ports/openmpi
ports/openmpi
1
1
Форкнуть 0
Код Релизы Активность
479f6c5e6b
openmpi/src/communicator/comm_cid.c

689 строки
22 KiB
C
Исходник Обычный вид История

adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
/*
* $HEADER$
*/
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
#include "ompi_config.h"
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
#include <string.h>
#include <stdio.h>
#include "mpi.h"
#include "communicator/communicator.h"
#include "op/op.h"
#include "proc/proc.h"
#include "include/constants.h"
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
#include "class/ompi_pointer_array.h"
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
#include "mca/pcm/pcm.h"
#include "mca/pml/pml.h"
#include "mca/coll/coll.h"
#include "mca/coll/base/base.h"
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
#include "mca/oob/oob.h"
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
#define OMPI_COLL_TAG_ALLREDUCE 31000
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
#define OMPI_MAX_COMM 32768
/**
* These functions make sure, that we determine the global result over
* an intra communicators (simple), an inter-communicator and a
* pseudo inter-communicator described by two separate intra-comms
* and a bridge-comm (intercomm-create scenario).
*/
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
typedef int ompi_comm_cid_allredfct (int *inbuf, int* outbuf,
int count, ompi_op_t *op,
ompi_communicator_t *comm,
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
ompi_communicator_t *bridgecomm,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
void* lleader, void* rleader,
int send_first );
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
static int ompi_comm_allreduce_intra (int *inbuf, int* outbuf,
int count, ompi_op_t *op,
ompi_communicator_t *intercomm,
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
ompi_communicator_t *bridgecomm,
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
void* local_leader,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
void* remote_ledaer,
int send_first );
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
static int ompi_comm_allreduce_inter (int *inbuf, int *outbuf,
int count, ompi_op_t *op,
ompi_communicator_t *intercomm,
ompi_communicator_t *bridgecomm,
void* local_leader,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
void* remote_leader,
int send_first );
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
static int ompi_comm_allreduce_intra_bridge(int *inbuf, int* outbuf,
int count, ompi_op_t *op,
ompi_communicator_t *intercomm,
- intra-comms tested, several minor bugs fixed - cid allocation for intra-comms is now enabled and working This commit was SVN r1750.
2004-07-16 00:55:15 +04:00
ompi_communicator_t *bridgecomm,
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
void* local_leader,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
void* remote_leader,
int send_first);
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
static int ompi_comm_allreduce_intra_oob (int *inbuf, int* outbuf,
int count, ompi_op_t *op,
ompi_communicator_t *intercomm,
- intra-comms tested, several minor bugs fixed - cid allocation for intra-comms is now enabled and working This commit was SVN r1750.
2004-07-16 00:55:15 +04:00
ompi_communicator_t *bridgecomm,
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
void* local_leader,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
void* remote_leader,
int send_first );
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
fixes to make comm_join work This commit was SVN r2745.
2004-09-17 20:28:58 +04:00
#ifdef MULTI_THREADS
static ompi_mutex_t ompi_cid_lock;
static volatile ompi_list_t registered_threads;
static volatile int running_cid=-1;
int ompi_comm_nextcid ( ompi_communicator_t* newcomm,
ompi_communicator_t* comm,
ompi_communicator_t* bridgecomm,
void* local_leader,
void* remote_leader,
int mode, int send_first )
{
int nextlocal_cid;
int nextcid;
int done=0;
int response=0, glresponse=0;
int flag;
int start=ompi_mpi_communicators.lowest_free;
int i;
ompi_comm_cid_allredfct* allredfnct;
/**
* Determine which implementation of allreduce we have to use
* for the current scenario
*/
switch (mode)
{
case OMPI_COMM_CID_INTRA:
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra;
break;
case OMPI_COMM_CID_INTER:
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_inter;
break;
case OMPI_COMM_CID_INTRA_BRIDGE:
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_bridge;
break;
case OMPI_COMM_CID_INTRA_OOB:
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_oob;
