sys/utils/mpits.c

#include <petscsys.h> /*I  "petscsys.h"  I*/
#include <petsc/private/petscimpl.h>

PetscLogEvent PETSC_BuildTwoSided;
PetscLogEvent PETSC_BuildTwoSidedF;

const char *const PetscBuildTwoSidedTypes[] = {"ALLREDUCE", "IBARRIER", "REDSCATTER", "PetscBuildTwoSidedType", "PETSC_BUILDTWOSIDED_", NULL};

static PetscBuildTwoSidedType _twosided_type = PETSC_BUILDTWOSIDED_NOTSET;

/*@
   PetscCommBuildTwoSidedSetType - set algorithm to use when building two-sided communication

   Logically Collective

   Input Parameters:
+  comm - `PETSC_COMM_WORLD`
-  twosided - algorithm to use in subsequent calls to `PetscCommBuildTwoSided()`

   Level: developer

   Note:
   This option is currently global, but could be made per-communicator.

.seealso: `PetscCommBuildTwoSided()`, `PetscCommBuildTwoSidedGetType()`, `PetscBuildTwoSidedType`
@*/
PetscErrorCode PetscCommBuildTwoSidedSetType(MPI_Comm comm, PetscBuildTwoSidedType twosided)
{
  PetscFunctionBegin;
  if (PetscDefined(USE_DEBUG)) { /* We don't have a PetscObject so can't use PetscValidLogicalCollectiveEnum */
    PetscMPIInt b1[2], b2[2];
    b1[0] = -(PetscMPIInt)twosided;
    b1[1] = (PetscMPIInt)twosided;
    PetscCall(MPIU_Allreduce(b1, b2, 2, MPI_INT, MPI_MAX, comm));
    PetscCheck(-b2[0] == b2[1], comm, PETSC_ERR_ARG_WRONG, "Enum value must be same on all processes");
  }
  _twosided_type = twosided;
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@
   PetscCommBuildTwoSidedGetType - get algorithm used when building two-sided communication

   Logically Collective

   Output Parameters:
+  comm - communicator on which to query algorithm
-  twosided - algorithm to use for `PetscCommBuildTwoSided()`

   Level: developer

.seealso: `PetscCommBuildTwoSided()`, `PetscCommBuildTwoSidedSetType()`, `PetscBuildTwoSidedType`
@*/
PetscErrorCode PetscCommBuildTwoSidedGetType(MPI_Comm comm, PetscBuildTwoSidedType *twosided)
{
  PetscMPIInt size;

  PetscFunctionBegin;
  *twosided = PETSC_BUILDTWOSIDED_NOTSET;
  if (_twosided_type == PETSC_BUILDTWOSIDED_NOTSET) {
    PetscCallMPI(MPI_Comm_size(comm, &size));
    _twosided_type = PETSC_BUILDTWOSIDED_ALLREDUCE; /* default for small comms, see https://gitlab.com/petsc/petsc/-/merge_requests/2611 */
#if defined(PETSC_HAVE_MPI_NONBLOCKING_COLLECTIVES)
    if (size > 1024) _twosided_type = PETSC_BUILDTWOSIDED_IBARRIER;
#endif
    PetscCall(PetscOptionsGetEnum(NULL, NULL, "-build_twosided", PetscBuildTwoSidedTypes, (PetscEnum *)&_twosided_type, NULL));
  }
  *twosided = _twosided_type;
  PetscFunctionReturn(PETSC_SUCCESS);
}

#if defined(PETSC_HAVE_MPI_NONBLOCKING_COLLECTIVES)
static PetscErrorCode PetscCommBuildTwoSided_Ibarrier(MPI_Comm comm, PetscMPIInt count, MPI_Datatype dtype, PetscMPIInt nto, const PetscMPIInt *toranks, const void *todata, PetscMPIInt *nfrom, PetscMPIInt **fromranks, void *fromdata)
{
  PetscMPIInt    nrecvs, tag, done, i;
  MPI_Aint       lb, unitbytes;
  char          *tdata;
  MPI_Request   *sendreqs, barrier;
  PetscSegBuffer segrank, segdata;
  PetscBool      barrier_started;

