int MPI_Ssend_init(void *buf, int count, MPI_Datatype type,
int dest, int tag, MPI_Comm comm,
MPI_Request *request)
{
int rc;
if ( MPI_PARAM_CHECK ) {
rc = MPI_SUCCESS;
OMPI_ERR_INIT_FINALIZE(FUNC_NAME);
if (ompi_comm_invalid(comm)) {
return OMPI_ERRHANDLER_INVOKE(MPI_COMM_WORLD, MPI_ERR_COMM, FUNC_NAME);
} else if (count < 0) {
rc = MPI_ERR_COUNT;
} else if (type == MPI_DATATYPE_NULL) {
rc = MPI_ERR_TYPE;
} else if (tag < 0 || tag > mca_pml.pml_max_tag) {
rc = MPI_ERR_TAG;
} else if (ompi_comm_peer_invalid(comm, dest) &&
(MPI_PROC_NULL != dest)) {
rc = MPI_ERR_RANK;
} else if (request == NULL) {
rc = MPI_ERR_REQUEST;
}
OMPI_ERRHANDLER_CHECK(rc, comm, rc, FUNC_NAME);
}
if (MPI_PROC_NULL == dest) {
*request = OBJ_NEW(ompi_request_t);
/* Other fields were initialized by the constructor for
ompi_request_t */
(*request)->req_type = OMPI_REQUEST_NOOP;
(*request)->req_status = ompi_request_empty.req_status;
(*request)->req_complete = true;
(*request)->req_state = OMPI_REQUEST_INACTIVE;
(*request)->req_persistent = true;
(*request)->req_free = ompi_request_persistent_proc_null_free;
return MPI_SUCCESS;
}
rc = MCA_PML_CALL(isend_init(buf,count,type,dest,tag,
MCA_PML_BASE_SEND_SYNCHRONOUS,comm,request));
OMPI_ERRHANDLER_RETURN(rc, comm, rc, FUNC_NAME);
}
/*
* Linear functions are copied from the basic coll module. For
* some small number of nodes and/or small data sizes they are just as
* fast as tuned/tree based segmenting operations and as such may be
* selected by the decision functions. These are copied into this module
* due to the way we select modules in V1. i.e. in V2 we will handle this
* differently and so will not have to duplicate code.
* GEF Oct05 after asking Jeff.
*/
int
ompi_coll_tuned_alltoallv_intra_basic_linear(void *sbuf, int *scounts, int *sdisps,
struct ompi_datatype_t *sdtype,
void *rbuf, int *rcounts, int *rdisps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, size, rank, err;
char *psnd, *prcv;
int nreqs;
ptrdiff_t sext, rext;
MPI_Request *preq;
mca_coll_tuned_module_t *tuned_module = (mca_coll_tuned_module_t*) module;
mca_coll_tuned_comm_t *data = tuned_module->tuned_data;
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,
"coll:tuned:alltoallv_intra_basic_linear rank %d", rank));
ompi_ddt_type_extent(sdtype, &sext);
ompi_ddt_type_extent(rdtype, &rext);
/* Simple optimization - handle send to self first */
psnd = ((char *) sbuf) + (sdisps[rank] * sext);
prcv = ((char *) rbuf) + (rdisps[rank] * rext);
if (0 != scounts[rank]) {
err = ompi_ddt_sndrcv(psnd, scounts[rank], sdtype,
prcv, rcounts[rank], rdtype);
if (MPI_SUCCESS != err) {
return err;
}
}
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Now, initiate all send/recv to/from others. */
nreqs = 0;
preq = data->mcct_reqs;
/* Post all receives first */
for (i = 0; i < size; ++i) {
if (i == rank || 0 == rcounts[i]) {
continue;
}
prcv = ((char *) rbuf) + (rdisps[i] * rext);
err = MCA_PML_CALL(irecv_init(prcv, rcounts[i], rdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV, comm,
preq++));
++nreqs;
if (MPI_SUCCESS != err) {
ompi_coll_tuned_free_reqs(data->mcct_reqs, nreqs);
return err;
}
}
/* Now post all sends */
for (i = 0; i < size; ++i) {
if (i == rank || 0 == scounts[i]) {
continue;
}
psnd = ((char *) sbuf) + (sdisps[i] * sext);
err = MCA_PML_CALL(isend_init(psnd, scounts[i], sdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV,
MCA_PML_BASE_SEND_STANDARD, comm,