break;
default:
return MPI_UNDEFINED;
break;
}
/* Lock the list
register threads in registered_threads
unlock
*/
while (!done) {
/* lock
if (!lowest_cid ) {
unlock;
continue;
}
unlock;
*/
/**
* This is the real algorithm described in the doc
*/
for (i=start; i<OMPI_MAX_COMM ;i++) {
flag=ompi_pointer_array_test_and_set_item(&ompi_mpi_communicators, i, comm);
if (true == flag) {
nextlocal_cid = i;
break;
}
}
(allredfnct)(&nextlocal_cid, &nextcid, 1, MPI_MAX, comm, bridgecomm,
local_leader, remote_leader, send_first );
if (nextcid == nextlocal_cid) {
response = 1; /* fine with me */
}
else {
ompi_pointer_array_set_item(&ompi_mpi_communicators,
nextlocal_cid, NULL);
flag = ompi_pointer_array_test_and_set_item(&ompi_mpi_communicators,
nextcid, comm );
if (true == flag) {
response = 1; /* works as well */
}
else {
response = 0; /* nope, not acceptable */
}
}
(allredfnct)(&response, &glresponse, 1, MPI_MIN, comm, bridgecomm,
local_leader, remote_leader, send_first );
if (1 == glresponse) {
done = 1; /* we are done */
break;
}
else if ( 0 == glresponse ) {
if ( 1 == response ) {
/* we could use that, but other don't agree */
ompi_pointer_array_set_item(&ompi_mpi_communicators,
nextcid, NULL);
}
start = nextcid+1; /* that's where we can start the next round */
}
}
/* set the according values to the newcomm */
newcomm->c_contextid = nextcid;
newcomm->c_f_to_c_index = newcomm->c_contextid;
ompi_pointer_array_set_item (&ompi_mpi_communicators, nextcid, newcomm);
/*
lock
unregister
unlock
*/
return (MPI_SUCCESS);
}
#endif
extracting the creation of comm_cid from the comm_set routine. The reason is, that we had to pass four arguments to comm_sdet, which were just passed to comm_cid. However, for the dynamic case, even these four arguments are not enough. So I extracted it. A typical sequence for a comm-creation will therefore be: newcomm = ompi_comm_set (...); ompi_comm_nextcid (newcomm, oldcomm,...); This commit was SVN r1312.
2004-06-16 19:40:52 +04:00
int ompi_comm_nextcid ( ompi_communicator_t* newcomm,
ompi_communicator_t* comm,
ompi_communicator_t* bridgecomm,
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
void* local_leader,
void* remote_leader,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
int mode, int send_first )
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
{
int nextlocal_cid;
int nextcid;
int done=0;
int response=0, glresponse=0;
int flag;
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
int start=ompi_mpi_communicators.lowest_free;
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
int i;
ompi_comm_cid_allredfct* allredfnct;
/**
* Determine which implementation of allreduce we have to use
* for the current scenario
*/
switch (mode)
{
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
case OMPI_COMM_CID_INTRA:
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra;
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
break;
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
case OMPI_COMM_CID_INTER:
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_inter;
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
break;
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
case OMPI_COMM_CID_INTRA_BRIDGE:
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_bridge;
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
break;
case OMPI_COMM_CID_INTRA_OOB:
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_oob;
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
break;
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
default:
return MPI_UNDEFINED;
break;
}
/**
* This is the real algorithm described in the doc
*/
while ( !done ){
for (i=start; i<OMPI_MAX_COMM ;i++) {
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
flag=ompi_pointer_array_test_and_set_item(&ompi_mpi_communicators, i, comm);
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
if (true == flag) {
nextlocal_cid = i;
break;
}
}
(allredfnct)(&nextlocal_cid, &nextcid, 1, MPI_MAX, comm, bridgecomm,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
local_leader, remote_leader, send_first );
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
if (nextcid == nextlocal_cid) {
response = 1; /* fine with me */
}
else {
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
ompi_pointer_array_set_item(&ompi_mpi_communicators,
nextlocal_cid, NULL);
fixing a stupid bug This commit was SVN r2711.
2004-09-16 16:16:21 +04:00
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
flag = ompi_pointer_array_test_and_set_item(&ompi_mpi_communicators,
fixing a stupid bug This commit was SVN r2711.