  PetscFunctionBegin;
  PetscCall(PetscCommDuplicate(comm, &comm, &tag));
  PetscCallMPI(MPI_Type_get_extent(dtype, &lb, &unitbytes));
  PetscCheck(lb == 0, comm, PETSC_ERR_SUP, "Datatype with nonzero lower bound %ld", (long)lb);
  tdata = (char *)todata;
  PetscCall(PetscMalloc1(nto, &sendreqs));
  for (i = 0; i < nto; i++) PetscCallMPI(MPI_Issend((void *)(tdata + count * unitbytes * i), count, dtype, toranks[i], tag, comm, sendreqs + i));
  PetscCall(PetscSegBufferCreate(sizeof(PetscMPIInt), 4, &segrank));
  PetscCall(PetscSegBufferCreate(unitbytes, 4 * count, &segdata));

  nrecvs  = 0;
  barrier = MPI_REQUEST_NULL;
  /* MPICH-3.2 sometimes does not create a request in some "optimized" cases.  This is arguably a standard violation,
   * but we need to work around it. */
  barrier_started = PETSC_FALSE;
  for (done = 0; !done;) {
    PetscMPIInt flag;
    MPI_Status  status;
    PetscCallMPI(MPI_Iprobe(MPI_ANY_SOURCE, tag, comm, &flag, &status));
    if (flag) { /* incoming message */
      PetscMPIInt *recvrank;
      void        *buf;
      PetscCall(PetscSegBufferGet(segrank, 1, &recvrank));
      PetscCall(PetscSegBufferGet(segdata, count, &buf));
      *recvrank = status.MPI_SOURCE;
      PetscCallMPI(MPI_Recv(buf, count, dtype, status.MPI_SOURCE, tag, comm, MPI_STATUS_IGNORE));
      nrecvs++;
    }
    if (!barrier_started) {
      PetscMPIInt sent, nsends;
      PetscCall(PetscMPIIntCast(nto, &nsends));
      PetscCallMPI(MPI_Testall(nsends, sendreqs, &sent, MPI_STATUSES_IGNORE));
      if (sent) {
        PetscCallMPI(MPI_Ibarrier(comm, &barrier));
        barrier_started = PETSC_TRUE;
        PetscCall(PetscFree(sendreqs));
      }
    } else {
      PetscCallMPI(MPI_Test(&barrier, &done, MPI_STATUS_IGNORE));
    }
  }
  *nfrom = nrecvs;
  PetscCall(PetscSegBufferExtractAlloc(segrank, fromranks));
  PetscCall(PetscSegBufferDestroy(&segrank));
  PetscCall(PetscSegBufferExtractAlloc(segdata, fromdata));
  PetscCall(PetscSegBufferDestroy(&segdata));
  PetscCall(PetscCommDestroy(&comm));
  PetscFunctionReturn(PETSC_SUCCESS);
}
#endif

static PetscErrorCode PetscCommBuildTwoSided_Allreduce(MPI_Comm comm, PetscMPIInt count, MPI_Datatype dtype, PetscMPIInt nto, const PetscMPIInt *toranks, const void *todata, PetscMPIInt *nfrom, PetscMPIInt **fromranks, void *fromdata)
{
  PetscMPIInt       size, rank, *iflags, nrecvs, tag, *franks, i, flg;
  MPI_Aint          lb, unitbytes;
  char             *tdata, *fdata;
  MPI_Request      *reqs, *sendreqs;
  MPI_Status       *statuses;
  PetscCommCounter *counter;