preq++));
++nreqs;
if (MPI_SUCCESS != err) {
ompi_coll_tuned_free_reqs(data->mcct_reqs, nreqs);
return err;
}
}
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, data->mcct_reqs));
/* Wait for them all. If there's an error, note that we don't care
* what the error was -- just that there *was* an error. The PML
* will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return the
* error after we free everything. */
err = ompi_request_wait_all(nreqs, data->mcct_reqs,
MPI_STATUSES_IGNORE);
/* Free the requests. */
ompi_coll_tuned_free_reqs(data->mcct_reqs, nreqs);
return err;
}
int ompi_coll_tuned_alltoall_intra_basic_linear(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm)
{
int i;
int rank;
int size;
int err;
int nreqs;
char *psnd;
char *prcv;
MPI_Aint lb;
MPI_Aint sndinc;
MPI_Aint rcvinc;
ompi_request_t **req;
ompi_request_t **sreq;
ompi_request_t **rreq;
/* Initialize. */
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_tuned_stream,"ompi_coll_tuned_alltoall_intra_basic_linear rank %d", rank));
err = ompi_ddt_get_extent(sdtype, &lb, &sndinc);
if (OMPI_SUCCESS != err) {
return err;
}
sndinc *= scount;
err = ompi_ddt_get_extent(rdtype, &lb, &rcvinc);
if (OMPI_SUCCESS != err) {
return err;
}
rcvinc *= rcount;
/* simple optimization */
psnd = ((char *) sbuf) + (rank * sndinc);
prcv = ((char *) rbuf) + (rank * rcvinc);
err = ompi_ddt_sndrcv(psnd, scount, sdtype, prcv, rcount, rdtype);
if (MPI_SUCCESS != err) {
return err;
}
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Initiate all send/recv to/from others. */
req = rreq = comm->c_coll_basic_data->mcct_reqs;
sreq = rreq + size - 1;
prcv = (char *) rbuf;
psnd = (char *) sbuf;
/* Post all receives first -- a simple optimization */
for (nreqs = 0, i = (rank + 1) % size; i != rank;
i = (i + 1) % size, ++rreq, ++nreqs) {
err =
MCA_PML_CALL(irecv_init
(prcv + (i * rcvinc), rcount, rdtype, i,
MCA_COLL_BASE_TAG_ALLTOALL, comm, rreq));
if (MPI_SUCCESS != err) {
ompi_coll_tuned_free_reqs(req, rreq - req);
return err;
}
}
/* Now post all sends in reverse order
- We would like to minimize the search time through message queue
when messages actually arrive in the order in which they were posted.
*/
for (nreqs = 0, i = (rank + size - 1) % size; i != rank;
i = (i + size - 1) % size, ++sreq, ++nreqs) {
err =
MCA_PML_CALL(isend_init
(psnd + (i * sndinc), scount, sdtype, i,
MCA_COLL_BASE_TAG_ALLTOALL,
MCA_PML_BASE_SEND_STANDARD, comm, sreq));
if (MPI_SUCCESS != err) {
ompi_coll_tuned_free_reqs(req, sreq - req);
return err;
}
}
nreqs = (size - 1) * 2;
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, req));
/* Wait for them all. If there's an error, note that we don't
* care what the error was -- just that there *was* an error. The
* PML will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return
* the error after we free everything. */
err = ompi_request_wait_all(nreqs, req, MPI_STATUSES_IGNORE);
/* Free the reqs */
ompi_coll_tuned_free_reqs(req, nreqs);
/* All done */
return err;
}
/**
* Linear functions are copied from the basic coll module. For
* some small number of nodes and/or small data sizes they are just as
* fast as base/tree based segmenting operations and as such may be
* selected by the decision functions. These are copied into this module
* due to the way we select modules in V1. i.e. in V2 we will handle this
* differently and so will not have to duplicate code.
