outbound edges to everyone in our communicator */
MPIU_CHKLMEM_MALLOC(reqs, MPI_Request *, 2*comm_size*sizeof(MPI_Request), mpi_errno, "temp request array");
for (i = 0; i < comm_size; ++i) {
if (rin_sizes[i]) {
/* send edges where i is a destination to process i */
mpi_errno = MPIC_Isend(&rin[i][0], rin_sizes[i], MPI_INT, i, MPIR_TOPO_A_TAG, comm_old, &reqs[idx++]);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
if (rout_sizes[i]) {
/* send edges where i is a source to process i */
mpi_errno = MPIC_Isend(&rout[i][0], rout_sizes[i], MPI_INT, i, MPIR_TOPO_B_TAG, comm_old, &reqs[idx++]);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
MPIU_Assert(idx <= (2 * comm_size));
/* Create the topology structure */
MPIU_CHKPMEM_MALLOC(topo_ptr, MPIR_Topology *, sizeof(MPIR_Topology), mpi_errno, "topo_ptr");
topo_ptr->kind = MPI_DIST_GRAPH;
dist_graph_ptr = &topo_ptr->topo.dist_graph;
dist_graph_ptr->indegree = 0;
dist_graph_ptr->in = NULL;
dist_graph_ptr->in_weights = NULL;
dist_graph_ptr->outdegree = 0;
dist_graph_ptr->out = NULL;
dist_graph_ptr->out_weights = NULL;
dist_graph_ptr->is_weighted = (weights != MPI_UNWEIGHTED);
/* can't use CHKPMEM macros for this b/c we need to realloc */
in_capacity = 10; /* arbitrary */
dist_graph_ptr->in = MPIU_Malloc(in_capacity*sizeof(int));
if (dist_graph_ptr->is_weighted)
dist_graph_ptr->in_weights = MPIU_Malloc(in_capacity*sizeof(int));
out_capacity = 10; /* arbitrary */
dist_graph_ptr->out = MPIU_Malloc(out_capacity*sizeof(int));
if (dist_graph_ptr->is_weighted)
dist_graph_ptr->out_weights = MPIU_Malloc(out_capacity*sizeof(int));
for (i = 0; i < in_out_peers[0]; ++i) {
MPI_Status status;
int count;
int *buf;
/* receive inbound edges */
mpi_errno = MPIC_Probe(MPI_ANY_SOURCE, MPIR_TOPO_A_TAG, comm_old, &status);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIR_Get_count_impl(&status, MPI_INT, &count);
/* can't use CHKLMEM macros b/c we are in a loop */
buf = MPIU_Malloc(count*sizeof(int));
MPIU_ERR_CHKANDJUMP(!buf, mpi_errno, MPIR_ERR_RECOVERABLE, "**nomem");
mpi_errno = MPIC_Recv(buf, count, MPI_INT, MPI_ANY_SOURCE, MPIR_TOPO_A_TAG, comm_old, MPI_STATUS_IGNORE);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
for (j = 0; j < count/2; ++j) {
int deg = dist_graph_ptr->indegree++;
if (deg >= in_capacity) {
in_capacity *= 2;
MPIU_REALLOC_ORJUMP(dist_graph_ptr->in, in_capacity*sizeof(int), mpi_errno);
if (dist_graph_ptr->is_weighted)
MPIU_REALLOC_ORJUMP(dist_graph_ptr->in_weights, in_capacity*sizeof(int), mpi_errno);
}
dist_graph_ptr->in[deg] = buf[2*j];
if (dist_graph_ptr->is_weighted)
dist_graph_ptr->in_weights[deg] = buf[2*j+1];
}
MPIU_Free(buf);
}
for (i = 0; i < in_out_peers[1]; ++i) {
MPI_Status status;
int count;
int *buf;
/* receive outbound edges */
mpi_errno = MPIC_Probe(MPI_ANY_SOURCE, MPIR_TOPO_B_TAG, comm_old, &status);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIR_Get_count_impl(&status, MPI_INT, &count);
/* can't use CHKLMEM macros b/c we are in a loop */
buf = MPIU_Malloc(count*sizeof(int));
MPIU_ERR_CHKANDJUMP(!buf, mpi_errno, MPIR_ERR_RECOVERABLE, "**nomem");
mpi_errno = MPIC_Recv(buf, count, MPI_INT, MPI_ANY_SOURCE, MPIR_TOPO_B_TAG, comm_old, MPI_STATUS_IGNORE);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
for (j = 0; j < count/2; ++j) {
int deg = dist_graph_ptr->outdegree++;
if (deg >= out_capacity) {
out_capacity *= 2;
MPIU_REALLOC_ORJUMP(dist_graph_ptr->out, out_capacity*sizeof(int), mpi_errno);
if (dist_graph_ptr->is_weighted)
MPIU_REALLOC_ORJUMP(dist_graph_ptr->out_weights, out_capacity*sizeof(int), mpi_errno);
}
dist_graph_ptr->out[deg] = buf[2*j];
if (dist_graph_ptr->is_weighted)
dist_graph_ptr->out_weights[deg] = buf[2*j+1];
}
MPIU_Free(buf);
}
mpi_errno = MPIR_Waitall_impl(idx, reqs, MPI_STATUSES_IGNORE);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* remove any excess memory allocation */
MPIU_REALLOC_ORJUMP(dist_graph_ptr->in, dist_graph_ptr->indegree*sizeof(int), mpi_errno);
MPIU_REALLOC_ORJUMP(dist_graph_ptr->out, dist_graph_ptr->outdegree*sizeof(int), mpi_errno);
if (dist_graph_ptr->is_weighted) {
MPIU_REALLOC_ORJUMP(dist_graph_ptr->in_weights, dist_graph_ptr->indegree*sizeof(int), mpi_errno);
MPIU_REALLOC_ORJUMP(dist_graph_ptr->out_weights, dist_graph_ptr->outdegree*sizeof(int), mpi_errno);
