sf/tutorials/ex1.c

static const char help[] = "Test star forest communication (PetscSF)\n\n";

/*
    Description: A star is a simple tree with one root and zero or more leaves.
    A star forest is a union of disjoint stars.
    Many common communication patterns can be expressed as updates of rootdata using leafdata and vice-versa.
    This example creates a star forest, communicates values using the graph (see options for types of communication), views the graph, then destroys it.
*/

/*
  Include petscsf.h so we can use PetscSF objects. Note that this automatically
  includes petscsys.h.
*/
#include <petscsf.h>
#include <petscviewer.h>

/* like PetscSFView() but with alternative array of local indices */
static PetscErrorCode PetscSFViewCustomLocals_Private(PetscSF sf, const PetscInt locals[], PetscViewer viewer)
{
  const PetscSFNode *iremote;
  PetscInt           i, nroots, nleaves, nranks;
  PetscMPIInt        rank;

  PetscFunctionBeginUser;
  PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank));
  PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, NULL, &iremote));
  PetscCall(PetscSFGetRootRanks(sf, &nranks, NULL, NULL, NULL, NULL));
  PetscCall(PetscViewerASCIIPushTab(viewer));
  PetscCall(PetscViewerASCIIPushSynchronized(viewer));
  PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Number of roots=%" PetscInt_FMT ", leaves=%" PetscInt_FMT ", remote ranks=%" PetscInt_FMT "\n", rank, nroots, nleaves, nranks));
  for (i = 0; i < nleaves; i++) PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] %" PetscInt_FMT " <- (%" PetscInt_FMT ",%" PetscInt_FMT ")\n", rank, locals[i], iremote[i].rank, iremote[i].index));
  PetscCall(PetscViewerFlush(viewer));
  PetscCall(PetscViewerASCIIPopTab(viewer));
  PetscCall(PetscViewerASCIIPopSynchronized(viewer));
  PetscFunctionReturn(PETSC_SUCCESS);
}

int main(int argc, char **argv)
{
  PetscInt     i, bs = 2, nroots, nrootsalloc, nleaves, nleavesalloc, *mine, stride;
  PetscSFNode *remote;
  PetscMPIInt  rank, size;
  PetscSF      sf, vsf;
  PetscBool    test_all, test_bcast, test_bcastop, test_reduce, test_degree, test_fetchandop, test_gather, test_scatter, test_embed, test_invert, test_sf_distribute, test_char, test_vector = PETSC_FALSE;
  MPI_Op       mop = MPI_OP_NULL; /* initialize to prevent compiler warnings with cxx_quad build */
  char         opstring[256];
  PetscBool    strflg;

  PetscFunctionBeginUser;
  PetscCall(PetscInitialize(&argc, &argv, (char *)0, help));
  PetscCallMPI(MPI_Comm_rank(PETSC_COMM_WORLD, &rank));
  PetscCallMPI(MPI_Comm_size(PETSC_COMM_WORLD, &size));

