int
shmem_transport_fini(void)
{
ptl_ct_event_t ct;
/* synchronize the atomic cache, if there is one */
PtlAtomicSync();
/* wait for remote completion (acks) of all pending events */
PtlCTWait(shmem_transport_portals4_put_ct_h,
shmem_transport_portals4_pending_put_counter, &ct);
if (shmem_transport_portals4_pending_put_counter != ct.success + ct.failure) {
fprintf(stderr, "[%03d] WARNING: put count mismatch: %ld, %ld\n",
shmem_internal_my_pe, (long) shmem_transport_portals4_pending_put_counter,
(long) (ct.success + ct.failure));
}
PtlCTWait(shmem_transport_portals4_get_ct_h,
shmem_transport_portals4_pending_get_counter, &ct);
if (shmem_transport_portals4_pending_get_counter != ct.success + ct.failure) {
fprintf(stderr, "[%03d] WARNING: get count mismatch: %ld, %ld\n",
shmem_internal_my_pe, (long) shmem_transport_portals4_pending_get_counter,
(long) (ct.success + ct.failure));
}
cleanup_handles();
PtlFini();
SHMEM_MUTEX_DESTROY(shmem_internal_mutex_ptl4_pt_state);
SHMEM_MUTEX_DESTROY(shmem_internal_mutex_ptl4_frag);
SHMEM_MUTEX_DESTROY(shmem_internal_mutex_ptl4_event_slots);
SHMEM_MUTEX_DESTROY(shmem_internal_mutex_ptl4_nb_fence);
return 0;
}
/*
** This is a function used to send the int value rpt to the other side
*/
void
SendRepeat(ArgStruct *p, int rpt)
{
int rc;
int index;
ptl_process_t dest;
ptl_ct_event_t cnt_value;
send_int= rpt;
index= PTL_SEND_INT_INDEX;
dest.rank= p->source_node;
rc= PtlPut(send_int_md_handle, 0, sizeof(int), PTL_NO_ACK_REQ, dest, index, 0, 0, NULL, 0);
LIBTEST_CHECK(rc, "PtlPut in SendRepeat()");
rc= PtlCTWait(send_int_ct_handle, total_int_sends, &cnt_value);
LIBTEST_CHECK(rc, "PtlCTWait in SendRepeat()");
if (cnt_value.failure != 0) {
fprintf(stderr, "SendRepeat() PtlPut failed %d (%d succeeded)\n",
(int)cnt_value.failure, (int)cnt_value.success);
}
total_int_sends++;
} /* end of SendRepeat() */
/*
** Send a buffer's worth of data
*/
void
SendData(ArgStruct *p)
{
int rc;
int index;
ptl_process_t dest;
ptl_ct_event_t cnt_value;
index= PTL_XMIT_INDEX;
dest.rank= p->source_node;
rc= PtlPut(md_handle, 0, p->bufflen, PTL_NO_ACK_REQ, dest, index, 0, 0, NULL, 0);
LIBTEST_CHECK(rc, "PtlPut in SendData()");
rc= PtlCTWait(send_ct_handle, total_sends, &cnt_value);
LIBTEST_CHECK(rc, "PtlCTWait in SendData()");
if (cnt_value.failure != 0) {
fprintf(stderr, "SendData() PtlPut failed %d (%d succeeded)\n",
(int)cnt_value.failure, (int)cnt_value.success);
}
total_sends++;
} /* end of SendData() */
/*
** Wait for a buffer's worth of data to arrive
*/
void
RecvData(ArgStruct *p)
{
int rc;
ptl_ct_event_t cnt_value;
rc= PtlCTWait(recv_ct_handle, total_recvs, &cnt_value);
LIBTEST_CHECK(rc, "PtlCTWait in RecvData");
if (cnt_value.failure != 0) {
fprintf(stderr, "RecvData() PtlPut failed %d (%d succeeded)\n",
(int)cnt_value.failure, (int)cnt_value.success);
}
total_recvs++;
} /* end of RecvData() */
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_logical;
ptl_pt_index_t logical_pt_index;
ptl_process_t myself;
struct timeval start, stop;
int potato = 0;
ENTRY_T potato_catcher;
HANDLE_T potato_catcher_handle;
ptl_md_t potato_launcher;
ptl_handle_md_t potato_launcher_handle;
int num_procs;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
num_procs = libtest_get_size();
if (NULL != getenv("MAKELEVEL") && num_procs > 2) {
return 77;
}
CHECK_RETURNVAL(PtlNIInit
(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL, PTL_PID_ANY,
NULL, NULL, &ni_logical));
CHECK_RETURNVAL(PtlSetMap(ni_logical, num_procs,
libtest_get_mapping(ni_logical)));
CHECK_RETURNVAL(PtlGetId(ni_logical, &myself));
CHECK_RETURNVAL(PtlPTAlloc
(ni_logical, 0, PTL_EQ_NONE, PTL_PT_ANY,