}
mpi_errno = MPIR_Topology_put(comm_dist_graph_ptr, topo_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
MPIU_CHKPMEM_COMMIT();
MPIU_OBJ_PUBLISH_HANDLE(*comm_dist_graph, comm_dist_graph_ptr->handle);
/* ... end of body of routine ... */
fn_exit:
for (i = 0; i < comm_size; ++i) {
if (rin[i])
MPIU_Free(rin[i]);
if (rout[i])
MPIU_Free(rout[i]);
}
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_DIST_GRAPH_CREATE);
MPIU_THREAD_CS_EXIT(ALLFUNC,);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
if (dist_graph_ptr && dist_graph_ptr->in)
MPIU_Free(dist_graph_ptr->in);
if (dist_graph_ptr && dist_graph_ptr->in_weights)
MPIU_Free(dist_graph_ptr->in_weights);
if (dist_graph_ptr && dist_graph_ptr->out)
MPIU_Free(dist_graph_ptr->out);
if (dist_graph_ptr && dist_graph_ptr->out_weights)
MPIU_Free(dist_graph_ptr->out_weights);
MPIU_CHKPMEM_REAP();
#ifdef HAVE_ERROR_CHECKING
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER,
"**mpi_dist_graph_create", "**mpi_dist_graph_create %C %d %p %p %p %p %I %d %p",
comm_old, n, sources, degrees, destinations, weights, info, reorder, comm_dist_graph);
#endif
mpi_errno = MPIR_Err_return_comm(comm_ptr, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_PG_Create(int vct_sz, void * pg_id, MPIDI_PG_t ** pg_ptr)
{
MPIDI_PG_t * pg = NULL, *pgnext;
int p;
int mpi_errno = MPI_SUCCESS;
MPIU_CHKPMEM_DECL(2);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_PG_CREATE);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_PG_CREATE);
MPIU_CHKPMEM_MALLOC(pg,MPIDI_PG_t*,sizeof(MPIDI_PG_t),mpi_errno,"pg");
MPIU_CHKPMEM_MALLOC(pg->vct,MPIDI_VC_t *,sizeof(MPIDI_VC_t)*vct_sz,
mpi_errno,"pg->vct");
if (verbose) {
fprintf( stdout, "Creating a process group of size %d\n", vct_sz );
fflush(stdout);
}
pg->handle = 0;
/* The reference count indicates the number of vc's that are or
have been in use and not disconnected. It starts at zero,
except for MPI_COMM_WORLD. */
MPIU_Object_set_ref(pg, 0);
pg->size = vct_sz;
pg->id = pg_id;
pg->finalize = 0;
/* Initialize the connection information to null. Use
the appropriate MPIDI_PG_InitConnXXX routine to set up these
fields */
pg->connData = 0;
pg->getConnInfo = 0;
pg->connInfoToString = 0;
pg->connInfoFromString = 0;
pg->freeConnInfo = 0;
for (p = 0; p < vct_sz; p++)
{
/* Initialize device fields in the VC object */
MPIDI_VC_Init(&pg->vct[p], pg, p);
}
/* We may first need to initialize the channel before calling the channel
VC init functions. This routine may be a no-op; look in the
ch3_init.c file in each channel */
MPIDI_CH3_PG_Init(pg);
/* These are now done in MPIDI_VC_Init */
#if 0
for (p = 0; p < vct_sz; p++)
{
/* Initialize the channel fields in the VC object */
MPIDI_CH3_VC_Init( &pg->vct[p] );
}
#endif
/* The first process group is always the world group */
if (!pg_world) { pg_world = pg; }
/* Add pg's at the tail so that comm world is always the first pg */
pg->next = 0;
if (!MPIDI_PG_list)
{
MPIDI_PG_list = pg;
}
else
{
pgnext = MPIDI_PG_list;
while (pgnext->next)
{
pgnext = pgnext->next;
}
pgnext->next = pg;
}
*pg_ptr = pg;
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_PG_CREATE);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
static int init_default_collops(void)
{
int mpi_errno = MPI_SUCCESS;
int i;
struct MPID_Collops *ops = NULL;
MPIU_CHKPMEM_DECL(MPID_HIERARCHY_SIZE + 1);
/* first initialize the intracomms */
for (i = 0; i < MPID_HIERARCHY_SIZE; ++i) {
MPIU_CHKPMEM_CALLOC(ops, struct MPID_Collops *, sizeof(struct MPID_Collops), mpi_errno,
"default intracomm collops");
ops->ref_count = 1; /* force existence until finalize time */
/* intracomm default defaults... */
ops->Ibcast_sched = &MPIR_Ibcast_intra;
ops->Ibarrier_sched = &MPIR_Ibarrier_intra;
ops->Ireduce_sched = &MPIR_Ireduce_intra;
ops->Ialltoall_sched = &MPIR_Ialltoall_intra;
ops->Ialltoallv_sched = &MPIR_Ialltoallv_intra;
ops->Ialltoallw_sched = &MPIR_Ialltoallw_intra;
ops->Iallreduce_sched = &MPIR_Iallreduce_intra;
ops->Igather_sched = &MPIR_Igather_intra;
ops->Igatherv_sched = &MPIR_Igatherv;
ops->Iscatter_sched = &MPIR_Iscatter_intra;
ops->Iscatterv_sched = &MPIR_Iscatterv;
ops->Ireduce_scatter_sched = &MPIR_Ireduce_scatter_intra;
ops->Ireduce_scatter_block_sched = &MPIR_Ireduce_scatter_block_intra;
ops->Iallgather_sched = &MPIR_Iallgather_intra;