  PetscOptionsBegin(PETSC_COMM_WORLD, "", "PetscSF Test Options", "none");
  test_all = PETSC_FALSE;
  PetscCall(PetscOptionsBool("-test_all", "Test all SF communications", "", test_all, &test_all, NULL));
  test_bcast = test_all;
  PetscCall(PetscOptionsBool("-test_bcast", "Test broadcast", "", test_bcast, &test_bcast, NULL));
  test_bcastop = test_all;
  PetscCall(PetscOptionsBool("-test_bcastop", "Test broadcast and reduce", "", test_bcastop, &test_bcastop, NULL));
  test_reduce = test_all;
  PetscCall(PetscOptionsBool("-test_reduce", "Test reduction", "", test_reduce, &test_reduce, NULL));
  test_char = test_all;
  PetscCall(PetscOptionsBool("-test_char", "Test signed char, unsigned char, and char", "", test_char, &test_char, NULL));
  mop = MPI_SUM;
  PetscCall(PetscStrncpy(opstring, "sum", sizeof(opstring)));
  PetscCall(PetscOptionsString("-test_op", "Designate which MPI_Op to use", "", opstring, opstring, sizeof(opstring), NULL));
  PetscCall(PetscStrcmp("sum", opstring, &strflg));
  if (strflg) mop = MPIU_SUM;
  PetscCall(PetscStrcmp("prod", opstring, &strflg));
  if (strflg) mop = MPI_PROD;
  PetscCall(PetscStrcmp("max", opstring, &strflg));
  if (strflg) mop = MPI_MAX;
  PetscCall(PetscStrcmp("min", opstring, &strflg));
  if (strflg) mop = MPI_MIN;
  PetscCall(PetscStrcmp("land", opstring, &strflg));
  if (strflg) mop = MPI_LAND;
  PetscCall(PetscStrcmp("band", opstring, &strflg));
  if (strflg) mop = MPI_BAND;
  PetscCall(PetscStrcmp("lor", opstring, &strflg));
  if (strflg) mop = MPI_LOR;
  PetscCall(PetscStrcmp("bor", opstring, &strflg));
  if (strflg) mop = MPI_BOR;
  PetscCall(PetscStrcmp("lxor", opstring, &strflg));
  if (strflg) mop = MPI_LXOR;
  PetscCall(PetscStrcmp("bxor", opstring, &strflg));
  if (strflg) mop = MPI_BXOR;
  test_degree = test_all;
  PetscCall(PetscOptionsBool("-test_degree", "Test computation of vertex degree", "", test_degree, &test_degree, NULL));
  test_fetchandop = test_all;
  PetscCall(PetscOptionsBool("-test_fetchandop", "Test atomic Fetch-And-Op", "", test_fetchandop, &test_fetchandop, NULL));
  test_gather = test_all;
  PetscCall(PetscOptionsBool("-test_gather", "Test point gather", "", test_gather, &test_gather, NULL));
  test_scatter = test_all;
  PetscCall(PetscOptionsBool("-test_scatter", "Test point scatter", "", test_scatter, &test_scatter, NULL));
  test_embed = test_all;
  PetscCall(PetscOptionsBool("-test_embed", "Test point embed", "", test_embed, &test_embed, NULL));
  test_invert = test_all;
  PetscCall(PetscOptionsBool("-test_invert", "Test point invert", "", test_invert, &test_invert, NULL));
  stride = 1;
  PetscCall(PetscOptionsInt("-stride", "Stride for leaf and root data", "", stride, &stride, NULL));
  test_sf_distribute = PETSC_FALSE;
  PetscCall(PetscOptionsBool("-test_sf_distribute", "Create an SF that 'distributes' to each process, like an alltoall", "", test_sf_distribute, &test_sf_distribute, NULL));
  PetscCall(PetscOptionsString("-test_op", "Designate which MPI_Op to use", "", opstring, opstring, sizeof(opstring), NULL));
  PetscCall(PetscOptionsInt("-bs", "Block size for vectorial SF", "", bs, &bs, NULL));
  PetscCall(PetscOptionsBool("-test_vector", "Run tests using the vectorial SF", "", test_vector, &test_vector, NULL));
  PetscOptionsEnd();

  if (test_sf_distribute) {
    nroots       = size;
    nrootsalloc  = size;
    nleaves      = size;
    nleavesalloc = size;
    mine         = NULL;
    PetscCall(PetscMalloc1(nleaves, &remote));
    for (i = 0; i < size; i++) {
      remote[i].rank  = i;
      remote[i].index = rank;
    }
  } else {
    nroots       = 2 + (PetscInt)(rank == 0);
    nrootsalloc  = nroots * stride;
    nleaves      = 2 + (PetscInt)(rank > 0);
    nleavesalloc = nleaves * stride;
    mine         = NULL;
    if (stride > 1) {
      PetscInt i;

