ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность

Merge pull request #1896 from PDeveze/Patchs-on-coll-portals4

Просмотр исходного кода 
Patchs on coll portals4
Этот коммит содержится в:
Todd Kordenbrock 2016-08-08 14:57:02 -05:00 коммит произвёл GitHub
родитель 5ced037488 a7e3de6c4f
Коммит 3be6052523
8 изменённых файлов: 324 добавлений и 103 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

3
ompi/mca/coll/portals4/coll_portals4.h
Просмотреть файл

@ -65,6 +65,7 @@ struct mca_coll_portals4_component_t {
opal_free_list_t requests; /* request free list for the i collectives */ opal_free_list_t requests; /* request free list for the i collectives */
ptl_ni_limits_t ni_limits; ptl_ni_limits_t ni_limits;
ptl_size_t portals_max_msg_size;
int use_binomial_gather_algorithm; int use_binomial_gather_algorithm;
@ -314,7 +315,7 @@ is_reduce_optimizable(struct ompi_datatype_t *dtype, size_t length, struct ompi_
} }
*ptl_dtype = ompi_coll_portals4_atomic_datatype[dtype->id]; *ptl_dtype = ompi_coll_portals4_atomic_datatype[dtype->id];
if (*ptl_dtype == COLL_PORTALS4_NO_DTYPE){ if (*ptl_dtype == COLL_PORTALS4_NO_DTYPE) {
opal_output_verbose(50, ompi_coll_base_framework.framework_output, opal_output_verbose(50, ompi_coll_base_framework.framework_output,
"datatype %d not supported\n", "datatype %d not supported\n",
dtype->id); dtype->id);

4
ompi/mca/coll/portals4/coll_portals4_allreduce.c
Просмотреть файл

@ -265,7 +265,7 @@ allreduce_kary_tree_top(const void *sendbuf, void *recvbuf, int count,
ompi_coll_portals4_get_peer(comm, child[i]), ompi_coll_portals4_get_peer(comm, child[i]),
mca_coll_portals4_component.pt_idx, mca_coll_portals4_component.pt_idx,
match_bits_rtr, 0, NULL, 0)) != PTL_OK) match_bits_rtr, 0, NULL, 0)) != PTL_OK)
return opal_stderr("Put RTR failed", __FILE__, __LINE__, ret); return opal_stderr("Put RTR failed %d", __FILE__, __LINE__, ret);
} }
} }
} }
@ -408,7 +408,7 @@ int ompi_coll_portals4_iallreduce_intra(const void* sendbuf, void* recvbuf, int
allreduce_kary_tree_top(sendbuf, recvbuf, count, allreduce_kary_tree_top(sendbuf, recvbuf, count,
dtype, op, comm, request, portals4_module); dtype, op, comm, request, portals4_module);
puts("iallreduce"); opal_output_verbose(10, ompi_coll_base_framework.framework_output, "iallreduce");
return (OMPI_SUCCESS); return (OMPI_SUCCESS);
} }

26
ompi/mca/coll/portals4/coll_portals4_barrier.c
Просмотреть файл

@ -147,9 +147,31 @@ barrier_hypercube_top(struct ompi_communicator_t *comm,
} }
if (is_sync) { if (is_sync) {
/* Send a put to self when we've received all our messages... */ /* Each process has a pending PtlTriggeredPut. To be sure this request will be triggered, we must
ret = PtlCTWait(request->u.barrier.rtr_ct_h, num_msgs, &event); call PtlTriggeredCTInc twice. Otherwise, we could free the CT too early and the Put wouldn't be triggered */
ptl_ct_event_t ct_inc;
ct_inc.success = 1;
ct_inc.failure = 0;
if ((ret = PtlTriggeredCTInc(request->u.barrier.rtr_ct_h, ct_inc,
request->u.barrier.rtr_ct_h, num_msgs)) != 0) {
return opal_stderr("PtlTriggeredCTInc failed", __FILE__, __LINE__, ret);
}
if ((ret = PtlTriggeredCTInc(request->u.barrier.rtr_ct_h, ct_inc,
request->u.barrier.rtr_ct_h, num_msgs + 1)) != 0) {
return opal_stderr("PtlTriggeredCTInc failed", __FILE__, __LINE__, ret);
}
ret = PtlCTWait(request->u.barrier.rtr_ct_h, num_msgs + 2, &event);
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;
}
} }
else { else {
/* Send a put to self when we've received all our messages... */ /* Send a put to self when we've received all our messages... */