ops->Iallgatherv_sched = &MPIR_Iallgatherv_intra;
ops->Iscan_sched = &MPIR_Iscan_rec_dbl;
ops->Iexscan_sched = &MPIR_Iexscan;
ops->Neighbor_allgather = &MPIR_Neighbor_allgather_default;
ops->Neighbor_allgatherv = &MPIR_Neighbor_allgatherv_default;
ops->Neighbor_alltoall = &MPIR_Neighbor_alltoall_default;
ops->Neighbor_alltoallv = &MPIR_Neighbor_alltoallv_default;
ops->Neighbor_alltoallw = &MPIR_Neighbor_alltoallw_default;
ops->Ineighbor_allgather = &MPIR_Ineighbor_allgather_default;
ops->Ineighbor_allgatherv = &MPIR_Ineighbor_allgatherv_default;
ops->Ineighbor_alltoall = &MPIR_Ineighbor_alltoall_default;
ops->Ineighbor_alltoallv = &MPIR_Ineighbor_alltoallv_default;
ops->Ineighbor_alltoallw = &MPIR_Ineighbor_alltoallw_default;
/* override defaults, such as for SMP */
switch (i) {
case MPID_HIERARCHY_FLAT:
break;
case MPID_HIERARCHY_PARENT:
ops->Ibcast_sched = &MPIR_Ibcast_SMP;
ops->Iscan_sched = &MPIR_Iscan_SMP;
ops->Iallreduce_sched = &MPIR_Iallreduce_SMP;
ops->Ireduce_sched = &MPIR_Ireduce_SMP;
break;
case MPID_HIERARCHY_NODE:
break;
case MPID_HIERARCHY_NODE_ROOTS:
break;
/* --BEGIN ERROR HANDLING-- */
default:
MPIU_Assertp(FALSE);
break;
/* --END ERROR HANDLING-- */
}
/* this is a default table, it's not overriding another table */
ops->prev_coll_fns = NULL;
default_collops[i] = ops;
}
/* now the intercomm table */
{
MPIU_CHKPMEM_CALLOC(ops, struct MPID_Collops *, sizeof(struct MPID_Collops), mpi_errno,
"default intercomm collops");
ops->ref_count = 1; /* force existence until finalize time */
/* intercomm defaults */
ops->Ibcast_sched = &MPIR_Ibcast_inter;
ops->Ibarrier_sched = &MPIR_Ibarrier_inter;
ops->Ireduce_sched = &MPIR_Ireduce_inter;
ops->Ialltoall_sched = &MPIR_Ialltoall_inter;
ops->Ialltoallv_sched = &MPIR_Ialltoallv_inter;
ops->Ialltoallw_sched = &MPIR_Ialltoallw_inter;
ops->Iallreduce_sched = &MPIR_Iallreduce_inter;
ops->Igather_sched = &MPIR_Igather_inter;
ops->Igatherv_sched = &MPIR_Igatherv;
ops->Iscatter_sched = &MPIR_Iscatter_inter;
ops->Iscatterv_sched = &MPIR_Iscatterv;
ops->Ireduce_scatter_sched = &MPIR_Ireduce_scatter_inter;
ops->Ireduce_scatter_block_sched = &MPIR_Ireduce_scatter_block_inter;
ops->Iallgather_sched = &MPIR_Iallgather_inter;
ops->Iallgatherv_sched = &MPIR_Iallgatherv_inter;
/* scan and exscan are not valid for intercommunicators, leave them NULL */
/* Ineighbor_all* routines are not valid for intercommunicators, leave
* them NULL */
/* this is a default table, it's not overriding another table */
ops->prev_coll_fns = NULL;
ic_default_collops = ops;
}
/* run after MPID_Finalize to permit collective usage during finalize */
MPIR_Add_finalize(cleanup_default_collops, NULL, MPIR_FINALIZE_CALLBACK_PRIO - 1);
MPIU_CHKPMEM_COMMIT();
fn_exit:
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIR_Graph_create( MPID_Comm *comm_ptr, int nnodes,
const int indx[], const int edges[], int reorder,
MPI_Comm *comm_graph)
{
int mpi_errno = MPI_SUCCESS;
int i, nedges;
MPID_Comm *newcomm_ptr = NULL;
MPIR_Topology *graph_ptr = NULL;
MPIU_CHKPMEM_DECL(3);
/* Set this to null in case there is an error */
*comm_graph = MPI_COMM_NULL;
/* Create a new communicator */
if (reorder) {
int nrank;
/* Allow the cart map routine to remap the assignment of ranks to
processes */
mpi_errno = MPIR_Graph_map_impl(comm_ptr, nnodes, indx, edges, &nrank);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* Create the new communicator with split, since we need to reorder
the ranks (including the related internals, such as the connection
tables */
mpi_errno = MPIR_Comm_split_impl( comm_ptr,
nrank == MPI_UNDEFINED ? MPI_UNDEFINED : 1,
nrank, &newcomm_ptr );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
} else {
/* Just use the first nnodes processes in the communicator */
mpi_errno = MPIR_Comm_copy( (MPID_Comm *)comm_ptr, nnodes,
&newcomm_ptr );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
/* If this process is not in the resulting communicator, return a
null communicator and exit */
if (!newcomm_ptr) {
*comm_graph = MPI_COMM_NULL;
goto fn_exit;
}
nedges = indx[nnodes-1];
MPIU_CHKPMEM_MALLOC(graph_ptr,MPIR_Topology*,sizeof(MPIR_Topology),
mpi_errno,"graph_ptr");
graph_ptr->kind = MPI_GRAPH;
graph_ptr->topo.graph.nnodes = nnodes;
graph_ptr->topo.graph.nedges = nedges;