&logical_pt_index));
assert(logical_pt_index == 0);
/* Now do the initial setup on ni_logical */
potato_catcher.start = &potato;
potato_catcher.length = sizeof(potato);
potato_catcher.uid = PTL_UID_ANY;
potato_catcher.options = OPTIONS;
#if INTERFACE == 1
potato_catcher.match_id.rank = PTL_RANK_ANY;
potato_catcher.match_bits = 1;
potato_catcher.ignore_bits = ~potato_catcher.match_bits;
#endif
CHECK_RETURNVAL(PtlCTAlloc(ni_logical, &potato_catcher.ct_handle));
CHECK_RETURNVAL(APPEND
(ni_logical, logical_pt_index, &potato_catcher,
PTL_PRIORITY_LIST, NULL, &potato_catcher_handle));
/* Now do a barrier (on ni_physical) to make sure that everyone has their
* logical interface set up */
libtest_barrier();
/* now I can communicate between ranks with ni_logical */
/* set up the potato launcher */
potato_launcher.start = &potato;
potato_launcher.length = sizeof(potato);
potato_launcher.options =
PTL_MD_EVENT_CT_ACK | PTL_MD_EVENT_CT_SEND;
potato_launcher.eq_handle = PTL_EQ_NONE; // i.e. don't queue send events
CHECK_RETURNVAL(PtlCTAlloc(ni_logical, &potato_launcher.ct_handle));
CHECK_RETURNVAL(PtlMDBind
(ni_logical, &potato_launcher, &potato_launcher_handle));
/* rank 0 starts the potato going */
if (myself.rank == 0) {
ptl_process_t nextrank;
nextrank.rank = myself.rank + 1;
nextrank.rank *= (nextrank.rank <= num_procs - 1);
gettimeofday(&start, NULL);
CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0, potato_launcher.length,
(LOOPS == 1) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
nextrank, logical_pt_index, 1, 0,
NULL, 1));
}
{ /* the potato-passing loop */
size_t waitfor;
ptl_ct_event_t ctc;
ptl_process_t nextrank;
nextrank.rank = myself.rank + 1;
nextrank.rank *= (nextrank.rank <= num_procs - 1);
for (waitfor = 1; waitfor <= LOOPS; ++waitfor) {
CHECK_RETURNVAL(PtlCTWait(potato_catcher.ct_handle, waitfor, &ctc)); // wait for potato
assert(ctc.failure == 0);
assert(ctc.success == waitfor);
/* I have the potato! */
++potato;
if (potato < LOOPS * (num_procs)) { // otherwise, the recipient may have exited
/* Bomb's away! */
if (myself.rank == 0) {
CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0,
potato_launcher.length,
(waitfor == (LOOPS - 1)) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
nextrank, logical_pt_index, 3, 0, NULL, 2));
} else {
CHECK_RETURNVAL(PtlPut(potato_launcher_handle, 0,
potato_launcher.length,
(waitfor == LOOPS) ? PTL_OC_ACK_REQ : PTL_NO_ACK_REQ,
nextrank, logical_pt_index, 3, 0, NULL, 2));
}
}
}
// make sure that last send completed before exiting
CHECK_RETURNVAL(PtlCTWait(potato_launcher.ct_handle, LOOPS+1, &ctc));
assert(ctc.failure == 0);
}
if (myself.rank == 0) {
double accumulate = 0.0;
gettimeofday(&stop, NULL);
accumulate =
(stop.tv_sec + stop.tv_usec * 1e-6) - (start.tv_sec +
start.tv_usec * 1e-6);
/* calculate the average time waiting */
printf("Total time: %g secs\n", accumulate);
accumulate /= LOOPS;
printf("Average time around the loop: %g microseconds\n",
accumulate * 1e6);
accumulate /= num_procs;
printf("Average catch-to-toss latency: %g microseconds\n",
accumulate * 1e6);
}
/* cleanup */
CHECK_RETURNVAL(PtlMDRelease(potato_launcher_handle));
CHECK_RETURNVAL(PtlCTFree(potato_launcher.ct_handle));
CHECK_RETURNVAL(UNLINK(potato_catcher_handle));
CHECK_RETURNVAL(PtlCTFree(potato_catcher.ct_handle));
/* major cleanup */
CHECK_RETURNVAL(PtlPTFree(ni_logical, logical_pt_index));
CHECK_RETURNVAL(PtlNIFini(ni_logical));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_handle;
ptl_process_t *procs;
int rank;
ptl_pt_index_t pt_index, signal_pt_index;
HANDLE_T signal_e_handle;
