static int send_sreq_data(MPIDI_VC_t *vc, MPID_Request *sreq, knem_cookie_t *s_cookiep)
{
int mpi_errno = MPI_SUCCESS;
int dt_contig;
MPI_Aint dt_true_lb;
MPIDI_msg_sz_t data_sz;
MPID_Datatype * dt_ptr;
/* MT: this code assumes only one thread can be at this point at a time */
if (knem_fd < 0) {
mpi_errno = open_knem_dev();
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
/* find out contig/noncontig, size, and lb for the datatype */
MPIDI_Datatype_get_info(sreq->dev.user_count, sreq->dev.datatype,
dt_contig, data_sz, dt_ptr, dt_true_lb);
if (dt_contig) {
/* handle the iov creation ourselves */
sreq->dev.iov[0].MPID_IOV_BUF = (char *)sreq->dev.user_buf + dt_true_lb;
sreq->dev.iov[0].MPID_IOV_LEN = data_sz;
sreq->dev.iov_count = 1;
}
else {
/* use the segment routines to handle the iovec creation */
if (sreq->dev.segment_ptr == NULL) {
sreq->dev.iov_count = MPID_IOV_LIMIT;
sreq->dev.iov_offset = 0;
/* segment_ptr may be non-null when this is a continuation of a
many-part message that we couldn't fit in one single flight of
iovs. */
sreq->dev.segment_ptr = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((sreq->dev.segment_ptr == NULL), mpi_errno,
MPI_ERR_OTHER, "**nomem",
"**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(sreq->dev.user_buf, sreq->dev.user_count,
sreq->dev.datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
/* FIXME we should write our own function that isn't dependent on
the in-request iov array. This will let us use IOVs that are
larger than MPID_IOV_LIMIT. */
mpi_errno = MPIDI_CH3U_Request_load_send_iov(sreq, &sreq->dev.iov[0],
&sreq->dev.iov_count);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
mpi_errno = do_dma_send(vc, sreq, sreq->dev.iov_count, sreq->dev.iov, s_cookiep);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
void
MPIDI_Win_datatype_map(MPIDI_Datatype * dt)
{
if (dt->contig)
{
dt->num_contig = 1;
dt->map = &dt->__map;
dt->map[0].DLOOP_VECTOR_BUF = (void*)(size_t)dt->true_lb;
dt->map[0].DLOOP_VECTOR_LEN = dt->size;
}
else
{
unsigned map_size = dt->pointer->max_contig_blocks*dt->count + 1;
dt->num_contig = map_size;
dt->map = (DLOOP_VECTOR*)MPIU_Malloc(map_size * sizeof(DLOOP_VECTOR));
MPID_assert(dt->map != NULL);
DLOOP_Offset last = dt->pointer->size*dt->count;
MPID_Segment seg;
MPID_Segment_init(NULL, dt->count, dt->type, &seg, 0);
MPID_Segment_pack_vector(&seg, 0, &last, dt->map, &dt->num_contig);
MPID_assert((unsigned)dt->num_contig <= map_size);
#ifdef TRACE_ON
TRACE_ERR("dt->pointer->size=%d num_contig: orig=%u new=%d\n", dt->pointer->size, map_size, dt->num_contig);
int i;
for(i=0; i<dt->num_contig; ++i)
TRACE_ERR(" %d: BUF=%zu LEN=%zu\n", i, (size_t)dt->map[i].DLOOP_VECTOR_BUF, (size_t)dt->map[i].DLOOP_VECTOR_LEN);
#endif
}
}
/* MPIDI_CH3_EagerNoncontigSend - Eagerly send noncontiguous data */
int MPIDI_CH3_EagerNoncontigSend( MPID_Request **sreq_p,
MPIDI_CH3_Pkt_type_t reqtype,
const void * buf, MPI_Aint count,
MPI_Datatype datatype, MPIDI_msg_sz_t data_sz,
int rank,
int tag, MPID_Comm * comm,
int context_offset )
{
int mpi_errno = MPI_SUCCESS;
MPIDI_VC_t * vc;
MPID_Request *sreq = *sreq_p;
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_eager_send_t * const eager_pkt = &upkt.eager_send;
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"sending non-contiguous eager message, data_sz=" MPIDI_MSG_SZ_FMT,
data_sz));
sreq->dev.OnDataAvail = 0;
sreq->dev.OnFinal = 0;
MPIDI_Pkt_init(eager_pkt, reqtype);
eager_pkt->match.parts.rank = comm->rank;
eager_pkt->match.parts.tag = tag;
eager_pkt->match.parts.context_id = comm->context_id + context_offset;
eager_pkt->sender_req_id = MPI_REQUEST_NULL;
eager_pkt->data_sz = data_sz;
MPIDI_Comm_get_vc_set_active(comm, rank, &vc);
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Pkt_set_seqnum(eager_pkt, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
MPIU_DBG_MSGPKT(vc,tag,eager_pkt->match.parts.context_id,rank,data_sz,
"Eager");
sreq->dev.segment_ptr = MPID_Segment_alloc( );
MPIR_ERR_CHKANDJUMP1((sreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
MPID_THREAD_CS_ENTER(POBJ, vc->pobj_mutex);
mpi_errno = vc->sendNoncontig_fn(vc, sreq, eager_pkt,
sizeof(MPIDI_CH3_Pkt_eager_send_t));
MPID_THREAD_CS_EXIT(POBJ, vc->pobj_mutex);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
fn_exit:
return mpi_errno;
fn_fail:
*sreq_p = NULL;
goto fn_exit;
}
int MPID_nem_ib_lmt_done_recv(struct MPIDI_VC *vc, struct MPID_Request *rreq)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_IB_LMT_DONE_RECV);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_IB_LMT_DONE_RECV);
dprintf("lmt_done_recv,enter,rreq=%p,head=%p\n", rreq, MPID_nem_ib_lmtq.head);
int is_contig;
MPID_Datatype_is_contig(rreq->dev.datatype, &is_contig);
if (!is_contig) {
dprintf("lmt_done_recv,copying noncontiguous data to user buffer\n");
/* see MPIDI_CH3U_Request_unpack_uebuf (in /src/mpid/ch3/src/ch3u_request.c) */
/* or MPIDI_CH3U_Receive_data_found (in src/mpid/ch3/src/ch3u_handle_recv_pkt.c) */
MPIDI_msg_sz_t unpack_sz = rreq->ch.lmt_data_sz;
MPID_Segment seg;
MPI_Aint last;
MPID_Segment_init(rreq->dev.user_buf, rreq->dev.user_count, rreq->dev.datatype, &seg, 0);
last = unpack_sz;
MPID_Segment_unpack(&seg, 0, &last, REQ_FIELD(rreq, lmt_pack_buf));
if (last != unpack_sz) {
/* --BEGIN ERROR HANDLING-- */
/* received data was not entirely consumed by unpack()
* because too few bytes remained to fill the next basic
* datatype */
MPIR_STATUS_SET_COUNT(rreq->status, last);
rreq->status.MPI_ERROR =
MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__,
MPI_ERR_TYPE, "**MPID_nem_ib_lmt_done_recv", 0);
/* --END ERROR HANDLING-- */
}
//MPIU_Free(REQ_FIELD(rreq, lmt_pack_buf));
MPID_nem_ib_stfree(REQ_FIELD(rreq, lmt_pack_buf), (size_t) rreq->ch.lmt_data_sz);
}
dprintf("lmt_done_recv,1,req=%p,pcc=%d\n", rreq, MPIDI_CH3I_progress_completion_count.v);
MPIDI_CH3U_Request_complete(rreq);
dprintf("lmt_done_recv,complete,req=%p\n", rreq);
dprintf("lmt_done_recv,2,pcc=%d\n", MPIDI_CH3I_progress_completion_count.v);
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_IB_LMT_DONE_RECV);
return mpi_errno;
//fn_fail:
goto fn_exit;
}
/* fills in req->dev.iov{,_offset,_count} based on the datatype info in the
request, creating a segment if necessary */
static int populate_iov_from_req(MPID_Request *req)
{
int mpi_errno = MPI_SUCCESS;
int dt_contig;
MPI_Aint dt_true_lb;
MPIDI_msg_sz_t data_sz;
MPID_Datatype * dt_ptr;
/* find out contig/noncontig, size, and lb for the datatype */
MPIDI_Datatype_get_info(req->dev.user_count, req->dev.datatype,
dt_contig, data_sz, dt_ptr, dt_true_lb);
if (dt_contig) {
/* handle the iov creation ourselves */
req->dev.iov[0].MPL_IOV_BUF = (char *)req->dev.user_buf + dt_true_lb;
req->dev.iov[0].MPL_IOV_LEN = data_sz;
req->dev.iov_count = 1;
}
else {
/* use the segment routines to handle the iovec creation */
MPIU_Assert(req->dev.segment_ptr == NULL);
req->dev.iov_count = MPL_IOV_LIMIT;
req->dev.iov_offset = 0;
/* XXX DJG FIXME where is this segment freed? */
req->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1((req->dev.segment_ptr == NULL), mpi_errno,
MPI_ERR_OTHER, "**nomem",
"**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(req->dev.user_buf, req->dev.user_count,
req->dev.datatype, req->dev.segment_ptr, 0);
req->dev.segment_first = 0;
req->dev.segment_size = data_sz;
/* FIXME we should write our own function that isn't dependent on
the in-request iov array. This will let us use IOVs that are
larger than MPL_IOV_LIMIT. */
mpi_errno = MPIDI_CH3U_Request_load_send_iov(req, &req->dev.iov[0],
&req->dev.iov_count);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
fn_fail:
return mpi_errno;
}
static int _mxm_process_rdtype(MPID_Request ** rreq_p, MPI_Datatype datatype,
MPID_Datatype * dt_ptr, MPIDI_msg_sz_t data_sz, const void *buf,
int count, mxm_req_buffer_t ** iov_buf, int *iov_count)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *rreq = *rreq_p;
MPIDI_msg_sz_t last;
MPL_IOV *iov;
int n_iov = 0;
int index;
if (rreq->dev.segment_ptr == NULL) {
rreq->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1((rreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem",
"**nomem %s", "MPID_Segment_alloc");
}
MPID_Segment_init(buf, count, datatype, rreq->dev.segment_ptr, 0);
rreq->dev.segment_first = 0;
rreq->dev.segment_size = data_sz;
last = rreq->dev.segment_size;
MPID_Segment_count_contig_blocks(rreq->dev.segment_ptr, rreq->dev.segment_first, &last,
(MPI_Aint *) & n_iov);
MPIU_Assert(n_iov > 0);
iov = MPIU_Malloc(n_iov * sizeof(*iov));
MPIU_Assert(iov);
last = rreq->dev.segment_size;
MPID_Segment_unpack_vector(rreq->dev.segment_ptr, rreq->dev.segment_first, &last, iov, &n_iov);
MPIU_Assert(last == rreq->dev.segment_size);
#if defined(MXM_DEBUG) && (MXM_DEBUG > 0)
_dbg_mxm_output(7, "Recv Noncontiguous data vector %i entries (free slots : %i)\n", n_iov,
MXM_REQ_DATA_MAX_IOV);
for (index = 0; index < n_iov; index++) {
_dbg_mxm_output(7, "======= Recv iov[%i] = ptr : %p, len : %i \n",
index, iov[index].MPL_IOV_BUF, iov[index].MPL_IOV_LEN);
}
#endif
if (n_iov <= MXM_REQ_DATA_MAX_IOV) {
if (n_iov > MXM_MPICH_MAX_IOV) {
*iov_buf = (mxm_req_buffer_t *) MPIU_Malloc(n_iov * sizeof(**iov_buf));
MPIU_Assert(*iov_buf);
}
for (index = 0; index < n_iov; index++) {
(*iov_buf)[index].ptr = iov[index].MPL_IOV_BUF;
(*iov_buf)[index].length = iov[index].MPL_IOV_LEN;
}
rreq->dev.tmpbuf = NULL;
rreq->dev.tmpbuf_sz = 0;
*iov_count = n_iov;
}
else {
MPI_Aint packsize = 0;
MPIR_Pack_size_impl(rreq->dev.user_count, rreq->dev.datatype, &packsize);
rreq->dev.tmpbuf = MPIU_Malloc((size_t) packsize);
MPIU_Assert(rreq->dev.tmpbuf);
rreq->dev.tmpbuf_sz = packsize;
(*iov_buf)[0].ptr = rreq->dev.tmpbuf;
(*iov_buf)[0].length = (size_t) packsize;
*iov_count = 1;
}
MPIU_Free(iov);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int
MPIDO_Allgather_simple(const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
int recvcount,
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
/* *********************************
* Check the nature of the buffers
* *********************************
*/
const struct MPIDI_Comm* const mpid = &(comm_ptr->mpid);
MPID_Datatype * dt_null = NULL;
void *snd_noncontig_buff = NULL, *rcv_noncontig_buff = NULL;
MPI_Aint send_true_lb = 0;
MPI_Aint recv_true_lb = 0;
int snd_data_contig = 1, rcv_data_contig = 1;
size_t send_size = 0;
size_t recv_size = 0;
MPID_Segment segment;
volatile unsigned allgather_active = 1;
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 pami_metadata_t *my_md;
char *rbuf = NULL, *sbuf = NULL;
if ((sendcount < 1 && sendbuf != MPI_IN_PLACE) || recvcount < 1)
return MPI_SUCCESS;
/* Gather datatype information */
MPIDI_Datatype_get_info(recvcount,
recvtype,
rcv_data_contig,
recv_size,
dt_null,
recv_true_lb);
send_size = recv_size;
if(MPIDI_Pamix_collsel_advise != NULL && mpid->collsel_fast_query != NULL)
{
advisor_algorithm_t advisor_algorithms[1];
int num_algorithms = MPIDI_Pamix_collsel_advise(mpid->collsel_fast_query, PAMI_XFER_ALLGATHER, send_size, advisor_algorithms, 1);
if(num_algorithms)
{
if(advisor_algorithms[0].algorithm_type == COLLSEL_EXTERNAL_ALGO)
{
return MPIR_Allgather(sendbuf, sendcount, sendtype,
recvbuf, recvcount, recvtype,
comm_ptr, mpierrno);
}
else if(advisor_algorithms[0].metadata && advisor_algorithms[0].metadata->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", advisor_algorithms[0].metadata->name);
MPIDO_Barrier(comm_ptr, &tmpmpierrno);
}
}
}
rbuf = (char *)recvbuf+recv_true_lb;
if(!rcv_data_contig)
{
rcv_noncontig_buff = MPL_malloc(recv_size * size);
rbuf = rcv_noncontig_buff;
if(rcv_noncontig_buff == NULL)
{
MPID_Abort(NULL, MPI_ERR_NO_SPACE, 1,
"Fatal: Cannot allocate pack buffer");
}
if(sendbuf == MPI_IN_PLACE)
{
sbuf = PAMI_IN_PLACE;
size_t extent;
MPID_Datatype_get_extent_macro(recvtype,extent);
MPIR_Localcopy(recvbuf + (rank*recvcount*extent), recvcount, recvtype,
rcv_noncontig_buff + (rank*recv_size), recv_size,MPI_CHAR);
}
}
if(sendbuf != MPI_IN_PLACE)
{
MPIDI_Datatype_get_info(sendcount,
sendtype,
snd_data_contig,
send_size,
dt_null,
send_true_lb);
sbuf = (char *)sendbuf+send_true_lb;
if(!snd_data_contig)
{
snd_noncontig_buff = MPL_malloc(send_size);
sbuf = snd_noncontig_buff;
if(snd_noncontig_buff == NULL)
{
MPID_Abort(NULL, MPI_ERR_NO_SPACE, 1,
"Fatal: Cannot allocate pack buffer");
}
DLOOP_Offset last = send_size;
MPID_Segment_init(sendbuf, sendcount, sendtype, &segment, 0);
MPID_Segment_pack(&segment, 0, &last, snd_noncontig_buff);
}
}
else
sbuf = PAMI_IN_PLACE;
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;/* stype is ignored when sndbuf == PAMI_IN_PLACE */
allgather.cmd.xfer_allgather.rtype = PAMI_TYPE_BYTE;
allgather.cmd.xfer_allgather.stypecount = send_size;
allgather.cmd.xfer_allgather.rtypecount = recv_size;
allgather.algorithm = mpid->coll_algorithm[PAMI_XFER_ALLGATHER][0][0];
my_md = &mpid->coll_metadata[PAMI_XFER_ALLGATHER][0][0];
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);
if(!rcv_data_contig)
{
MPIR_Localcopy(rcv_noncontig_buff, recv_size * size, MPI_CHAR,
recvbuf, recvcount, recvtype);
MPL_free(rcv_noncontig_buff);
}
if(!snd_data_contig) MPL_free(snd_noncontig_buff);
TRACE_ERR("Allgather done\n");
return MPI_SUCCESS;
}
int MPID_nem_ib_lmt_switch_send(struct MPIDI_VC *vc, struct MPID_Request *req)