  PetscFunctionBegin;
  PetscCallMPI(MPI_Comm_size(comm, &size));
  PetscCallMPI(MPI_Comm_rank(comm, &rank));
  PetscCall(PetscCommDuplicate(comm, &comm, &tag));
  PetscCallMPI(MPI_Comm_get_attr(comm, Petsc_Counter_keyval, &counter, &flg));
  PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Inner PETSc communicator does not have its tag/name counter attribute set");
  if (!counter->iflags) {
    PetscCall(PetscCalloc1(size, &counter->iflags));
    iflags = counter->iflags;
  } else {
    iflags = counter->iflags;
    PetscCall(PetscArrayzero(iflags, size));
  }
  for (i = 0; i < nto; i++) iflags[toranks[i]] = 1;
  PetscCall(MPIU_Allreduce(MPI_IN_PLACE, iflags, size, MPI_INT, MPI_SUM, comm));
  nrecvs = iflags[rank];
  PetscCallMPI(MPI_Type_get_extent(dtype, &lb, &unitbytes));
  PetscCheck(lb == 0, comm, PETSC_ERR_SUP, "Datatype with nonzero lower bound %ld", (long)lb);
  PetscCall(PetscMalloc(nrecvs * count * unitbytes, &fdata));
  tdata = (char *)todata;
  PetscCall(PetscMalloc2(nto + nrecvs, &reqs, nto + nrecvs, &statuses));
  sendreqs = reqs + nrecvs;
  for (i = 0; i < nrecvs; i++) PetscCallMPI(MPI_Irecv((void *)(fdata + count * unitbytes * i), count, dtype, MPI_ANY_SOURCE, tag, comm, reqs + i));
  for (i = 0; i < nto; i++) PetscCallMPI(MPI_Isend((void *)(tdata + count * unitbytes * i), count, dtype, toranks[i], tag, comm, sendreqs + i));
  PetscCallMPI(MPI_Waitall(nto + nrecvs, reqs, statuses));
  PetscCall(PetscMalloc1(nrecvs, &franks));
  for (i = 0; i < nrecvs; i++) franks[i] = statuses[i].MPI_SOURCE;
  PetscCall(PetscFree2(reqs, statuses));
  PetscCall(PetscCommDestroy(&comm));

  *nfrom             = nrecvs;
  *fromranks         = franks;
  *(void **)fromdata = fdata;
  PetscFunctionReturn(PETSC_SUCCESS);
}

#if defined(PETSC_HAVE_MPI_REDUCE_SCATTER_BLOCK)
static PetscErrorCode PetscCommBuildTwoSided_RedScatter(MPI_Comm comm, PetscMPIInt count, MPI_Datatype dtype, PetscMPIInt nto, const PetscMPIInt *toranks, const void *todata, PetscMPIInt *nfrom, PetscMPIInt **fromranks, void *fromdata)
{
  PetscMPIInt       size, *iflags, nrecvs, tag, *franks, i, flg;
  MPI_Aint          lb, unitbytes;
  char             *tdata, *fdata;
  MPI_Request      *reqs, *sendreqs;
  MPI_Status       *statuses;
  PetscCommCounter *counter;

  PetscFunctionBegin;
  PetscCallMPI(MPI_Comm_size(comm, &size));
  PetscCall(PetscCommDuplicate(comm, &comm, &tag));
  PetscCallMPI(MPI_Comm_get_attr(comm, Petsc_Counter_keyval, &counter, &flg));
  PetscCheck(flg, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Inner PETSc communicator does not have its tag/name counter attribute set");
  if (!counter->iflags) {
    PetscCall(PetscCalloc1(size, &counter->iflags));
    iflags = counter->iflags;
  } else {
    iflags = counter->iflags;
    PetscCall(PetscArrayzero(iflags, size));
  }
  for (i = 0; i < nto; i++) iflags[toranks[i]] = 1;
  PetscCallMPI(MPI_Reduce_scatter_block(iflags, &nrecvs, 1, MPI_INT, MPI_SUM, comm));
  PetscCallMPI(MPI_Type_get_extent(dtype, &lb, &unitbytes));
  PetscCheck(lb == 0, comm, PETSC_ERR_SUP, "Datatype with nonzero lower bound %ld", (long)lb);
  PetscCall(PetscMalloc(nrecvs * count * unitbytes, &fdata));
  tdata = (char *)todata;
  PetscCall(PetscMalloc2(nto + nrecvs, &reqs, nto + nrecvs, &statuses));
  sendreqs = reqs + nrecvs;
  for (i = 0; i < nrecvs; i++) PetscCallMPI(MPI_Irecv((void *)(fdata + count * unitbytes * i), count, dtype, MPI_ANY_SOURCE, tag, comm, reqs + i));
  for (i = 0; i < nto; i++) PetscCallMPI(MPI_Isend((void *)(tdata + count * unitbytes * i), count, dtype, toranks[i], tag, comm, sendreqs + i));
  PetscCallMPI(MPI_Waitall(nto + nrecvs, reqs, statuses));
  PetscCall(PetscMalloc1(nrecvs, &franks));
  for (i = 0; i < nrecvs; i++) franks[i] = statuses[i].MPI_SOURCE;
  PetscCall(PetscFree2(reqs, statuses));
  PetscCall(PetscCommDestroy(&comm));