2004-09-16 16:16:21 +04:00
nextcid, comm );
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
if (true == flag) {
response = 1; /* works as well */
}
else {
response = 0; /* nope, not acceptable */
}
}
(allredfnct)(&response, &glresponse, 1, MPI_MIN, comm, bridgecomm,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
local_leader, remote_leader, send_first );
fixing a stupid bug in comm_nextcid: forgot to free the value in the pointer_array in case the others did not agree with my suggestion for the cid. This commit was SVN r2631.
2004-09-13 19:14:09 +04:00
if (1 == glresponse) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
done = 1; /* we are done */
break;
}
fixing a stupid bug in comm_nextcid: forgot to free the value in the pointer_array in case the others did not agree with my suggestion for the cid. This commit was SVN r2631.
2004-09-13 19:14:09 +04:00
else if ( 0 == glresponse ) {
fixing a stupid bug This commit was SVN r2711.
2004-09-16 16:16:21 +04:00
if ( 1 == response ) {
/* we could use that, but other don't agree */
ompi_pointer_array_set_item(&ompi_mpi_communicators,
nextcid, NULL);
}
start = nextcid+1; /* that's where we can start the next round */
fixing a stupid bug in comm_nextcid: forgot to free the value in the pointer_array in case the others did not agree with my suggestion for the cid. This commit was SVN r2631.
2004-09-13 19:14:09 +04:00
}
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
}
extracting the creation of comm_cid from the comm_set routine. The reason is, that we had to pass four arguments to comm_sdet, which were just passed to comm_cid. However, for the dynamic case, even these four arguments are not enough. So I extracted it. A typical sequence for a comm-creation will therefore be: newcomm = ompi_comm_set (...); ompi_comm_nextcid (newcomm, oldcomm,...); This commit was SVN r1312.
2004-06-16 19:40:52 +04:00
/* set the according values to the newcomm */
newcomm->c_contextid = nextcid;
- intra-comms tested, several minor bugs fixed - cid allocation for intra-comms is now enabled and working This commit was SVN r1750.
2004-07-16 00:55:15 +04:00
newcomm->c_f_to_c_index = newcomm->c_contextid;
extracting the creation of comm_cid from the comm_set routine. The reason is, that we had to pass four arguments to comm_sdet, which were just passed to comm_cid. However, for the dynamic case, even these four arguments are not enough. So I extracted it. A typical sequence for a comm-creation will therefore be: newcomm = ompi_comm_set (...); ompi_comm_nextcid (newcomm, oldcomm,...); This commit was SVN r1312.
2004-06-16 19:40:52 +04:00
ompi_pointer_array_set_item (&ompi_mpi_communicators, nextcid, newcomm);
return (MPI_SUCCESS);
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
}
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
/**************************************************************************/
/**************************************************************************/
/**************************************************************************/
/* This routine serves two purposes:
* - the allreduce acts as a kind of Barrier,
* which avoids, that we have incoming fragments
* on the new communicator before everybody has set
* up the comm structure.
* - some components (e.g. the collective MagPIe component
* might want to generate new communicators and communicate
* using the new comm. Thus, it can just be called after
* the 'barrier'.
*
* The reason that this routine is in comm_cid and not in
* comm.c is, that this file contains the allreduce implementations
* which are required, and thus we avoid having duplicate code...
*/
int ompi_comm_activate ( ompi_communicator_t* newcomm,
ompi_communicator_t* comm,
ompi_communicator_t* bridgecomm,
void* local_leader,
void* remote_leader,
int mode,
int send_first,
mca_base_component_t *collcomponent )
{
initializing two variables used in a dummy allreduce This commit was SVN r1926.
2004-08-06 19:46:44 +04:00
int ok=0, gok=0;
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
ompi_comm_cid_allredfct* allredfnct;
/* Step 1: the barrier, after which it is allowed to
* send messages over the new communicator
*/
switch (mode)
{
case OMPI_COMM_CID_INTRA:
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra;
break;
case OMPI_COMM_CID_INTER:
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_inter;
break;
case OMPI_COMM_CID_INTRA_BRIDGE:
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_bridge;
break;
case OMPI_COMM_CID_INTRA_OOB:
allredfnct=(ompi_comm_cid_allredfct*)ompi_comm_allreduce_intra_oob;
break;
default:
return MPI_UNDEFINED;
break;
}
(allredfnct)(&ok, &gok, 1, MPI_MIN, comm, bridgecomm,
local_leader, remote_leader, send_first );
/* Step 2: call all functions, which might use the new communicator
* already.