HANDLE_T signal_e2_handle;
int num_procs;
ptl_handle_eq_t eq_handle;
ptl_handle_ct_t ct_handle;
ptl_handle_md_t md_handle;
ptl_ni_limits_t limits_reqd, limits_actual;
ENTRY_T value_e;
limits_reqd.max_entries = 1024;
limits_reqd.max_unexpected_headers = ITERS*2;
limits_reqd.max_mds = 1024;
limits_reqd.max_eqs = 1024;
limits_reqd.max_cts = 1024;
limits_reqd.max_pt_index = 64;
limits_reqd.max_iovecs = 1024;
limits_reqd.max_list_size = 1024;
limits_reqd.max_triggered_ops = 1024;
limits_reqd.max_msg_size = 1048576;
limits_reqd.max_atomic_size = 1048576;
limits_reqd.max_fetch_atomic_size = 1048576;
limits_reqd.max_waw_ordered_size = 1048576;
limits_reqd.max_war_ordered_size = 1048576;
limits_reqd.max_volatile_size = 1048576;
limits_reqd.features = 0;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
rank = libtest_get_rank();
num_procs = libtest_get_size();
if (num_procs < 2) {
fprintf(stderr, "test_flowctl_noeq requires at least two processes\n");
return 77;
}
int iters;
if (num_procs < ITERS)
iters = ITERS*2+1;
else
iters = ITERS;
CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
PTL_PID_ANY, &limits_reqd, &limits_actual, &ni_handle));
procs = libtest_get_mapping(ni_handle);
CHECK_RETURNVAL(PtlSetMap(ni_handle, num_procs, procs));
if (0 == rank) {
/* create data PT space */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, (num_procs - 1) * iters + 64, &eq_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, PTL_PT_FLOWCTRL, eq_handle, 5,
&pt_index));
/* create signal ME */
CHECK_RETURNVAL(PtlCTAlloc(ni_handle, &ct_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, 1, eq_handle, 6,
&signal_pt_index));
value_e.start = NULL;
value_e.length = 0;
value_e.ct_handle = ct_handle;
value_e.uid = PTL_UID_ANY;
value_e.options = OPTIONS | PTL_LE_EVENT_CT_COMM;
#if INTERFACE == 1
value_e.match_id.rank = PTL_RANK_ANY;
value_e.match_bits = 0;
value_e.ignore_bits = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_OVERFLOW_LIST, NULL, &signal_e_handle));
} else {
ptl_md_t md;
/* 16 extra just in case... */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, iters*2 + 16, &eq_handle));
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = eq_handle;
md.ct_handle = PTL_CT_NONE;
CHECK_RETURNVAL(PtlMDBind(ni_handle, &md, &md_handle));
}
fprintf(stderr,"at barrier \n");
libtest_barrier();
if (0 == rank) {
ptl_ct_event_t ct;
ptl_event_t ev;
int ret, count = 0, saw_flowctl = 0;
fprintf(stderr,"begin ctwait \n");
/* wait for signal counts */
CHECK_RETURNVAL(PtlCTWait(ct_handle, iters / 2 , &ct));
if (ct.success != iters / 2 || ct.failure != 0) {
return 1;
}
fprintf(stderr,"done CT wait \n");
/* wait for event entries */
while (1) {
ret = PtlEQGet(eq_handle, &ev);
if (PTL_OK == ret) {
count++;
fprintf(stderr, "found EQ value \n");
} else if (ret == PTL_EQ_EMPTY) {
continue;
} else {
fprintf(stderr, "0: Unexpected return code from EQGet: %d\n", ret);
return 1;
}
if (ev.type == PTL_EVENT_PT_DISABLED) {
saw_flowctl++;
break;
}
}
fprintf(stderr, "0: Saw %d flowctl\n", saw_flowctl);
if (saw_flowctl == 0) {
return 1;
}
/* Now clear out all of the unexpected messages so we can clean up everything */
CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_PRIORITY_LIST, NULL, &signal_e2_handle));
ret = PTL_OK;
while (ret != PTL_EQ_EMPTY)
ret = PtlEQGet(eq_handle, &ev);
} else {
ptl_process_t target;
ptl_event_t ev;
int ret, count = 0, fails = 0;
int i;
target.rank = 0;
printf("beginning puts \n");
for (i = 0 ; i < iters ; ++i) {
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_ACK_REQ,
target,
5,
0,
0,
NULL,
0));
usleep(100);
}
while (count < iters) {