MPIU_CHKPMEM_MALLOC(graph_ptr->topo.graph.index,int*,
nnodes*sizeof(int),mpi_errno,"graph.index");
MPIU_CHKPMEM_MALLOC(graph_ptr->topo.graph.edges,int*,
nedges*sizeof(int),mpi_errno,"graph.edges");
for (i=0; i<nnodes; i++)
graph_ptr->topo.graph.index[i] = indx[i];
for (i=0; i<nedges; i++)
graph_ptr->topo.graph.edges[i] = edges[i];
/* Finally, place the topology onto the new communicator and return the
handle */
mpi_errno = MPIR_Topology_put( newcomm_ptr, graph_ptr );
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
MPIU_OBJ_PUBLISH_HANDLE(*comm_graph, newcomm_ptr->handle);
/* ... end of body of routine ... */
fn_exit:
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
MPIU_CHKPMEM_REAP();
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER,
"**mpi_graph_create",
"**mpi_graph_create %C %d %p %p %d %p", comm_ptr->handle,
nnodes, indx,
edges, reorder, comm_graph);
}
# endif
mpi_errno = MPIR_Err_return_comm( (MPID_Comm*)comm_ptr, FCNAME, mpi_errno );
goto fn_exit;
/* --END ERROR HANDLING-- */
}
PMPI_LOCAL int MPIR_Comm_create_calculate_mapping(MPID_Group *group_ptr,
MPID_Comm *comm_ptr,
MPID_VCR **mapping_vcr_out,
int **mapping_out)
{
int mpi_errno = MPI_SUCCESS;
int subsetOfWorld = 0;
int i, j;
int n;
int *mapping=0;
int vcr_size;
MPID_VCR *vcr;
MPIU_CHKPMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_CREATE_CALCULATE_MAPPING);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_CREATE_CALCULATE_MAPPING);
*mapping_out = NULL;
*mapping_vcr_out = NULL;
/* N.B. For intracomms only the comm_ptr->vcr is valid and populated,
* however local_size and remote_size are always set to the same value for
* intracomms. For intercomms both are valid and populated, with the
* local_vcr holding VCs corresponding to the local_group, local_comm, and
* local_size.
*
* For this mapping calculation we always want the logically local vcr,
* regardless of whether it is stored in the "plain" vcr or local_vcr. */
if (comm_ptr->comm_kind == MPID_INTERCOMM) {
vcr = comm_ptr->local_vcr;
vcr_size = comm_ptr->local_size;
}
else {
vcr = comm_ptr->vcr;
vcr_size = comm_ptr->remote_size;
}
n = group_ptr->size;
MPIU_CHKPMEM_MALLOC(mapping,int*,n*sizeof(int),mpi_errno,"mapping");
/* Make sure that the processes for this group are contained within
the input communicator. Also identify the mapping from the ranks of
the old communicator to the new communicator.
We do this by matching the lpids of the members of the group
with the lpids of the members of the input communicator.
It is an error if the group contains a reference to an lpid that
does not exist in the communicator.
An important special case is groups (and communicators) that
are subsets of MPI_COMM_WORLD. In this case, the lpids are
exactly the same as the ranks in comm world.
*/
/* we examine the group's lpids in both the intracomm and non-comm_world cases */
MPIR_Group_setup_lpid_list( group_ptr );
/* Optimize for groups contained within MPI_COMM_WORLD. */
if (comm_ptr->comm_kind == MPID_INTRACOMM) {
int wsize;
subsetOfWorld = 1;
wsize = MPIR_Process.comm_world->local_size;
for (i=0; i<n; i++) {
int g_lpid = group_ptr->lrank_to_lpid[i].lpid;
/* This mapping is relative to comm world */
MPIU_DBG_MSG_FMT(COMM,VERBOSE,
(MPIU_DBG_FDEST,
"comm-create - mapping into world[%d] = %d\n",
i, g_lpid ));
if (g_lpid < wsize) {
mapping[i] = g_lpid;
}
else {
subsetOfWorld = 0;
break;
}
}
}
MPIU_DBG_MSG_D(COMM,VERBOSE, "subsetOfWorld=%d", subsetOfWorld );
if (subsetOfWorld) {
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
int idx;
mpi_errno = MPIR_GroupCheckVCRSubset( group_ptr, vcr_size, vcr, &idx );
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Override the vcr to be used with the mapping array. */
vcr = MPIR_Process.comm_world->vcr;
vcr_size = MPIR_Process.comm_world->local_size;
}
else {
for (i=0; i<n; i++) {
/* mapping[i] is the rank in the communicator of the process
that is the ith element of the group */
/* FIXME : BUBBLE SORT */
mapping[i] = -1;
for (j=0; j<vcr_size; j++) {
int comm_lpid;
MPID_VCR_Get_lpid( vcr[j], &comm_lpid );
if (comm_lpid == group_ptr->lrank_to_lpid[i].lpid) {
mapping[i] = j;
break;
}
}
MPIU_ERR_CHKANDJUMP1(mapping[i] == -1,mpi_errno,MPI_ERR_GROUP,
"**groupnotincomm", "**groupnotincomm %d", i );
}
}
MPIU_Assert(vcr != NULL);