      PetscCall(PetscMalloc1(nleaves, &mine));
      for (i = 0; i < nleaves; i++) mine[i] = stride * i;
    }
    PetscCall(PetscMalloc1(nleaves, &remote));
    /* Left periodic neighbor */
    remote[0].rank  = (rank + size - 1) % size;
    remote[0].index = 1 * stride;
    /* Right periodic neighbor */
    remote[1].rank  = (rank + 1) % size;
    remote[1].index = 0 * stride;
    if (rank > 0) { /* All processes reference rank 0, index 1 */
      remote[2].rank  = 0;
      remote[2].index = 2 * stride;
    }
  }

  /* Create a star forest for communication */
  PetscCall(PetscSFCreate(PETSC_COMM_WORLD, &sf));
  PetscCall(PetscSFSetFromOptions(sf));
  PetscCall(PetscSFSetGraph(sf, nrootsalloc, nleaves, mine, PETSC_OWN_POINTER, remote, PETSC_OWN_POINTER));
  PetscCall(PetscSFSetUp(sf));

  /* We can also obtain a strided SF to communicate interleaved blocks of data */
  PetscCall(PetscSFCreateStridedSF(sf, bs, PETSC_DECIDE, PETSC_DECIDE, &vsf));
  if (test_vector) { /* perform all tests below using the vectorial SF */
    PetscSF t = sf;

    sf  = vsf;
    vsf = t;

    nroots *= bs;
    nleaves *= bs;
    nrootsalloc *= bs;
    nleavesalloc *= bs;
  }
  PetscCall(PetscSFDestroy(&vsf));

  /* View graph, mostly useful for debugging purposes. */
  PetscCall(PetscViewerPushFormat(PETSC_VIEWER_STDOUT_WORLD, PETSC_VIEWER_ASCII_INFO_DETAIL));
  PetscCall(PetscSFView(sf, PETSC_VIEWER_STDOUT_WORLD));
  PetscCall(PetscViewerPopFormat(PETSC_VIEWER_STDOUT_WORLD));

  if (test_bcast) { /* broadcast rootdata into leafdata */
    PetscInt *rootdata, *leafdata;
    /* Allocate space for send and receive buffers. This example communicates PetscInt, but other types, including
     * user-defined structures, could also be used. */
    PetscCall(PetscMalloc2(nrootsalloc, &rootdata, nleavesalloc, &leafdata));
    /* Set rootdata buffer to be broadcast */
    for (i = 0; i < nrootsalloc; i++) rootdata[i] = -1;
    for (i = 0; i < nroots; i++) rootdata[i * stride] = 100 * (rank + 1) + i;
    /* Initialize local buffer, these values are never used. */
    for (i = 0; i < nleavesalloc; i++) leafdata[i] = -1;
    /* Broadcast entries from rootdata to leafdata. Computation or other communication can be performed between the begin and end calls. */
    PetscCall(PetscSFBcastBegin(sf, MPIU_INT, rootdata, leafdata, MPI_REPLACE));
    PetscCall(PetscSFBcastEnd(sf, MPIU_INT, rootdata, leafdata, MPI_REPLACE));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Bcast Rootdata\n"));
    PetscCall(PetscIntView(nrootsalloc, rootdata, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Bcast Leafdata\n"));
    PetscCall(PetscIntView(nleavesalloc, leafdata, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscFree2(rootdata, leafdata));
  }

  if (test_bcast && test_char) { /* Bcast with char */
    PetscInt len;
    char     buf[256];
    char    *rootdata, *leafdata;
    PetscCall(PetscMalloc2(nrootsalloc, &rootdata, nleavesalloc, &leafdata));
    /* Set rootdata buffer to be broadcast */
    for (i = 0; i < nrootsalloc; i++) rootdata[i] = '*';
    for (i = 0; i < nroots; i++) rootdata[i * stride] = 'A' + rank * 3 + i; /* rank is very small, so it is fine to compute a char */
    /* Initialize local buffer, these values are never used. */
    for (i = 0; i < nleavesalloc; i++) leafdata[i] = '?';