148
ompi/mca/coll/portals4/coll_portals4_bcast.c
Просмотреть файл

@ -89,12 +89,20 @@ static int prepare_bcast_data (struct ompi_communicator_t *comm,
} }
/* Number of segments */ /* Number of segments */
request->u.bcast.segment_nb = (request->u.bcast.tmpsize > COLL_PORTALS4_MAX_BW) ? {
(((request->u.bcast.tmpsize + COLL_PORTALS4_MAX_BW -1) / COLL_PORTALS4_MAX_BW) < COLL_PORTALS4_MAX_SEGMENT ? size_t max_msg_size = (COLL_PORTALS4_MAX_BW > mca_coll_portals4_component.ni_limits.max_msg_size) ?
((request->u.bcast.tmpsize + COLL_PORTALS4_MAX_BW -1) / COLL_PORTALS4_MAX_BW) : mca_coll_portals4_component.ni_limits.max_msg_size :
COLL_PORTALS4_MAX_SEGMENT) : COLL_PORTALS4_MAX_BW;
//TODO : Either make compatible Portals size limits and COLL_PORTALS4_MAX_SEGMENT or remove COLL_PORTALS4_MAX_SEGMENT
request->u.bcast.segment_nb = (request->u.bcast.tmpsize > max_msg_size) ?
(((request->u.bcast.tmpsize + max_msg_size -1) / max_msg_size) < COLL_PORTALS4_MAX_SEGMENT ?
((request->u.bcast.tmpsize + max_msg_size -1) / max_msg_size) : COLL_PORTALS4_MAX_SEGMENT) :
1; 1;
OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"seg_number=%d , seg_size_max=%lu", request->u.bcast.segment_nb, max_msg_size));
}
if (request->u.bcast.segment_nb > COLL_PORTALS4_BCAST_ALGO_THRESHOLD) { if (request->u.bcast.segment_nb > COLL_PORTALS4_BCAST_ALGO_THRESHOLD) {
request->u.bcast.algo = OMPI_COLL_PORTALS4_BCAST_PIPELINE_ALGO; request->u.bcast.algo = OMPI_COLL_PORTALS4_BCAST_PIPELINE_ALGO;
} }
@ -137,9 +145,9 @@ bcast_kary_tree_top(void *buff, int count,
mca_coll_portals4_module_t *portals4_module) mca_coll_portals4_module_t *portals4_module)
{ {
bool is_sync = request->is_sync; bool is_sync = request->is_sync;
int ret, seg; int ret;
unsigned int i; unsigned int i, seg, seg_size, nb_long;
int segment_nb = request->u.bcast.segment_nb; unsigned int segment_nb = request->u.bcast.segment_nb;
unsigned int child_nb; unsigned int child_nb;
int size = ompi_comm_size(comm); int size = ompi_comm_size(comm);
int rank = ompi_comm_rank(comm); int rank = ompi_comm_rank(comm);
@ -201,15 +209,22 @@ bcast_kary_tree_top(void *buff, int count,
COLL_PORTALS4_SET_BITS(match_bits, ompi_comm_get_cid(comm), 0, 0, COLL_PORTALS4_SET_BITS(match_bits, ompi_comm_get_cid(comm), 0, 0,
COLL_PORTALS4_BCAST, 0, internal_count); COLL_PORTALS4_BCAST, 0, internal_count);
/* The data will be cut in segment_nb segments.
* nb_long segments will have a size of (seg_size + 1)
* and (segment_nb - nb_long) segments will have a size of seg_size
*/
seg_size = request->u.bcast.tmpsize / segment_nb;
nb_long = request->u.bcast.tmpsize % segment_nb;
opal_output_verbose(10, ompi_coll_base_framework.framework_output, "seg_size=%d nb_long=%d segment_nb=%d", seg_size, nb_long, segment_nb);
if (rank != root) { if (rank != root) {
for (seg = 1, offset = 0, length = 0 ; for (seg = 1, offset = 0, length = 0 ;
seg <= segment_nb ; seg <= segment_nb ;
seg++, offset += length) { seg++, offset += length) {
/* Divide buffer into segments */ /* Divide buffer into segments */
length = (seg < segment_nb) ? if (seg <= nb_long) length = seg_size + 1;
(request->u.bcast.tmpsize + segment_nb - 1) / segment_nb : else length = seg_size;
request->u.bcast.tmpsize - ((request->u.bcast.tmpsize + segment_nb - 1) / segment_nb) * (segment_nb - 1);
/* /*
** Prepare Data ME ** Prepare Data ME
@ -352,13 +367,14 @@ bcast_kary_tree_top(void *buff, int count,
seg++, offset += length) { seg++, offset += length) {
/* Divide buffer into segments */ /* Divide buffer into segments */
length = (seg < segment_nb) ? if (seg <= nb_long) length = seg_size + 1;
(request->u.bcast.tmpsize + segment_nb - 1) / segment_nb : else length = seg_size;
request->u.bcast.tmpsize - ((request->u.bcast.tmpsize + segment_nb - 1) / segment_nb) * (segment_nb - 1); opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"bcast with k-ary tree : segment of size %ld", length);
/* compute the triggering threshold to send data to the children */ /* compute the triggering threshold to send data to the children */
trig_thr = (rank == root) ? (segment_nb) : trig_thr = segment_nb + seg - 1; /* To be sure the set of PtlTriggeredPut of DATA will be executed in order */
(segment_nb + seg); if (rank != root) trig_thr ++;
/* /*
** Send Data to children ** Send Data to children
@ -381,6 +397,17 @@ bcast_kary_tree_top(void *buff, int count,
} }
} }
if (rank == root) {
trig_thr = segment_nb;
ct_inc.success = segment_nb;
ct_inc.failure = 0;
if ((ret = PtlTriggeredCTInc(request->u.bcast.trig_ct_h, ct_inc,
request->u.bcast.trig_ct_h, trig_thr)) != 0) {
return opal_stderr("PtlTriggeredCTInc failed", __FILE__, __LINE__, ret);
}
}
ack_thr = child_nb; ack_thr = child_nb;
if (is_sync) { if (is_sync) {
@ -409,9 +436,28 @@ bcast_kary_tree_top(void *buff, int count,
*/ */
if (rank != root) { if (rank != root) {
ack_thr = segment_nb; trig_thr = segment_nb;
if (is_sync) { if (is_sync) {
if ((ret = PtlCTWait(request->u.bcast.trig_ct_h, ack_thr, &ct)) != 0) { /* Each leaf has a pending PtlTriggeredPut (to send the final ACK). We must call PtlTriggeredCTInc twice.
Otherwise, we could pass the PtlCTWait and then free the CT too early and the Put wouldn't be triggered.
This is necessary because portals4 does not insure the order in the triggered operations associated
with the same threshold. In the case where PtlCTWait is not called (else case), this is not necessary. */
ct_inc.success = 1;
ct_inc.failure = 0;
if ((ret = PtlTriggeredCTInc(request->u.bcast.trig_ct_h, ct_inc,
request->u.bcast.trig_ct_h, trig_thr)) != 0) {
return opal_stderr("PtlTriggeredCTInc failed", __FILE__, __LINE__, ret);
}
if ((ret = PtlTriggeredCTInc(request->u.bcast.trig_ct_h, ct_inc,
request->u.bcast.trig_ct_h, trig_thr + 1)) != 0) {
return opal_stderr("PtlTriggeredCTInc failed", __FILE__, __LINE__, ret);
}
if ((ret = PtlCTWait(request->u.bcast.trig_ct_h, trig_thr + 2, &ct)) != 0) {
opal_stderr("PtlCTWait failed", __FILE__, __LINE__, ret); opal_stderr("PtlCTWait failed", __FILE__, __LINE__, ret);
} }
} }
@ -421,7 +467,7 @@ bcast_kary_tree_top(void *buff, int count,