*/
/* Initialize the coll components */
/* Let the collectives components fight over who will do
collective on this new comm. */
if (OMPI_ERROR == mca_coll_base_comm_select(newcomm, collcomponent)) {
return OMPI_ERROR;
}
return OMPI_SUCCESS;
}
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
/**************************************************************************/
/**************************************************************************/
/**************************************************************************/
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
/* Arguments not used in this implementation:
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
* - bridgecomm
* - local_leader
* - remote_leader
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
* - send_first
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
*/
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
static int ompi_comm_allreduce_intra ( int *inbuf, int *outbuf,
int count, ompi_op_t *op,
ompi_communicator_t *comm,
ompi_communicator_t *bridgecomm,
void* local_leader,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
void* remote_leader,
int send_first )
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
{
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
return comm->c_coll.coll_allreduce ( inbuf, outbuf, count, MPI_INT,
op,comm );
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
}
/* Arguments not used in this implementation:
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
* - bridgecomm
* - local_leader
* - remote_leader
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
* - send_first
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
*/
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
static int ompi_comm_allreduce_inter ( int *inbuf, int *outbuf,
int count, ompi_op_t *op,
ompi_communicator_t *intercomm,
ompi_communicator_t *bridgecomm,
void* local_leader,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
void* remote_leader,
int send_first )
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
{
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
int local_rank, rsize;
int i, rc;
int *sbuf;
int *tmpbuf=NULL;
int *rcounts=NULL, scount=0;
int *rdisps=NULL;
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( &ompi_mpi_op_sum != op && &ompi_mpi_op_prod != op &&
&ompi_mpi_op_max != op && &ompi_mpi_op_min != op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
return MPI_ERR_OP;
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( !OMPI_COMM_IS_INTER (intercomm)) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
return MPI_ERR_COMM;
}
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
/* Allocate temporary arrays */
rsize = ompi_comm_remote_size (intercomm);
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
local_rank = ompi_comm_rank ( intercomm );
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
tmpbuf = (int *) malloc ( count * sizeof(int));
rdisps = (int *) calloc ( rsize, sizeof(int));
rcounts = (int *) calloc ( rsize, sizeof(int) );
if ( NULL == tmpbuf || NULL == rdisps || NULL == rcounts ) {
return OMPI_ERR_OUT_OF_RESOURCE;
}
/* Execute the inter-allreduce: the result of our group will
be in the buffer of the remote group */
Lots of changes in this commit, mostly having to do with the first real commit of the collectives. MPI_SCAN and MPI_EXSCAN are still not implemented, but lots of other things are in the critical path and holding up other people, so it's ok to commit without them: - better checks for sizes in configure, and add defaults for fortran sizes if we don't have a fortran compiler - fix some logic that was accidentally broken for size checks for the file type offset_t - add some C equivalent types for fortran's complex and double complex (for use in internal reduction/op functions) - additionals and slight reorganization of ompi_mpi_init() ompi_mpi_finalize() - fully implement all top-level MPI collective calls, including all param checking for both intra- and inter-communicators (woof) - change the communicator_t type for stuff that we need in coll, and update all references throughout the code base to match - all kinds of updates to the coll framework base - next cut of the basic coll module -- has all intracommunicator collectives implemented except scan and exscan (see note above). All intercommunicator functions return ERR_NOT_IMPLEMENTED. - MPI_Op is a fixed implementation -- not component-ized yet. So there are generic C loops for all implementations. This commit was SVN r1491.