MPIU_Assert(mapping != NULL);
*mapping_vcr_out = vcr;
*mapping_out = mapping;
MPL_VG_CHECK_MEM_IS_DEFINED(*mapping_vcr_out, vcr_size * sizeof(**mapping_vcr_out));
MPL_VG_CHECK_MEM_IS_DEFINED(*mapping_out, n * sizeof(**mapping_out));
MPIU_CHKPMEM_COMMIT();
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_CREATE_CALCULATE_MAPPING);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
int MPIDI_CH3U_Win_gather_info(void *base, MPI_Aint size, int disp_unit,
MPID_Info * info, MPID_Comm * comm_ptr, MPID_Win ** win_ptr)
{
int mpi_errno = MPI_SUCCESS, i, k, comm_size, rank;
MPI_Aint *tmp_buf;
MPIR_Errflag_t errflag = MPIR_ERR_NONE;
MPIU_CHKPMEM_DECL(1);
MPIU_CHKLMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_WIN_GATHER_INFO);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_WIN_GATHER_INFO);
comm_size = (*win_ptr)->comm_ptr->local_size;
rank = (*win_ptr)->comm_ptr->rank;
MPIR_T_PVAR_TIMER_START(RMA, rma_wincreate_allgather);
/* allocate memory for the base addresses, disp_units, and
* completion counters of all processes */
MPIU_CHKPMEM_MALLOC((*win_ptr)->basic_info_table, MPIDI_Win_basic_info_t *,
comm_size * sizeof(MPIDI_Win_basic_info_t),
mpi_errno, "(*win_ptr)->basic_info_table");
/* get the addresses of the windows, window objects, and completion
* counters of all processes. allocate temp. buffer for communication */
MPIU_CHKLMEM_MALLOC(tmp_buf, MPI_Aint *, 4 * comm_size * sizeof(MPI_Aint),
mpi_errno, "tmp_buf");
/* FIXME: This needs to be fixed for heterogeneous systems */
/* FIXME: If we wanted to validate the transfer as within range at the
* origin, we'd also need the window size. */
tmp_buf[4 * rank] = MPIU_PtrToAint(base);
tmp_buf[4 * rank + 1] = size;
tmp_buf[4 * rank + 2] = (MPI_Aint) disp_unit;
tmp_buf[4 * rank + 3] = (MPI_Aint) (*win_ptr)->handle;
mpi_errno = MPIR_Allgather_impl(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL,
tmp_buf, 4, MPI_AINT, (*win_ptr)->comm_ptr, &errflag);
MPIR_T_PVAR_TIMER_END(RMA, rma_wincreate_allgather);
if (mpi_errno) {
MPIR_ERR_POP(mpi_errno);
}
MPIR_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
k = 0;
for (i = 0; i < comm_size; i++) {
(*win_ptr)->basic_info_table[i].base_addr = MPIU_AintToPtr(tmp_buf[k++]);
(*win_ptr)->basic_info_table[i].size = tmp_buf[k++];
(*win_ptr)->basic_info_table[i].disp_unit = (int) tmp_buf[k++];
(*win_ptr)->basic_info_table[i].win_handle = (MPI_Win) tmp_buf[k++];
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_WIN_GATHER_INFO);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_CH3U_Win_create_gather( void *base, MPI_Aint size, int disp_unit,
MPID_Info *info, MPID_Comm *comm_ptr, MPID_Win **win_ptr )
{
int mpi_errno=MPI_SUCCESS, i, k, comm_size, rank;
MPI_Aint *tmp_buf;
int errflag = FALSE;
MPIU_CHKPMEM_DECL(5);
MPIU_CHKLMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_WIN_CREATE_GATHER);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_WIN_CREATE_GATHER);
comm_size = (*win_ptr)->comm_ptr->local_size;
rank = (*win_ptr)->comm_ptr->rank;
/* RMA handlers should be set before calling this function */
mpi_errno = (*win_ptr)->RMAFns.Win_set_info(*win_ptr, info);
MPIU_INSTR_DURATION_START(wincreate_allgather);
/* allocate memory for the base addresses, disp_units, and
completion counters of all processes */
MPIU_CHKPMEM_MALLOC((*win_ptr)->base_addrs, void **,
comm_size*sizeof(void *),
mpi_errno, "(*win_ptr)->base_addrs");
MPIU_CHKPMEM_MALLOC((*win_ptr)->sizes, MPI_Aint *, comm_size*sizeof(MPI_Aint),
mpi_errno, "(*win_ptr)->sizes");
MPIU_CHKPMEM_MALLOC((*win_ptr)->disp_units, int *, comm_size*sizeof(int),
mpi_errno, "(*win_ptr)->disp_units");
MPIU_CHKPMEM_MALLOC((*win_ptr)->all_win_handles, MPI_Win *,
comm_size*sizeof(MPI_Win),
mpi_errno, "(*win_ptr)->all_win_handles");
MPIU_CHKPMEM_MALLOC((*win_ptr)->pt_rma_puts_accs, int *,
comm_size*sizeof(int),
mpi_errno, "(*win_ptr)->pt_rma_puts_accs");
for (i=0; i<comm_size; i++) (*win_ptr)->pt_rma_puts_accs[i] = 0;
/* get the addresses of the windows, window objects, and completion
counters of all processes. allocate temp. buffer for communication */
MPIU_CHKLMEM_MALLOC(tmp_buf, MPI_Aint *, 4*comm_size*sizeof(MPI_Aint),
mpi_errno, "tmp_buf");
/* FIXME: This needs to be fixed for heterogeneous systems */