  *nfrom             = nrecvs;
  *fromranks         = franks;
  *(void **)fromdata = fdata;
  PetscFunctionReturn(PETSC_SUCCESS);
}
#endif

/*@C
   PetscCommBuildTwoSided - discovers communicating ranks given one-sided information, moving constant-sized data in the process (often message lengths)

   Collective

   Input Parameters:
+  comm - communicator
.  count - number of entries to send/receive (must match on all ranks)
.  dtype - datatype to send/receive from each rank (must match on all ranks)
.  nto - number of ranks to send data to
.  toranks - ranks to send to (array of length nto)
-  todata - data to send to each rank (packed)

   Output Parameters:
+  nfrom - number of ranks receiving messages from
.  fromranks - ranks receiving messages from (length nfrom; caller should `PetscFree()`)
-  fromdata - packed data from each rank, each with count entries of type dtype (length nfrom, caller responsible for `PetscFree()`)

   Level: developer

   Options Database Key:
.  -build_twosided <allreduce|ibarrier|redscatter> - algorithm to set up two-sided communication. Default is allreduce for communicators with <= 1024 ranks, otherwise ibarrier.

   Notes:
   This memory-scalable interface is an alternative to calling `PetscGatherNumberOfMessages()` and
   `PetscGatherMessageLengths()`, possibly with a subsequent round of communication to send other constant-size data.

   Basic data types as well as contiguous types are supported, but non-contiguous (e.g., strided) types are not.

   References:
.  * - Hoefler, Siebert and Lumsdaine, The MPI_Ibarrier implementation uses the algorithm in
   Scalable communication protocols for dynamic sparse data exchange, 2010.

.seealso: `PetscGatherNumberOfMessages()`, `PetscGatherMessageLengths()`, `PetscCommBuildTwoSidedSetType()`, `PetscCommBuildTwoSidedType`
@*/
PetscErrorCode PetscCommBuildTwoSided(MPI_Comm comm, PetscMPIInt count, MPI_Datatype dtype, PetscMPIInt nto, const PetscMPIInt *toranks, const void *todata, PetscMPIInt *nfrom, PetscMPIInt **fromranks, void *fromdata)
{
  PetscBuildTwoSidedType buildtype = PETSC_BUILDTWOSIDED_NOTSET;