ret = PtlEQGet(eq_handle, &ev);
if (PTL_EQ_EMPTY == ret) {
continue;
} else if (PTL_OK != ret) {
fprintf(stderr, "%d: PtlEQGet returned %d\n", rank, ret);
return 1;
}
if (ev.ni_fail_type == PTL_NI_OK) {
if (ev.type == PTL_EVENT_SEND) {
continue;
} else if (ev.type == PTL_EVENT_ACK) {
count++;
} else {
fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
}
} else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
count++;
fails++;
} else if (ev.ni_fail_type == PTL_EQ_EMPTY) {
continue;
} else if (ev.ni_fail_type == PTL_EQ_DROPPED) {
continue;
} else {
fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
return 1;
}
}
fprintf(stderr, "%d: Saw %d of %d ACKs as fails\n", rank, fails, count);
}
fprintf(stderr,"at final barrier \n");
libtest_barrier();
if (0 == rank) {
CHECK_RETURNVAL(UNLINK(signal_e_handle));
CHECK_RETURNVAL(UNLINK(signal_e2_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, signal_pt_index));
CHECK_RETURNVAL(PtlCTFree(ct_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, pt_index));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
} else {
CHECK_RETURNVAL(PtlMDRelease(md_handle));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
}
fprintf(stderr,"final cleanup \n");
CHECK_RETURNVAL(PtlNIFini(ni_handle));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
int main(int argc,
char *argv[])
{
ptl_handle_ni_t ni_handle;
ptl_process_t *procs;
int rank;
ptl_pt_index_t pt_index, signal_pt_index;
HANDLE_T value_e_handle, signal_e_handle;
int num_procs;
ptl_handle_eq_t eq_handle;
ptl_handle_ct_t ct_handle;
ptl_handle_md_t md_handle;
CHECK_RETURNVAL(PtlInit());
CHECK_RETURNVAL(libtest_init());
rank = libtest_get_rank();
num_procs = libtest_get_size();
if (num_procs < 2) {
fprintf(stderr, "test_flowctl_noeq requires at least two processes\n");
return 77;
}
CHECK_RETURNVAL(PtlNIInit(PTL_IFACE_DEFAULT, NI_TYPE | PTL_NI_LOGICAL,
PTL_PID_ANY, NULL, NULL, &ni_handle));
procs = libtest_get_mapping(ni_handle);
CHECK_RETURNVAL(PtlSetMap(ni_handle, num_procs, procs));
if (0 == rank) {
ENTRY_T value_e;
/* create data ME */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, (num_procs - 1) * ITERS / 2, &eq_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, PTL_PT_FLOWCTRL, eq_handle, 5,
&pt_index));
value_e.start = NULL;
value_e.length = 0;
value_e.ct_handle = PTL_CT_NONE;
value_e.uid = PTL_UID_ANY;
value_e.options = OPTIONS;
#if INTERFACE == 1
value_e.match_id.rank = PTL_RANK_ANY;
value_e.match_bits = 0;
value_e.ignore_bits = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_handle, 5, &value_e, PTL_PRIORITY_LIST, NULL, &value_e_handle));
/* create signal ME */
CHECK_RETURNVAL(PtlCTAlloc(ni_handle, &ct_handle));
CHECK_RETURNVAL(PtlPTAlloc(ni_handle, 0, PTL_EQ_NONE, 6,
&signal_pt_index));
value_e.start = NULL;
value_e.length = 0;
value_e.ct_handle = ct_handle;
value_e.uid = PTL_UID_ANY;
value_e.options = OPTIONS | PTL_LE_EVENT_SUCCESS_DISABLE | PTL_LE_EVENT_CT_COMM;
#if INTERFACE == 1
value_e.match_id.rank = PTL_RANK_ANY;
value_e.match_bits = 0;
value_e.ignore_bits = 0;
#endif
CHECK_RETURNVAL(APPEND(ni_handle, 6, &value_e, PTL_PRIORITY_LIST, NULL, &signal_e_handle));
} else {
ptl_md_t md;
/* 16 extra just in case... */
CHECK_RETURNVAL(PtlEQAlloc(ni_handle, ITERS * 2 + 16, &eq_handle));
md.start = NULL;
md.length = 0;
md.options = 0;
md.eq_handle = eq_handle;
md.ct_handle = PTL_CT_NONE;
CHECK_RETURNVAL(PtlMDBind(ni_handle, &md, &md_handle));
}
libtest_barrier();
if (0 == rank) {
ptl_ct_event_t ct;
ptl_event_t ev;
int ret, count = 0, saw_dropped = 0, saw_flowctl = 0;
/* wait for signal counts */