2004-06-29 04:02:25 +04:00
rc = intercomm->c_coll.coll_allreduce ( inbuf, tmpbuf, count, MPI_INT,
op, intercomm );
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( OMPI_SUCCESS != rc ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
goto exit;
}
if ( 0 == local_rank ) {
MPI_Request req;
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
/* for the allgatherv later */
scount = count;
/* local leader exchange their data and determine the overall result
for both groups */
rc = mca_pml.pml_irecv (outbuf, count, MPI_INT, 0,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
OMPI_COLL_TAG_ALLREDUCE
, intercomm, &req );
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( OMPI_SUCCESS != rc ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
goto exit;
}
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
rc = mca_pml.pml_send (tmpbuf, count, MPI_INT, 0,
OMPI_COLL_TAG_ALLREDUCE,
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
MCA_PML_BASE_SEND_STANDARD, intercomm );
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( OMPI_SUCCESS != rc ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
goto exit;
}
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
rc = mca_pml.pml_wait_all ( 1, &req, MPI_STATUS_IGNORE );
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( OMPI_SUCCESS != rc ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
goto exit;
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( &ompi_mpi_op_max == op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
for ( i = 0 ; i < count; i++ ) {
if (tmpbuf[i] > outbuf[i]) outbuf[i] = tmpbuf[i];
}
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
else if ( &ompi_mpi_op_min == op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
for ( i = 0 ; i < count; i++ ) {
if (tmpbuf[i] < outbuf[i]) outbuf[i] = tmpbuf[i];
}
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
else if ( &ompi_mpi_op_sum == op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
for ( i = 0 ; i < count; i++ ) {
outbuf[i] += tmpbuf[i];
}
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
else if ( &ompi_mpi_op_prod == op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
for ( i = 0 ; i < count; i++ ) {
outbuf[i] *= tmpbuf[i];
}
}
}
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
/* distribute the overall result to all processes in the other group.
Instead of using bcast, we are using here allgatherv, to avoid the
possible deadlock. Else, we need an algorithm to determine,
which group sends first in the inter-bcast and which receives
the result first.
*/
rcounts[0] = count;
sbuf = outbuf;
rc = intercomm->c_coll.coll_allgatherv (sbuf, scount, MPI_INT, outbuf,
rcounts, rdisps, MPI_INT,
intercomm);
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
exit:
if ( NULL != tmpbuf ) {
free ( tmpbuf );
}
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
if ( NULL != rcounts ) {
free ( rcounts );
}
if ( NULL != rdisps ) {
free ( rdisps );
}
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
return (rc);
}
/* Arguments not used in this implementation:
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
* - send_first
*/
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
static int ompi_comm_allreduce_intra_bridge (int *inbuf, int *outbuf,
int count, ompi_op_t *op,
ompi_communicator_t *comm,
ompi_communicator_t *bcomm,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
void* lleader, void* rleader,
int send_first )
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
{
int *tmpbuf=NULL;
int local_rank;
int i;
int rc;
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
int local_leader, remote_leader;
local_leader = (*((int*)lleader));
remote_leader = (*((int*)rleader));
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( &ompi_mpi_op_sum != op && &ompi_mpi_op_prod != op &&
&ompi_mpi_op_max != op && &ompi_mpi_op_min != op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
return MPI_ERR_OP;
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
local_rank = ompi_comm_rank ( comm );
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
tmpbuf = (int *) malloc ( count * sizeof(int));
if ( NULL == tmpbuf ) {
return MPI_ERR_INTERN;
}
/* Intercomm_create */
Lots of changes in this commit, mostly having to do with the first real commit of the collectives. MPI_SCAN and MPI_EXSCAN are still not implemented, but lots of other things are in the critical path and holding up other people, so it's ok to commit without them: - better checks for sizes in configure, and add defaults for fortran sizes if we don't have a fortran compiler - fix some logic that was accidentally broken for size checks for the file type offset_t - add some C equivalent types for fortran's complex and double complex (for use in internal reduction/op functions) - additionals and slight reorganization of ompi_mpi_init() ompi_mpi_finalize() - fully implement all top-level MPI collective calls, including all param checking for both intra- and inter-communicators (woof) - change the communicator_t type for stuff that we need in coll, and update all references throughout the code base to match - all kinds of updates to the coll framework base - next cut of the basic coll module -- has all intracommunicator collectives implemented except scan and exscan (see note above). All intercommunicator functions return ERR_NOT_IMPLEMENTED. - MPI_Op is a fixed implementation -- not component-ized yet. So there are generic C loops for all implementations. This commit was SVN r1491.