*/
int
ompi_coll_base_alltoallv_intra_basic_linear(const void *sbuf, const int *scounts, const int *sdisps,
struct ompi_datatype_t *sdtype,
void *rbuf, const int *rcounts, const int *rdisps,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, size, rank, err, nreqs;
char *psnd, *prcv;
ptrdiff_t sext, rext;
ompi_request_t **preq, **reqs;
mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module;
mca_coll_base_comm_t *data = base_module->base_data;
if (MPI_IN_PLACE == sbuf) {
return mca_coll_base_alltoallv_intra_basic_inplace (rbuf, rcounts, rdisps,
rdtype, comm, module);
}
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"coll:base:alltoallv_intra_basic_linear rank %d", rank));
ompi_datatype_type_extent(sdtype, &sext);
ompi_datatype_type_extent(rdtype, &rext);
/* Simple optimization - handle send to self first */
psnd = ((char *) sbuf) + (ptrdiff_t)sdisps[rank] * sext;
prcv = ((char *) rbuf) + (ptrdiff_t)rdisps[rank] * rext;
if (0 != scounts[rank]) {
err = ompi_datatype_sndrcv(psnd, scounts[rank], sdtype,
prcv, rcounts[rank], rdtype);
if (MPI_SUCCESS != err) {
return err;
}
}
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Now, initiate all send/recv to/from others. */
nreqs = 0;
reqs = preq = coll_base_comm_get_reqs(data, 2 * size);
if( NULL == reqs ) { err = OMPI_ERR_OUT_OF_RESOURCE; goto err_hndl; }
/* Post all receives first */
for (i = 0; i < size; ++i) {
if (i == rank || 0 == rcounts[i]) {
continue;
}
++nreqs;
prcv = ((char *) rbuf) + (ptrdiff_t)rdisps[i] * rext;
err = MCA_PML_CALL(irecv_init(prcv, rcounts[i], rdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV, comm,
preq++));
if (MPI_SUCCESS != err) { goto err_hndl; }
}
/* Now post all sends */
for (i = 0; i < size; ++i) {
if (i == rank || 0 == scounts[i]) {
continue;
}
++nreqs;
psnd = ((char *) sbuf) + (ptrdiff_t)sdisps[i] * sext;
err = MCA_PML_CALL(isend_init(psnd, scounts[i], sdtype,
i, MCA_COLL_BASE_TAG_ALLTOALLV,
MCA_PML_BASE_SEND_STANDARD, comm,
preq++));
if (MPI_SUCCESS != err) { goto err_hndl; }
}
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, reqs));
/* Wait for them all. If there's an error, note that we don't care
* what the error was -- just that there *was* an error. The PML
* will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return the
* error after we free everything. */
err = ompi_request_wait_all(nreqs, reqs, MPI_STATUSES_IGNORE);
err_hndl:
/* Free the requests in all cases as they are persistent */
ompi_coll_base_free_reqs(reqs, nreqs);
return err;
}
/*
* bcast_lin_intra
*
* Function: - broadcast using O(N) algorithm
* Accepts: - same arguments as MPI_Bcast()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_basic_bcast_lin_intra(void *buff, int count,
struct ompi_datatype_t *datatype, int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i;