CHECK_RETURNVAL(PtlCTWait(ct_handle, num_procs - 1, &ct));
if (ct.success != num_procs - 1 || ct.failure != 0) {
return 1;
}
/* wait for event entries */
while (count < ITERS * (num_procs - 1)) {
ret = PtlEQWait(eq_handle, &ev);
if (PTL_OK == ret) {
;
} else if (PTL_EQ_DROPPED == ret) {
saw_dropped++;
if (ev.type == PTL_EVENT_PT_DISABLED){
saw_flowctl++;
CHECK_RETURNVAL(PtlPTEnable(ni_handle, pt_index));
}
break;
} else {
fprintf(stderr, "0: Unexpected return code from EQWait: %d\n", ret);
return 1;
}
if (ev.type == PTL_EVENT_PT_DISABLED) {
CHECK_RETURNVAL(PtlPTEnable(ni_handle, pt_index));
saw_flowctl++;
} else {
count++;
}
}
fprintf(stderr, "0: Saw %d dropped, %d flowctl\n", saw_dropped, saw_flowctl);
if (saw_flowctl == 0) {
return 1;
}
} else {
ptl_process_t target;
ptl_event_t ev;
int ret, count = 0, fails = 0;
int i;
int *fail_seen;
fail_seen = malloc(sizeof(int) * ITERS);
if (NULL == fail_seen) {
fprintf(stderr, "%d: malloc failed\n", rank);
return 1;
}
memset(fail_seen, 0, sizeof(int) * ITERS);
target.rank = 0;
for (i = 0 ; i < ITERS ; ++i) {
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_ACK_REQ,
target,
5,
0,
0,
(void*)(size_t)i,
0));
usleep(100);
}
while (count < ITERS) {
ret = PtlEQGet(eq_handle, &ev);
if (PTL_EQ_EMPTY == ret) {
continue;
} else if (PTL_OK != ret) {
fprintf(stderr, "%d: PtlEQGet returned %d\n", rank, ret);
return 1;
}
if (ev.ni_fail_type == PTL_NI_OK) {
if (ev.type == PTL_EVENT_SEND) {
continue;
} else if (ev.type == PTL_EVENT_ACK) {
count++;
} else {
fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
}
} else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
int iter = (size_t) ev.user_ptr;
if (fail_seen[iter]++ > 0) {
fprintf(stderr, "%d: Double report of PT_DISABLED for "
"iteration %d\n", rank, iter);
return 1;
}
count++;
fails++;
} else {
fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
return 1;
}
}
fprintf(stderr, "%d: Saw %d of %d events as fails\n", rank, fails, count);
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_NO_ACK_REQ,
target,
6,
0,
0,
NULL,
0));
/* wait for the send event on the last put */
CHECK_RETURNVAL(PtlEQWait(eq_handle, &ev));
while (fails > 0) {
CHECK_RETURNVAL(PtlPut(md_handle,
0,
0,
PTL_ACK_REQ,
target,
5,
0,
0,
NULL,
0));
while (1) {
ret = PtlEQWait(eq_handle, &ev);
if (PTL_OK != ret) {
fprintf(stderr, "%d: PtlEQWait returned %d\n", rank, ret);
return 1;
}
if (ev.ni_fail_type == PTL_NI_OK) {
if (ev.type == PTL_EVENT_SEND) {
continue;
} else if (ev.type == PTL_EVENT_ACK) {
fails--;
break;
} else {
fprintf(stderr, "%d: Unexpected event type %d\n", rank, ev.type);
}
} else if (ev.ni_fail_type == PTL_NI_PT_DISABLED) {
break;
} else {
fprintf(stderr, "%d: Unexpected fail type: %d\n", rank, ev.ni_fail_type);
return 1;
}
}
}
}
libtest_barrier();
if (0 == rank) {
CHECK_RETURNVAL(UNLINK(signal_e_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, signal_pt_index));
CHECK_RETURNVAL(PtlCTFree(ct_handle));
CHECK_RETURNVAL(UNLINK(value_e_handle));
CHECK_RETURNVAL(PtlPTFree(ni_handle, pt_index));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
} else {
CHECK_RETURNVAL(PtlMDRelease(md_handle));
CHECK_RETURNVAL(PtlEQFree(eq_handle));
}
CHECK_RETURNVAL(PtlNIFini(ni_handle));
CHECK_RETURNVAL(libtest_fini());
PtlFini();
return 0;
}
int
ompi_coll_portals4_barrier_intra(struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
mca_coll_portals4_module_t *portals4_module = (mca_coll_portals4_module_t*) module;
int ret, i, dim, hibit, mask, num_msgs;
int size = ompi_comm_size(comm);
int rank = ompi_comm_rank(comm);
ptl_ct_event_t ct;
ptl_handle_ct_t ct_h;
ptl_handle_me_t me_h;
ptl_me_t me;