  PetscFunctionBegin;
  PetscCall(PetscSysInitializePackage());
  PetscCall(PetscLogEventSync(PETSC_BuildTwoSided, comm));
  PetscCall(PetscLogEventBegin(PETSC_BuildTwoSided, 0, 0, 0, 0));
  PetscCall(PetscCommBuildTwoSidedGetType(comm, &buildtype));
  switch (buildtype) {
  case PETSC_BUILDTWOSIDED_IBARRIER:
#if defined(PETSC_HAVE_MPI_NONBLOCKING_COLLECTIVES)
    PetscCall(PetscCommBuildTwoSided_Ibarrier(comm, count, dtype, nto, toranks, todata, nfrom, fromranks, fromdata));
    break;
#else
    SETERRQ(comm, PETSC_ERR_PLIB, "MPI implementation does not provide MPI_Ibarrier (part of MPI-3)");
#endif
  case PETSC_BUILDTWOSIDED_ALLREDUCE:
    PetscCall(PetscCommBuildTwoSided_Allreduce(comm, count, dtype, nto, toranks, todata, nfrom, fromranks, fromdata));
    break;
  case PETSC_BUILDTWOSIDED_REDSCATTER:
#if defined(PETSC_HAVE_MPI_REDUCE_SCATTER_BLOCK)
    PetscCall(PetscCommBuildTwoSided_RedScatter(comm, count, dtype, nto, toranks, todata, nfrom, fromranks, fromdata));
    break;
#else
    SETERRQ(comm, PETSC_ERR_PLIB, "MPI implementation does not provide MPI_Reduce_scatter_block (part of MPI-2.2)");
#endif
  default:
    SETERRQ(comm, PETSC_ERR_PLIB, "Unknown method for building two-sided communication");
  }
  PetscCall(PetscLogEventEnd(PETSC_BuildTwoSided, 0, 0, 0, 0));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode PetscCommBuildTwoSidedFReq_Reference(MPI_Comm comm, PetscMPIInt count, MPI_Datatype dtype, PetscMPIInt nto, const PetscMPIInt *toranks, const void *todata, PetscMPIInt *nfrom, PetscMPIInt **fromranks, void *fromdata, PetscMPIInt ntags, MPI_Request **toreqs, MPI_Request **fromreqs, PetscErrorCode (*send)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, PetscMPIInt, void *, MPI_Request[], void *), PetscErrorCode (*recv)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, void *, MPI_Request[], void *), void *ctx)
{
  PetscMPIInt  i, *tag;
  MPI_Aint     lb, unitbytes;
  MPI_Request *sendreq, *recvreq;

  PetscFunctionBegin;
  PetscCall(PetscMalloc1(ntags, &tag));
  if (ntags > 0) PetscCall(PetscCommDuplicate(comm, &comm, &tag[0]));
  for (i = 1; i < ntags; i++) PetscCall(PetscCommGetNewTag(comm, &tag[i]));

  /* Perform complete initial rendezvous */
  PetscCall(PetscCommBuildTwoSided(comm, count, dtype, nto, toranks, todata, nfrom, fromranks, fromdata));

  PetscCall(PetscMalloc1(nto * ntags, &sendreq));
  PetscCall(PetscMalloc1(*nfrom * ntags, &recvreq));

  PetscCallMPI(MPI_Type_get_extent(dtype, &lb, &unitbytes));
  PetscCheck(lb == 0, comm, PETSC_ERR_SUP, "Datatype with nonzero lower bound %ld", (long)lb);
  for (i = 0; i < nto; i++) {
    PetscMPIInt k;
    for (k = 0; k < ntags; k++) sendreq[i * ntags + k] = MPI_REQUEST_NULL;
    PetscCall((*send)(comm, tag, i, toranks[i], ((char *)todata) + count * unitbytes * i, sendreq + i * ntags, ctx));
  }
  for (i = 0; i < *nfrom; i++) {
    void       *header = (*(char **)fromdata) + count * unitbytes * i;
    PetscMPIInt k;
    for (k = 0; k < ntags; k++) recvreq[i * ntags + k] = MPI_REQUEST_NULL;
    PetscCall((*recv)(comm, tag, (*fromranks)[i], header, recvreq + i * ntags, ctx));
  }
  PetscCall(PetscFree(tag));
  PetscCall(PetscCommDestroy(&comm));
  *toreqs   = sendreq;
  *fromreqs = recvreq;
  PetscFunctionReturn(PETSC_SUCCESS);
}

#if defined(PETSC_HAVE_MPI_NONBLOCKING_COLLECTIVES)

static PetscErrorCode PetscCommBuildTwoSidedFReq_Ibarrier(MPI_Comm comm, PetscMPIInt count, MPI_Datatype dtype, PetscMPIInt nto, const PetscMPIInt *toranks, const void *todata, PetscMPIInt *nfrom, PetscMPIInt **fromranks, void *fromdata, PetscMPIInt ntags, MPI_Request **toreqs, MPI_Request **fromreqs, PetscErrorCode (*send)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, PetscMPIInt, void *, MPI_Request[], void *), PetscErrorCode (*recv)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, void *, MPI_Request[], void *), void *ctx)
{
  PetscMPIInt    nrecvs, tag, *tags, done, i;
  MPI_Aint       lb, unitbytes;
  char          *tdata;
  MPI_Request   *sendreqs, *usendreqs, *req, barrier;
  PetscSegBuffer segrank, segdata, segreq;
  PetscBool      barrier_started;