{
int mpi_errno = MPI_SUCCESS;
int dt_contig;
MPIDI_msg_sz_t data_sz;
MPID_Datatype *dt_ptr;
MPI_Aint dt_true_lb;
MPID_IOV r_cookie = req->ch.lmt_tmp_cookie;
MPID_nem_ib_lmt_cookie_t *r_cookie_buf = r_cookie.iov_base;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_IB_LMT_SWITCH_SEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_IB_LMT_SWITCH_SEND);
MPIDI_Datatype_get_info(req->dev.user_count, req->dev.datatype, dt_contig, data_sz, dt_ptr,
dt_true_lb);
void *write_from_buf;
if (dt_contig) {
write_from_buf = req->dev.user_buf;
}
else {
/* see MPIDI_CH3_EagerNoncontigSend (in ch3u_eager.c) */
req->dev.segment_ptr = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP((req->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER,
"**outofmemory");
MPID_Segment_init(req->dev.user_buf, req->dev.user_count, req->dev.datatype,
req->dev.segment_ptr, 0);
req->dev.segment_first = 0;
req->dev.segment_size = data_sz;
MPIDI_msg_sz_t last;
last = req->dev.segment_size; /* segment_size is byte offset */
MPIU_Assert(last > 0);
REQ_FIELD(req, lmt_pack_buf) = MPIU_Malloc(data_sz);
MPIU_ERR_CHKANDJUMP(!REQ_FIELD(req, lmt_pack_buf), mpi_errno, MPI_ERR_OTHER,
"**outofmemory");
MPID_Segment_pack(req->dev.segment_ptr, req->dev.segment_first, &last,
(char *) (REQ_FIELD(req, lmt_pack_buf)));
MPIU_Assert(last == req->dev.segment_size);
write_from_buf = REQ_FIELD(req, lmt_pack_buf);
}
//assert(dt_true_lb == 0);
uint8_t *tailp =
(uint8_t *) ((uint8_t *) write_from_buf /*+ dt_true_lb */ + data_sz - sizeof(uint8_t));
#if 0
*is_end_flag_same = (r_cookie_buf->tail == *tailp) ? 1 : 0;
#else
REQ_FIELD(req, lmt_receiver_tail) = r_cookie_buf->tail;
REQ_FIELD(req, lmt_sender_tail) = *tailp;
dprintf("lmt_switch_send,tail on sender=%02x,tail onreceiver=%02x,req=%p\n", *tailp,
r_cookie_buf->tail, req);
#ifdef MPID_NEM_IB_DEBUG_LMT
uint8_t *tail_wordp = (uint8_t *) ((uint8_t *) write_from_buf + data_sz - sizeof(uint32_t) * 2);
#endif
dprintf("lmt_switch_send,tail on sender=%d\n", *tail_wordp);
fflush(stdout);
#endif
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_IB_LMT_SWITCH_SEND);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_tcp_module_lmt_start_send (MPIDI_VC_t *vc, MPID_Request *req, MPID_IOV r_cookie)
{
int mpi_errno = MPI_SUCCESS;
int ret;
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype * dt_ptr;
MPIDI_msg_sz_t last;
int nb;
int s_len = 0;
int r_len;
int r_port;
char *r_hostname;
MPIDI_CH3I_VC *vc_ch = (MPIDI_CH3I_VC *)vc->channel_private;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_TCP_MODULE_LMT_START_SEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_TCP_MODULE_LMT_START_SEND);
mpi_errno = read_r_cookie (r_cookie, &r_hostname, &r_port, &r_len);
if (mpi_errno) MPIU_ERR_POP (mpi_errno);
free_cookie (vc_ch->net.tcp.lmt_cookie);
if (!vc_ch->net.tcp.lmt_connected)
{
struct sockaddr_in saddr;
struct hostent *hp;
vc_ch->net.tcp.lmt_desc = socket (AF_INET, SOCK_STREAM, 0);
MPIU_ERR_CHKANDJUMP2 (vc_ch->net.tcp.lmt_desc == -1, mpi_errno, MPI_ERR_OTHER, "**sock_create", "**sock_create %s %d", strerror (errno), errno);
// ret = fcntl (vc_ch->net.tcp.lmt_desc, F_SETFL, O_NONBLOCK);
// MPIU_ERR_CHKANDJUMP2 (ret == -1, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %s %d", strerror (errno), errno);
hp = gethostbyname (r_hostname);
MPIU_ERR_CHKANDJUMP2 (hp == NULL, mpi_errno, MPI_ERR_OTHER, "**gethostbyname", "**gethostbyname %s %d", hstrerror (h_errno), h_errno);
memset (&saddr, sizeof(saddr), 0);
saddr.sin_family = AF_INET;
saddr.sin_port = htons (r_port);
MPIU_Memcpy (&saddr.sin_addr, hp->h_addr, hp->h_length);
set_sockopts (vc_ch->net.tcp.lmt_desc);
ret = connect (vc_ch->net.tcp.lmt_desc, (struct sockaddr *)&saddr, sizeof(saddr));
MPIU_ERR_CHKANDJUMP2 (ret == -1, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %s %d", strerror (errno), errno);
vc_ch->net.tcp.lmt_connected = 1;
}
MPIDI_Datatype_get_info (req->dev.user_count, req->dev.datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
if (r_len < data_sz)
{
/* message will be truncated */
s_len = data_sz;
data_sz = r_len;
req->status.MPI_ERROR = MPIU_ERR_SET2 (mpi_errno, MPI_ERR_TRUNCATE, "**truncate", "**truncate %d %d", s_len, r_len);
}
MPID_Segment_init (req->dev.user_buf, req->dev.user_count, req->dev.datatype, &req->dev.segment, 0);
req->dev.segment_first = 0;
req->dev.segment_size = data_sz;
req->dev.iov_count = MPID_IOV_LIMIT;
req->dev.iov_offset = 0;
last = data_sz;
do
{
int iov_offset;
int left_to_send;
MPID_Segment_pack_vector (&req->dev.segment, req->dev.segment_first, &last, req->dev.iov, &req->dev.iov_count);
left_to_send = last - req->dev.segment_first;
iov_offset = 0;
#ifdef TESTING_CHUNKING
{
char *buf = req->dev.iov[0].MPID_IOV_BUF;
int l;
while (left_to_send)
{
if (left_to_send > CHUNK)
l = CHUNK;
else
l = left_to_send;
do
nb = write (vc_ch->net.tcp.lmt_desc, buf, l);
while (nb == -1 && errno == EINTR);
MPIU_ERR_CHKANDJUMP (nb == -1, mpi_errno, MPI_ERR_OTHER, "**sock_writev");
left_to_send -= nb;
buf += nb;
}
MPIDI_CH3U_Request_complete (req);
goto fn_exit;
}
#endif
do
nb = writev (vc_ch->net.tcp.lmt_desc, &req->dev.iov[iov_offset], req->dev.iov_count - iov_offset);
while (nb == -1 && errno == EINTR);
MPIU_ERR_CHKANDJUMP (nb == -1, mpi_errno, MPI_ERR_OTHER, "**sock_writev");
left_to_send -= nb;
while (left_to_send)
{ /* send rest of iov */
while (nb >= req->dev.iov[iov_offset].MPID_IOV_LEN)
{ /* update iov to reflect sent bytes */
nb -= req->dev.iov[iov_offset].MPID_IOV_LEN;
++iov_offset;
}
req->dev.iov[iov_offset].MPID_IOV_BUF = (char *)req->dev.iov[iov_offset].MPID_IOV_BUF + nb;
req->dev.iov[iov_offset].MPID_IOV_LEN -= nb;
do
nb = writev (vc_ch->net.tcp.lmt_desc, &req->dev.iov[iov_offset], req->dev.iov_count - iov_offset);
while (nb == -1 && errno == EINTR);
MPIU_ERR_CHKANDJUMP (nb == -1, mpi_errno, MPI_ERR_OTHER, "**sock_writev");
left_to_send -= nb;
}
}
while (last < data_sz);
MPIDI_CH3U_Request_complete (req);
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_TCP_MODULE_LMT_START_SEND);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDO_Gatherv_simple(const void *sendbuf,
int sendcount,
MPI_Datatype sendtype,
void *recvbuf,
const int *recvcounts,
const int *displs,
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
TRACE_ERR("Entering MPIDO_Gatherv_optimized\n");
int snd_contig = 1, rcv_contig = 1;
void *snd_noncontig_buff = NULL, *rcv_noncontig_buff = NULL;
void *sbuf = NULL, *rbuf = NULL;
int *rcounts = NULL;
int *rdispls = NULL;
int send_size = 0;
int recv_size = 0;
int rcvlen = 0;
int totalrecvcount = 0;
pami_type_t rtype = PAMI_TYPE_NULL;
MPID_Segment segment;
MPID_Datatype *data_ptr = NULL;
int send_true_lb, recv_true_lb = 0;
int i, tmp;
volatile unsigned gatherv_active = 1;
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);
int recvok=PAMI_SUCCESS, recvcontinuous=0;
if(sendbuf != MPI_IN_PLACE)
{
MPIDI_Datatype_get_info(sendcount, sendtype, snd_contig,
send_size, data_ptr, send_true_lb);
if(MPIDI_Pamix_collsel_advise != NULL && mpid->collsel_fast_query != NULL)
{
advisor_algorithm_t advisor_algorithms[1];
int num_algorithms = MPIDI_Pamix_collsel_advise(mpid->collsel_fast_query, PAMI_XFER_GATHERV_INT, 64, advisor_algorithms, 1);
if(num_algorithms)
{
if(advisor_algorithms[0].algorithm_type == COLLSEL_EXTERNAL_ALGO)
{
return MPIR_Gatherv(sendbuf, sendcount, sendtype,
recvbuf, recvcounts, displs, recvtype,
root, comm_ptr, mpierrno);
}
else if(advisor_algorithms[0].metadata && advisor_algorithms[0].metadata->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", advisor_algorithms[0].metadata->name);
MPIDO_Barrier(comm_ptr, &tmpmpierrno);
}
}
}
sbuf = (char *)sendbuf + send_true_lb;
if(!snd_contig)
{
snd_noncontig_buff = MPL_malloc(send_size);
sbuf = snd_noncontig_buff;
if(snd_noncontig_buff == NULL)
{
MPID_Abort(NULL, MPI_ERR_NO_SPACE, 1,
"Fatal: Cannot allocate pack buffer");
}
DLOOP_Offset last = send_size;
MPID_Segment_init(sendbuf, sendcount, sendtype, &segment, 0);
MPID_Segment_pack(&segment, 0, &last, snd_noncontig_buff);
}
}
else
{
MPIDI_Datatype_get_info(1, recvtype, rcv_contig,
rcvlen, data_ptr, recv_true_lb);
if(MPIDI_Pamix_collsel_advise != NULL && mpid->collsel_fast_query != NULL)
{
advisor_algorithm_t advisor_algorithms[1];
int num_algorithms = MPIDI_Pamix_collsel_advise(mpid->collsel_fast_query, PAMI_XFER_GATHERV_INT, 64, advisor_algorithms, 1);
if(num_algorithms)
{
if(advisor_algorithms[0].algorithm_type == COLLSEL_EXTERNAL_ALGO)
{
return MPIR_Gatherv(sendbuf, sendcount, sendtype,
recvbuf, recvcounts, displs, recvtype,
root, comm_ptr, mpierrno);
}
else if(advisor_algorithms[0].metadata && advisor_algorithms[0].metadata->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", advisor_algorithms[0].metadata->name);
MPIDO_Barrier(comm_ptr, &tmpmpierrno);
}
}
}
}
pami_xfer_t gatherv;
rbuf = (char *)recvbuf + recv_true_lb;
rcounts = (int*)recvcounts;
rdispls = (int*)displs;
if(rank == root)
{
if((recvok = MPIDI_Datatype_to_pami(recvtype, &rtype, -1, NULL, &tmp)) != MPI_SUCCESS)
{
MPIDI_Datatype_get_info(1, recvtype, rcv_contig,
rcvlen, data_ptr, recv_true_lb);
totalrecvcount = recvcounts[0];
recvcontinuous = displs[0] == 0? 1 : 0 ;
rcounts = (int*)MPL_malloc(size);
rdispls = (int*)MPL_malloc(size);
rdispls[0] = 0;
rcounts[0] = rcvlen * recvcounts[0];
for(i = 1; i < size; i++)
{
rdispls[i]= rcvlen * totalrecvcount;
totalrecvcount += recvcounts[i];
if(displs[i] != (displs[i-1] + recvcounts[i-1]))
recvcontinuous = 0;
rcounts[i] = rcvlen * recvcounts[i];
}
recv_size = rcvlen * totalrecvcount;
rcv_noncontig_buff = MPL_malloc(recv_size);
rbuf = rcv_noncontig_buff;
rtype = PAMI_TYPE_BYTE;
if(rcv_noncontig_buff == NULL)
{
MPID_Abort(NULL, MPI_ERR_NO_SPACE, 1,
"Fatal: Cannot allocate pack buffer");
}
if(sendbuf == MPI_IN_PLACE)
{
size_t extent;
MPID_Datatype_get_extent_macro(recvtype,extent);
MPIR_Localcopy(recvbuf + displs[rank]*extent, recvcounts[rank], recvtype,
rcv_noncontig_buff + rdispls[rank], rcounts[rank],MPI_CHAR);
}
}
if(sendbuf == MPI_IN_PLACE)
{
gatherv.cmd.xfer_gatherv_int.sndbuf = PAMI_IN_PLACE;
}
else
{
gatherv.cmd.xfer_gatherv_int.sndbuf = sbuf;
}
gatherv.cmd.xfer_gatherv_int.stype = PAMI_TYPE_BYTE;/* stype is ignored when sndbuf == PAMI_IN_PLACE */
gatherv.cmd.xfer_gatherv_int.stypecount = send_size;
}
else
{
gatherv.cmd.xfer_gatherv_int.sndbuf = sbuf;
gatherv.cmd.xfer_gatherv_int.stype = PAMI_TYPE_BYTE;
gatherv.cmd.xfer_gatherv_int.stypecount = send_size;
}
gatherv.cb_done = cb_gatherv;
gatherv.cookie = (void *)&gatherv_active;
gatherv.cmd.xfer_gatherv_int.root = MPIDI_Task_to_endpoint(MPID_VCR_GET_LPID(comm_ptr->vcr, root), 0);
gatherv.cmd.xfer_gatherv_int.rcvbuf = rbuf;
gatherv.cmd.xfer_gatherv_int.rtype = rtype;
gatherv.cmd.xfer_gatherv_int.rtypecounts = (int *) rcounts;
gatherv.cmd.xfer_gatherv_int.rdispls = (int *) rdispls;
const pami_metadata_t *my_gatherv_md;
gatherv.algorithm = mpid->coll_algorithm[PAMI_XFER_GATHERV_INT][0][0];
my_gatherv_md = &mpid->coll_metadata[PAMI_XFER_GATHERV_INT][0][0];
MPIDI_Update_last_algorithm(comm_ptr, my_gatherv_md->name);
MPIDI_Post_coll_t gatherv_post;
TRACE_ERR("%s gatherv\n", MPIDI_Process.context_post.active>0?"Posting":"Invoking");
MPIDI_Context_post(MPIDI_Context[0], &gatherv_post.state,
MPIDI_Pami_post_wrapper, (void *)&gatherv);
TRACE_ERR("Gatherv %s\n", MPIDI_Process.context_post.active>0?"posted":"invoked");
TRACE_ERR("Waiting on active %d\n", gatherv_active);
MPID_PROGRESS_WAIT_WHILE(gatherv_active);
if(!rcv_contig || recvok != PAMI_SUCCESS)
{
if(recvcontinuous)
{
MPIR_Localcopy(rcv_noncontig_buff, recv_size, MPI_CHAR,
recvbuf, totalrecvcount, recvtype);
}
else
{
size_t extent;
MPID_Datatype_get_extent_macro(recvtype,extent);
for(i=0; i<size; ++i)
{
char* scbuf = (char*)rcv_noncontig_buff+ rdispls[i];
char* rcbuf = (char*)recvbuf + displs[i]*extent;
MPIR_Localcopy(scbuf, rcounts[i], MPI_CHAR,
rcbuf, recvcounts[i], recvtype);
TRACE_ERR("Pack recv src extent %zu, displ[%zu]=%zu, count[%zu]=%zu buf[%zu]=%u\n",
(size_t)extent, (size_t)i,(size_t)precvdispls[i],(size_t)i,(size_t)precvcounts[i],(size_t)precvdispls[i], *(int*)scbuf);
TRACE_ERR("Pack recv dest extent %zu, displ[%zu]=%zu, count[%zu]=%zu buf[%zu]=%u\n",
(size_t)extent, (size_t)i,(size_t)displs[i],(size_t)i,(size_t)recvcounts[i],(size_t)displs[i], *(int*)rcbuf);
}
}
MPL_free(rcv_noncontig_buff);
if(rank == root)
{
MPL_free(rcounts);
MPL_free(rdispls);
}
}
if(!snd_contig) MPL_free(snd_noncontig_buff);
TRACE_ERR("Leaving MPIDO_Gatherv_optimized\n");
return MPI_SUCCESS;
}
int MPID_nem_lmt_dma_start_recv(MPIDI_VC_t *vc, MPID_Request *rreq, MPID_IOV s_cookie)
{
int mpi_errno = MPI_SUCCESS;
int nodma;
int dt_contig;
MPI_Aint dt_true_lb;
MPIDI_msg_sz_t data_sz;
MPID_Datatype * dt_ptr;
volatile knem_status_t *status;
knem_status_t current_status;
struct lmt_dma_node *node = NULL;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_LMT_DMA_START_RECV);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_LMT_DMA_START_RECV);
/* MT: this code assumes only one thread can be at this point at a time */
if (knem_fd < 0) {
mpi_errno = open_knem_dev();
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
/* find out contig/noncontig, size, and lb for the datatype */
MPIDI_Datatype_get_info(rreq->dev.user_count, rreq->dev.datatype,
dt_contig, data_sz, dt_ptr, dt_true_lb);