size_t count;
ptl_match_bits_t match_bits;
ptl_handle_md_t md_h;
void *base;
ompi_coll_portals4_get_md(0, &md_h, &base);
count = opal_atomic_add_size_t(&portals4_module->barrier_count, 1);
ret = PtlCTAlloc(mca_coll_portals4_component.ni_h,
&ct_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTAlloc failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERR_TEMP_OUT_OF_RESOURCE;
}
COLL_PORTALS4_SET_BITS(match_bits, ompi_comm_get_cid(comm),
0, COLL_PORTALS4_BARRIER, count);
/* Build "tree" out of hypercube */
dim = comm->c_cube_dim;
hibit = opal_hibit(rank, dim);
--dim;
/* receive space */
me.start = NULL;
me.length = 0;
me.ct_handle = ct_h;
me.min_free = 0;
me.uid = mca_coll_portals4_component.uid;
me.options = PTL_ME_OP_PUT | PTL_ME_EVENT_SUCCESS_DISABLE |
PTL_ME_EVENT_LINK_DISABLE | PTL_ME_EVENT_UNLINK_DISABLE |
PTL_ME_EVENT_CT_COMM | PTL_ME_EVENT_CT_OVERFLOW;
me.match_id.phys.nid = PTL_NID_ANY;
me.match_id.phys.pid = PTL_PID_ANY;
me.match_bits = match_bits;
me.ignore_bits = 0;
ret = PtlMEAppend(mca_coll_portals4_component.ni_h,
mca_coll_portals4_component.pt_idx,
&me,
PTL_PRIORITY_LIST,
NULL,
&me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEAppend failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* calculate number of children to receive from */
num_msgs = ompi_coll_portals4_get_nchildren(dim + 1, hibit, rank, size);
/* send to parent when children have sent to us */
if (rank > 0) {
int parent = rank & ~(1 << hibit);
ret = PtlTriggeredPut(md_h,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, parent),
mca_coll_portals4_component.pt_idx,
match_bits,
0,
NULL,
0,
ct_h,
num_msgs);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlTriggeredPut failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* we'll need to wait for the parent response before the next set of comms */
num_msgs++;
}
/* send to children when parent (or all children if root) has sent to us */
for (i = hibit + 1, mask = 1 << i; i <= dim; ++i, mask <<= 1) {
int peer = rank | mask;
if (peer < size) {
ret = PtlTriggeredPut(md_h,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, peer),
mca_coll_portals4_component.pt_idx,
match_bits,
0,
NULL,
0,
ct_h,
num_msgs);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlTriggeredPut failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
}
}
/* Wait for all incoming messages */
ret = PtlCTWait(ct_h, num_msgs, &ct);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTWait failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
/* cleanup */
ret = PtlMEUnlink(me_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlMEUnlink failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
ret = PtlCTFree(ct_h);
if (PTL_OK != ret) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"%s:%d: PtlCTFree failed: %d\n",
__FILE__, __LINE__, ret);
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
static int
ompi_coll_portals4_scatter_intra_linear_top(const void *sbuf, int scount, struct ompi_datatype_t *sdtype,
void *rbuf, int rcount, struct ompi_datatype_t *rdtype,
int root,
struct ompi_communicator_t *comm,
ompi_coll_portals4_request_t *request,
mca_coll_base_module_t *module)
{
mca_coll_portals4_module_t *portals4_module = (mca_coll_portals4_module_t*) module;
int ret, line;
ptl_ct_event_t ct;
ptl_ct_event_t sync_incr_event;
int8_t i_am_root;
int32_t expected_rtrs = 0;
int32_t expected_puts = 0;
int32_t expected_acks = 0;
int32_t expected_ops = 0;
int32_t expected_chained_rtrs = 0;
int32_t expected_chained_acks = 0;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:portals4:scatter_intra_linear_top enter rank %d", request->u.scatter.my_rank));