  PetscFunctionBegin;
  PetscCall(PetscCommDuplicate(comm, &comm, &tag));
  PetscCall(PetscMalloc1(ntags, &tags));
  for (i = 0; i < ntags; i++) PetscCall(PetscCommGetNewTag(comm, &tags[i]));
  PetscCallMPI(MPI_Type_get_extent(dtype, &lb, &unitbytes));
  PetscCheck(lb == 0, comm, PETSC_ERR_SUP, "Datatype with nonzero lower bound %ld", (long)lb);
  tdata = (char *)todata;
  PetscCall(PetscMalloc1(nto, &sendreqs));
  PetscCall(PetscMalloc1(nto * ntags, &usendreqs));
  /* Post synchronous sends */
  for (i = 0; i < nto; i++) PetscCallMPI(MPI_Issend((void *)(tdata + count * unitbytes * i), count, dtype, toranks[i], tag, comm, sendreqs + i));
  /* Post actual payloads.  These are typically larger messages.  Hopefully sending these later does not slow down the
   * synchronous messages above. */
  for (i = 0; i < nto; i++) {
    PetscMPIInt k;
    for (k = 0; k < ntags; k++) usendreqs[i * ntags + k] = MPI_REQUEST_NULL;
    PetscCall((*send)(comm, tags, i, toranks[i], tdata + count * unitbytes * i, usendreqs + i * ntags, ctx));
  }

  PetscCall(PetscSegBufferCreate(sizeof(PetscMPIInt), 4, &segrank));
  PetscCall(PetscSegBufferCreate(unitbytes, 4 * count, &segdata));
  PetscCall(PetscSegBufferCreate(sizeof(MPI_Request), 4, &segreq));

  nrecvs  = 0;
  barrier = MPI_REQUEST_NULL;
  /* MPICH-3.2 sometimes does not create a request in some "optimized" cases.  This is arguably a standard violation,
   * but we need to work around it. */
  barrier_started = PETSC_FALSE;
  for (done = 0; !done;) {
    PetscMPIInt flag;
    MPI_Status  status;
    PetscCallMPI(MPI_Iprobe(MPI_ANY_SOURCE, tag, comm, &flag, &status));
    if (flag) { /* incoming message */
      PetscMPIInt *recvrank, k;
      void        *buf;
      PetscCall(PetscSegBufferGet(segrank, 1, &recvrank));
      PetscCall(PetscSegBufferGet(segdata, count, &buf));
      *recvrank = status.MPI_SOURCE;
      PetscCallMPI(MPI_Recv(buf, count, dtype, status.MPI_SOURCE, tag, comm, MPI_STATUS_IGNORE));
      PetscCall(PetscSegBufferGet(segreq, ntags, &req));
      for (k = 0; k < ntags; k++) req[k] = MPI_REQUEST_NULL;
      PetscCall((*recv)(comm, tags, status.MPI_SOURCE, buf, req, ctx));
      nrecvs++;
    }
    if (!barrier_started) {
      PetscMPIInt sent, nsends;
      PetscCall(PetscMPIIntCast(nto, &nsends));
      PetscCallMPI(MPI_Testall(nsends, sendreqs, &sent, MPI_STATUSES_IGNORE));
      if (sent) {
        PetscCallMPI(MPI_Ibarrier(comm, &barrier));
        barrier_started = PETSC_TRUE;
      }
    } else {
      PetscCallMPI(MPI_Test(&barrier, &done, MPI_STATUS_IGNORE));
    }
  }
  *nfrom = nrecvs;
  PetscCall(PetscSegBufferExtractAlloc(segrank, fromranks));
  PetscCall(PetscSegBufferDestroy(&segrank));
  PetscCall(PetscSegBufferExtractAlloc(segdata, fromdata));
  PetscCall(PetscSegBufferDestroy(&segdata));
  *toreqs = usendreqs;
  PetscCall(PetscSegBufferExtractAlloc(segreq, fromreqs));
  PetscCall(PetscSegBufferDestroy(&segreq));
  PetscCall(PetscFree(sendreqs));
  PetscCall(PetscFree(tags));
  PetscCall(PetscCommDestroy(&comm));
  PetscFunctionReturn(PETSC_SUCCESS);
}
#endif