nodma = !knem_has_dma || data_sz < MPIR_CVAR_NEMESIS_LMT_DMA_THRESHOLD;
if (dt_contig) {
/* handle the iov creation ourselves */
rreq->dev.iov[0].MPID_IOV_BUF = (char *)rreq->dev.user_buf + dt_true_lb;
rreq->dev.iov[0].MPID_IOV_LEN = data_sz;
rreq->dev.iov_count = 1;
}
else {
if (rreq->dev.segment_ptr == NULL) {
/* segment_ptr may be non-null when this is a continuation of a
many-part message that we couldn't fit in one single flight of
iovs. */
MPIU_Assert(rreq->dev.segment_ptr == NULL);
rreq->dev.segment_ptr = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((rreq->dev.segment_ptr == NULL), mpi_errno,
MPI_ERR_OTHER, "**nomem",
"**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(rreq->dev.user_buf, rreq->dev.user_count,
rreq->dev.datatype, rreq->dev.segment_ptr, 0);
rreq->dev.segment_first = 0;
rreq->dev.segment_size = data_sz;
/* see load_send_iov FIXME above */
mpi_errno = MPIDI_CH3U_Request_load_recv_iov(rreq);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
}
MPIU_Assert(s_cookie.MPID_IOV_LEN == sizeof(knem_cookie_t));
MPIU_Assert(s_cookie.MPID_IOV_BUF != NULL);
mpi_errno = do_dma_recv(rreq->dev.iov_count, rreq->dev.iov,
*((knem_cookie_t *)s_cookie.MPID_IOV_BUF), nodma,
&status, ¤t_status);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* TODO refactor this block and MPID_nem_lmt_dma_progress (and anywhere
* else) to share a common function. This advancement/completion code is
* duplication. */
if (current_status != KNEM_STATUS_PENDING) {
/* complete the request if all data has been sent, remove it from the list */
int complete = 0;
MPIU_ERR_CHKANDJUMP1(current_status == KNEM_STATUS_FAILED, mpi_errno, MPI_ERR_OTHER,
"**recv_status", "**recv_status %d", current_status);
mpi_errno = check_req_complete(vc, rreq, &complete);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
free_status_index(status - knem_status);
if (complete) {
/* request was completed by the OnDataAvail fn */
MPID_nem_lmt_send_DONE(vc, rreq); /* tell the other side to complete its request */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, ".... complete");
}
else {
/* There is more data to send. We must inform the sender that we have
completely received the current batch and that the next batch should
be sent. */
MPID_nem_lmt_send_COOKIE(vc, rreq, NULL, 0);
}
}
/* XXX DJG FIXME this looks like it always pushes! */
/* push request if not complete for progress checks later */
node = MPIU_Malloc(sizeof(struct lmt_dma_node));
node->vc = vc;
node->req = rreq;
node->status_p = status;
node->next = outstanding_head;
outstanding_head = node;
++MPID_nem_local_lmt_pending;
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_LMT_DMA_START_RECV);
return mpi_errno;
fn_fail:
goto fn_exit;
}
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 MPID_nem_newmad_process_rdtype(MPID_Request **rreq_p, MPID_Datatype * dt_ptr, MPIDI_msg_sz_t data_sz, struct iovec *newmad_iov[], int *num_iov)
{
MPID_Request *rreq = *rreq_p;
MPIDI_msg_sz_t last;
MPID_IOV *iov;
int n_iov = 0;
int mpi_errno = MPI_SUCCESS;
int index;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_NEWMAD_PROCESS_RDTYPE);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_NEWMAD_PROCESS_RDTYPE);
if (rreq->dev.segment_ptr == NULL)
{
rreq->dev.segment_ptr = MPID_Segment_alloc( );
MPIU_ERR_CHKANDJUMP1((rreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
}
MPID_Segment_init(rreq->dev.user_buf, rreq->dev.user_count, rreq->dev.datatype, rreq->dev.segment_ptr, 0);
rreq->dev.segment_first = 0;
rreq->dev.segment_size = data_sz;
last = rreq->dev.segment_size;
MPID_Segment_count_contig_blocks(rreq->dev.segment_ptr,rreq->dev.segment_first,&last,&n_iov);
MPIU_Assert(n_iov > 0);
iov = MPIU_Malloc(n_iov*sizeof(MPID_IOV));
MPID_Segment_unpack_vector(rreq->dev.segment_ptr, rreq->dev.segment_first, &last,iov, &n_iov);
MPIU_Assert(last == rreq->dev.segment_size);
#ifdef DEBUG
for(index = 0; index < n_iov ; index++)
{
fprintf(stdout,"======================\n");
fprintf(stdout,"RECV iov[%i]: [base %p][len %i]\n",index,
iov[index].MPID_IOV_BUF,iov[index].MPID_IOV_LEN);
}
#endif
if(n_iov <= NMAD_IOV_MAX_DEPTH)
{
for(index=0; index < n_iov ; index++)
{
(*newmad_iov)[index].iov_base = iov[index].MPID_IOV_BUF;
(*newmad_iov)[index].iov_len = iov[index].MPID_IOV_LEN;
}
rreq->dev.tmpbuf = NULL;
*num_iov = n_iov;
}
else
{
int packsize = 0;
MPIR_Pack_size_impl(rreq->dev.user_count, rreq->dev.datatype, &packsize);
rreq->dev.tmpbuf = MPIU_Malloc((size_t) packsize);
MPIU_Assert(rreq->dev.tmpbuf);
rreq->dev.tmpbuf_sz = packsize;
(*newmad_iov)[0].iov_base = (char *) rreq->dev.tmpbuf;
(*newmad_iov)[0].iov_len = (uint32_t) packsize;
*num_iov = 1 ;
}
MPIU_Free(iov);
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_NEWMAD_PROCESS_RDTYPE);
return mpi_errno;
fn_fail: ATTRIBUTE((unused))
goto fn_exit;
}
int MPIR_Localcopy(const void *sendbuf, MPI_Aint sendcount, MPI_Datatype sendtype,
void *recvbuf, MPI_Aint recvcount, MPI_Datatype recvtype)
{
int mpi_errno = MPI_SUCCESS;
int sendtype_iscontig, recvtype_iscontig;
MPI_Aint sendsize, recvsize, sdata_sz, rdata_sz, copy_sz;
MPI_Aint true_extent, sendtype_true_lb, recvtype_true_lb;
MPIU_CHKLMEM_DECL(1);
MPID_MPI_STATE_DECL(MPID_STATE_MPIR_LOCALCOPY);
MPID_MPI_FUNC_ENTER(MPID_STATE_MPIR_LOCALCOPY);
MPID_Datatype_get_size_macro(sendtype, sendsize);
MPID_Datatype_get_size_macro(recvtype, recvsize);
sdata_sz = sendsize * sendcount;
rdata_sz = recvsize * recvcount;
/* if there is no data to copy, bail out */
if (!sdata_sz || !rdata_sz)
goto fn_exit;
#if defined(HAVE_ERROR_CHECKING)
if (sdata_sz > rdata_sz) {
MPIU_ERR_SET2(mpi_errno, MPI_ERR_TRUNCATE, "**truncate", "**truncate %d %d", sdata_sz, rdata_sz);
copy_sz = rdata_sz;
}
else
#endif /* HAVE_ERROR_CHECKING */
copy_sz = sdata_sz;
/* Builtin types is the common case; optimize for it */
if ((HANDLE_GET_KIND(sendtype) == HANDLE_KIND_BUILTIN) &&
HANDLE_GET_KIND(recvtype) == HANDLE_KIND_BUILTIN) {
MPIU_Memcpy(recvbuf, sendbuf, copy_sz);
goto fn_exit;
}
MPIR_Datatype_iscontig(sendtype, &sendtype_iscontig);
MPIR_Datatype_iscontig(recvtype, &recvtype_iscontig);
MPIR_Type_get_true_extent_impl(sendtype, &sendtype_true_lb, &true_extent);
MPIR_Type_get_true_extent_impl(recvtype, &recvtype_true_lb, &true_extent);
if (sendtype_iscontig && recvtype_iscontig)
{
#if defined(HAVE_ERROR_CHECKING)
MPIU_ERR_CHKMEMCPYANDJUMP(mpi_errno,
((char *)recvbuf + recvtype_true_lb),
((char *)sendbuf + sendtype_true_lb),
copy_sz);
#endif
MPIU_Memcpy(((char *) recvbuf + recvtype_true_lb),
((char *) sendbuf + sendtype_true_lb),
copy_sz);
}
else if (sendtype_iscontig)
{
MPID_Segment seg;
MPI_Aint last;
MPID_Segment_init(recvbuf, recvcount, recvtype, &seg, 0);
last = copy_sz;
MPID_Segment_unpack(&seg, 0, &last, (char*)sendbuf + sendtype_true_lb);
MPIU_ERR_CHKANDJUMP(last != copy_sz, mpi_errno, MPI_ERR_TYPE, "**dtypemismatch");
}
else if (recvtype_iscontig)
{
MPID_Segment seg;
MPI_Aint last;
MPID_Segment_init(sendbuf, sendcount, sendtype, &seg, 0);
last = copy_sz;
MPID_Segment_pack(&seg, 0, &last, (char*)recvbuf + recvtype_true_lb);
MPIU_ERR_CHKANDJUMP(last != copy_sz, mpi_errno, MPI_ERR_TYPE, "**dtypemismatch");
}
else
{
char * buf;
MPIDI_msg_sz_t buf_off;
MPID_Segment sseg;
MPIDI_msg_sz_t sfirst;
MPID_Segment rseg;
MPIDI_msg_sz_t rfirst;
MPIU_CHKLMEM_MALLOC(buf, char *, COPY_BUFFER_SZ, mpi_errno, "buf");
MPID_Segment_init(sendbuf, sendcount, sendtype, &sseg, 0);
MPID_Segment_init(recvbuf, recvcount, recvtype, &rseg, 0);
sfirst = 0;
rfirst = 0;
buf_off = 0;
while (1)
{
MPI_Aint last;
char * buf_end;
if (copy_sz - sfirst > COPY_BUFFER_SZ - buf_off)
{
last = sfirst + (COPY_BUFFER_SZ - buf_off);
}
else
{
last = copy_sz;
}
MPID_Segment_pack(&sseg, sfirst, &last, buf + buf_off);
MPIU_Assert(last > sfirst);
buf_end = buf + buf_off + (last - sfirst);
sfirst = last;
MPID_Segment_unpack(&rseg, rfirst, &last, buf);
MPIU_Assert(last > rfirst);
rfirst = last;
if (rfirst == copy_sz)
{
/* successful completion */
break;
}
/* if the send side finished, but the recv side couldn't unpack it, there's a datatype mismatch */
MPIU_ERR_CHKANDJUMP(sfirst == copy_sz, mpi_errno, MPI_ERR_TYPE, "**dtypemismatch");
/* if not all data was unpacked, copy it to the front of the buffer for next time */
buf_off = sfirst - rfirst;
if (buf_off > 0)
{
memmove(buf, buf_end - buf_off, buf_off);
}
}
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_LOCALCOPY);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDO_Bcast_simple(void *buffer,
int count,
MPI_Datatype datatype,
int root,
MPID_Comm *comm_ptr,
int *mpierrno)
{
TRACE_ERR("Entering MPIDO_Bcast_optimized\n");
int data_contig;
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;
/* 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);
if(MPIDI_Pamix_collsel_advise != NULL && mpid->collsel_fast_query != NULL)
{
advisor_algorithm_t advisor_algorithms[1];
int num_algorithms = MPIDI_Pamix_collsel_advise(mpid->collsel_fast_query, PAMI_XFER_BROADCAST, data_size_one * count, advisor_algorithms, 1);
if(num_algorithms)
{
if(advisor_algorithms[0].algorithm_type == COLLSEL_EXTERNAL_ALGO)
{
return MPIR_Bcast_intra(buffer, count, datatype, root, comm_ptr, mpierrno);
}
}
}
const int data_size = data_size_one*(size_t)count;
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;
const pami_metadata_t *my_bcast_md;
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->coll_algorithm[PAMI_XFER_BROADCAST][0][0];
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;
my_bcast_md = &mpid->coll_metadata[PAMI_XFER_BROADCAST][0][0];
MPIDI_Context_post(MPIDI_Context[0], &bcast_post.state, MPIDI_Pami_post_wrapper, (void *)&bcast);
MPIDI_Update_last_algorithm(comm_ptr, my_bcast_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("Exiting MPIDO_Bcast_optimized\n");
return 0;
}
int MPIDI_Get_accumulate(const void *origin_addr, int origin_count,
MPI_Datatype origin_datatype, void *result_addr, int result_count,
MPI_Datatype result_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op, MPID_Win *win_ptr)
{
int mpi_errno = MPI_SUCCESS;
MPIDI_msg_sz_t data_sz;
int rank, origin_predefined, result_predefined, target_predefined;
int shm_locked = 0;
int dt_contig ATTRIBUTE((unused));
MPI_Aint dt_true_lb ATTRIBUTE((unused));
MPID_Datatype *dtp;
MPIU_CHKLMEM_DECL(2);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_GET_ACCUMULATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_GET_ACCUMULATE);
if (target_rank == MPI_PROC_NULL) {
goto fn_exit;
}
if (win_ptr->epoch_state == MPIDI_EPOCH_NONE && win_ptr->fence_issued) {
win_ptr->epoch_state = MPIDI_EPOCH_FENCE;
}
MPIU_ERR_CHKANDJUMP(win_ptr->epoch_state == MPIDI_EPOCH_NONE,
mpi_errno, MPI_ERR_RMA_SYNC, "**rmasync");
MPIDI_Datatype_get_info(target_count, target_datatype, dt_contig, data_sz,
dtp, dt_true_lb);
if (data_sz == 0) {
goto fn_exit;
}
rank = win_ptr->myrank;
origin_predefined = TRUE; /* quiet uninitialized warnings (b/c goto) */
if (op != MPI_NO_OP) {
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, origin_predefined);
}
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(result_datatype, result_predefined);
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, target_predefined);
/* Do =! rank first (most likely branch?) */
if (target_rank == rank || win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED)
{
MPI_User_function *uop;
void *base;
int disp_unit;
if (win_ptr->create_flavor == MPI_WIN_FLAVOR_SHARED) {
base = win_ptr->shm_base_addrs[target_rank];
disp_unit = win_ptr->disp_units[target_rank];
MPIDI_CH3I_SHM_MUTEX_LOCK(win_ptr);
shm_locked = 1;
}
else {
base = win_ptr->base;
disp_unit = win_ptr->disp_unit;
}
/* Perform the local get first, then the accumulate */
mpi_errno = MPIR_Localcopy((char *) base + disp_unit * target_disp,
target_count, target_datatype,
result_addr, result_count, result_datatype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
/* NO_OP: Don't perform the accumulate */
if (op == MPI_NO_OP) {
if (shm_locked) {
MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr);
shm_locked = 0;
}
goto fn_exit;
}
if (op == MPI_REPLACE) {
mpi_errno = MPIR_Localcopy(origin_addr, origin_count, origin_datatype,
(char *) base + disp_unit * target_disp,
target_count, target_datatype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
if (shm_locked) {
MPIDI_CH3I_SHM_MUTEX_UNLOCK(win_ptr);
shm_locked = 0;
}
goto fn_exit;
}
MPIU_ERR_CHKANDJUMP1((HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN),
mpi_errno, MPI_ERR_OP, "**opnotpredefined",
"**opnotpredefined %d", op );
/* get the function by indexing into the op table */
uop = MPIR_OP_HDL_TO_FN(op);
if (origin_predefined && target_predefined) {
/* Cast away const'ness for origin_address in order to
* avoid changing the prototype for MPI_User_function */
(*uop)((void *) origin_addr, (char *) base + disp_unit*target_disp,
&target_count, &target_datatype);
}
else {
/* derived datatype */
MPID_Segment *segp;
DLOOP_VECTOR *dloop_vec;
MPI_Aint first, last;
int vec_len, i, type_size, count;
MPI_Datatype type;
MPI_Aint true_lb, true_extent, extent;
void *tmp_buf=NULL, *target_buf;
const void *source_buf;
if (origin_datatype != target_datatype) {
/* first copy the data into a temporary buffer with
the same datatype as the target. Then do the
accumulate operation. */