2004-06-29 04:02:25 +04:00
rc = comm->c_coll.coll_allreduce ( inbuf, tmpbuf, count, MPI_INT,
op, comm );
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( OMPI_SUCCESS != rc ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
goto exit;
}
if (local_rank == local_leader ) {
MPI_Request req;
rc = mca_pml.pml_irecv ( outbuf, count, MPI_INT, remote_leader,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
OMPI_COLL_TAG_ALLREDUCE,
bcomm, &req );
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( OMPI_SUCCESS != rc ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
goto exit;
}
rc = mca_pml.pml_send (tmpbuf, count, MPI_INT, remote_leader,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
OMPI_COLL_TAG_ALLREDUCE,
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
MCA_PML_BASE_SEND_STANDARD, bcomm );
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( OMPI_SUCCESS != rc ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
goto exit;
}
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
rc = mca_pml.pml_wait_all ( 1, &req, MPI_STATUS_IGNORE);
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( OMPI_SUCCESS != rc ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
goto exit;
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
if ( &ompi_mpi_op_max == op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
for ( i = 0 ; i < count; i++ ) {
if (tmpbuf[i] > outbuf[i]) outbuf[i] = tmpbuf[i];
}
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
else if ( &ompi_mpi_op_min == op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
for ( i = 0 ; i < count; i++ ) {
if (tmpbuf[i] < outbuf[i]) outbuf[i] = tmpbuf[i];
}
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
else if ( &ompi_mpi_op_sum == op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
for ( i = 0 ; i < count; i++ ) {
outbuf[i] += tmpbuf[i];
}
}
First pass of lam -> ompi conversion This commit was SVN r1191.
2004-06-07 19:33:53 +04:00
else if ( &ompi_mpi_op_prod == op ) {
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
for ( i = 0 ; i < count; i++ ) {
outbuf[i] *= tmpbuf[i];
}
}
}
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
rc = comm->c_coll.coll_bcast ( outbuf, count, MPI_INT, local_leader,
comm);
adding most of the required functionality for handling MPI-1 and most of MPI-2 communicator functions (except dynamic process management). The cid are currently not yet calculated properly, although the functions are checked in. Still to do: - make the CID allocation routine thread safe - add the ACK in lam_comm_free - fix a bug in lam_comm_split for inter-communicators ( in this case we can not have allgather_intra and allgather_inter at the same time at this communicators, that's however what the current implemention assumes). Reviewed by Jeff, Rich and Tim. This commit was SVN r1148.
2004-05-21 23:36:19 +04:00
exit:
if (NULL != tmpbuf ) {
free (tmpbuf);
}
return (rc);
}
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
/* Arguments not used in this implementation:
* - bridgecomm
*
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
* lleader is the local rank of root in comm
* rleader is the OOB contact information of the
* root processes in the other world.
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
*/
restructuring some of the communicator code. with a couple of internal tricks, intercomm_create works now. This commit was SVN r1855.
2004-08-04 02:07:45 +04:00
static int ompi_comm_allreduce_intra_oob (int *inbuf, int *outbuf,
int count, ompi_op_t *op,
ompi_communicator_t *comm,
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
ompi_communicator_t *bridgecomm,
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
void* lleader, void* rleader,
int send_first )
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
{
int *tmpbuf=NULL;
int i;
int rc;
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
int local_leader, local_rank;
ompi_process_name_t *remote_leader=NULL;
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
local_leader = (*((int*)lleader));
remote_leader = (ompi_process_name_t*)rleader;
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
if ( &ompi_mpi_op_sum != op && &ompi_mpi_op_prod != op &&
&ompi_mpi_op_max != op && &ompi_mpi_op_min != op ) {
return MPI_ERR_OP;
}
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
local_rank = ompi_comm_rank ( comm );
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
tmpbuf = (int *) malloc ( count * sizeof(int));
if ( NULL == tmpbuf ) {
return MPI_ERR_INTERN;
}
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
/* comm is an intra-communicator */
rc = comm->c_coll.coll_allreduce(inbuf,tmpbuf,count,MPI_INT,op, comm );
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
if ( OMPI_SUCCESS != rc ) {
goto exit;
}
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
if (local_rank == local_leader ) {
OOB API changes: - removed send_hton/recv_ntoh routines as we now have send_packed/recv_packed - removed constness from apis - adding flag (in work) to allow recv to allocate and return recv buffer - updated edgars communicator code to use pack routines rather than ntoh routines This commit was SVN r2095.