int size;
int rank;
int err;
ompi_request_t **preq;
mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module;
ompi_request_t **reqs = basic_module->mccb_reqs;
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
/* Non-root receive the data. */
if (rank != root) {
return MCA_PML_CALL(recv(buff, count, datatype, root,
MCA_COLL_BASE_TAG_BCAST, comm,
MPI_STATUS_IGNORE));
}
/* Root sends data to all others. */
for (i = 0, preq = reqs; i < size; ++i) {
if (i == rank) {
continue;
}
err = MCA_PML_CALL(isend_init(buff, count, datatype, i,
MCA_COLL_BASE_TAG_BCAST,
MCA_PML_BASE_SEND_STANDARD,
comm, preq++));
if (MPI_SUCCESS != err) {
return err;
}
}
--i;
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(i, reqs));
/* Wait for them all. If there's an error, note that we don't
* care what the error was -- just that there *was* an error. The
* PML will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return
* the error after we free everything. */
err = ompi_request_wait_all(i, reqs, MPI_STATUSES_IGNORE);
/* Free the reqs */
mca_coll_basic_free_reqs(reqs, i);
/* All done */
return err;
}
/*
* bcast_log_intra
*
* Function: - broadcast using O(log(N)) algorithm
* Accepts: - same arguments as MPI_Bcast()
* Returns: - MPI_SUCCESS or error code
*/
int
mca_coll_basic_bcast_log_intra(void *buff, int count,
struct ompi_datatype_t *datatype, int root,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i;
int size;
int rank;
int vrank;
int peer;
int dim;
int hibit;
int mask;
int err;
int nreqs;
ompi_request_t **preq;
mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module;
ompi_request_t **reqs = basic_module->mccb_reqs;
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
vrank = (rank + size - root) % size;
dim = comm->c_cube_dim;
hibit = opal_hibit(vrank, dim);
--dim;
/* Receive data from parent in the tree. */
if (vrank > 0) {
peer = ((vrank & ~(1 << hibit)) + root) % size;
err = MCA_PML_CALL(recv(buff, count, datatype, peer,
MCA_COLL_BASE_TAG_BCAST,
comm, MPI_STATUS_IGNORE));
if (MPI_SUCCESS != err) {
return err;
}
}
/* Send data to the children. */
err = MPI_SUCCESS;
preq = reqs;
nreqs = 0;
for (i = hibit + 1, mask = 1 << i; i <= dim; ++i, mask <<= 1) {
peer = vrank | mask;
if (peer < size) {
peer = (peer + root) % size;
++nreqs;
err = MCA_PML_CALL(isend_init(buff, count, datatype, peer,
MCA_COLL_BASE_TAG_BCAST,
MCA_PML_BASE_SEND_STANDARD,
comm, preq++));
if (MPI_SUCCESS != err) {
mca_coll_basic_free_reqs(reqs, nreqs);
return err;
}
}
}
/* Start and wait on all requests. */
if (nreqs > 0) {
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, reqs));
/* Wait for them all. If there's an error, note that we don't
* care what the error was -- just that there *was* an error.