MPIR_Type_get_true_extent_impl(target_datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(target_datatype, extent);
MPIU_CHKLMEM_MALLOC(tmp_buf, void *,
target_count * (MPIR_MAX(extent,true_extent)),
mpi_errno, "temporary buffer");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
mpi_errno = MPIR_Localcopy(origin_addr, origin_count,
origin_datatype, tmp_buf,
target_count, target_datatype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
if (target_predefined) {
/* target predefined type, origin derived datatype */
(*uop)(tmp_buf, (char *) base + disp_unit * target_disp,
&target_count, &target_datatype);
}
else {
segp = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((!segp), mpi_errno, MPI_ERR_OTHER,
"**nomem","**nomem %s","MPID_Segment_alloc");
MPID_Segment_init(NULL, target_count, target_datatype, segp, 0);
first = 0;
last = SEGMENT_IGNORE_LAST;
MPID_Datatype_get_ptr(target_datatype, dtp);
vec_len = dtp->max_contig_blocks * target_count + 1;
/* +1 needed because Rob says so */
MPIU_CHKLMEM_MALLOC(dloop_vec, DLOOP_VECTOR *,
vec_len * sizeof(DLOOP_VECTOR),
mpi_errno, "dloop vector");
MPID_Segment_pack_vector(segp, first, &last, dloop_vec, &vec_len);
source_buf = (tmp_buf != NULL) ? tmp_buf : origin_addr;
target_buf = (char *) base + disp_unit * target_disp;
type = dtp->eltype;
type_size = MPID_Datatype_get_basic_size(type);
for (i=0; i<vec_len; i++) {
count = (dloop_vec[i].DLOOP_VECTOR_LEN)/type_size;
(*uop)((char *)source_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF),
(char *)target_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF),
&count, &type);
}
MPID_Segment_free(segp);
}
}
int MPIDI_CH3_PktHandler_RndvClrToSend( MPIDI_VC_t *vc, MPIDI_CH3_Pkt_t *pkt,
MPIDI_msg_sz_t *buflen, MPID_Request **rreqp )
{
MPIDI_CH3_Pkt_rndv_clr_to_send_t * cts_pkt = &pkt->rndv_clr_to_send;
MPID_Request * sreq;
MPID_Request * rts_sreq;
MPIDI_CH3_Pkt_t upkt;
MPIDI_CH3_Pkt_rndv_send_t * rs_pkt = &upkt.rndv_send;
int dt_contig;
MPI_Aint dt_true_lb;
MPIDI_msg_sz_t data_sz;
MPID_Datatype * dt_ptr;
int mpi_errno = MPI_SUCCESS;
MPIU_DBG_MSG(CH3_OTHER,VERBOSE,"received rndv CTS pkt");
MPID_Request_get_ptr(cts_pkt->sender_req_id, sreq);
MPIU_DBG_PRINTF(("received cts, count=%d\n", sreq->dev.user_count));
sreq->dev.OnDataAvail = 0;
sreq->dev.OnFinal = 0;
/* Release the RTS request if one exists.
MPID_Request_fetch_and_clear_rts_sreq() needs to be atomic to
prevent
cancel send from cancelling the wrong (future) request.
If MPID_Request_fetch_and_clear_rts_sreq() returns a NULL
rts_sreq, then MPID_Cancel_send() is responsible for releasing
the RTS request object. */
MPIDI_Request_fetch_and_clear_rts_sreq(sreq, &rts_sreq);
if (rts_sreq != NULL)
{
MPID_Request_release(rts_sreq);
}
*buflen = sizeof(MPIDI_CH3_Pkt_t);
MPIDI_Pkt_init(rs_pkt, MPIDI_CH3_PKT_RNDV_SEND);
rs_pkt->receiver_req_id = cts_pkt->receiver_req_id;
MPIDI_Datatype_get_info(sreq->dev.user_count, sreq->dev.datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
if (dt_contig)
{
MPID_IOV iov[MPID_IOV_LIMIT];
MPIU_DBG_MSG_FMT(CH3_OTHER,VERBOSE,(MPIU_DBG_FDEST,
"sending contiguous rndv data, data_sz=" MPIDI_MSG_SZ_FMT,
data_sz));
iov[0].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)rs_pkt;
iov[0].MPID_IOV_LEN = sizeof(*rs_pkt);
iov[1].MPID_IOV_BUF = (MPID_IOV_BUF_CAST)((char *)sreq->dev.user_buf + dt_true_lb);
iov[1].MPID_IOV_LEN = data_sz;
MPIU_THREAD_CS_ENTER(CH3COMM,vc);
mpi_errno = MPIU_CALL(MPIDI_CH3,iSendv(vc, sreq, iov, 2));
MPIU_THREAD_CS_EXIT(CH3COMM,vc);
MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|senddata");
}
else
{
sreq->dev.segment_ptr = MPID_Segment_alloc( );
MPIU_ERR_CHKANDJUMP1((sreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(sreq->dev.user_buf, sreq->dev.user_count,
sreq->dev.datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
MPIU_THREAD_CS_ENTER(CH3COMM,vc);
mpi_errno = vc->sendNoncontig_fn(vc, sreq, rs_pkt, sizeof(*rs_pkt));
MPIU_THREAD_CS_EXIT(CH3COMM,vc);
MPIU_ERR_CHKANDJUMP(mpi_errno, mpi_errno, MPI_ERR_OTHER, "**ch3|senddata");
}
*rreqp = NULL;
fn_fail:
return mpi_errno;
}
static int send_msg(ptl_hdr_data_t ssend_flag, struct MPIDI_VC *vc, const void *buf, MPI_Aint count, MPI_Datatype datatype, int dest,
int tag, MPID_Comm *comm, int context_offset, struct MPID_Request **request)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
int ret;
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype *dt_ptr;
MPID_Request *sreq = NULL;
ptl_me_t me;
int initial_iov_count, remaining_iov_count;
ptl_md_t md;
MPI_Aint last;
MPIU_CHKPMEM_DECL(2);
MPIDI_STATE_DECL(MPID_STATE_SEND_MSG);
MPIDI_FUNC_ENTER(MPID_STATE_SEND_MSG);
MPID_nem_ptl_request_create_sreq(sreq, mpi_errno, comm);
sreq->dev.match.parts.rank = dest;
sreq->dev.match.parts.tag = tag;
sreq->dev.match.parts.context_id = comm->context_id + context_offset;
sreq->ch.vc = vc;
if (!vc_ptl->id_initialized) {
mpi_errno = MPID_nem_ptl_init_id(vc);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE, (MPIU_DBG_FDEST, "count="MPI_AINT_FMT_DEC_SPEC" datatype=%#x contig=%d data_sz=%lu", count, datatype, dt_contig, data_sz));
if (data_sz <= PTL_LARGE_THRESHOLD) {
/* Small message. Send all data eagerly */
if (dt_contig) {
void *start = (char *)buf + dt_true_lb;
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Small contig message");
REQ_PTL(sreq)->event_handler = handler_send;
MPIU_DBG_MSG_P(CH3_CHANNEL, VERBOSE, "&REQ_PTL(sreq)->event_handler = %p", &(REQ_PTL(sreq)->event_handler));
if (start == NULL)
ret = MPID_nem_ptl_rptl_put(MPIDI_nem_ptl_global_md, (ptl_size_t)&dummy, data_sz, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag, data_sz));
else
ret = MPID_nem_ptl_rptl_put(MPIDI_nem_ptl_global_md, (ptl_size_t)start, data_sz, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("global", data_sz, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag, data_sz));
MPIU_DBG_MSG_D(CH3_CHANNEL, VERBOSE, "id.nid = %#x", vc_ptl->id.phys.nid);
MPIU_DBG_MSG_D(CH3_CHANNEL, VERBOSE, "id.pid = %#x", vc_ptl->id.phys.pid);
MPIU_DBG_MSG_P(CH3_CHANNEL, VERBOSE, "sreq = %p", sreq);
MPIU_DBG_MSG_D(CH3_CHANNEL, VERBOSE, "vc_ptl->pt = %d", vc_ptl->pt);
MPIU_DBG_MSG_P(CH3_CHANNEL, VERBOSE, "REQ_PTL(sreq)->event_handler = %p", REQ_PTL(sreq)->event_handler);
goto fn_exit;
}
/* noncontig data */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Small noncontig message");
sreq->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1(sreq->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
last = sreq->dev.segment_size;
sreq->dev.iov_count = MPL_IOV_LIMIT;
MPID_Segment_pack_vector(sreq->dev.segment_ptr, sreq->dev.segment_first, &last, sreq->dev.iov, &sreq->dev.iov_count);
if (last == sreq->dev.segment_size) {
/* IOV is able to describe entire message */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " entire message fits in IOV");
md.start = sreq->dev.iov;
md.length = sreq->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(sreq)->md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdbind", "**ptlmdbind %s", MPID_nem_ptl_strerror(ret));
REQ_PTL(sreq)->event_handler = handler_send;
ret = MPID_nem_ptl_rptl_put(REQ_PTL(sreq)->md, 0, data_sz, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("sreq", data_sz, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag, data_sz));
goto fn_exit;
}
/* IOV is not long enough to describe entire message */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " IOV too long: using bounce buffer");
MPIU_CHKPMEM_MALLOC(REQ_PTL(sreq)->chunk_buffer[0], void *, data_sz, mpi_errno, "chunk_buffer");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
last = data_sz;
MPID_Segment_pack(sreq->dev.segment_ptr, sreq->dev.segment_first, &last, REQ_PTL(sreq)->chunk_buffer[0]);
MPIU_Assert(last == sreq->dev.segment_size);
REQ_PTL(sreq)->event_handler = handler_send;
ret = MPID_nem_ptl_rptl_put(MPIDI_nem_ptl_global_md, (ptl_size_t)REQ_PTL(sreq)->chunk_buffer[0], data_sz, PTL_NO_ACK_REQ,
vc_ptl->id, vc_ptl->pt, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("global", data_sz, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag, data_sz));
goto fn_exit;
}
/* Large message. Send first chunk of data and let receiver get the rest */
if (dt_contig) {
/* create ME for buffer so receiver can issue a GET for the data */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Large contig message");
big_meappend((char *)buf + dt_true_lb + PTL_LARGE_THRESHOLD, data_sz - PTL_LARGE_THRESHOLD, vc,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), sreq);
REQ_PTL(sreq)->event_handler = handler_send;
ret = MPID_nem_ptl_rptl_put(MPIDI_nem_ptl_global_md, (ptl_size_t)((char *)buf + dt_true_lb), PTL_LARGE_THRESHOLD, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag | NPTL_LARGE, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("global", PTL_LARGE_THRESHOLD, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag | NPTL_LARGE, data_sz));
goto fn_exit;
}
/* Large noncontig data */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Large noncontig message");
sreq->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1(sreq->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
last = PTL_LARGE_THRESHOLD;
sreq->dev.iov_count = MPL_IOV_LIMIT;
MPID_Segment_pack_vector(sreq->dev.segment_ptr, sreq->dev.segment_first, &last, sreq->dev.iov, &sreq->dev.iov_count);
initial_iov_count = sreq->dev.iov_count;
sreq->dev.segment_first = last;
if (last == PTL_LARGE_THRESHOLD) {
/* first chunk of message fits into IOV */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " first chunk fits in IOV");
if (initial_iov_count < MPL_IOV_LIMIT) {
/* There may be space for the rest of the message in this IOV */
sreq->dev.iov_count = MPL_IOV_LIMIT - sreq->dev.iov_count;
last = sreq->dev.segment_size;
MPID_Segment_pack_vector(sreq->dev.segment_ptr, sreq->dev.segment_first, &last,
&sreq->dev.iov[initial_iov_count], &sreq->dev.iov_count);
remaining_iov_count = sreq->dev.iov_count;
if (last == sreq->dev.segment_size && last <= MPIDI_nem_ptl_ni_limits.max_msg_size + PTL_LARGE_THRESHOLD) {
/* Entire message fit in one IOV */
int was_incomplete;
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " rest of message fits in one IOV");
/* Create ME for remaining data */
me.start = &sreq->dev.iov[initial_iov_count];
me.length = remaining_iov_count;
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = ( PTL_ME_OP_PUT | PTL_ME_OP_GET | PTL_ME_USE_ONCE | PTL_ME_IS_ACCESSIBLE | PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE | PTL_IOVEC );
me.match_id = vc_ptl->id;
me.match_bits = NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank);
me.ignore_bits = 0;
me.min_free = 0;
MPIU_CHKPMEM_MALLOC(REQ_PTL(sreq)->get_me_p, ptl_handle_me_t *, sizeof(ptl_handle_me_t), mpi_errno, "get_me_p");
ret = PtlMEAppend(MPIDI_nem_ptl_ni, MPIDI_nem_ptl_get_pt, &me, PTL_PRIORITY_LIST, sreq,
&REQ_PTL(sreq)->get_me_p[0]);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmeappend", "**ptlmeappend %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_MEAPPEND("CTL", vc->pg_rank, me, sreq);
/* increment the cc for the get operation */
MPIDI_CH3U_Request_increment_cc(sreq, &was_incomplete);
MPIU_Assert(was_incomplete);
/* Create MD for first chunk */
md.start = sreq->dev.iov;
md.length = initial_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(sreq)->md);
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlmdbind", "**ptlmdbind %s", MPID_nem_ptl_strerror(ret));
REQ_PTL(sreq)->event_handler = handler_send;
ret = MPID_nem_ptl_rptl_put(REQ_PTL(sreq)->md, 0, PTL_LARGE_THRESHOLD, PTL_NO_ACK_REQ, vc_ptl->id, vc_ptl->pt,
NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), 0, sreq,
NPTL_HEADER(ssend_flag | NPTL_LARGE, data_sz));
MPIR_ERR_CHKANDJUMP1(ret, mpi_errno, MPI_ERR_OTHER, "**ptlput", "**ptlput %s", MPID_nem_ptl_strerror(ret));
DBG_MSG_PUT("req", PTL_LARGE_THRESHOLD, vc->pg_rank, NPTL_MATCH(tag, comm->context_id + context_offset, comm->rank), NPTL_HEADER(ssend_flag | NPTL_LARGE, data_sz));
goto fn_exit;
}
int MPIR_Pack_impl(const void *inbuf,
int incount,
MPI_Datatype datatype,
void *outbuf,
MPI_Aint outsize,
MPI_Aint *position)
{
int mpi_errno = MPI_SUCCESS;
MPI_Aint first, last;
MPID_Segment *segp;
int contig;
MPI_Aint dt_true_lb;
MPI_Aint data_sz;
if (incount == 0) {
goto fn_exit;
}
/* Handle contig case quickly */
if (HANDLE_GET_KIND(datatype) == HANDLE_KIND_BUILTIN) {
contig = TRUE;
dt_true_lb = 0;
data_sz = incount * MPID_Datatype_get_basic_size(datatype);
} else {
MPID_Datatype *dt_ptr;
MPID_Datatype_get_ptr(datatype, dt_ptr);
contig = dt_ptr->is_contig;
dt_true_lb = dt_ptr->true_lb;
data_sz = incount * dt_ptr->size;
}
if (contig) {
MPIU_Memcpy((char *) outbuf + *position, (char *)inbuf + dt_true_lb, data_sz);
*position = (int)((MPI_Aint)*position + data_sz);
goto fn_exit;
}
/* non-contig case */
/* TODO: CHECK RETURN VALUES?? */
/* TODO: SHOULD THIS ALL BE IN A MPID_PACK??? */
segp = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1(segp == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment");
mpi_errno = MPID_Segment_init(inbuf, incount, datatype, segp, 0);
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* NOTE: the use of buffer values and positions in MPI_Pack and in
* MPID_Segment_pack are quite different. See code or docs or something.