/* FIXME: If we wanted to validate the transfer as within range at the
origin, we'd also need the window size. */
tmp_buf[4*rank] = MPIU_PtrToAint(base);
tmp_buf[4*rank+1] = size;
tmp_buf[4*rank+2] = (MPI_Aint) disp_unit;
tmp_buf[4*rank+3] = (MPI_Aint) (*win_ptr)->handle;
mpi_errno = MPIR_Allgather_impl(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL,
tmp_buf, 4, MPI_AINT,
(*win_ptr)->comm_ptr, &errflag);
MPIU_INSTR_DURATION_END(wincreate_allgather);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
k = 0;
for (i=0; i<comm_size; i++)
{
(*win_ptr)->base_addrs[i] = MPIU_AintToPtr(tmp_buf[k++]);
(*win_ptr)->sizes[i] = tmp_buf[k++];
(*win_ptr)->disp_units[i] = (int) tmp_buf[k++];
(*win_ptr)->all_win_handles[i] = (MPI_Win) tmp_buf[k++];
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_WIN_CREATE_GATHER);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
static int handler_recv_dequeue_large(const ptl_event_t *e)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *const rreq = e->user_ptr;
MPIDI_VC_t *vc;
MPID_nem_ptl_vc_area *vc_ptl;
int ret;
int dt_contig;
MPIDI_msg_sz_t data_sz;
MPID_Datatype *dt_ptr;
MPI_Aint dt_true_lb;
MPI_Aint last;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_HANDLER_RECV_DEQUEUE_LARGE);
MPIDI_FUNC_ENTER(MPID_STATE_HANDLER_RECV_DEQUEUE_LARGE);
MPIU_Assert(e->type == PTL_EVENT_PUT || e->type == PTL_EVENT_PUT_OVERFLOW);
MPIDI_Comm_get_vc(rreq->comm, NPTL_MATCH_GET_RANK(e->match_bits), &vc);
vc_ptl = VC_PTL(vc);
dequeue_req(e);
MPIDI_Datatype_get_info(rreq->dev.user_count, rreq->dev.datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
/* unpack data from unexpected buffer first */
if (e->type == PTL_EVENT_PUT_OVERFLOW) {
if (dt_contig) {
MPIU_Memcpy((char *)rreq->dev.user_buf + dt_true_lb, e->start, e->mlength);
} else {
last = e->mlength;
MPID_Segment_unpack(rreq->dev.segment_ptr, 0, &last, e->start);
MPIU_Assert(last == e->mlength);
rreq->dev.segment_first = e->mlength;
}
}
if (!(e->hdr_data & NPTL_LARGE)) {
/* all data has already been received; we're done */
mpi_errno = handler_recv_complete(e);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
goto fn_exit;
}
MPIU_Assert (e->mlength == PTL_LARGE_THRESHOLD);
/* we need to GET the rest of the data from the sender's buffer */
if (dt_contig) {
big_get((char *)rreq->dev.user_buf + dt_true_lb + PTL_LARGE_THRESHOLD, data_sz - PTL_LARGE_THRESHOLD,
vc, e->match_bits, rreq);
goto fn_exit;
}
/* noncontig recv buffer */
last = rreq->dev.segment_size;
rreq->dev.iov_count = MPL_IOV_LIMIT;
MPID_Segment_pack_vector(rreq->dev.segment_ptr, rreq->dev.segment_first, &last, rreq->dev.iov, &rreq->dev.iov_count);
if (last == rreq->dev.segment_size && rreq->dev.segment_size <= MPIDI_nem_ptl_ni_limits.max_msg_size + PTL_LARGE_THRESHOLD) {
/* Rest of message fits in one IOV */
ptl_md_t md;
md.start = rreq->dev.iov;
md.length = rreq->dev.iov_count;
md.options = PTL_IOVEC;
md.eq_handle = MPIDI_nem_ptl_origin_eq;
md.ct_handle = PTL_CT_NONE;
ret = PtlMDBind(MPIDI_nem_ptl_ni, &md, &REQ_PTL(rreq)->md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdbind", "**ptlmdbind %s", MPID_nem_ptl_strerror(ret));
REQ_PTL(rreq)->event_handler = handler_recv_complete;
ret = MPID_nem_ptl_rptl_get(REQ_PTL(rreq)->md, 0, rreq->dev.segment_size - rreq->dev.segment_first, vc_ptl->id, vc_ptl->ptg,
e->match_bits, 0, rreq);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlget", "**ptlget %s", MPID_nem_ptl_strerror(ret));
goto fn_exit;
}
/* message won't fit in a single IOV, allocate buffer and unpack when received */
/* FIXME: For now, allocate a single large buffer to hold entire message */
MPIU_CHKPMEM_MALLOC(REQ_PTL(rreq)->chunk_buffer[0], void *, data_sz - PTL_LARGE_THRESHOLD,
mpi_errno, "chunk_buffer");
big_get(REQ_PTL(rreq)->chunk_buffer[0], data_sz - PTL_LARGE_THRESHOLD, vc, e->match_bits, rreq);
fn_exit:
MPIU_CHKPMEM_COMMIT();
fn_exit2:
MPIDI_FUNC_EXIT(MPID_STATE_HANDLER_RECV_DEQUEUE_LARGE);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit2;
}
int MPIDI_Win_create(void *base, MPI_Aint size, int disp_unit, MPID_Info *info,
MPID_Comm *comm_ptr, MPID_Win **win_ptr )
{
int mpi_errno=MPI_SUCCESS, i, k, comm_size, rank;
MPI_Aint *tmp_buf;
MPID_Comm *win_comm_ptr;
int errflag = FALSE;
MPIU_CHKPMEM_DECL(4);