mca_coll_portals4_component.finish_pt_idx, mca_coll_portals4_component.finish_pt_idx,
0, 0, NULL, (uintptr_t) request, 0, 0, NULL, (uintptr_t) request,
request->u.bcast.trig_ct_h, request->u.bcast.trig_ct_h,
ack_thr)) != 0) { trig_thr)) != 0) {
return opal_stderr("PtlTriggeredPut failed", __FILE__, __LINE__, ret); return opal_stderr("PtlTriggeredPut failed", __FILE__, __LINE__, ret);
} }
@ -440,8 +486,9 @@ bcast_pipeline_top(void *buff, int count,
mca_coll_portals4_module_t *portals4_module) mca_coll_portals4_module_t *portals4_module)
{ {
bool is_sync = request->is_sync; bool is_sync = request->is_sync;
int ret, seg; int ret;
int segment_nb = request->u.bcast.segment_nb; unsigned int seg, seg_size, nb_long;
unsigned int segment_nb = request->u.bcast.segment_nb;
int size = ompi_comm_size(comm); int size = ompi_comm_size(comm);
int rank = ompi_comm_rank(comm); int rank = ompi_comm_rank(comm);
ptl_rank_t parent, child; ptl_rank_t parent, child;
@ -492,6 +539,13 @@ bcast_pipeline_top(void *buff, int count,
COLL_PORTALS4_SET_BITS(match_bits, ompi_comm_get_cid(comm), 0, 0, COLL_PORTALS4_SET_BITS(match_bits, ompi_comm_get_cid(comm), 0, 0,
COLL_PORTALS4_BCAST, 0, internal_count); COLL_PORTALS4_BCAST, 0, internal_count);
/* The data will be cut in segment_nb segments.
* nb_long segments will have a size of (seg_size + 1)
* and (segment_nb - nb_long) segments will have a size of seg_size
*/
seg_size = request->u.bcast.tmpsize / segment_nb;
nb_long = request->u.bcast.tmpsize % segment_nb;
opal_output_verbose(10, ompi_coll_base_framework.framework_output, "seg_size=%d nb_long=%d", seg_size, nb_long);
if (rank != root) { if (rank != root) {
for (seg = 1, offset = 0, length = 0 ; for (seg = 1, offset = 0, length = 0 ;
@ -499,9 +553,8 @@ bcast_pipeline_top(void *buff, int count,
seg++, offset += length) { seg++, offset += length) {
/* Divide buffer into segments */ /* Divide buffer into segments */
length = (seg < segment_nb) ? if (seg <= nb_long) length = seg_size + 1;
(request->u.bcast.tmpsize + segment_nb - 1) / segment_nb : else length = seg_size;
request->u.bcast.tmpsize - ((request->u.bcast.tmpsize + segment_nb - 1) / segment_nb) * (segment_nb - 1);
/* /*
** Prepare Data ME ** Prepare Data ME
@ -642,13 +695,14 @@ bcast_pipeline_top(void *buff, int count,
seg++, offset += length) { seg++, offset += length) {
/* Divide buffer into segments */ /* Divide buffer into segments */
length = (seg < segment_nb) ? if (seg <= nb_long) length = seg_size + 1;
(request->u.bcast.tmpsize + segment_nb - 1) / segment_nb : else length = seg_size;
request->u.bcast.tmpsize - ((request->u.bcast.tmpsize + segment_nb - 1) / segment_nb) * (segment_nb - 1); opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"bcast with pipeline : segment of size %ld \n", length);
/* compute the triggering threshold to send data to the children */ /* compute the triggering threshold to send data to the children */
trig_thr = (rank == root) ? (segment_nb) : trig_thr = segment_nb + seg - 1; /* To be sure the PtlTriggeredPut will be executed in order */
(segment_nb + seg); if (rank != root) trig_thr ++;
/* /*
** Send Data to children ** Send Data to children
@ -668,6 +722,16 @@ bcast_pipeline_top(void *buff, int count,
} }
} }
} }
if (rank == root) {
trig_thr = segment_nb;
ct_inc.success = segment_nb;
ct_inc.failure = 0;
if ((ret = PtlTriggeredCTInc(request->u.bcast.trig_ct_h, ct_inc,
request->u.bcast.trig_ct_h, trig_thr)) != 0) {
return opal_stderr("PtlTriggeredCTInc failed", __FILE__, __LINE__, ret);
}
}
if (is_sync) { if (is_sync) {
if ((ret = PtlCTWait(request->u.bcast.ack_ct_h, 1, &ct)) != 0) { if ((ret = PtlCTWait(request->u.bcast.ack_ct_h, 1, &ct)) != 0) {
@ -696,8 +760,29 @@ bcast_pipeline_top(void *buff, int count,
*/ */
if (rank != root) { if (rank != root) {
trig_thr = segment_nb;
if (is_sync) { if (is_sync) {
if ((ret = PtlCTWait(request->u.bcast.trig_ct_h, segment_nb, &ct)) != 0) { /* Each leaf has a pending PtlTriggeredPut (to send the final ACK). We must call PtlTriggeredCTInc twice.
Otherwise, we could pass the PtlCTWait and then free the CT too early and the Put wouldn't be triggered.
This is necessary because portals4 does not insure the order in the triggered operations associated
with the same threshold. In the case where PtlCTWait is not called (else case), this is not necessary. */
ct_inc.success = 1;
ct_inc.failure = 0;
if ((ret = PtlTriggeredCTInc(request->u.bcast.trig_ct_h, ct_inc,
request->u.bcast.trig_ct_h, trig_thr)) != 0) {
return opal_stderr("PtlTriggeredCTInc failed", __FILE__, __LINE__, ret);
}
if ((ret = PtlTriggeredCTInc(request->u.bcast.trig_ct_h, ct_inc,
request->u.bcast.trig_ct_h, trig_thr + 1)) != 0) {
return opal_stderr("PtlTriggeredCTInc failed", __FILE__, __LINE__, ret);
}
if ((ret = PtlCTWait(request->u.bcast.trig_ct_h, trig_thr + 2, &ct)) != 0) {
opal_stderr("PtlCTWait failed", __FILE__, __LINE__, ret); opal_stderr("PtlCTWait failed", __FILE__, __LINE__, ret);
} }
} }
@ -707,7 +792,7 @@ bcast_pipeline_top(void *buff, int count,
mca_coll_portals4_component.finish_pt_idx, mca_coll_portals4_component.finish_pt_idx,
0, 0, NULL, (uintptr_t) request, 0, 0, NULL, (uintptr_t) request,
request->u.bcast.trig_ct_h, request->u.bcast.trig_ct_h,
segment_nb)) != 0) { trig_thr)) != 0) {
return opal_stderr("PtlTriggeredPut failed", __FILE__, __LINE__, ret); return opal_stderr("PtlTriggeredPut failed", __FILE__, __LINE__, ret);
} }
} }
@ -831,7 +916,7 @@ ompi_coll_portals4_ibcast_intra(void *buff, int count,
return OMPI_ERROR; return OMPI_ERROR;
} }
puts("ibcast"); opal_output_verbose(10, ompi_coll_base_framework.framework_output, "ibcast_intra");
return (OMPI_SUCCESS); return (OMPI_SUCCESS);
} }
@ -860,5 +945,6 @@ ompi_coll_portals4_ibcast_intra_fini(ompi_coll_portals4_request_t *request)
ompi_request_complete(&request->super, true); ompi_request_complete(&request->super, true);
OPAL_THREAD_UNLOCK(&ompi_request_lock); OPAL_THREAD_UNLOCK(&ompi_request_lock);
opal_output_verbose(10, ompi_coll_base_framework.framework_output, "ibcast_intra_fini");
return (OMPI_SUCCESS); return (OMPI_SUCCESS);
} }