* The PML will finish all requests, even if one or more of them
* fail. i.e., by the end of this call, all the requests are
* free-able. So free them anyway -- even if there was an
* error, and return the error after we free everything. */
err = ompi_request_wait_all(nreqs, reqs, MPI_STATUSES_IGNORE);
/* Free the reqs */
mca_coll_basic_free_reqs(reqs, nreqs);
}
/* All done */
return err;
}
/*
* alltoallw_inter
*
* Function: - MPI_Alltoallw
* Accepts: - same as MPI_Alltoallw()
* Returns: - MPI_SUCCESS or an MPI error code
*/
int
mca_coll_basic_alltoallw_inter(void *sbuf, int *scounts, int *sdisps,
struct ompi_datatype_t **sdtypes,
void *rbuf, int *rcounts, int *rdisps,
struct ompi_datatype_t **rdtypes,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i;
int size;
int err;
char *psnd;
char *prcv;
int nreqs;
MPI_Request *preq;
mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module;
/* Initialize. */
size = ompi_comm_remote_size(comm);
/* Initiate all send/recv to/from others. */
nreqs = size * 2;
preq = basic_module->mccb_reqs;
/* Post all receives first -- a simple optimization */
for (i = 0; i < size; ++i) {
prcv = ((char *) rbuf) + rdisps[i];
err = MCA_PML_CALL(irecv_init(prcv, rcounts[i], rdtypes[i],
i, MCA_COLL_BASE_TAG_ALLTOALLW,
comm, preq++));
if (OMPI_SUCCESS != err) {
mca_coll_basic_free_reqs(basic_module->mccb_reqs,
nreqs);
return err;
}
}
/* Now post all sends */
for (i = 0; i < size; ++i) {
psnd = ((char *) sbuf) + sdisps[i];
err = MCA_PML_CALL(isend_init(psnd, scounts[i], sdtypes[i],
i, MCA_COLL_BASE_TAG_ALLTOALLW,
MCA_PML_BASE_SEND_STANDARD, comm,
preq++));
if (OMPI_SUCCESS != err) {
mca_coll_basic_free_reqs(basic_module->mccb_reqs,
nreqs);
return err;
}
}
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, basic_module->mccb_reqs));
/* Wait for them all. If there's an error, note that we don't care
* what the error was -- just that there *was* an error. The PML
* will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return the
* error after we free everything. */
err = ompi_request_wait_all(nreqs, basic_module->mccb_reqs,
MPI_STATUSES_IGNORE);
/* Free the requests. */
mca_coll_basic_free_reqs(basic_module->mccb_reqs, nreqs);
/* All done */
return err;
}
int ompi_coll_base_alltoall_intra_basic_linear(const void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void* rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i, rank, size, err, line;
int nreqs = 0;
char *psnd, *prcv;
MPI_Aint lb, sndinc, rcvinc;
ompi_request_t **req, **sreq, **rreq;
mca_coll_base_module_t *base_module = (mca_coll_base_module_t*) module;
mca_coll_base_comm_t *data = base_module->base_data;
if (MPI_IN_PLACE == sbuf) {
return mca_coll_base_alltoall_intra_basic_inplace (rbuf, rcount, rdtype,
comm, module);
}
/* Initialize. */
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
OPAL_OUTPUT((ompi_coll_base_framework.framework_output,
"ompi_coll_base_alltoall_intra_basic_linear rank %d", rank));
err = ompi_datatype_get_extent(sdtype, &lb, &sndinc);
if (OMPI_SUCCESS != err) {
return err;
}
sndinc *= scount;
err = ompi_datatype_get_extent(rdtype, &lb, &rcvinc);
if (OMPI_SUCCESS != err) {
return err;
}
rcvinc *= rcount;
/* simple optimization */
psnd = ((char *) sbuf) + (ptrdiff_t)rank * sndinc;
prcv = ((char *) rbuf) + (ptrdiff_t)rank * rcvinc;
err = ompi_datatype_sndrcv(psnd, scount, sdtype, prcv, rcount, rdtype);
if (MPI_SUCCESS != err) {
return err;
}
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Initiate all send/recv to/from others. */
req = rreq = coll_base_comm_get_reqs(data, (size - 1) * 2);
if (NULL == req) { err = OMPI_ERR_OUT_OF_RESOURCE; line = __LINE__; goto err_hndl; }
prcv = (char *) rbuf;
psnd = (char *) sbuf;
/* Post all receives first -- a simple optimization */
for (nreqs = 0, i = (rank + 1) % size; i != rank;
i = (i + 1) % size, ++rreq) {
nreqs++;
err = MCA_PML_CALL(irecv_init
(prcv + (ptrdiff_t)i * rcvinc, rcount, rdtype, i,
MCA_COLL_BASE_TAG_ALLTOALL, comm, rreq));
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
}
/* Now post all sends in reverse order
- We would like to minimize the search time through message queue
when messages actually arrive in the order in which they were posted.