MPIU_CHKLMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_WIN_CREATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_WIN_CREATE);
/* FIXME: There should be no unreferenced args */
MPIU_UNREFERENCED_ARG(info);
if(initRMAoptions) {
int rc;
MPIU_THREADSAFE_INIT_BLOCK_BEGIN(initRMAoptions);
/* Default is to enable the use of the immediate accumulate feature */
if (!MPL_env2bool( "MPICH_RMA_ACC_IMMED", &rc ))
rc = 1;
MPIDI_CH3_RMA_SetAccImmed(rc);
#ifdef USE_MPIU_INSTR
/* Define all instrumentation handle used in the CH3 RMA here*/
MPIU_INSTR_DURATION_INIT(wincreate_allgather,0,"WIN_CREATE:Allgather");
MPIU_INSTR_DURATION_INIT(winfree_rs,0,"WIN_FREE:ReduceScatterBlock");
MPIU_INSTR_DURATION_INIT(winfree_complete,0,"WIN_FREE:Complete");
MPIU_INSTR_DURATION_INIT(rmaqueue_alloc,0,"Allocate RMA Queue element");
MPIDI_CH3_RMA_InitInstr();
#endif
MPIU_THREADSAFE_INIT_CLEAR(initRMAoptions);
MPIU_THREADSAFE_INIT_BLOCK_END(initRMAoptions);
}
comm_size = comm_ptr->local_size;
rank = comm_ptr->rank;
*win_ptr = (MPID_Win *)MPIU_Handle_obj_alloc( &MPID_Win_mem );
MPIU_ERR_CHKANDJUMP1(!(*win_ptr),mpi_errno,MPI_ERR_OTHER,"**nomem",
"**nomem %s","MPID_Win_mem");
MPIU_Object_set_ref(*win_ptr, 1);
(*win_ptr)->fence_cnt = 0;
(*win_ptr)->base = base;
(*win_ptr)->size = size;
(*win_ptr)->disp_unit = disp_unit;
(*win_ptr)->start_group_ptr = NULL;
(*win_ptr)->start_assert = 0;
(*win_ptr)->attributes = NULL;
(*win_ptr)->rma_ops_list_head = NULL;
(*win_ptr)->rma_ops_list_tail = NULL;
(*win_ptr)->lock_granted = 0;
(*win_ptr)->current_lock_type = MPID_LOCK_NONE;
(*win_ptr)->shared_lock_ref_cnt = 0;
(*win_ptr)->lock_queue = NULL;
(*win_ptr)->my_counter = 0;
(*win_ptr)->my_pt_rma_puts_accs = 0;
mpi_errno = MPIR_Comm_dup_impl(comm_ptr, &win_comm_ptr);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
(*win_ptr)->comm_ptr = win_comm_ptr;
(*win_ptr)->myrank = rank;
MPIU_INSTR_DURATION_START(wincreate_allgather);
/* allocate memory for the base addresses, disp_units, and
completion counters of all processes */
MPIU_CHKPMEM_MALLOC((*win_ptr)->base_addrs, void **,
comm_size*sizeof(void *),
mpi_errno, "(*win_ptr)->base_addrs");
MPIU_CHKPMEM_MALLOC((*win_ptr)->disp_units, int *, comm_size*sizeof(int),
mpi_errno, "(*win_ptr)->disp_units");
MPIU_CHKPMEM_MALLOC((*win_ptr)->all_win_handles, MPI_Win *,
comm_size*sizeof(MPI_Win),
mpi_errno, "(*win_ptr)->all_win_handles");
MPIU_CHKPMEM_MALLOC((*win_ptr)->pt_rma_puts_accs, int *,
comm_size*sizeof(int),
mpi_errno, "(*win_ptr)->pt_rma_puts_accs");
for (i=0; i<comm_size; i++) (*win_ptr)->pt_rma_puts_accs[i] = 0;
/* get the addresses of the windows, window objects, and completion
counters of all processes. allocate temp. buffer for communication */
MPIU_CHKLMEM_MALLOC(tmp_buf, MPI_Aint *, 3*comm_size*sizeof(MPI_Aint),
mpi_errno, "tmp_buf");
/* FIXME: This needs to be fixed for heterogeneous systems */
tmp_buf[3*rank] = MPIU_PtrToAint(base);
tmp_buf[3*rank+1] = (MPI_Aint) disp_unit;
tmp_buf[3*rank+2] = (MPI_Aint) (*win_ptr)->handle;
mpi_errno = MPIR_Allgather_impl(MPI_IN_PLACE, 0, MPI_DATATYPE_NULL,
tmp_buf, 3 * sizeof(MPI_Aint), MPI_BYTE,
comm_ptr, &errflag);
MPIU_INSTR_DURATION_END(wincreate_allgather);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
MPIU_ERR_CHKANDJUMP(errflag, mpi_errno, MPI_ERR_OTHER, "**coll_fail");
k = 0;
for (i=0; i<comm_size; i++)
{
(*win_ptr)->base_addrs[i] = MPIU_AintToPtr(tmp_buf[k++]);
(*win_ptr)->disp_units[i] = (int) tmp_buf[k++];
(*win_ptr)->all_win_handles[i] = (MPI_Win) tmp_buf[k++];
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_WIN_CREATE);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIR_Comm_create_calculate_mapping(MPID_Group *group_ptr,
MPID_Comm *comm_ptr,
int **mapping_out,
MPID_Comm **mapping_comm)
{
int mpi_errno = MPI_SUCCESS;
int subsetOfWorld = 0;
int i, j;
int n;
int *mapping=0;
MPIU_CHKPMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_COMM_CREATE_CALCULATE_MAPPING);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_COMM_CREATE_CALCULATE_MAPPING);
*mapping_out = NULL;
*mapping_comm = comm_ptr;
n = group_ptr->size;
MPIU_CHKPMEM_MALLOC(mapping,int*,n*sizeof(int),mpi_errno,"mapping");
/* Make sure that the processes for this group are contained within
the input communicator. Also identify the mapping from the ranks of
the old communicator to the new communicator.