/*@C
   PetscCommBuildTwoSidedF - discovers communicating ranks given one-sided information, calling user-defined functions during rendezvous

   Collective

   Input Parameters:
+  comm - communicator
.  count - number of entries to send/receive in initial rendezvous (must match on all ranks)
.  dtype - datatype to send/receive from each rank (must match on all ranks)
.  nto - number of ranks to send data to
.  toranks - ranks to send to (array of length nto)
.  todata - data to send to each rank (packed)
.  ntags - number of tags needed by send/recv callbacks
.  send - callback invoked on sending process when ready to send primary payload
.  recv - callback invoked on receiving process after delivery of rendezvous message
-  ctx - context for callbacks

   Output Parameters:
+  nfrom - number of ranks receiving messages from
.  fromranks - ranks receiving messages from (length nfrom; caller should `PetscFree()`)
-  fromdata - packed data from each rank, each with count entries of type dtype (length nfrom, caller responsible for `PetscFree()`)

   Level: developer

   Notes:
   This memory-scalable interface is an alternative to calling `PetscGatherNumberOfMessages()` and
   `PetscGatherMessageLengths()`, possibly with a subsequent round of communication to send other data.

   Basic data types as well as contiguous types are supported, but non-contiguous (e.g., strided) types are not.

   References:
.  * - Hoefler, Siebert and Lumsdaine, The MPI_Ibarrier implementation uses the algorithm in
   Scalable communication protocols for dynamic sparse data exchange, 2010.

.seealso: `PetscCommBuildTwoSided()`, `PetscCommBuildTwoSidedFReq()`, `PetscGatherNumberOfMessages()`, `PetscGatherMessageLengths()`
@*/
PetscErrorCode PetscCommBuildTwoSidedF(MPI_Comm comm, PetscMPIInt count, MPI_Datatype dtype, PetscMPIInt nto, const PetscMPIInt *toranks, const void *todata, PetscMPIInt *nfrom, PetscMPIInt **fromranks, void *fromdata, PetscMPIInt ntags, PetscErrorCode (*send)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, PetscMPIInt, void *, MPI_Request[], void *), PetscErrorCode (*recv)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, void *, MPI_Request[], void *), void *ctx)
{
  MPI_Request *toreqs, *fromreqs;

  PetscFunctionBegin;
  PetscCall(PetscCommBuildTwoSidedFReq(comm, count, dtype, nto, toranks, todata, nfrom, fromranks, fromdata, ntags, &toreqs, &fromreqs, send, recv, ctx));
  PetscCallMPI(MPI_Waitall(nto * ntags, toreqs, MPI_STATUSES_IGNORE));
  PetscCallMPI(MPI_Waitall(*nfrom * ntags, fromreqs, MPI_STATUSES_IGNORE));
  PetscCall(PetscFree(toreqs));
  PetscCall(PetscFree(fromreqs));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@C
   PetscCommBuildTwoSidedFReq - discovers communicating ranks given one-sided information, calling user-defined functions during rendezvous, returns requests

   Collective

   Input Parameters:
+  comm - communicator
.  count - number of entries to send/receive in initial rendezvous (must match on all ranks)
.  dtype - datatype to send/receive from each rank (must match on all ranks)
.  nto - number of ranks to send data to
.  toranks - ranks to send to (array of length nto)
.  todata - data to send to each rank (packed)
.  ntags - number of tags needed by send/recv callbacks
.  send - callback invoked on sending process when ready to send primary payload
.  recv - callback invoked on receiving process after delivery of rendezvous message
-  ctx - context for callbacks

   Output Parameters:
+  nfrom - number of ranks receiving messages from
.  fromranks - ranks receiving messages from (length nfrom; caller should `PetscFree()`)
.  fromdata - packed data from each rank, each with count entries of type dtype (length nfrom, caller responsible for `PetscFree()`)
.  toreqs - array of nto*ntags sender requests (caller must wait on these, then `PetscFree()`)
-  fromreqs - array of nfrom*ntags receiver requests (caller must wait on these, then `PetscFree()`)

   Level: developer

   Notes:
   This memory-scalable interface is an alternative to calling `PetscGatherNumberOfMessages()` and
   `PetscGatherMessageLengths()`, possibly with a subsequent round of communication to send other data.