*/
first = 0;
last = SEGMENT_IGNORE_LAST;
/* Ensure that pointer increment fits in a pointer */
MPID_Ensure_Aint_fits_in_pointer((MPI_VOID_PTR_CAST_TO_MPI_AINT outbuf) +
(MPI_Aint) *position);
MPID_Segment_pack(segp,
first,
&last,
(void *) ((char *) outbuf + *position));
/* Ensure that calculation fits into an int datatype. */
MPID_Ensure_Aint_fits_in_int((MPI_Aint)*position + last);
*position = (int)((MPI_Aint)*position + last);
MPID_Segment_free(segp);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
/**
* \brief MPID buffer copy
*
* Implements non-contiguous buffers correctly.
*
* \param[in] sbuf The address of the input buffer
* \param[in] scount The number of elements in that buffer
* \param[in] sdt The datatype of those elements
* \param[out] smpi_errno Returns errors
* \param[in] rbuf The address of the output buffer
* \param[out] rcount The number of elements in that buffer
* \param[in] rdt The datatype of those elements
* \param[out] rsz The size of the ouput data
* \param[out] rmpi_errno Returns errors
*/
void MPIDI_Buffer_copy(
const void * const sbuf, MPI_Aint scount, MPI_Datatype sdt, int * smpi_errno,
void * const rbuf, MPI_Aint rcount, MPI_Datatype rdt, MPIDI_msg_sz_t * rsz, int * rmpi_errno)
{
int sdt_contig;
int rdt_contig;
MPI_Aint sdt_true_lb, rdt_true_lb;
MPIDI_msg_sz_t sdata_sz;
MPIDI_msg_sz_t rdata_sz;
MPID_Datatype * sdt_ptr;
MPID_Datatype * rdt_ptr;
MPI_Aint sdt_extent;
MPI_Aint rdt_extent;
*smpi_errno = MPI_SUCCESS;
*rmpi_errno = MPI_SUCCESS;
/* printf("bufcopy: src count=%d dt=%d\n", scount, sdt); */
/* printf("bufcopy: dst count=%d dt=%d\n", rcount, rdt); */
MPIDI_Datatype_get_info(scount, sdt, sdt_contig, sdata_sz, sdt_ptr, sdt_true_lb);
MPIDI_Datatype_get_info(rcount, rdt, rdt_contig, rdata_sz, rdt_ptr, rdt_true_lb);
/* --BEGIN ERROR HANDLING-- */
if (sdata_sz > rdata_sz)
{
*rmpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, __FUNCTION__, __LINE__, MPI_ERR_TRUNCATE, "**truncate", "**truncate %d %d", sdata_sz, rdata_sz );
sdata_sz = rdata_sz;
}
/* --END ERROR HANDLING-- */
if (sdata_sz == 0)
{
*rsz = 0;
goto fn_exit;
}
if (sdt_contig && rdt_contig)
{
#if CUDA_AWARE_SUPPORT
if(MPIDI_Process.cuda_aware_support_on && MPIDI_cuda_is_device_buf(rbuf))
{
cudaError_t cudaerr = CudaMemcpy(rbuf + rdt_true_lb, sbuf + sdt_true_lb, sdata_sz, cudaMemcpyHostToDevice);
}
else
#endif
memcpy((char*)rbuf + rdt_true_lb, (const char *)sbuf + sdt_true_lb, sdata_sz);
*rsz = sdata_sz;
}
else if (sdt_contig)
{
#if CUDA_AWARE_SUPPORT
// This will need to be done in two steps:
// 1 - Allocate a temp buffer which is the same size as user buffer and unpack in it.
// 2 - Copy unpacked data into user buffer from temp buffer.
if(MPIDI_Process.cuda_aware_support_on && MPIDI_cuda_is_device_buf(rbuf))
{
MPID_Datatype_get_extent_macro(rdt, rdt_extent);
char *buf = MPL_malloc(rdt_extent * rcount);
memset(buf, 0, rdt_extent * rcount);
MPID_Segment seg;
DLOOP_Offset last;
MPID_Segment_init(buf, rcount, rdt, &seg, 0);
last = sdata_sz;
MPID_Segment_unpack(&seg, 0, &last, (char*)sbuf + sdt_true_lb);
/* --BEGIN ERROR HANDLING-- */
if (last != sdata_sz)
{
*rmpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, __FUNCTION__, __LINE__, MPI_ERR_TYPE, "**dtypemismatch", 0);
}
/* --END ERROR HANDLING-- */
*rsz = last;
cudaError_t cudaerr = CudaMemcpy(rbuf + rdt_true_lb, buf, rdt_extent * rcount, cudaMemcpyHostToDevice);
MPL_free(buf);
goto fn_exit;
}
#endif
MPID_Segment seg;
DLOOP_Offset last;
MPID_Segment_init(rbuf, rcount, rdt, &seg, 0);
last = sdata_sz;
MPID_Segment_unpack(&seg, 0, &last, (char*)sbuf + sdt_true_lb);
/* --BEGIN ERROR HANDLING-- */
if (last != sdata_sz)
{
*rmpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, __FUNCTION__, __LINE__, MPI_ERR_TYPE, "**dtypemismatch", 0);
}
/* --END ERROR HANDLING-- */
*rsz = last;
}
else if (rdt_contig)
{
MPID_Segment seg;
DLOOP_Offset last;
MPID_Segment_init(sbuf, scount, sdt, &seg, 0);
last = sdata_sz;
MPID_Segment_pack(&seg, 0, &last, (char*)rbuf + rdt_true_lb);
/* --BEGIN ERROR HANDLING-- */
if (last != sdata_sz)
{
*rmpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, __FUNCTION__, __LINE__, MPI_ERR_TYPE, "**dtypemismatch", 0);
}
/* --END ERROR HANDLING-- */
*rsz = last;
}
else
{
char * buf;
MPIDI_msg_sz_t buf_off;
MPID_Segment sseg;
MPIDI_msg_sz_t sfirst;
MPID_Segment rseg;
MPIDI_msg_sz_t rfirst;
buf = MPL_malloc(MPIDI_COPY_BUFFER_SZ);
/* --BEGIN ERROR HANDLING-- */
if (buf == NULL)
{
*smpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_FATAL, __FUNCTION__, __LINE__, MPI_ERR_OTHER, "**nomem", 0);
*rmpi_errno = *smpi_errno;
*rsz = 0;
goto fn_exit;
}
/* --END ERROR HANDLING-- */
MPID_Segment_init(sbuf, scount, sdt, &sseg, 0);
MPID_Segment_init(rbuf, rcount, rdt, &rseg, 0);
sfirst = 0;
rfirst = 0;
buf_off = 0;
for(;;)
{
DLOOP_Offset last;
char * buf_end;
if (sdata_sz - sfirst > MPIDI_COPY_BUFFER_SZ - buf_off)
{
last = sfirst + (MPIDI_COPY_BUFFER_SZ - buf_off);
}
else
{
last = sdata_sz;
}
MPID_Segment_pack(&sseg, sfirst, &last, buf + buf_off);
/* --BEGIN ERROR HANDLING-- */
MPID_assert(last > sfirst);
/* --END ERROR HANDLING-- */
buf_end = buf + buf_off + (last - sfirst);
sfirst = last;
MPID_Segment_unpack(&rseg, rfirst, &last, buf);
/* --BEGIN ERROR HANDLING-- */
MPID_assert(last > rfirst);
/* --END ERROR HANDLING-- */
rfirst = last;
if (rfirst == sdata_sz)
{
/* successful completion */
break;
}
/* --BEGIN ERROR HANDLING-- */
if (sfirst == sdata_sz)
{
/* datatype mismatch -- remaining bytes could not be unpacked */
*rmpi_errno = MPIR_Err_create_code(MPI_SUCCESS, MPIR_ERR_RECOVERABLE, __FUNCTION__, __LINE__, MPI_ERR_TYPE, "**dtypemismatch", 0);
break;
}
/* --END ERROR HANDLING-- */
buf_off = sfirst - rfirst;
if (buf_off > 0)
{
memmove(buf, buf_end - buf_off, buf_off);
}
}
*rsz = rfirst;
MPL_free(buf);
}
fn_exit:
return;
}
int MPIDO_Scatterv_simple(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 snd_contig = 1;
int rcv_contig = 1;
int send_size = 0, recv_size = 0;
int ssize = 0;
MPID_Datatype *dt_ptr = NULL;
MPI_Aint send_true_lb=0, recv_true_lb=0;
void *snd_noncontig_buff = NULL, *rcv_noncontig_buff = NULL;
void *sbuf = NULL, *rbuf = NULL;
int *sdispls = NULL, *scounts = NULL;
int sndcount = 0;
MPID_Segment segment;
int tmp, i;
pami_type_t stype = PAMI_TYPE_NULL;
const int rank = comm_ptr->rank;
const int size = comm_ptr->local_size;
const struct MPIDI_Comm* const mpid = &(comm_ptr->mpid);
if (rank == root && sendtype != MPI_DATATYPE_NULL && sendcounts[0] >= 0)
{
MPIDI_Datatype_get_info(1, sendtype, snd_contig, ssize, dt_ptr, send_true_lb);
if(MPIDI_Pamix_collsel_advise != NULL && mpid->collsel_fast_query != NULL)
{
advisor_algorithm_t advisor_algorithms[1];
int num_algorithms = MPIDI_Pamix_collsel_advise(mpid->collsel_fast_query, PAMI_XFER_SCATTERV_INT, 64, advisor_algorithms, 1);
if(num_algorithms)
{
if(advisor_algorithms[0].algorithm_type == COLLSEL_EXTERNAL_ALGO)
{
return MPIR_Scatterv(sendbuf, sendcounts, displs, sendtype,
recvbuf, recvcount, recvtype,
root, comm_ptr, mpierrno);
}
else if(advisor_algorithms[0].metadata && advisor_algorithms[0].metadata->check_correct.values.asyncflowctl && !(--(comm_ptr->mpid.num_requests)))
{
comm_ptr->mpid.num_requests = MPIDI_Process.optimized.num_requests;
int tmpmpierrno;
MPIDO_Barrier(comm_ptr, &tmpmpierrno);
}
}
}
}
if (recvtype != MPI_DATATYPE_NULL && recvcount >= 0)
{
MPIDI_Datatype_get_info(recvcount, recvtype, rcv_contig,
recv_size, dt_ptr, recv_true_lb);
if(MPIDI_Pamix_collsel_advise != NULL && mpid->collsel_fast_query != NULL)
{
advisor_algorithm_t advisor_algorithms[1];
int num_algorithms = MPIDI_Pamix_collsel_advise(mpid->collsel_fast_query, PAMI_XFER_SCATTERV_INT, 64, advisor_algorithms, 1);
if(num_algorithms)
{
if(advisor_algorithms[0].algorithm_type == COLLSEL_EXTERNAL_ALGO)
{
return MPIR_Scatterv(sendbuf, sendcounts, displs, sendtype,
recvbuf, recvcount, recvtype,
root, comm_ptr, mpierrno);
}
else if(advisor_algorithms[0].metadata && advisor_algorithms[0].metadata->check_correct.values.asyncflowctl && !(--(comm_ptr->mpid.num_requests)))
{
comm_ptr->mpid.num_requests = MPIDI_Process.optimized.num_requests;
int tmpmpierrno;
MPIDO_Barrier(comm_ptr, &tmpmpierrno);
}
}
}
}
pami_xfer_t scatterv;
const pami_metadata_t *my_scatterv_md;
volatile unsigned scatterv_active = 1;
sbuf = (char *)sendbuf + send_true_lb;
rbuf = (char *)recvbuf + recv_true_lb;
scounts = (int*)sendcounts;
sdispls = (int*)displs;
if(rank == root)
{
if(MPIDI_Datatype_to_pami(sendtype, &stype, -1, NULL, &tmp) != MPI_SUCCESS)
{
if (!snd_contig)
{
scounts = (int*)MPIU_Malloc(size);
sdispls = (int*)MPIU_Malloc(size);
for(i = 0; i < size; i++)
{
scounts[i] = ssize * sendcounts[i];
sdispls[i] = ssize * displs[i];
send_size += scounts[i];
sndcount += sendcounts[i];
}
snd_noncontig_buff = MPIU_Malloc(send_size);
sbuf = snd_noncontig_buff;
stype = PAMI_TYPE_BYTE;
if(snd_noncontig_buff == NULL)
{
MPID_Abort(NULL, MPI_ERR_NO_SPACE, 1,
"Fatal: Cannot allocate pack buffer");
}
DLOOP_Offset last = send_size;
MPID_Segment_init(sendbuf, sndcount, sendtype, &segment, 0);
MPID_Segment_pack(&segment, 0, &last, snd_noncontig_buff);
}
}
if(recvbuf == MPI_IN_PLACE)
{
rbuf = PAMI_IN_PLACE;
}
}
if(recvbuf != MPI_IN_PLACE)
{
if (!rcv_contig)
{
rcv_noncontig_buff = MPIU_Malloc(recv_size);
rbuf = rcv_noncontig_buff;
if(rcv_noncontig_buff == NULL)
{
MPID_Abort(NULL, MPI_ERR_NO_SPACE, 1,
"Fatal: Cannot allocate pack buffer");
}
}
}
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 = mpid->coll_algorithm[PAMI_XFER_SCATTERV_INT][0][0];
my_scatterv_md = &mpid->coll_metadata[PAMI_XFER_SCATTERV_INT][0][0];
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 = PAMI_TYPE_BYTE;/* rtype is ignored when rcvbuf == PAMI_IN_PLACE */
scatterv.cmd.xfer_scatterv_int.stypecounts = (int *) scounts;
scatterv.cmd.xfer_scatterv_int.rtypecount = recv_size;