16
ompi/mca/coll/portals4/coll_portals4_component.c
Просмотреть файл

@ -211,6 +211,16 @@ portals4_register(void)
MCA_BASE_VAR_SCOPE_READONLY, MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_portals4_component.use_binomial_gather_algorithm); &mca_coll_portals4_component.use_binomial_gather_algorithm);
mca_coll_portals4_component.portals_max_msg_size = PTL_SIZE_MAX;
(void) mca_base_component_var_register(&mca_coll_portals4_component.super.collm_version,
"max_msg_size",
"Max size supported by portals4 (above that, a message is cut into messages less than that size)",
MCA_BASE_VAR_TYPE_UNSIGNED_LONG,
NULL, 0, 0,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_coll_portals4_component.portals_max_msg_size);
return OMPI_SUCCESS; return OMPI_SUCCESS;
} }
@ -369,7 +379,13 @@ portals4_init_query(bool enable_progress_threads,
__FILE__, __LINE__, ret); __FILE__, __LINE__, ret);
return OMPI_ERROR; return OMPI_ERROR;
} }
opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"ni_limits.max_atomic_size=%ld", mca_coll_portals4_component.ni_limits.max_atomic_size);
if (mca_coll_portals4_component.portals_max_msg_size < mca_coll_portals4_component.ni_limits.max_msg_size)
mca_coll_portals4_component.ni_limits.max_msg_size = mca_coll_portals4_component.portals_max_msg_size;
opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"ni_limits.max_msg_size=%lu", mca_coll_portals4_component.ni_limits.max_msg_size);
ret = PtlGetId(mca_coll_portals4_component.ni_h, &mca_coll_portals4_component.id); ret = PtlGetId(mca_coll_portals4_component.ni_h, &mca_coll_portals4_component.id);
if (PTL_OK != ret) { if (PTL_OK != ret) {