We do this by matching the lpids of the members of the group
with the lpids of the members of the input communicator.
It is an error if the group contains a reference to an lpid that
does not exist in the communicator.
An important special case is groups (and communicators) that
are subsets of MPI_COMM_WORLD. In this case, the lpids are
exactly the same as the ranks in comm world.
*/
/* we examine the group's lpids in both the intracomm and non-comm_world cases */
MPIR_Group_setup_lpid_list( group_ptr );
/* Optimize for groups contained within MPI_COMM_WORLD. */
if (comm_ptr->comm_kind == MPID_INTRACOMM) {
int wsize;
subsetOfWorld = 1;
wsize = MPIR_Process.comm_world->local_size;
for (i=0; i<n; i++) {
int g_lpid = group_ptr->lrank_to_lpid[i].lpid;
/* This mapping is relative to comm world */
MPL_DBG_MSG_FMT(MPIR_DBG_COMM,VERBOSE,
(MPL_DBG_FDEST,
"comm-create - mapping into world[%d] = %d",
i, g_lpid ));
if (g_lpid < wsize) {
mapping[i] = g_lpid;
}
else {
subsetOfWorld = 0;
break;
}
}
}
MPL_DBG_MSG_D(MPIR_DBG_COMM,VERBOSE, "subsetOfWorld=%d", subsetOfWorld );
if (subsetOfWorld) {
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
mpi_errno = MPIR_Group_check_subset( group_ptr, comm_ptr );
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
MPID_END_ERROR_CHECKS;
}
# endif
/* Override the comm to be used with the mapping array. */
*mapping_comm = MPIR_Process.comm_world;
}
else {
for (i=0; i<n; i++) {
/* mapping[i] is the rank in the communicator of the process
that is the ith element of the group */
/* FIXME : BUBBLE SORT */
mapping[i] = -1;
for (j=0; j<comm_ptr->local_size; j++) {
int comm_lpid;
MPID_Comm_get_lpid( comm_ptr, j, &comm_lpid, FALSE );
if (comm_lpid == group_ptr->lrank_to_lpid[i].lpid) {
mapping[i] = j;
break;
}
}
MPIR_ERR_CHKANDJUMP1(mapping[i] == -1,mpi_errno,MPI_ERR_GROUP,
"**groupnotincomm", "**groupnotincomm %d", i );
}
}
MPIU_Assert(mapping != NULL);
*mapping_out = mapping;
MPL_VG_CHECK_MEM_IS_DEFINED(*mapping_out, n * sizeof(**mapping_out));
MPIU_CHKPMEM_COMMIT();
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_COMM_CREATE_CALCULATE_MAPPING);
return mpi_errno;
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
}
int MPIDI_Get(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_ptr)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_msg_sz_t data_sz;
int dt_contig, rank, predefined;
MPI_Aint dt_true_lb;
MPIDI_RMA_ops *new_ptr;
MPID_Datatype *dtp;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_GET);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_GET);
MPIDI_Datatype_get_info(origin_count, origin_datatype,
dt_contig, data_sz, dtp, dt_true_lb);
if ((data_sz == 0) || (target_rank == MPI_PROC_NULL))
{
goto fn_exit;
}
rank = win_ptr->myrank;
/* If the get is a local operation, do it here */
if (target_rank == rank)
{
mpi_errno = MPIR_Localcopy((char *) win_ptr->base +
win_ptr->disp_unit * target_disp,
target_count, target_datatype,
origin_addr, origin_count,
origin_datatype);
}
else
{
/* queue it up */
MPIU_INSTR_DURATION_START(rmaqueue_alloc);
MPIU_CHKPMEM_MALLOC(new_ptr, MPIDI_RMA_ops *, sizeof(MPIDI_RMA_ops),
mpi_errno, "RMA operation entry");
MPIU_INSTR_DURATION_END(rmaqueue_alloc);
if (win_ptr->rma_ops_list_tail)
win_ptr->rma_ops_list_tail->next = new_ptr;
else
win_ptr->rma_ops_list_head = new_ptr;
win_ptr->rma_ops_list_tail = new_ptr;
/* FIXME: For contig and very short operations, use a streamlined op */
new_ptr->next = NULL;
new_ptr->type = MPIDI_RMA_GET;
new_ptr->origin_addr = origin_addr;
new_ptr->origin_count = origin_count;
new_ptr->origin_datatype = origin_datatype;
new_ptr->target_rank = target_rank;
new_ptr->target_disp = target_disp;
new_ptr->target_count = target_count;
new_ptr->target_datatype = target_datatype;
/* if source or target datatypes are derived, increment their
reference counts */
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, predefined);
if (!predefined)
{
MPID_Datatype_get_ptr(origin_datatype, dtp);
MPID_Datatype_add_ref(dtp);
}
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, predefined);
if (!predefined)
{
MPID_Datatype_get_ptr(target_datatype, dtp);
MPID_Datatype_add_ref(dtp);
}
}
fn_exit:
MPIDI_RMA_FUNC_EXIT(MPID_STATE_MPIDI_GET);
return mpi_errno;
/* --BEGIN ERROR HANDLING-- */
fn_fail:
MPIU_CHKPMEM_REAP();
goto fn_exit;
/* --END ERROR HANDLING-- */
}