scatterv.cmd.xfer_scatterv_int.sdispls = (int *) sdispls;
MPIDI_Update_last_algorithm(comm_ptr, my_scatterv_md->name);
MPIDI_Post_coll_t scatterv_post;
TRACE_ERR("%s scatterv\n", MPIDI_Process.context_post.active>0?"Posting":"Invoking");
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);
if(!rcv_contig)
{
MPIR_Localcopy(rcv_noncontig_buff, recv_size, MPI_CHAR,
recvbuf, recvcount, recvtype);
MPIU_Free(rcv_noncontig_buff);
}
if(!snd_contig)
{
MPIU_Free(snd_noncontig_buff);
MPIU_Free(scounts);
MPIU_Free(sdispls);
}
TRACE_ERR("Leaving MPIDO_Scatterv_optimized\n");
return MPI_SUCCESS;
}
/* create a request */
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIU_Assert(sreq != NULL);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SEND);
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(datatype, sreq->dev.datatype_ptr);
MPID_Datatype_add_ref(sreq->dev.datatype_ptr);
}
sreq->partner_request = NULL;
sreq->dev.OnDataAvail = NULL;
sreq->dev.tmpbuf = NULL;
sreq->ch.vc = vc;
sreq->ch.noncontig = FALSE;
_dbg_mxm_output(5,
"isSend ========> Sending USER msg for req %p (context %d to %d tag %d size %d) \n",
sreq, comm->context_id + context_offset, rank, tag, data_sz);
vc_area = VC_BASE(vc);
req_area = REQ_BASE(sreq);
req_area-> ctx = sreq;
req_area->iov_buf = req_area->tmp_buf;
req_area->iov_count = 0;
req_area->iov_buf[0].ptr = NULL;
req_area->iov_buf[0].length = 0;
if (data_sz) {
if (dt_contig) {
req_area->iov_count = 1;
req_area->iov_buf[0].ptr = (char *) (buf) + dt_true_lb;
req_area->iov_buf[0].length = data_sz;
}
else {
MPIDI_msg_sz_t last;
MPI_Aint packsize = 0;
sreq->ch.noncontig = TRUE;
sreq->dev.segment_ptr = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((sreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER,
"**nomem", "**nomem %s", "MPID_Segment_alloc");
MPIR_Pack_size_impl(count, datatype, &packsize);
last = data_sz;
if (packsize > 0) {
sreq->dev.tmpbuf = MPIU_Malloc((size_t) packsize);
MPIU_Assert(sreq->dev.tmpbuf);
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
MPID_Segment_pack(sreq->dev.segment_ptr, 0, &last, sreq->dev.tmpbuf);
req_area->iov_count = 1;
req_area->iov_buf[0].ptr = sreq->dev.tmpbuf;
req_area->iov_buf[0].length = last;
}
}
}
vc_area->pending_sends += 1;
mpi_errno = _mxm_isend(vc_area->mxm_ep, req_area, MXM_MPICH_ISEND_SYNC,
(mxm_mq_h) comm->dev.ch.netmod_priv, comm->rank, tag, _mxm_tag_mpi2mxm(tag,
comm->context_id
+
context_offset),
0);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
_dbg_mxm_out_req(sreq);
fn_exit:
*sreq_ptr = sreq;
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_MXM_ISSEND);
return mpi_errno;
fn_fail:
goto fn_exit;
}
static int _mxm_handle_sreq(MPID_Request * req)
{
int complete = FALSE;
int (*reqFn) (MPIDI_VC_t *, MPID_Request *, int *);
MPID_nem_mxm_vc_area *vc_area = NULL;
MPID_nem_mxm_req_area *req_area = NULL;
vc_area = VC_BASE(req->ch.vc);
req_area = REQ_BASE(req);
_dbg_mxm_out_buf(req_area->iov_buf[0].ptr,
(req_area->iov_buf[0].length >
16 ? 16 : req_area->iov_buf[0].length));
vc_area->pending_sends -= 1;
if (((req->dev.datatype_ptr != NULL) && (req->dev.tmpbuf != NULL))) {
MPIU_Free(req->dev.tmpbuf);
}
if (req_area->iov_count > MXM_MPICH_MAX_IOV) {
MPIU_Free(req_area->iov_buf);
req_area->iov_buf = req_area->tmp_buf;
req_area->iov_count = 0;
}
reqFn = req->dev.OnDataAvail;
if (!reqFn) {
MPIDI_CH3U_Request_complete(req);
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, ".... complete");
}
else {
MPIDI_VC_t *vc = req->ch.vc;
reqFn(vc, req, &complete);
if (!complete) {
MPIU_Assert(complete == TRUE);
}
}
return complete;
}
static void _mxm_send_completion_cb(void *context)
{
MPID_Request *req = (MPID_Request *) context;
MPID_nem_mxm_vc_area *vc_area = NULL;
MPID_nem_mxm_req_area *req_area = NULL;
MPIU_Assert(req);
_dbg_mxm_out_req(req);
vc_area = VC_BASE(req->ch.vc);
req_area = REQ_BASE(req);
_mxm_to_mpi_status(req_area->mxm_req->item.base.error, &req->status);
list_enqueue(&vc_area->mxm_ep->free_queue, &req_area->mxm_req->queue);
_dbg_mxm_output(5, "========> %s SEND req %p status %d\n",
(MPIR_STATUS_GET_CANCEL_BIT(req->status) ? "Canceling" : "Completing"),
req, req->status.MPI_ERROR);
if (likely(!MPIR_STATUS_GET_CANCEL_BIT(req->status))) {
_mxm_handle_sreq(req);
}
}
static int _mxm_isend(MPID_nem_mxm_ep_t * ep, MPID_nem_mxm_req_area * req,
int type, mxm_mq_h mxm_mq, int mxm_rank, int id, mxm_tag_t mxm_tag, int block)
{
int mpi_errno = MPI_SUCCESS;
mxm_error_t ret = MXM_OK;
mxm_send_req_t *mxm_sreq;
list_head_t *free_queue = NULL;
MPIU_Assert(ep);
MPIU_Assert(req);
free_queue = &ep->free_queue;
req->mxm_req = list_dequeue_mxm_req(free_queue);
if (!req->mxm_req) {
list_grow_mxm_req(free_queue);
req->mxm_req = list_dequeue_mxm_req(free_queue);
if (!req->mxm_req) {
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "empty free queue");
mpi_errno = MPI_ERR_OTHER;
goto fn_fail;
}
}
mxm_sreq = &(req->mxm_req->item.send);
mxm_sreq->base.state = MXM_REQ_NEW;
mxm_sreq->base.mq = mxm_mq;
mxm_sreq->base.conn = ep->mxm_conn;
mxm_sreq->base.completed_cb = _mxm_send_completion_cb;
mxm_sreq->base.context = req->ctx;
if (type == MXM_MPICH_ISEND_AM) {
mxm_sreq->opcode = MXM_REQ_OP_AM;
mxm_sreq->flags = 0;
mxm_sreq->op.am.hid = id;
mxm_sreq->op.am.imm_data = mxm_rank;
}
else if (type == MXM_MPICH_ISEND_SYNC) {
mxm_sreq->opcode = MXM_REQ_OP_SEND_SYNC;
mxm_sreq->flags = 0;
mxm_sreq->op.send.tag = mxm_tag;
mxm_sreq->op.send.imm_data = mxm_rank;
}
else {
mxm_sreq->opcode = MXM_REQ_OP_SEND;
mxm_sreq->flags = 0;
mxm_sreq->op.send.tag = mxm_tag;
mxm_sreq->op.send.imm_data = mxm_rank;
}
if (likely(req->iov_count == 1)) {
mxm_sreq->base.data_type = MXM_REQ_DATA_BUFFER;
mxm_sreq->base.data.buffer.ptr = req->iov_buf[0].ptr;
mxm_sreq->base.data.buffer.length = req->iov_buf[0].length;
}
else {
mxm_sreq->base.data_type = MXM_REQ_DATA_IOV;
mxm_sreq->base.data.iov.vector = req->iov_buf;
mxm_sreq->base.data.iov.count = req->iov_count;
}
ret = mxm_req_send(mxm_sreq);
if (MXM_OK != ret) {
list_enqueue(free_queue, &req->mxm_req->queue);
mpi_errno = MPI_ERR_OTHER;
goto fn_fail;
}
if (block)
_mxm_req_wait(&mxm_sreq->base);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
#if 0 /* Consider using this function in case non contiguous data */
static int _mxm_process_sdtype(MPID_Request ** sreq_p, MPI_Datatype datatype,
MPID_Datatype * dt_ptr, MPIDI_msg_sz_t data_sz, const void *buf,
int count, mxm_req_buffer_t ** iov_buf, int *iov_count)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *sreq = *sreq_p;
MPIDI_msg_sz_t last;
MPID_IOV *iov;
int n_iov = 0;
int index;
int size_to_copy = 0;
sreq->dev.segment_ptr = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((sreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem",
"**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
sreq->dev.segment_first = 0;
sreq->dev.segment_size = data_sz;
last = sreq->dev.segment_size;
MPID_Segment_count_contig_blocks(sreq->dev.segment_ptr, sreq->dev.segment_first, &last,
(MPI_Aint *) & n_iov);
MPIU_Assert(n_iov > 0);
iov = MPIU_Malloc(n_iov * sizeof(*iov));
MPIU_Assert(iov);
last = sreq->dev.segment_size;
MPID_Segment_pack_vector(sreq->dev.segment_ptr, sreq->dev.segment_first, &last, iov, &n_iov);
MPIU_Assert(last == sreq->dev.segment_size);
#if defined(MXM_DEBUG) && (MXM_DEBUG > 0)
_dbg_mxm_output(7, "Send Noncontiguous data vector %i entries (free slots : %i)\n", n_iov, MXM_REQ_DATA_MAX_IOV);
for(index = 0; index < n_iov; index++) {
_dbg_mxm_output(7, "======= Recv iov[%i] = ptr : %p, len : %i \n",
index, iov[index].MPID_IOV_BUF, iov[index].MPID_IOV_LEN);
}
#endif
if (n_iov > MXM_MPICH_MAX_IOV) {
*iov_buf = (mxm_req_buffer_t *) MPIU_Malloc(n_iov * sizeof(**iov_buf));
MPIU_Assert(*iov_buf);
}
for (index = 0; index < n_iov; index++) {
if (index < (MXM_REQ_DATA_MAX_IOV - 1)) {
(*iov_buf)[index].ptr = iov[index].MPID_IOV_BUF;
(*iov_buf)[index].length = iov[index].MPID_IOV_LEN;
}
else {
size_to_copy += iov[index].MPID_IOV_LEN;
}
}
if (size_to_copy == 0) {
sreq->dev.tmpbuf = NULL;
sreq->dev.tmpbuf_sz = 0;
*iov_count = n_iov;
}
else {
int offset = 0;
sreq->dev.tmpbuf = MPIU_Malloc(size_to_copy);
sreq->dev.tmpbuf_sz = size_to_copy;
MPIU_Assert(sreq->dev.tmpbuf);
for (index = (MXM_REQ_DATA_MAX_IOV - 1); index < n_iov; index++) {
MPIU_Memcpy((char *) (sreq->dev.tmpbuf) + offset, iov[index].MPID_IOV_BUF,
iov[index].MPID_IOV_LEN);
offset += iov[index].MPID_IOV_LEN;
}
(*iov_buf)[MXM_REQ_DATA_MAX_IOV - 1].ptr = sreq->dev.tmpbuf;
(*iov_buf)[MXM_REQ_DATA_MAX_IOV - 1].length = size_to_copy;
*iov_count = MXM_REQ_DATA_MAX_IOV;
}
MPIU_Free(iov);
fn_exit:
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPIDI_CH3U_Request_unpack_uebuf(MPID_Request * rreq)
{
int dt_contig;
MPI_Aint dt_true_lb;
MPIDI_msg_sz_t userbuf_sz;
MPID_Datatype * dt_ptr;
MPIDI_msg_sz_t unpack_sz;
int mpi_errno = MPI_SUCCESS;
MPIDI_STATE_DECL(MPID_STATE_MPIDI_CH3U_REQUEST_UNPACK_UEBUF);
MPIDI_STATE_DECL(MPID_STATE_MEMCPY);
MPIDI_FUNC_ENTER(MPID_STATE_MPIDI_CH3U_REQUEST_UNPACK_UEBUF);
MPIDI_Datatype_get_info(rreq->dev.user_count, rreq->dev.datatype,
dt_contig, userbuf_sz, dt_ptr, dt_true_lb);
if (rreq->dev.recv_data_sz <= userbuf_sz)
{
unpack_sz = rreq->dev.recv_data_sz;
}
else
{
/* --BEGIN ERROR HANDLING-- */
MPIU_DBG_MSG_FMT(CH3_CHANNEL,VERBOSE,(MPIU_DBG_FDEST,
"receive buffer overflow; message truncated, msg_sz=" MPIDI_MSG_SZ_FMT
", buf_sz=" MPIDI_MSG_SZ_FMT,
rreq->dev.recv_data_sz, userbuf_sz));
unpack_sz = userbuf_sz;
MPIR_STATUS_SET_COUNT(rreq->status, userbuf_sz);
rreq->status.MPI_ERROR = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_TRUNCATE,
"**truncate", "**truncate %d %d",
rreq->dev.recv_data_sz, userbuf_sz);
/* --END ERROR HANDLING-- */
}
if (unpack_sz > 0)
{
if (dt_contig)
{
/* TODO - check that amount of data is consistent with datatype.