117
ompi/mca/coll/portals4/coll_portals4_gather.c
Просмотреть файл

@ -21,6 +21,7 @@
#include "coll_portals4.h" #include "coll_portals4.h"
#include "coll_portals4_request.h" #include "coll_portals4_request.h"
#include <string.h> // included for ffs in get_tree_numdescendants_of
#undef RTR_USES_TRIGGERED_PUT #undef RTR_USES_TRIGGERED_PUT
@ -55,6 +56,22 @@
* | * |
* 15 * 15
*/ */
static int32_t get_tree_numdescendants_of(struct ompi_communicator_t* comm,
int vrank)
{
int max;
int size = ompi_comm_size(comm);
if (0 == vrank) {
return size - 1;
} else {
max = 1 << ffs(vrank - 1);
return ((vrank + max <= size ) ? max : size - vrank) -1;
}
}
static ompi_coll_portals4_tree_t* static ompi_coll_portals4_tree_t*
ompi_coll_portals4_build_in_order_bmtree( struct ompi_communicator_t* comm, ompi_coll_portals4_build_in_order_bmtree( struct ompi_communicator_t* comm,
int root ) int root )
@ -506,8 +523,10 @@ ompi_coll_portals4_gather_intra_binomial_top(const void *sbuf, int scount, struc
int32_t expected_ops =0; int32_t expected_ops =0;
int32_t expected_acks=0; int32_t expected_acks=0;
ptl_size_t number_of_fragment_gathered = 0;
ptl_size_t number_of_fragment_send = 1;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra_binomial_top enter rank %d", request->u.gather.my_rank)); "coll:portals4:gather_intra_binomial_top enter rank %d", request->u.gather.my_rank));
request->type = OMPI_COLL_PORTALS4_TYPE_GATHER; request->type = OMPI_COLL_PORTALS4_TYPE_GATHER;
@ -579,6 +598,23 @@ ompi_coll_portals4_gather_intra_binomial_top(const void *sbuf, int scount, struc
ret = setup_sync_handles(comm, request, portals4_module); ret = setup_sync_handles(comm, request, portals4_module);
if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; } if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; }
OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"%s:%d: packed_size=%lu, fragment_size=%lu",
__FILE__, __LINE__, request->u.gather.packed_size, mca_coll_portals4_component.ni_limits.max_msg_size));
for (int i =0; i < bmtree->tree_nextsize; i++) {
int child_vrank = VRANK(bmtree->tree_next[i], request->u.gather.root_rank, request->u.gather.size);
int sub_tree_size = get_tree_numdescendants_of(comm, child_vrank) + 1;
ptl_size_t local_number_of_fragment = ((sub_tree_size * request->u.gather.packed_size) + mca_coll_portals4_component.ni_limits.max_msg_size -1) / mca_coll_portals4_component.ni_limits.max_msg_size;
OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"%s:%d: %d is child of %d(%d) with %d descendants (nb_frag += %lu)",
__FILE__, __LINE__, bmtree->tree_next[i], vrank, request->u.gather.root_rank , sub_tree_size, local_number_of_fragment));
number_of_fragment_gathered += local_number_of_fragment;
}
number_of_fragment_send = (request->u.gather.gather_bytes + mca_coll_portals4_component.ni_limits.max_msg_size -1) / mca_coll_portals4_component.ni_limits.max_msg_size;
/***********************************************/ /***********************************************/
/* Chain the RTR and Recv-ACK to the Gather CT */ /* Chain the RTR and Recv-ACK to the Gather CT */
/***********************************************/ /***********************************************/
@ -603,7 +639,7 @@ ompi_coll_portals4_gather_intra_binomial_top(const void *sbuf, int scount, struc
if (vrank == 0) { if (vrank == 0) {
/* root, so do nothing */ /* root, so do nothing */
expected_ops=bmtree->tree_nextsize; /* gather put from each child */ expected_ops=number_of_fragment_gathered ; /* gather put from each child */
expected_acks=0; expected_acks=0;
} else { } else {
@ -617,22 +653,32 @@ ompi_coll_portals4_gather_intra_binomial_top(const void *sbuf, int scount, struc
__FILE__, __LINE__, vrank, __FILE__, __LINE__, vrank,
remote_offset, vrank, vparent, request->u.gather.packed_size); remote_offset, vrank, vparent, request->u.gather.packed_size);
expected_ops=bmtree->tree_nextsize + 1; /* gather put from each child + a chained RTR */ expected_ops=number_of_fragment_gathered + 1; /* gather puts from each child + a chained RTR */
expected_acks=1; /* Recv-ACK from parent */ expected_acks=1; /* Recv-ACK from parent */
ptl_size_t size_sent = 0;
ptl_size_t size_left = request->u.gather.gather_bytes;
for (ptl_size_t i = 0 ; i < number_of_fragment_send; i++) {
ptl_size_t frag_size = (size_left > mca_coll_portals4_component.ni_limits.max_msg_size) ?
mca_coll_portals4_component.ni_limits.max_msg_size:
size_left;
ret = PtlTriggeredPut(request->u.gather.gather_mdh, ret = PtlTriggeredPut(request->u.gather.gather_mdh,
request->u.gather.gather_offset, request->u.gather.gather_offset + size_sent,
request->u.gather.gather_bytes, frag_size,
PTL_NO_ACK_REQ, PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, parent), ompi_coll_portals4_get_peer(comm, parent),
mca_coll_portals4_component.pt_idx, mca_coll_portals4_component.pt_idx,
request->u.gather.gather_match_bits, request->u.gather.gather_match_bits,
remote_offset, remote_offset + size_sent,
NULL, NULL,
0, 0,
request->u.gather.gather_cth, request->u.gather.gather_cth,
expected_ops); expected_ops);
if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; } if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; }
size_left -= frag_size;
size_sent += frag_size;
}
} }
/************************************/ /************************************/
@ -734,7 +780,7 @@ ompi_coll_portals4_gather_intra_binomial_top(const void *sbuf, int scount, struc
ompi_coll_portals4_destroy_tree(&(portals4_module->cached_in_order_bmtree)); ompi_coll_portals4_destroy_tree(&(portals4_module->cached_in_order_bmtree));
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra_binomial_top exit rank %d", request->u.gather.my_rank)); "coll:portals4:gather_intra_binomial_top exit rank %d", request->u.gather.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -773,8 +819,9 @@ ompi_coll_portals4_gather_intra_linear_top(const void *sbuf, int scount, struct
int32_t expected_ops =0; int32_t expected_ops =0;
int32_t expected_acks=0; int32_t expected_acks=0;
ptl_size_t number_of_fragment = 1;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra_linear_top enter rank %d", request->u.gather.my_rank)); "coll:portals4:gather_intra_linear_top enter rank %d", request->u.gather.my_rank));
request->type = OMPI_COLL_PORTALS4_TYPE_GATHER; request->type = OMPI_COLL_PORTALS4_TYPE_GATHER;
@ -843,6 +890,13 @@ ompi_coll_portals4_gather_intra_linear_top(const void *sbuf, int scount, struct
ret = setup_sync_handles(comm, request, portals4_module); ret = setup_sync_handles(comm, request, portals4_module);
if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; } if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; }
number_of_fragment = (request->u.gather.packed_size > mca_coll_portals4_component.ni_limits.max_msg_size) ?
(request->u.gather.packed_size + mca_coll_portals4_component.ni_limits.max_msg_size - 1) / mca_coll_portals4_component.ni_limits.max_msg_size :
1;
opal_output_verbose(90, ompi_coll_base_framework.framework_output,
"%s:%d:rank %d:number_of_fragment = %lu",
__FILE__, __LINE__, request->u.gather.my_rank, number_of_fragment);
/***********************************************/ /***********************************************/
/* Chain the RTR and Recv-ACK to the Gather CT */ /* Chain the RTR and Recv-ACK to the Gather CT */