*/
sreq = rreq;
for (i = (rank + size - 1) % size; i != rank;
i = (i + size - 1) % size, ++sreq) {
nreqs++;
err = MCA_PML_CALL(isend_init
(psnd + (ptrdiff_t)i * sndinc, scount, sdtype, i,
MCA_COLL_BASE_TAG_ALLTOALL,
MCA_PML_BASE_SEND_STANDARD, comm, sreq));
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
}
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, req));
/* Wait for them all. If there's an error, note that we don't
* care what the error was -- just that there *was* an error. The
* PML will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return
* the error after we free everything. */
err = ompi_request_wait_all(nreqs, req, MPI_STATUSES_IGNORE);
if (MPI_SUCCESS != err) { line = __LINE__; goto err_hndl; }
err_hndl:
if( MPI_SUCCESS != err ) {
OPAL_OUTPUT( (ompi_coll_base_framework.framework_output,"%s:%4d\tError occurred %d, rank %2d",
__FILE__, line, err, rank) );
(void)line; // silence compiler warning
}
/* Free the reqs in all cases as they are persistent requests */
ompi_coll_base_free_reqs(req, nreqs);
/* All done */
return err;
}
/*
* alltoall_intra
*
* Function: - MPI_Alltoall
* Accepts: - same as MPI_Alltoall()
* Returns: - MPI_SUCCESS or an MPI error code
*/
int
mca_coll_basic_alltoall_intra(void *sbuf, int scount,
struct ompi_datatype_t *sdtype,
void *rbuf, int rcount,
struct ompi_datatype_t *rdtype,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i;
int rank;
int size;
int err;
int nreqs;
char *psnd;
char *prcv;
MPI_Aint lb;
MPI_Aint sndinc;
MPI_Aint rcvinc;
ompi_request_t **req;
ompi_request_t **sreq;
ompi_request_t **rreq;
mca_coll_basic_module_t *basic_module = (mca_coll_basic_module_t*) module;
/* Initialize. */
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
err = ompi_datatype_get_extent(sdtype, &lb, &sndinc);
if (OMPI_SUCCESS != err) {
return err;
}
sndinc *= scount;
err = ompi_datatype_get_extent(rdtype, &lb, &rcvinc);
if (OMPI_SUCCESS != err) {
return err;
}
rcvinc *= rcount;
/* simple optimization */
psnd = ((char *) sbuf) + (rank * sndinc);
prcv = ((char *) rbuf) + (rank * rcvinc);
err = ompi_datatype_sndrcv(psnd, scount, sdtype, prcv, rcount, rdtype);
if (MPI_SUCCESS != err) {
return err;
}
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Initiate all send/recv to/from others. */
req = rreq = basic_module->mccb_reqs;
sreq = rreq + size - 1;
prcv = (char *) rbuf;
psnd = (char *) sbuf;
/* Post all receives first -- a simple optimization */
for (nreqs = 0, i = (rank + 1) % size; i != rank; i = (i + 1) % size, ++rreq, ++nreqs) {
err =
MCA_PML_CALL(irecv_init
(prcv + (i * rcvinc), rcount, rdtype, i,
MCA_COLL_BASE_TAG_ALLTOALL, comm, rreq));
if (MPI_SUCCESS != err) {
mca_coll_basic_free_reqs(req, nreqs);
return err;
}
}
/* Now post all sends */
for (nreqs = 0, i = (rank + 1) % size; i != rank; i = (i + 1) % size, ++sreq, ++nreqs) {
err =
MCA_PML_CALL(isend_init
(psnd + (i * sndinc), scount, sdtype, i,
MCA_COLL_BASE_TAG_ALLTOALL,
MCA_PML_BASE_SEND_STANDARD, comm, sreq));
if (MPI_SUCCESS != err) {
mca_coll_basic_free_reqs(req, nreqs);
return err;
}
}
nreqs = (size - 1) * 2;
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, req));