In other words, if we were to use Segment_unpack()
would last = unpack? If not we should return an error
(unless configured with --enable-fast) */
MPIDI_FUNC_ENTER(MPID_STATE_MEMCPY);
MPIU_Memcpy((char *)rreq->dev.user_buf + dt_true_lb, rreq->dev.tmpbuf,
unpack_sz);
MPIDI_FUNC_EXIT(MPID_STATE_MEMCPY);
}
else
{
MPID_Segment seg;
MPI_Aint last;
MPID_Segment_init(rreq->dev.user_buf, rreq->dev.user_count,
rreq->dev.datatype, &seg, 0);
last = unpack_sz;
MPID_Segment_unpack(&seg, 0, &last, rreq->dev.tmpbuf);
if (last != unpack_sz)
{
/* --BEGIN ERROR HANDLING-- */
/* received data was not entirely consumed by unpack()
because too few bytes remained to fill the next basic
datatype */
MPIR_STATUS_SET_COUNT(rreq->status, last);
rreq->status.MPI_ERROR = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_TYPE,
"**dtypemismatch", 0);
/* --END ERROR HANDLING-- */
}
}
}
MPIDI_FUNC_EXIT(MPID_STATE_MPIDI_CH3U_REQUEST_UNPACK_UEBUF);
return mpi_errno;
}
int MPID_nem_tcp_module_lmt_start_recv (MPIDI_VC_t *vc, MPID_Request *req)
{
int mpi_errno = MPI_SUCCESS;
int ret;
MPIDI_msg_sz_t data_sz;
int dt_contig;
MPI_Aint dt_true_lb;
MPID_Datatype * dt_ptr;
MPIDI_msg_sz_t last;
int nb;
int r_len;
MPIDI_CH3I_VC *vc_ch = (MPIDI_CH3I_VC *)vc->channel_private;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_TCP_MODULE_LMT_START_RECV);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_TCP_MODULE_LMT_START_RECV);
free_cookie (vc_ch->net.tcp.lmt_cookie);
if (!vc_ch->net.tcp.lmt_connected)
{
int len;
struct sockaddr_in saddr;
int connfd;
len = sizeof (saddr);
connfd = accept (vc_ch->net.tcp.lmt_desc, (struct sockaddr *)&saddr, &len);
MPIU_ERR_CHKANDJUMP2 (connfd == -1, mpi_errno, MPI_ERR_OTHER, "**sock|poll|accept", "**sock|poll|accept %d %s", errno, strerror (errno));
/* close listen fd */
do
ret = close (vc_ch->net.tcp.lmt_desc);
while (ret == -1 && errno == EINTR);
MPIU_ERR_CHKANDJUMP2 (ret == -1, mpi_errno, MPI_ERR_OTHER, "**closesocket", "**closesocket %s %d", strerror (errno), errno);
/* set lmt_desc to new connected fd */
vc_ch->net.tcp.lmt_desc = connfd;
vc_ch->net.tcp.lmt_connected = 1;
// ret = fcntl (vc_ch->net.tcp.lmt_desc, F_SETFL, O_NONBLOCK);
// MPIU_ERR_CHKANDJUMP2 (ret == -1, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %s %d", strerror (errno), errno);
}
MPIDI_Datatype_get_info (req->dev.user_count, req->dev.datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
if (data_sz > vc_ch->net.tcp.lmt_s_len)
{
data_sz = vc_ch->net.tcp.lmt_s_len;
}
else if (data_sz < vc_ch->net.tcp.lmt_s_len)
{
/* message will be truncated */
r_len = data_sz;
req->status.MPI_ERROR = MPIU_ERR_SET2 (mpi_errno, MPI_ERR_TRUNCATE, "**truncate", "**truncate %d %d", vc_ch->net.tcp.lmt_s_len, r_len);
}
MPID_Segment_init (req->dev.user_buf, req->dev.user_count, req->dev.datatype, &req->dev.segment, 0);
req->dev.segment_first = 0;
req->dev.segment_size = data_sz;
req->dev.iov_count = MPID_IOV_LIMIT;
req->dev.iov_offset = 0;
last = data_sz;
do
{
int iov_offset;
int left_to_recv;
MPID_Segment_unpack_vector (&req->dev.segment, req->dev.segment_first, &last, req->dev.iov, &req->dev.iov_count);
left_to_recv = last - req->dev.segment_first;
iov_offset = 0;
#ifdef TESTING_CHUNKING
{
char *buf = req->dev.iov[0].MPID_IOV_BUF;
int l;
while (left_to_recv)
{
if (left_to_recv > CHUNK)
l = CHUNK;
else
l = left_to_recv;
do
nb = read (vc_ch->net.tcp.lmt_desc, buf, l);
while (nb == -1 && errno == EINTR);
MPIU_ERR_CHKANDJUMP (nb == -1, mpi_errno, MPI_ERR_OTHER, "**sock_writev");
left_to_recv -= nb;
buf += nb;
}
MPIDI_CH3U_Request_complete (req);
goto fn_exit;
}
#endif
do
nb = readv (vc_ch->net.tcp.lmt_desc, &req->dev.iov[iov_offset], req->dev.iov_count - iov_offset);
while (nb == -1 && errno == EINTR);
MPIU_ERR_CHKANDJUMP2 (nb == -1, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %s %d", strerror (errno), errno);
MPIU_ERR_CHKANDJUMP (nb == 0, mpi_errno, MPI_ERR_OTHER, "**fail");
left_to_recv -= nb;
while (left_to_recv)
{ /* recv rest of iov */
while (nb >= req->dev.iov[iov_offset].MPID_IOV_LEN)
{ /* update iov to reflect sent bytes */
nb -= req->dev.iov[iov_offset].MPID_IOV_LEN;
++iov_offset;
}
req->dev.iov[iov_offset].MPID_IOV_BUF = (char *)req->dev.iov[iov_offset].MPID_IOV_BUF + nb;
req->dev.iov[iov_offset].MPID_IOV_LEN -= nb;
do
nb = readv (vc_ch->net.tcp.lmt_desc, &req->dev.iov[iov_offset], req->dev.iov_count - iov_offset);
while (nb == -1 && errno == EINTR);
MPIU_ERR_CHKANDJUMP2 (nb == -1, mpi_errno, MPI_ERR_OTHER, "**fail", "**fail %s %d", strerror (errno), errno);
MPIU_ERR_CHKANDJUMP (nb == 0, mpi_errno, MPI_ERR_OTHER, "**fail");
left_to_recv -= nb;
}
}
while (last < data_sz);
MPIDI_CH3U_Request_complete (req);
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_TCP_MODULE_LMT_START_RECV);
return mpi_errno;
fn_fail:
goto fn_exit;
}
/*@
MPI_Unpack_external - Unpack a buffer (packed with MPI_Pack_external)
according to a datatype into contiguous memory
Input Parameters:
+ datarep - data representation (string)
. inbuf - input buffer start (choice)
. insize - input buffer size, in bytes (address integer)
. outcount - number of output data items (integer)
. datatype - datatype of output data item (handle)
Input/Output Parameters:
. position - current position in buffer, in bytes (address integer)
Output Parameters:
. outbuf - output buffer start (choice)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_ARG
@*/
int MPI_Unpack_external(const char datarep[],
const void *inbuf,
MPI_Aint insize,
MPI_Aint *position,
void *outbuf,
int outcount,
MPI_Datatype datatype)
{
static const char FCNAME[] = "MPI_Unpack_external";
int mpi_errno = MPI_SUCCESS;
MPI_Aint first, last;
MPID_Segment *segp;
MPIR_FUNC_TERSE_STATE_DECL(MPID_STATE_MPI_UNPACK_EXTERNAL);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPIR_FUNC_TERSE_ENTER(MPID_STATE_MPI_UNPACK_EXTERNAL);
/* Validate parameters, especially handles needing to be converted */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
if (insize > 0) {
MPIR_ERRTEST_ARGNULL(inbuf, "input buffer", mpi_errno);
}
/* NOTE: outbuf could be MPI_BOTTOM; don't test for NULL */
MPIR_ERRTEST_COUNT(insize, mpi_errno);
MPIR_ERRTEST_COUNT(outcount, mpi_errno);
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
if (datatype != MPI_DATATYPE_NULL && HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPIR_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPIR_Datatype_valid_ptr(datatype_ptr, mpi_errno);
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
}
/* If datatye_ptr is not valid, it will be reset to null */
if (mpi_errno) goto fn_fail;
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
if (insize == 0) {
goto fn_exit;
}
segp = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1((segp == NULL), mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
mpi_errno = MPID_Segment_init(outbuf, outcount, datatype, segp, 1);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* NOTE: buffer values and positions in MPI_Unpack_external are used very
* differently from use in MPID_Segment_unpack_external...
*/
first = 0;
last = SEGMENT_IGNORE_LAST;
/* Ensure that pointer increment fits in a pointer */
MPIR_Ensure_Aint_fits_in_pointer((MPIR_VOID_PTR_CAST_TO_MPI_AINT inbuf) + *position);
MPID_Segment_unpack_external32(segp,
first,
&last,
(void *) ((char *) inbuf + *position));
*position += last;
MPID_Segment_free(segp);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* ... end of body of routine ... */
fn_exit:
MPIR_FUNC_TERSE_EXIT(MPID_STATE_MPI_UNPACK_EXTERNAL);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_unpack_external",
"**mpi_unpack_external %s %p %d %p %p %d %D", datarep, inbuf, insize, position, outbuf, outcount, datatype);
}
# endif
mpi_errno = MPIR_Err_return_comm(NULL, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPIDI_Accumulate(void *origin_addr, int origin_count, MPI_Datatype
origin_datatype, int target_rank, MPI_Aint target_disp,
int target_count, MPI_Datatype target_datatype, MPI_Op op,
MPID_Win *win_ptr)
{
int mpi_errno=MPI_SUCCESS;
MPIDI_msg_sz_t data_sz;
int dt_contig, rank, origin_predefined, target_predefined;
MPI_Aint dt_true_lb;
MPIDI_RMA_ops *new_ptr;
MPID_Datatype *dtp;
MPIU_CHKLMEM_DECL(2);
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPIDI_ACCUMULATE);
MPIDI_RMA_FUNC_ENTER(MPID_STATE_MPIDI_ACCUMULATE);
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;
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(origin_datatype, origin_predefined);
MPIDI_CH3I_DATATYPE_IS_PREDEFINED(target_datatype, target_predefined);
/* Do =! rank first (most likely branch?) */
if (target_rank == rank)
{
MPI_User_function *uop;
if (op == MPI_REPLACE)
{
mpi_errno = MPIR_Localcopy(origin_addr, origin_count,
origin_datatype,
(char *) win_ptr->base + win_ptr->disp_unit *
target_disp, target_count, target_datatype);
goto fn_exit;
}
MPIU_ERR_CHKANDJUMP1((HANDLE_GET_KIND(op) != HANDLE_KIND_BUILTIN),
mpi_errno, MPI_ERR_OP, "**opnotpredefined",
"**opnotpredefined %d", op );
/* get the function by indexing into the op table */
uop = MPIR_Op_table[((op)&0xf) - 1];
if (origin_predefined && target_predefined)
{
(*uop)(origin_addr, (char *) win_ptr->base + win_ptr->disp_unit *
target_disp, &target_count, &target_datatype);
}
else
{
/* derived datatype */
MPID_Segment *segp;
DLOOP_VECTOR *dloop_vec;
MPI_Aint first, last;
int vec_len, i, type_size, count;
MPI_Datatype type;
MPI_Aint true_lb, true_extent, extent;
void *tmp_buf=NULL, *source_buf, *target_buf;
if (origin_datatype != target_datatype)
{
/* first copy the data into a temporary buffer with
the same datatype as the target. Then do the
accumulate operation. */
MPIR_Type_get_true_extent_impl(target_datatype, &true_lb, &true_extent);
MPID_Datatype_get_extent_macro(target_datatype, extent);
MPIU_CHKLMEM_MALLOC(tmp_buf, void *,
target_count * (MPIR_MAX(extent,true_extent)),
mpi_errno, "temporary buffer");
/* adjust for potential negative lower bound in datatype */
tmp_buf = (void *)((char*)tmp_buf - true_lb);
mpi_errno = MPIR_Localcopy(origin_addr, origin_count,
origin_datatype, tmp_buf,
target_count, target_datatype);
if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
}
if (target_predefined) {
/* target predefined type, origin derived datatype */
(*uop)(tmp_buf, (char *) win_ptr->base + win_ptr->disp_unit *
target_disp, &target_count, &target_datatype);
}
else {
segp = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((!segp), mpi_errno, MPI_ERR_OTHER,
"**nomem","**nomem %s","MPID_Segment_alloc");
MPID_Segment_init(NULL, target_count, target_datatype, segp, 0);
first = 0;
last = SEGMENT_IGNORE_LAST;
MPID_Datatype_get_ptr(target_datatype, dtp);
vec_len = dtp->max_contig_blocks * target_count + 1;
/* +1 needed because Rob says so */
MPIU_CHKLMEM_MALLOC(dloop_vec, DLOOP_VECTOR *,
vec_len * sizeof(DLOOP_VECTOR),
mpi_errno, "dloop vector");
MPID_Segment_pack_vector(segp, first, &last, dloop_vec, &vec_len);
source_buf = (tmp_buf != NULL) ? tmp_buf : origin_addr;
target_buf = (char *) win_ptr->base +
win_ptr->disp_unit * target_disp;
type = dtp->eltype;
type_size = MPID_Datatype_get_basic_size(type);
for (i=0; i<vec_len; i++)
{
count = (dloop_vec[i].DLOOP_VECTOR_LEN)/type_size;
(*uop)((char *)source_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF),
(char *)target_buf + MPIU_PtrToAint(dloop_vec[i].DLOOP_VECTOR_BUF),
&count, &type);
}
MPID_Segment_free(segp);
}
}
int MPID_nem_mxm_issend(MPIDI_VC_t * vc, const void *buf, int count, MPI_Datatype datatype,
int rank, int tag, MPID_Comm * comm, int context_offset,
MPID_Request ** sreq_ptr)
{
int mpi_errno = MPI_SUCCESS;
MPID_Request *sreq = NULL;
MPID_Datatype *dt_ptr;
int dt_contig;
MPIDI_msg_sz_t data_sz;
MPI_Aint dt_true_lb;
MPID_nem_mxm_vc_area *vc_area = NULL;
MPID_nem_mxm_req_area *req_area = NULL;
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_MXM_ISSEND);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_MXM_ISSEND);
MPIDI_Datatype_get_info(count, datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
/* create a request */
MPIDI_Request_create_sreq(sreq, mpi_errno, goto fn_exit);
MPIU_Assert(sreq != NULL);
MPIDI_Request_set_type(sreq, MPIDI_REQUEST_TYPE_SEND);
MPIDI_VC_FAI_send_seqnum(vc, seqnum);
MPIDI_Request_set_seqnum(sreq, seqnum);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype_get_ptr(datatype, sreq->dev.datatype_ptr);
MPID_Datatype_add_ref(sreq->dev.datatype_ptr);
}
sreq->partner_request = NULL;
sreq->dev.OnDataAvail = NULL;
sreq->dev.tmpbuf = NULL;
sreq->ch.vc = vc;
sreq->ch.noncontig = FALSE;
_dbg_mxm_output(5,
"isSend ========> Sending USER msg for req %p (context %d to %d tag %d size %d) \n",
sreq, comm->context_id + context_offset, rank, tag, data_sz);
vc_area = VC_BASE(vc);
req_area = REQ_BASE(sreq);
req_area-> ctx = sreq;
req_area->iov_buf = req_area->tmp_buf;
req_area->iov_count = 0;
req_area->iov_buf[0].ptr = NULL;
req_area->iov_buf[0].length = 0;
if (data_sz) {
if (dt_contig) {
req_area->iov_count = 1;
req_area->iov_buf[0].ptr = (char *) (buf) + dt_true_lb;
req_area->iov_buf[0].length = data_sz;
}
else {
MPIDI_msg_sz_t last;