2004-08-13 02:41:42 +04:00
ompi_buffer_t sbuf;
ompi_buffer_t rbuf;
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
OOB API changes: - removed send_hton/recv_ntoh routines as we now have send_packed/recv_packed - removed constness from apis - adding flag (in work) to allow recv to allocate and return recv buffer - updated edgars communicator code to use pack routines rather than ntoh routines This commit was SVN r2095.
2004-08-13 02:41:42 +04:00
ompi_buffer_init(&sbuf, count * sizeof(int));
ompi_pack(sbuf, tmpbuf, count, OMPI_INT32);
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
OOB API changes: - removed send_hton/recv_ntoh routines as we now have send_packed/recv_packed - removed constness from apis - adding flag (in work) to allow recv to allocate and return recv buffer - updated edgars communicator code to use pack routines rather than ntoh routines This commit was SVN r2095.
2004-08-13 02:41:42 +04:00
if ( send_first ) {
rc = mca_oob_send_packed(remote_leader, sbuf, 0, 0);
rc = mca_oob_recv_packed (remote_leader, &rbuf, NULL);
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
}
else {
OOB API changes: - removed send_hton/recv_ntoh routines as we now have send_packed/recv_packed - removed constness from apis - adding flag (in work) to allow recv to allocate and return recv buffer - updated edgars communicator code to use pack routines rather than ntoh routines This commit was SVN r2095.
2004-08-13 02:41:42 +04:00
rc = mca_oob_recv_packed(remote_leader, &rbuf, NULL);
rc = mca_oob_send_packed(remote_leader, sbuf, 0, 0);
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
}
OOB API changes: - removed send_hton/recv_ntoh routines as we now have send_packed/recv_packed - removed constness from apis - adding flag (in work) to allow recv to allocate and return recv buffer - updated edgars communicator code to use pack routines rather than ntoh routines This commit was SVN r2095.
2004-08-13 02:41:42 +04:00
ompi_unpack(rbuf, outbuf, count, OMPI_INT32);
ompi_buffer_free(sbuf);
ompi_buffer_free(rbuf);
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
if ( &ompi_mpi_op_max == op ) {
for ( i = 0 ; i < count; i++ ) {
if (tmpbuf[i] > outbuf[i]) outbuf[i] = tmpbuf[i];
}
}
else if ( &ompi_mpi_op_min == op ) {
for ( i = 0 ; i < count; i++ ) {
if (tmpbuf[i] < outbuf[i]) outbuf[i] = tmpbuf[i];
}
}
else if ( &ompi_mpi_op_sum == op ) {
for ( i = 0 ; i < count; i++ ) {
outbuf[i] += tmpbuf[i];
}
}
else if ( &ompi_mpi_op_prod == op ) {
for ( i = 0 ; i < count; i++ ) {
outbuf[i] *= tmpbuf[i];
}
}
}
a couple of changes: - because we had the scenario, that fragments for a communicator, which was not not yet set up on all procs, arrived and caused problem, we introduced a comm_activate function call, which executes a kind of barrier (using the allreduce functions used for the comm_cid allocation). Setting up the coll-component has moved *after* this barrier, since some coll-modules (e.g. the MagPIe component) might want to communicate using this communicator already (e.g. a comm_split). - adding a new file comm_dyn.c, which basically abstracts the required functionality for connect-accept, and therefore is the 'magic' code (from the MPI point of view) for all dynamically created communicators. This commit was SVN r1900.
2004-08-05 20:31:30 +04:00
rc = comm->c_coll.coll_bcast (outbuf, count, MPI_INT,
local_leader, comm);
preparing the comm_cid allocation routine for dynamic process management This commit was SVN r1336.
2004-06-17 02:37:03 +04:00
exit:
if (NULL != tmpbuf ) {
free (tmpbuf);
}
return (rc);
}
Копировать постоянную ссылку