/***********************************************/ /***********************************************/
@ -867,11 +921,13 @@ ompi_coll_portals4_gather_intra_linear_top(const void *sbuf, int scount, struct
if (i_am_root) { if (i_am_root) {
/* root, so do nothing */ /* root, so do nothing */
expected_ops=request->u.gather.size-1; /* gather put from all other ranks */ expected_ops=(request->u.gather.size-1) * number_of_fragment; /* gather put from all other ranks */
expected_acks=0; expected_acks=0;
} else { } else {
ptl_size_t remote_offset=request->u.gather.my_rank * request->u.gather.packed_size; ptl_size_t remote_offset=request->u.gather.my_rank * request->u.gather.packed_size;
ptl_size_t split_offset = 0;
ptl_size_t size_left = request->u.gather.gather_bytes;
opal_output_verbose(30, ompi_coll_base_framework.framework_output, opal_output_verbose(30, ompi_coll_base_framework.framework_output,
"%s:%d:rank(%d): remote_offset(%lu)=rank(%d) * packed_size(%ld)", "%s:%d:rank(%d): remote_offset(%lu)=rank(%d) * packed_size(%ld)",
@ -881,19 +937,34 @@ ompi_coll_portals4_gather_intra_linear_top(const void *sbuf, int scount, struct
expected_ops=1; /* chained RTR */ expected_ops=1; /* chained RTR */
expected_acks=1; /* Recv-ACK from root */ expected_acks=1; /* Recv-ACK from root */
for (ptl_size_t j=0; j<number_of_fragment; j++) {
ptl_size_t frag_size = (size_left > mca_coll_portals4_component.ni_limits.max_msg_size) ?
mca_coll_portals4_component.ni_limits.max_msg_size :
size_left;
opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"%s:%d:rank(%d): frag(%lu),offset_frag (%lu) frag_size(%lu)",
__FILE__, __LINE__, request->u.gather.my_rank,
j, split_offset, frag_size);
ret = PtlTriggeredPut(request->u.gather.gather_mdh, ret = PtlTriggeredPut(request->u.gather.gather_mdh,
request->u.gather.gather_offset, request->u.gather.gather_offset + split_offset,
request->u.gather.gather_bytes, frag_size,
PTL_NO_ACK_REQ, PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, request->u.gather.root_rank), ompi_coll_portals4_get_peer(comm, request->u.gather.root_rank),
mca_coll_portals4_component.pt_idx, mca_coll_portals4_component.pt_idx,
request->u.gather.gather_match_bits, request->u.gather.gather_match_bits,
remote_offset, remote_offset + split_offset,
NULL, NULL,
0, 0,
request->u.gather.gather_cth, request->u.gather.gather_cth,
expected_ops); expected_ops);
if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; } if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; }
size_left -= frag_size;
split_offset += frag_size;
}
} }
/*****************************************/ /*****************************************/
@ -997,7 +1068,7 @@ ompi_coll_portals4_gather_intra_linear_top(const void *sbuf, int scount, struct
"completed CTWait(expected_ops=%d)\n", expected_ops); "completed CTWait(expected_ops=%d)\n", expected_ops);
} }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra_linear_top exit rank %d", request->u.gather.my_rank)); "coll:portals4:gather_intra_linear_top exit rank %d", request->u.gather.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -1020,7 +1091,7 @@ ompi_coll_portals4_gather_intra_binomial_bottom(struct ompi_communicator_t *comm
int ret, line; int ret, line;
int i; int i;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra_binomial_bottom enter rank %d", request->u.gather.my_rank)); "coll:portals4:gather_intra_binomial_bottom enter rank %d", request->u.gather.my_rank));
ret = cleanup_gather_handles(request); ret = cleanup_gather_handles(request);
@ -1065,7 +1136,7 @@ ompi_coll_portals4_gather_intra_binomial_bottom(struct ompi_communicator_t *comm
ompi_request_complete(&request->super, true); ompi_request_complete(&request->super, true);
OPAL_THREAD_UNLOCK(&ompi_request_lock); OPAL_THREAD_UNLOCK(&ompi_request_lock);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra_binomial_bottom exit rank %d", request->u.gather.my_rank)); "coll:portals4:gather_intra_binomial_bottom exit rank %d", request->u.gather.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -1090,7 +1161,7 @@ ompi_coll_portals4_gather_intra_linear_bottom(struct ompi_communicator_t *comm,
int ret, line; int ret, line;
int i; int i;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra_linear_bottom enter rank %d", request->u.gather.my_rank)); "coll:portals4:gather_intra_linear_bottom enter rank %d", request->u.gather.my_rank));
ret = cleanup_gather_handles(request); ret = cleanup_gather_handles(request);
@ -1128,7 +1199,7 @@ ompi_coll_portals4_gather_intra_linear_bottom(struct ompi_communicator_t *comm,
ompi_request_complete(&request->super, true); ompi_request_complete(&request->super, true);
OPAL_THREAD_UNLOCK(&ompi_request_lock); OPAL_THREAD_UNLOCK(&ompi_request_lock);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra_linear_bottom exit rank %d", request->u.gather.my_rank)); "coll:portals4:gather_intra_linear_bottom exit rank %d", request->u.gather.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -1157,7 +1228,7 @@ ompi_coll_portals4_gather_intra(const void *sbuf, int scount, struct ompi_dataty
ompi_coll_portals4_request_t *request; ompi_coll_portals4_request_t *request;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra enter rank %d", ompi_comm_rank(comm))); "coll:portals4:gather_intra enter rank %d", ompi_comm_rank(comm)));
/* /*
@ -1204,7 +1275,7 @@ ompi_coll_portals4_gather_intra(const void *sbuf, int scount, struct ompi_dataty
*/ */
OMPI_COLL_PORTALS4_REQUEST_RETURN(request); OMPI_COLL_PORTALS4_REQUEST_RETURN(request);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:gather_intra exit rank %d", request->u.gather.my_rank)); "coll:portals4:gather_intra exit rank %d", request->u.gather.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -1230,7 +1301,7 @@ ompi_coll_portals4_igather_intra(const void *sbuf, int scount, struct ompi_datat
ompi_coll_portals4_request_t *request; ompi_coll_portals4_request_t *request;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:igather_intra enter rank %d", ompi_comm_rank(comm))); "coll:portals4:igather_intra enter rank %d", ompi_comm_rank(comm)));
/* /*
@ -1267,7 +1338,7 @@ ompi_coll_portals4_igather_intra(const void *sbuf, int scount, struct ompi_datat
if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; } if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; }
} }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:igather_intra exit rank %d", request->u.gather.my_rank)); "coll:portals4:igather_intra exit rank %d", request->u.gather.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -1286,7 +1357,7 @@ ompi_coll_portals4_igather_intra_fini(ompi_coll_portals4_request_t *request)
{ {
int ret, line; int ret, line;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:igather_intra_fini enter rank %d", request->u.gather.my_rank)); "coll:portals4:igather_intra_fini enter rank %d", request->u.gather.my_rank));
/* /*
@ -1300,7 +1371,7 @@ ompi_coll_portals4_igather_intra_fini(ompi_coll_portals4_request_t *request)
if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; } if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; }
} }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:igather_intra_fini exit rank %d", request->u.gather.my_rank)); "coll:portals4:igather_intra_fini exit rank %d", request->u.gather.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;