request->type = OMPI_COLL_PORTALS4_TYPE_SCATTER;
request->u.scatter.scatter_buf = NULL;
request->u.scatter.scatter_mdh = PTL_INVALID_HANDLE;
request->u.scatter.scatter_cth = PTL_INVALID_HANDLE;
request->u.scatter.scatter_meh = PTL_INVALID_HANDLE;
request->u.scatter.sync_mdh = PTL_INVALID_HANDLE;
request->u.scatter.sync_cth = PTL_INVALID_HANDLE;
request->u.scatter.sync_meh = PTL_INVALID_HANDLE;
request->u.scatter.my_rank = ompi_comm_rank(comm);
request->u.scatter.size = ompi_comm_size(comm);
request->u.scatter.root_rank = root;
request->u.scatter.sbuf = sbuf;
request->u.scatter.rbuf = rbuf;
request->u.scatter.pack_src_buf = sbuf;
request->u.scatter.pack_src_count = scount;
request->u.scatter.pack_src_dtype = sdtype;
ompi_datatype_get_extent(request->u.scatter.pack_src_dtype,
&request->u.scatter.pack_src_lb,
&request->u.scatter.pack_src_extent);
ompi_datatype_get_true_extent(request->u.scatter.pack_src_dtype,
&request->u.scatter.pack_src_true_lb,
&request->u.scatter.pack_src_true_extent);
if ((root == request->u.scatter.my_rank) && (rbuf == MPI_IN_PLACE)) {
request->u.scatter.unpack_dst_buf = NULL;
request->u.scatter.unpack_dst_count = 0;
request->u.scatter.unpack_dst_dtype = MPI_DATATYPE_NULL;
} else {
request->u.scatter.unpack_dst_buf = rbuf;
request->u.scatter.unpack_dst_count = rcount;
request->u.scatter.unpack_dst_dtype = rdtype;
request->u.scatter.unpack_dst_offset = 0;
ompi_datatype_get_extent(request->u.scatter.unpack_dst_dtype,
&request->u.scatter.unpack_dst_lb,
&request->u.scatter.unpack_dst_extent);
ompi_datatype_get_true_extent(request->u.scatter.unpack_dst_dtype,
&request->u.scatter.unpack_dst_true_lb,
&request->u.scatter.unpack_dst_true_extent);
}
opal_output_verbose(30, ompi_coll_base_framework.framework_output,
"%s:%d:rank(%d): request->u.scatter.unpack_dst_offset(%lu)",
__FILE__, __LINE__, request->u.scatter.my_rank,
request->u.scatter.unpack_dst_offset);
/**********************************/
/* Setup Common Parameters */
/**********************************/
i_am_root = (request->u.scatter.my_rank == request->u.scatter.root_rank);
request->u.scatter.coll_count = opal_atomic_add_size_t(&portals4_module->coll_count, 1);
ret = setup_scatter_buffers_linear(comm, request, portals4_module);
if (MPI_SUCCESS != ret) {
line = __LINE__;
goto err_hdlr;
}
ret = setup_scatter_handles(comm, request, portals4_module);
if (MPI_SUCCESS != ret) {
line = __LINE__;
goto err_hdlr;
}
ret = setup_sync_handles(comm, request, portals4_module);
if (MPI_SUCCESS != ret) {
line = __LINE__;
goto err_hdlr;
}
/**********************************/
/* do the scatter */
/**********************************/
if (i_am_root) {
/* operations on the sync counter */
expected_rtrs = request->u.scatter.size - 1; /* expect RTRs from non-root ranks */
expected_acks = request->u.scatter.size - 1; /* expect Recv-ACKs from non-root ranks */
/* operations on the scatter counter */
expected_puts = 0;
expected_chained_rtrs = 1;
expected_chained_acks = 1;
/* Chain the RTR and Recv-ACK to the Scatter CT */
sync_incr_event.success=1;
sync_incr_event.failure=0;
ret = PtlTriggeredCTInc(request->u.scatter.scatter_cth,
sync_incr_event,
request->u.scatter.sync_cth,
expected_rtrs);
if (PTL_OK != ret) {
ret = OMPI_ERROR;
line = __LINE__;
goto err_hdlr;
}
ret = PtlTriggeredCTInc(request->u.scatter.scatter_cth,
sync_incr_event,
request->u.scatter.sync_cth,
expected_rtrs + expected_acks);
if (PTL_OK != ret) {
ret = OMPI_ERROR;
line = __LINE__;
goto err_hdlr;
}
/* root, so put packed bytes to other ranks */
for (int32_t i=0; i<request->u.scatter.size; i++) {