MPI_Aint packsize = 0;
sreq->ch.noncontig = TRUE;
sreq->dev.segment_ptr = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP1((sreq->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER,
"**nomem", "**nomem %s", "MPID_Segment_alloc");
MPIR_Pack_size_impl(count, datatype, &packsize);
last = data_sz;
if (packsize > 0) {
sreq->dev.tmpbuf = MPIU_Malloc((size_t) packsize);
MPIU_Assert(sreq->dev.tmpbuf);
MPID_Segment_init(buf, count, datatype, sreq->dev.segment_ptr, 0);
MPID_Segment_pack(sreq->dev.segment_ptr, 0, &last, sreq->dev.tmpbuf);
req_area->iov_count = 1;
req_area->iov_buf[0].ptr = sreq->dev.tmpbuf;
req_area->iov_buf[0].length = last;
}
}
}
vc_area->pending_sends += 1;
mpi_errno = _mxm_isend(vc_area->mxm_ep, req_area, MXM_MPICH_ISEND_SYNC,
(mxm_mq_h) comm->dev.ch.netmod_priv, comm->rank, tag, _mxm_tag_mpi2mxm(tag,
comm->context_id
+
context_offset),
0);
if (mpi_errno)
MPIU_ERR_POP(mpi_errno);
_dbg_mxm_out_req(sreq);
fn_exit:
*sreq_ptr = sreq;
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_MXM_ISSEND);
return mpi_errno;
fn_fail:
goto fn_exit;
}
int MPID_nem_ptl_recv_posted(MPIDI_VC_t *vc, MPID_Request *rreq)
{
int mpi_errno = MPI_SUCCESS;
MPID_nem_ptl_vc_area *const vc_ptl = VC_PTL(vc);
ptl_me_t me;
int dt_contig;
MPIDI_msg_sz_t data_sz;
MPID_Datatype *dt_ptr;
MPI_Aint dt_true_lb;
MPI_Aint last;
ptl_process_t id_any;
int ret;
MPIU_CHKPMEM_DECL(1);
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_PTL_RECV_POSTED);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_PTL_RECV_POSTED);
id_any.phys.nid = PTL_NID_ANY;
id_any.phys.pid = PTL_PID_ANY;
MPID_nem_ptl_init_req(rreq);
me.ct_handle = PTL_CT_NONE;
me.uid = PTL_UID_ANY;
me.options = ( PTL_ME_OP_PUT | PTL_ME_IS_ACCESSIBLE | PTL_ME_EVENT_LINK_DISABLE |
PTL_ME_EVENT_UNLINK_DISABLE | PTL_ME_USE_ONCE );
if (vc == NULL) {
/* MPI_ANY_SOURCE receive */
me.match_id = id_any;
} else {
if (!vc_ptl->id_initialized) {
mpi_errno = MPID_nem_ptl_init_id(vc);
if (mpi_errno) MPIR_ERR_POP(mpi_errno);
}
me.match_id = vc_ptl->id;
}
MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE, (MPIU_DBG_FDEST, "tag=%#x ctx=%#x rank=%#x", rreq->dev.match.parts.tag, rreq->dev.match.parts.context_id, rreq->dev.match.parts.rank));
me.match_bits = NPTL_MATCH(rreq->dev.match.parts.tag, rreq->dev.match.parts.context_id,
rreq->dev.match.parts.rank);
if (rreq->dev.match.parts.tag == MPI_ANY_TAG)
me.ignore_bits = NPTL_MATCH_IGNORE_ANY_TAG;
else
me.ignore_bits = NPTL_MATCH_IGNORE;
me.min_free = 0;
MPIDI_Datatype_get_info(rreq->dev.user_count, rreq->dev.datatype, dt_contig, data_sz, dt_ptr, dt_true_lb);
MPIU_DBG_MSG_FMT(CH3_CHANNEL, VERBOSE, (MPIU_DBG_FDEST, "count="MPI_AINT_FMT_DEC_SPEC" datatype=%#x contig=%d data_sz=%lu", rreq->dev.user_count, rreq->dev.datatype, dt_contig, data_sz));
if (data_sz <= PTL_LARGE_THRESHOLD) {
if (dt_contig) {
/* small contig message */
void *start = (char *)rreq->dev.user_buf + dt_true_lb;
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Small contig message");
if (start == NULL)
me.start = &dummy;
else
me.start = start;
me.length = data_sz;
REQ_PTL(rreq)->event_handler = handler_recv_dequeue_complete;
} else {
/* small noncontig */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, "Small noncontig message");
rreq->dev.segment_ptr = MPID_Segment_alloc();
MPIR_ERR_CHKANDJUMP1(rreq->dev.segment_ptr == NULL, mpi_errno, MPI_ERR_OTHER, "**nomem", "**nomem %s", "MPID_Segment_alloc");
MPID_Segment_init(rreq->dev.user_buf, rreq->dev.user_count, rreq->dev.datatype, rreq->dev.segment_ptr, 0);
rreq->dev.segment_first = 0;
rreq->dev.segment_size = data_sz;
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) {
/* entire message fits in IOV */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " entire message fits in IOV");
me.start = rreq->dev.iov;
me.length = rreq->dev.iov_count;
me.options |= PTL_IOVEC;
REQ_PTL(rreq)->event_handler = handler_recv_dequeue_complete;
} else {
/* IOV is not long enough to describe entire message: recv into
buffer and unpack later */
MPIU_DBG_MSG(CH3_CHANNEL, VERBOSE, " IOV too long: using bounce buffer");
MPIU_CHKPMEM_MALLOC(REQ_PTL(rreq)->chunk_buffer[0], void *, data_sz, mpi_errno, "chunk_buffer");
me.start = REQ_PTL(rreq)->chunk_buffer[0];
me.length = data_sz;
REQ_PTL(rreq)->event_handler = handler_recv_dequeue_unpack_complete;
}
}
} else {
/* Large message: Create an ME for the first chunk of data, then do a GET for the rest */
if (dt_contig) {
/*@
MPI_Pack_external - Packs a datatype into contiguous memory, using the
external32 format
Input Parameters:
+ datarep - data representation (string)
. inbuf - input buffer start (choice)
. incount - number of input data items (integer)
. datatype - datatype of each input data item (handle)
- outsize - output buffer size, in bytes (address integer)
Output Parameters:
. outbuf - output buffer start (choice)
Input/Output Parameters:
. position - current position in buffer, in bytes (address integer)
.N ThreadSafe
.N Fortran
.N Errors
.N MPI_SUCCESS
.N MPI_ERR_TYPE
.N MPI_ERR_ARG
.N MPI_ERR_COUNT
@*/
int MPI_Pack_external(const char datarep[],
const void *inbuf,
int incount,
MPI_Datatype datatype,
void *outbuf,
MPI_Aint outsize,
MPI_Aint *position)
{
static const char FCNAME[] = "MPI_Pack_external";
int mpi_errno = MPI_SUCCESS;
MPI_Aint first, last;
MPID_Segment *segp;
MPID_MPI_STATE_DECL(MPID_STATE_MPI_PACK_EXTERNAL);
MPIR_ERRTEST_INITIALIZED_ORDIE();
MPID_MPI_FUNC_ENTER(MPID_STATE_MPI_PACK_EXTERNAL);
/* Validate parameters and objects (post conversion) */
# ifdef HAVE_ERROR_CHECKING
{
MPID_BEGIN_ERROR_CHECKS;
{
MPIR_ERRTEST_COUNT(incount, mpi_errno);
MPIR_ERRTEST_COUNT(outsize, mpi_errno);
/* NOTE: inbuf could be null (MPI_BOTTOM) */
if (incount > 0) {
MPIR_ERRTEST_ARGNULL(outbuf, "output buffer", mpi_errno);
}
MPIR_ERRTEST_ARGNULL(position, "position", mpi_errno);
MPIR_ERRTEST_DATATYPE(datatype, "datatype", mpi_errno);
if (HANDLE_GET_KIND(datatype) != HANDLE_KIND_BUILTIN) {
MPID_Datatype *datatype_ptr = NULL;
MPID_Datatype_get_ptr(datatype, datatype_ptr);
MPID_Datatype_valid_ptr(datatype_ptr, mpi_errno);
MPID_Datatype_committed_ptr(datatype_ptr, mpi_errno);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
}
}
MPID_END_ERROR_CHECKS;
}
# endif /* HAVE_ERROR_CHECKING */
/* ... body of routine ... */
if (incount == 0) {
goto fn_exit;
}
segp = MPID_Segment_alloc();
/* --BEGIN ERROR HANDLING-- */
if (segp == NULL)
{
mpi_errno = MPIR_Err_create_code(MPI_SUCCESS,
MPIR_ERR_RECOVERABLE,
FCNAME,
__LINE__,
MPI_ERR_OTHER,
"**nomem",
"**nomem %s",
"MPID_Segment");
goto fn_fail;
}
/* --END ERROR HANDLING-- */
mpi_errno = MPID_Segment_init(inbuf, incount, datatype, segp, 1);
if (mpi_errno != MPI_SUCCESS) goto fn_fail;
/* NOTE: the use of buffer values and positions in MPI_Pack_external and
* in MPID_Segment_pack_external are quite different. See code or docs
* or something.
*/
first = 0;
last = SEGMENT_IGNORE_LAST;
/* Ensure that pointer increment fits in a pointer */
MPID_Ensure_Aint_fits_in_pointer((MPI_VOID_PTR_CAST_TO_MPI_AINT outbuf) + *position);
MPID_Segment_pack_external32(segp,
first,
&last,
(void *)((char *) outbuf + *position));
*position += last;
MPID_Segment_free(segp);
/* ... end of body of routine ... */
fn_exit:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPI_PACK_EXTERNAL);
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
# ifdef HAVE_ERROR_CHECKING
{
mpi_errno = MPIR_Err_create_code(
mpi_errno, MPIR_ERR_RECOVERABLE, FCNAME, __LINE__, MPI_ERR_OTHER, "**mpi_pack_external",
"**mpi_pack_external %s %p %d %D %p %d %p", datarep, inbuf, incount, datatype, outbuf, outsize, position);
}
# endif
mpi_errno = MPIR_Err_return_comm(0, FCNAME, mpi_errno);
goto fn_exit;
/* --END ERROR HANDLING-- */
}
int MPID_nem_ib_lmt_initiate_lmt(struct MPIDI_VC *vc, union MPIDI_CH3_Pkt *rts_pkt,
struct MPID_Request *req)
{
int mpi_errno = MPI_SUCCESS;
int dt_contig;
MPIDI_msg_sz_t data_sz;
MPID_Datatype *dt_ptr;
MPI_Aint dt_true_lb;
#if 0
MPID_nem_ib_vc_area *vc_ib = VC_IB(vc);
#endif
MPIDI_STATE_DECL(MPID_STATE_MPID_NEM_IB_LMT_INITIATE_LMT);
MPIDI_FUNC_ENTER(MPID_STATE_MPID_NEM_IB_LMT_INITIATE_LMT);
dprintf("lmt_initiate_lmt,enter,%d->%d,req=%p\n", MPID_nem_ib_myrank, vc->pg_rank, req);
/* obtain dt_true_lb */
/* see MPIDI_Datatype_get_info(in, in, out, out, out, out) (in src/mpid/ch3/include/mpidimpl.h) */
MPIDI_Datatype_get_info(req->dev.user_count, req->dev.datatype, dt_contig, data_sz, dt_ptr,
dt_true_lb);
/* FIXME: who frees s_cookie_buf? */
/* malloc memory area for cookie. auto variable is NG because isend does not copy payload */
MPID_nem_ib_lmt_cookie_t *s_cookie_buf =
(MPID_nem_ib_lmt_cookie_t *) MPIU_Malloc(sizeof(MPID_nem_ib_lmt_cookie_t));
/* remember address to "free" when receiving DONE from receiver */
req->ch.s_cookie = s_cookie_buf;
/* see MPIDI_CH3_PktHandler_RndvClrToSend (in src/mpid/ch3/src/ch3u_rndv.c) */
//assert(dt_true_lb == 0);
void *write_from_buf;
if (dt_contig) {
write_from_buf = (void *) ((char *) req->dev.user_buf + dt_true_lb);
}
else {
/* see MPIDI_CH3_EagerNoncontigSend (in ch3u_eager.c) */
req->dev.segment_ptr = MPID_Segment_alloc();
MPIU_ERR_CHKANDJUMP((req->dev.segment_ptr == NULL), mpi_errno, MPI_ERR_OTHER,
"**outofmemory");
MPID_Segment_init(req->dev.user_buf, req->dev.user_count, req->dev.datatype,
req->dev.segment_ptr, 0);
req->dev.segment_first = 0;
req->dev.segment_size = data_sz;
MPIDI_msg_sz_t last;
last = req->dev.segment_size; /* segment_size is byte offset */
MPIU_Assert(last > 0);
REQ_FIELD(req, lmt_pack_buf) = MPIU_Malloc((size_t) req->dev.segment_size);
MPIU_ERR_CHKANDJUMP(!REQ_FIELD(req, lmt_pack_buf), mpi_errno, MPI_ERR_OTHER,
"**outofmemory");
MPID_Segment_pack(req->dev.segment_ptr, req->dev.segment_first, &last,
(char *) (REQ_FIELD(req, lmt_pack_buf)));
MPIU_Assert(last == req->dev.segment_size);
write_from_buf = REQ_FIELD(req, lmt_pack_buf);
}
dprintf
("lmt_initate_lmt,dt_contig=%d,write_from_buf=%p,req->dev.user_buf=%p,REQ_FIELD(req, lmt_pack_buf)=%p\n",
dt_contig, write_from_buf, req->dev.user_buf, REQ_FIELD(req, lmt_pack_buf));
#ifdef HAVE_LIBDCFA
#else
s_cookie_buf->addr = write_from_buf;
#endif
/* put sz, see MPID_nem_lmt_RndvSend (in src/mpid/ch3/channels/nemesis/src/mpid_nem_lmt.c) */
/* TODO remove sz field
* pkt_RTS_handler (in src/mpid/ch3/channels/nemesis/src/mpid_nem_lmt.c)
* rreq->ch.lmt_data_sz = rts_pkt->data_sz; */
//s_cookie_buf->sz = (uint32_t)((MPID_nem_pkt_lmt_rts_t*)rts_pkt)->data_sz;
/* preserve and put tail, because tail magic is written on the tail of payload
* because we don't want to add another SGE or RDMA command */
MPIU_Assert(((MPID_nem_pkt_lmt_rts_t *) rts_pkt)->data_sz == data_sz);
s_cookie_buf->tail = *((uint8_t *) ((uint8_t *) write_from_buf + data_sz - sizeof(uint8_t)));
/* prepare magic */
//*((uint32_t*)(write_from_buf + data_sz - sizeof(tailmagic_t))) = MPID_NEM_IB_COM_MAGIC;
#if 0 /* moving to packet header */ /* embed RDMA-write-to buffer occupancy information */
dprintf("lmt_initiate_lmt,rsr_seq_num_tail=%d\n", vc_ib->ibcom->rsr_seq_num_tail);
/* embed RDMA-write-to buffer occupancy information */
s_cookie_buf->seq_num_tail = vc_ib->ibcom->rsr_seq_num_tail;
/* remember the last one sent */
vc_ib->ibcom->rsr_seq_num_tail_last_sent = vc_ib->ibcom->rsr_seq_num_tail;
#endif
int post_num;
uint32_t max_msg_sz;
MPID_nem_ib_vc_area *vc_ib = VC_IB(vc);
MPID_nem_ib_com_get_info_conn(vc_ib->sc->fd, MPID_NEM_IB_COM_INFOKEY_PATTR_MAX_MSG_SZ,
&max_msg_sz, sizeof(uint32_t));
/* Type of max_msg_sz is uint32_t. */
post_num = (data_sz + (long) max_msg_sz - 1) / (long) max_msg_sz;
s_cookie_buf->max_msg_sz = max_msg_sz;
s_cookie_buf->seg_seq_num = 1;
s_cookie_buf->seg_num = post_num;
REQ_FIELD(req, buf.from) = write_from_buf;
REQ_FIELD(req, data_sz) = data_sz;
REQ_FIELD(req, seg_seq_num) = 1; // only send 1st-segment, even if there are some segments.
REQ_FIELD(req, seg_num) = post_num;
REQ_FIELD(req, max_msg_sz) = max_msg_sz;
long length;
if (post_num > 1) {
length = max_msg_sz;
}
else {
length = data_sz;
}
/* put IB rkey */
struct MPID_nem_ib_com_reg_mr_cache_entry_t *mr_cache =
MPID_nem_ib_com_reg_mr_fetch(write_from_buf, length, 0, MPID_NEM_IB_COM_REG_MR_GLOBAL);
MPIU_ERR_CHKANDJUMP(!mr_cache, mpi_errno, MPI_ERR_OTHER, "**MPID_nem_ib_com_reg_mr_fetch");
struct ibv_mr *mr = mr_cache->mr;
REQ_FIELD(req, lmt_mr_cache) = (void *) mr_cache;
#ifdef HAVE_LIBDCFA
s_cookie_buf->addr = (void *) mr->host_addr;
dprintf("lmt_initiate_lmt,s_cookie_buf->addr=%p\n", s_cookie_buf->addr);
#endif
s_cookie_buf->rkey = mr->rkey;
dprintf("lmt_initiate_lmt,tail=%02x,mem-tail=%p,%02x,sz=%ld,raddr=%p,rkey=%08x\n",
s_cookie_buf->tail, write_from_buf + data_sz - sizeof(uint8_t),
*((uint8_t *) (write_from_buf + data_sz - sizeof(uint8_t))), data_sz,
s_cookie_buf->addr, s_cookie_buf->rkey);
/* send cookie. rts_pkt as the MPI-header, s_cookie_buf as the payload */
MPID_nem_lmt_send_RTS(vc, (MPID_nem_pkt_lmt_rts_t *) rts_pkt, s_cookie_buf,
sizeof(MPID_nem_ib_lmt_cookie_t));
fn_exit:
MPIDI_FUNC_EXIT(MPID_STATE_MPID_NEM_IB_LMT_INITIATE_LMT);
return mpi_errno;
fn_fail:
goto fn_exit;
}