/* Wait for them all. If there's an error, note that we don't
* care what the error was -- just that there *was* an error. The
* PML will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return
* the error after we free everything. */
err = ompi_request_wait_all(nreqs, req, MPI_STATUSES_IGNORE);
/* Free the reqs */
mca_coll_basic_free_reqs(req, nreqs);
/* All done */
return err;
}
/*
* alltoallw_intra
*
* Function: - MPI_Alltoallw
* Accepts: - same as MPI_Alltoallw()
* Returns: - MPI_SUCCESS or an MPI error code
*/
int
mca_coll_basic_alltoallw_intra(const void *sbuf, const int *scounts, const int *sdisps,
struct ompi_datatype_t * const *sdtypes,
void *rbuf, const int *rcounts, const int *rdisps,
struct ompi_datatype_t * const *rdtypes,
struct ompi_communicator_t *comm,
mca_coll_base_module_t *module)
{
int i;
int size;
int rank;
int err;
char *psnd;
char *prcv;
int nreqs;
MPI_Request *preq, *reqs;
/* Initialize. */
if (MPI_IN_PLACE == sbuf) {
return mca_coll_basic_alltoallw_intra_inplace (rbuf, rcounts, rdisps,
rdtypes, comm, module);
}
size = ompi_comm_size(comm);
rank = ompi_comm_rank(comm);
/* simple optimization */
psnd = ((char *) sbuf) + sdisps[rank];
prcv = ((char *) rbuf) + rdisps[rank];
err = ompi_datatype_sndrcv(psnd, scounts[rank], sdtypes[rank],
prcv, rcounts[rank], rdtypes[rank]);
if (MPI_SUCCESS != err) {
return err;
}
/* If only one process, we're done. */
if (1 == size) {
return MPI_SUCCESS;
}
/* Initiate all send/recv to/from others. */
nreqs = 0;
reqs = preq = coll_base_comm_get_reqs(module->base_data, 2 * size);
/* Post all receives first -- a simple optimization */
for (i = 0; i < size; ++i) {
size_t msg_size;
ompi_datatype_type_size(rdtypes[i], &msg_size);
msg_size *= rcounts[i];
if (i == rank || 0 == msg_size)
continue;
prcv = ((char *) rbuf) + rdisps[i];
err = MCA_PML_CALL(irecv_init(prcv, rcounts[i], rdtypes[i],
i, MCA_COLL_BASE_TAG_ALLTOALLW, comm,
preq++));
++nreqs;
if (MPI_SUCCESS != err) {
ompi_coll_base_free_reqs(reqs, nreqs);
return err;
}
}
/* Now post all sends */
for (i = 0; i < size; ++i) {
size_t msg_size;
ompi_datatype_type_size(sdtypes[i], &msg_size);
msg_size *= scounts[i];
if (i == rank || 0 == msg_size)
continue;
psnd = ((char *) sbuf) + sdisps[i];
err = MCA_PML_CALL(isend_init(psnd, scounts[i], sdtypes[i],
i, MCA_COLL_BASE_TAG_ALLTOALLW,
MCA_PML_BASE_SEND_STANDARD, comm,
preq++));
++nreqs;
if (MPI_SUCCESS != err) {
ompi_coll_base_free_reqs(reqs, nreqs);
return err;
}
}
/* Start your engines. This will never return an error. */
MCA_PML_CALL(start(nreqs, reqs));
/* Wait for them all. If there's an error, note that we don't care
* what the error was -- just that there *was* an error. The PML
* will finish all requests, even if one or more of them fail.
* i.e., by the end of this call, all the requests are free-able.
* So free them anyway -- even if there was an error, and return the
* error after we free everything. */
err = ompi_request_wait_all(nreqs, reqs, MPI_STATUSES_IGNORE);
/* Free the requests in all cases as they are persistent */
ompi_coll_base_free_reqs(reqs, nreqs);
/* All done */
return err;
}