/* do not put to my scatter_buf. my data gets unpacked into my out buffer in linear_bottom(). */
if (i == request->u.scatter.my_rank) {
continue;
}
ptl_size_t offset = request->u.scatter.packed_size * i;
opal_output_verbose(30, ompi_coll_base_framework.framework_output,
"%s:%d:rank(%d): offset(%lu)=rank(%d) * packed_size(%ld)",
__FILE__, __LINE__, request->u.scatter.my_rank,
offset, i, request->u.scatter.packed_size);
ret = PtlTriggeredPut(request->u.scatter.scatter_mdh,
(ptl_size_t)request->u.scatter.scatter_buf + offset,
request->u.scatter.packed_size,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, i),
mca_coll_portals4_component.pt_idx,
request->u.scatter.scatter_match_bits,
0,
NULL,
0,
request->u.scatter.scatter_cth,
expected_chained_rtrs);
if (PTL_OK != ret) {
ret = OMPI_ERROR;
line = __LINE__;
goto err_hdlr;
}
}
} else {
/* non-root, so do nothing */
/* operations on the sync counter */
expected_rtrs = 0;
expected_acks = 0;
/* operations on the scatter counter */
expected_puts = 1; /* scatter put from root */
expected_chained_rtrs = 0;
expected_chained_acks = 0;
}
expected_ops = expected_chained_rtrs + expected_puts;
/**********************************************/
/* only non-root ranks are PUT to, so only */
/* non-root ranks must PUT a Recv-ACK to root */
/**********************************************/
if (!i_am_root) {
ret = PtlTriggeredPut(request->u.scatter.sync_mdh,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, request->u.scatter.root_rank),
mca_coll_portals4_component.pt_idx,
request->u.scatter.sync_match_bits,
0,
NULL,
0,
request->u.scatter.scatter_cth,
expected_ops);
if (PTL_OK != ret) {
ret = OMPI_ERROR;
line = __LINE__;
goto err_hdlr;
}
}
expected_ops += expected_chained_acks;
if (!request->u.scatter.is_sync) {
/******************************************/
/* put to finish pt when all ops complete */
/******************************************/
ret = PtlTriggeredPut(mca_coll_portals4_component.zero_md_h,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, request->u.scatter.my_rank),
mca_coll_portals4_component.finish_pt_idx,
0,
0,
NULL,
(uintptr_t) request,
request->u.scatter.scatter_cth,
expected_ops);
if (PTL_OK != ret) {
ret = OMPI_ERROR;
line = __LINE__;
goto err_hdlr;
}
}
/**************************************/
/* all non-root ranks put RTR to root */
/**************************************/
if (!i_am_root) {
ret = PtlPut(request->u.scatter.sync_mdh,
0,
0,
PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, request->u.scatter.root_rank),
mca_coll_portals4_component.pt_idx,
request->u.scatter.sync_match_bits,
0,
NULL,
0);
if (PTL_OK != ret) {
ret = OMPI_ERROR;
line = __LINE__;
goto err_hdlr;
}
}
if (request->u.scatter.is_sync) {
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"calling CTWait(expected_ops=%d)\n", expected_ops);
/********************************/
/* Wait for all ops to complete */
/********************************/
ret = PtlCTWait(request->u.scatter.scatter_cth, expected_ops, &ct);
if (PTL_OK != ret) {
ret = OMPI_ERROR;
line = __LINE__;
goto err_hdlr;
}
opal_output_verbose(1, ompi_coll_base_framework.framework_output,
"completed CTWait(expected_ops=%d)\n", expected_ops);
}
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:portals4:scatter_intra_linear_top exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS;
err_hdlr:
if (NULL != request->u.scatter.scatter_buf)
free(request->u.scatter.scatter_buf);
opal_output(ompi_coll_base_framework.framework_output,
"%s:%4d:%4d\tError occurred ret=%d, rank %2d",
__FILE__, __LINE__, line, ret, request->u.scatter.my_rank);
return ret;
}