2
ompi/mca/coll/portals4/coll_portals4_reduce.c
Просмотреть файл

@ -428,7 +428,7 @@ ompi_coll_portals4_ireduce_intra(const void* sendbuf, void* recvbuf, int count,
OMPI_COLL_PORTALS4_REQUEST_RETURN(request); OMPI_COLL_PORTALS4_REQUEST_RETURN(request);
} }
puts("ireduce"); opal_output_verbose(10, ompi_coll_base_framework.framework_output, "ireduce");
return (OMPI_SUCCESS); return (OMPI_SUCCESS);
} }

75
ompi/mca/coll/portals4/coll_portals4_scatter.c
Просмотреть файл

@ -127,7 +127,7 @@ setup_scatter_handles(struct ompi_communicator_t *comm,
ptl_me_t me; ptl_me_t me;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:setup_scatter_handles enter rank %d", request->u.scatter.my_rank)); "coll:portals4:setup_scatter_handles enter rank %d", request->u.scatter.my_rank));
/**********************************/ /**********************************/
@ -136,7 +136,7 @@ setup_scatter_handles(struct ompi_communicator_t *comm,
COLL_PORTALS4_SET_BITS(request->u.scatter.scatter_match_bits, ompi_comm_get_cid(comm), COLL_PORTALS4_SET_BITS(request->u.scatter.scatter_match_bits, ompi_comm_get_cid(comm),
0, 0, COLL_PORTALS4_SCATTER, 0, request->u.scatter.coll_count); 0, 0, COLL_PORTALS4_SCATTER, 0, request->u.scatter.coll_count);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:setup_scatter_handles rank(%d) scatter_match_bits(0x%016lX)", "coll:portals4:setup_scatter_handles rank(%d) scatter_match_bits(0x%016lX)",
request->u.scatter.my_rank, request->u.scatter.scatter_match_bits)); request->u.scatter.my_rank, request->u.scatter.scatter_match_bits));
@ -166,7 +166,7 @@ setup_scatter_handles(struct ompi_communicator_t *comm,
&request->u.scatter.scatter_meh); &request->u.scatter.scatter_meh);
if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; } if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:setup_scatter_handles exit rank %d", request->u.scatter.my_rank)); "coll:portals4:setup_scatter_handles exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -188,7 +188,7 @@ setup_sync_handles(struct ompi_communicator_t *comm,
ptl_me_t me; ptl_me_t me;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:setup_sync_handles enter rank %d", request->u.scatter.my_rank)); "coll:portals4:setup_sync_handles enter rank %d", request->u.scatter.my_rank));
/**********************************/ /**********************************/
@ -197,7 +197,7 @@ setup_sync_handles(struct ompi_communicator_t *comm,
COLL_PORTALS4_SET_BITS(request->u.scatter.sync_match_bits, ompi_comm_get_cid(comm), COLL_PORTALS4_SET_BITS(request->u.scatter.sync_match_bits, ompi_comm_get_cid(comm),
0, 1, COLL_PORTALS4_SCATTER, 0, request->u.scatter.coll_count); 0, 1, COLL_PORTALS4_SCATTER, 0, request->u.scatter.coll_count);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:setup_sync_handles rank(%d) sync_match_bits(0x%016lX)", "coll:portals4:setup_sync_handles rank(%d) sync_match_bits(0x%016lX)",
request->u.scatter.my_rank, request->u.scatter.sync_match_bits)); request->u.scatter.my_rank, request->u.scatter.sync_match_bits));
@ -227,7 +227,7 @@ setup_sync_handles(struct ompi_communicator_t *comm,
&request->u.scatter.sync_meh); &request->u.scatter.sync_meh);
if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; } if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:setup_sync_handles exit rank %d", request->u.scatter.my_rank)); "coll:portals4:setup_sync_handles exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -245,7 +245,7 @@ cleanup_scatter_handles(ompi_coll_portals4_request_t *request)
{ {
int ret, line; int ret, line;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:cleanup_scatter_handles enter rank %d", request->u.scatter.my_rank)); "coll:portals4:cleanup_scatter_handles enter rank %d", request->u.scatter.my_rank));
/**********************************/ /**********************************/
@ -265,7 +265,7 @@ cleanup_scatter_handles(ompi_coll_portals4_request_t *request)
ret = PtlCTFree(request->u.scatter.scatter_cth); ret = PtlCTFree(request->u.scatter.scatter_cth);
if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; } if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:cleanup_scatter_handles exit rank %d", request->u.scatter.my_rank)); "coll:portals4:cleanup_scatter_handles exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -284,7 +284,7 @@ cleanup_sync_handles(ompi_coll_portals4_request_t *request)
int ret, line; int ret, line;
int ptl_ret; int ptl_ret;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:cleanup_sync_handles enter rank %d", request->u.scatter.my_rank)); "coll:portals4:cleanup_sync_handles enter rank %d", request->u.scatter.my_rank));
/**********************************/ /**********************************/
@ -304,7 +304,7 @@ cleanup_sync_handles(ompi_coll_portals4_request_t *request)
ret = PtlCTFree(request->u.scatter.sync_cth); ret = PtlCTFree(request->u.scatter.sync_cth);
if (PTL_OK != ret) { ptl_ret = ret; ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; } if (PTL_OK != ret) { ptl_ret = ret; ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:cleanup_sync_handles exit rank %d", request->u.scatter.my_rank)); "coll:portals4:cleanup_sync_handles exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -341,8 +341,9 @@ ompi_coll_portals4_scatter_intra_linear_top(const void *sbuf, int scount, struct
int32_t expected_chained_rtrs = 0; int32_t expected_chained_rtrs = 0;
int32_t expected_chained_acks = 0; int32_t expected_chained_acks = 0;
ptl_size_t number_of_fragment = 1;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:scatter_intra_linear_top enter rank %d", request->u.scatter.my_rank)); "coll:portals4:scatter_intra_linear_top enter rank %d", request->u.scatter.my_rank));
request->type = OMPI_COLL_PORTALS4_TYPE_SCATTER; request->type = OMPI_COLL_PORTALS4_TYPE_SCATTER;
@ -409,6 +410,13 @@ ompi_coll_portals4_scatter_intra_linear_top(const void *sbuf, int scount, struct
ret = setup_sync_handles(comm, request, portals4_module); ret = setup_sync_handles(comm, request, portals4_module);
if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; } if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; }
number_of_fragment = (request->u.scatter.packed_size > mca_coll_portals4_component.ni_limits.max_msg_size) ?
(request->u.scatter.packed_size + mca_coll_portals4_component.ni_limits.max_msg_size - 1) / mca_coll_portals4_component.ni_limits.max_msg_size :
1;
opal_output_verbose(90, ompi_coll_base_framework.framework_output,
"%s:%d:rank %d:number_of_fragment = %lu",
__FILE__, __LINE__, request->u.scatter.my_rank, number_of_fragment);
/**********************************/ /**********************************/