   Basic data types as well as contiguous types are supported, but non-contiguous (e.g., strided) types are not.

   References:
.  * - Hoefler, Siebert and Lumsdaine, The MPI_Ibarrier implementation uses the algorithm in
   Scalable communication protocols for dynamic sparse data exchange, 2010.

.seealso: `PetscCommBuildTwoSided()`, `PetscCommBuildTwoSidedF()`, `PetscGatherNumberOfMessages()`, `PetscGatherMessageLengths()`
@*/
PetscErrorCode PetscCommBuildTwoSidedFReq(MPI_Comm comm, PetscMPIInt count, MPI_Datatype dtype, PetscMPIInt nto, const PetscMPIInt *toranks, const void *todata, PetscMPIInt *nfrom, PetscMPIInt **fromranks, void *fromdata, PetscMPIInt ntags, MPI_Request **toreqs, MPI_Request **fromreqs, PetscErrorCode (*send)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, PetscMPIInt, void *, MPI_Request[], void *), PetscErrorCode (*recv)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, void *, MPI_Request[], void *), void *ctx)
{
  PetscErrorCode (*f)(MPI_Comm, PetscMPIInt, MPI_Datatype, PetscMPIInt, const PetscMPIInt[], const void *, PetscMPIInt *, PetscMPIInt **, void *, PetscMPIInt, MPI_Request **, MPI_Request **, PetscErrorCode (*send)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, PetscMPIInt, void *, MPI_Request[], void *), PetscErrorCode (*recv)(MPI_Comm, const PetscMPIInt[], PetscMPIInt, void *, MPI_Request[], void *), void *ctx);
  PetscBuildTwoSidedType buildtype = PETSC_BUILDTWOSIDED_NOTSET;
  PetscMPIInt            i, size;

  PetscFunctionBegin;
  PetscCall(PetscSysInitializePackage());
  PetscCallMPI(MPI_Comm_size(comm, &size));
  for (i = 0; i < nto; i++) PetscCheck(toranks[i] >= 0 && size > toranks[i], comm, PETSC_ERR_ARG_OUTOFRANGE, "toranks[%d] %d not in comm size %d", i, toranks[i], size);
  PetscCall(PetscLogEventSync(PETSC_BuildTwoSidedF, comm));
  PetscCall(PetscLogEventBegin(PETSC_BuildTwoSidedF, 0, 0, 0, 0));
  PetscCall(PetscCommBuildTwoSidedGetType(comm, &buildtype));
  switch (buildtype) {
  case PETSC_BUILDTWOSIDED_IBARRIER:
#if defined(PETSC_HAVE_MPI_NONBLOCKING_COLLECTIVES)
    f = PetscCommBuildTwoSidedFReq_Ibarrier;
    break;
#else
    SETERRQ(comm, PETSC_ERR_PLIB, "MPI implementation does not provide MPI_Ibarrier (part of MPI-3)");
#endif
  case PETSC_BUILDTWOSIDED_ALLREDUCE:
  case PETSC_BUILDTWOSIDED_REDSCATTER:
    f = PetscCommBuildTwoSidedFReq_Reference;
    break;
  default:
    SETERRQ(comm, PETSC_ERR_PLIB, "Unknown method for building two-sided communication");
  }
  PetscCall((*f)(comm, count, dtype, nto, toranks, todata, nfrom, fromranks, fromdata, ntags, toreqs, fromreqs, send, recv, ctx));
  PetscCall(PetscLogEventEnd(PETSC_BuildTwoSidedF, 0, 0, 0, 0));
  PetscFunctionReturn(PETSC_SUCCESS);
}