/* do the scatter */ /* do the scatter */
/**********************************/ /**********************************/
@ -445,25 +453,42 @@ ompi_coll_portals4_scatter_intra_linear_top(const void *sbuf, int scount, struct
} }
ptl_size_t offset = request->u.scatter.packed_size * i; ptl_size_t offset = request->u.scatter.packed_size * i;
ptl_size_t size_sent = 0;
ptl_size_t size_left = request->u.scatter.packed_size;
opal_output_verbose(30, ompi_coll_base_framework.framework_output, opal_output_verbose(10, ompi_coll_base_framework.framework_output,
"%s:%d:rank(%d): offset(%lu)=rank(%d) * packed_size(%ld)", "%s:%d:rank(%d): offset(%lu)=rank(%d) * packed_size(%ld)",
__FILE__, __LINE__, request->u.scatter.my_rank, __FILE__, __LINE__, request->u.scatter.my_rank,
offset, i, request->u.scatter.packed_size); offset, i, request->u.scatter.packed_size);
for (ptl_size_t j=0; j<number_of_fragment; j++) {
ptl_size_t frag_size = (size_left > mca_coll_portals4_component.ni_limits.max_msg_size) ?
mca_coll_portals4_component.ni_limits.max_msg_size :
size_left;
OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"%s:%d:rank(%d): frag(%lu),offset_frag (%lu) frag_size(%lu)",
__FILE__, __LINE__, request->u.scatter.my_rank,
j, size_sent, frag_size));
ret = PtlTriggeredPut(request->u.scatter.scatter_mdh, ret = PtlTriggeredPut(request->u.scatter.scatter_mdh,
(ptl_size_t)request->u.scatter.scatter_buf + offset, (ptl_size_t)request->u.scatter.scatter_buf + offset + size_sent,
request->u.scatter.packed_size, frag_size,
PTL_NO_ACK_REQ, PTL_NO_ACK_REQ,
ompi_coll_portals4_get_peer(comm, i), ompi_coll_portals4_get_peer(comm, i),
mca_coll_portals4_component.pt_idx, mca_coll_portals4_component.pt_idx,
request->u.scatter.scatter_match_bits, request->u.scatter.scatter_match_bits,
0, size_sent,
NULL, NULL,
0, 0,
request->u.scatter.scatter_cth, request->u.scatter.scatter_cth,
expected_chained_rtrs); expected_chained_rtrs);
if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; } if (PTL_OK != ret) { ret = OMPI_ERROR; line = __LINE__; goto err_hdlr; }
size_left -= frag_size;
size_sent += frag_size;
}
} }
} else { } else {
/* non-root, so do nothing */ /* non-root, so do nothing */
@ -473,7 +498,7 @@ ompi_coll_portals4_scatter_intra_linear_top(const void *sbuf, int scount, struct
expected_acks = 0; expected_acks = 0;
/* operations on the scatter counter */ /* operations on the scatter counter */
expected_puts = 1; /* scatter put from root */ expected_puts = number_of_fragment; /* scatter put from root */
expected_chained_rtrs = 0; expected_chained_rtrs = 0;
expected_chained_acks = 0; expected_chained_acks = 0;
} }
@ -552,7 +577,7 @@ ompi_coll_portals4_scatter_intra_linear_top(const void *sbuf, int scount, struct
"completed CTWait(expected_ops=%d)\n", expected_ops); "completed CTWait(expected_ops=%d)\n", expected_ops);
} }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:scatter_intra_linear_top exit rank %d", request->u.scatter.my_rank)); "coll:portals4:scatter_intra_linear_top exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -574,7 +599,7 @@ ompi_coll_portals4_scatter_intra_linear_bottom(struct ompi_communicator_t *comm,
{ {
int ret, line; int ret, line;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:scatter_intra_linear_bottom enter rank %d", request->u.scatter.my_rank)); "coll:portals4:scatter_intra_linear_bottom enter rank %d", request->u.scatter.my_rank));
ret = cleanup_scatter_handles(request); ret = cleanup_scatter_handles(request);
@ -616,7 +641,7 @@ ompi_coll_portals4_scatter_intra_linear_bottom(struct ompi_communicator_t *comm,
ompi_request_complete(&request->super, true); ompi_request_complete(&request->super, true);
OPAL_THREAD_UNLOCK(&ompi_request_lock); OPAL_THREAD_UNLOCK(&ompi_request_lock);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:scatter_intra_linear_bottom exit rank %d", request->u.scatter.my_rank)); "coll:portals4:scatter_intra_linear_bottom exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -645,7 +670,7 @@ ompi_coll_portals4_scatter_intra(const void *sbuf, int scount, struct ompi_datat
ompi_coll_portals4_request_t *request; ompi_coll_portals4_request_t *request;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:scatter_intra enter rank %d", ompi_comm_rank(comm))); "coll:portals4:scatter_intra enter rank %d", ompi_comm_rank(comm)));
/* /*
@ -679,7 +704,7 @@ ompi_coll_portals4_scatter_intra(const void *sbuf, int scount, struct ompi_datat
*/ */
OMPI_COLL_PORTALS4_REQUEST_RETURN(request); OMPI_COLL_PORTALS4_REQUEST_RETURN(request);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:scatter_intra exit rank %d", request->u.scatter.my_rank)); "coll:portals4:scatter_intra exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -705,7 +730,7 @@ ompi_coll_portals4_iscatter_intra(const void *sbuf, int scount, struct ompi_data
ompi_coll_portals4_request_t *request; ompi_coll_portals4_request_t *request;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:iscatter_intra enter rank %d", ompi_comm_rank(comm))); "coll:portals4:iscatter_intra enter rank %d", ompi_comm_rank(comm)));
/* /*
@ -732,7 +757,7 @@ ompi_coll_portals4_iscatter_intra(const void *sbuf, int scount, struct ompi_data
module); module);
if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; } if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:iscatter_intra exit rank %d", request->u.scatter.my_rank)); "coll:portals4:iscatter_intra exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;
@ -751,7 +776,7 @@ ompi_coll_portals4_iscatter_intra_fini(ompi_coll_portals4_request_t *request)
{ {
int ret, line; int ret, line;
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:iscatter_intra_fini enter rank %d", request->u.scatter.my_rank)); "coll:portals4:iscatter_intra_fini enter rank %d", request->u.scatter.my_rank));
/* /*
@ -760,7 +785,7 @@ ompi_coll_portals4_iscatter_intra_fini(ompi_coll_portals4_request_t *request)
ret = ompi_coll_portals4_scatter_intra_linear_bottom(request->super.req_mpi_object.comm, request); ret = ompi_coll_portals4_scatter_intra_linear_bottom(request->super.req_mpi_object.comm, request);
if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; } if (MPI_SUCCESS != ret) { line = __LINE__; goto err_hdlr; }
OPAL_OUTPUT((ompi_coll_base_framework.framework_output, OPAL_OUTPUT_VERBOSE((10, ompi_coll_base_framework.framework_output,
"coll:portals4:iscatter_intra_fini exit rank %d", request->u.scatter.my_rank)); "coll:portals4:iscatter_intra_fini exit rank %d", request->u.scatter.my_rank));
return OMPI_SUCCESS; return OMPI_SUCCESS;