    PetscCall(PetscSFBcastBegin(sf, MPI_CHAR, rootdata, leafdata, MPI_REPLACE));
    PetscCall(PetscSFBcastEnd(sf, MPI_CHAR, rootdata, leafdata, MPI_REPLACE));

    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Bcast Rootdata in type of char\n"));
    len = 0;
    PetscCall(PetscSNPrintf(buf, 256, "%4d:", rank));
    len += 5;
    for (i = 0; i < nrootsalloc; i++) {
      PetscCall(PetscSNPrintf(buf + len, 256 - len, "%5c", rootdata[i]));
      len += 5;
    }
    PetscCall(PetscSynchronizedPrintf(PETSC_COMM_WORLD, "%s\n", buf));
    PetscCall(PetscSynchronizedFlush(PETSC_COMM_WORLD, PETSC_STDOUT));

    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Bcast Leafdata in type of char\n"));
    len = 0;
    PetscCall(PetscSNPrintf(buf, 256, "%4d:", rank));
    len += 5;
    for (i = 0; i < nleavesalloc; i++) {
      PetscCall(PetscSNPrintf(buf + len, 256 - len, "%5c", leafdata[i]));
      len += 5;
    }
    PetscCall(PetscSynchronizedPrintf(PETSC_COMM_WORLD, "%s\n", buf));
    PetscCall(PetscSynchronizedFlush(PETSC_COMM_WORLD, PETSC_STDOUT));

    PetscCall(PetscFree2(rootdata, leafdata));
  }

  if (test_bcastop) { /* Reduce rootdata into leafdata */
    PetscInt *rootdata, *leafdata;
    /* Allocate space for send and receive buffers. This example communicates PetscInt, but other types, including
     * user-defined structures, could also be used. */
    PetscCall(PetscMalloc2(nrootsalloc, &rootdata, nleavesalloc, &leafdata));
    /* Set rootdata buffer to be broadcast */
    for (i = 0; i < nrootsalloc; i++) rootdata[i] = -1;
    for (i = 0; i < nroots; i++) rootdata[i * stride] = 100 * (rank + 1) + i;
    /* Set leaf values to reduce with */
    for (i = 0; i < nleavesalloc; i++) leafdata[i] = -10 * (rank + 1) - i;
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Pre-BcastAndOp Leafdata\n"));
    PetscCall(PetscIntView(nleavesalloc, leafdata, PETSC_VIEWER_STDOUT_WORLD));
    /* Broadcast entries from rootdata to leafdata. Computation or other communication can be performed between the begin and end calls. */
    PetscCall(PetscSFBcastBegin(sf, MPIU_INT, rootdata, leafdata, mop));
    PetscCall(PetscSFBcastEnd(sf, MPIU_INT, rootdata, leafdata, mop));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## BcastAndOp Rootdata\n"));
    PetscCall(PetscIntView(nrootsalloc, rootdata, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## BcastAndOp Leafdata\n"));
    PetscCall(PetscIntView(nleavesalloc, leafdata, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscFree2(rootdata, leafdata));
  }

  if (test_reduce) { /* Reduce leafdata into rootdata */
    PetscInt *rootdata, *leafdata;
    PetscCall(PetscMalloc2(nrootsalloc, &rootdata, nleavesalloc, &leafdata));
    /* Initialize rootdata buffer in which the result of the reduction will appear. */
    for (i = 0; i < nrootsalloc; i++) rootdata[i] = -1;
    for (i = 0; i < nroots; i++) rootdata[i * stride] = 100 * (rank + 1) + i;
    /* Set leaf values to reduce. */
    for (i = 0; i < nleavesalloc; i++) leafdata[i] = -1;
    for (i = 0; i < nleaves; i++) leafdata[i * stride] = 1000 * (rank + 1) + 10 * i;
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Pre-Reduce Rootdata\n"));
    PetscCall(PetscIntView(nrootsalloc, rootdata, PETSC_VIEWER_STDOUT_WORLD));
    /* Perform reduction. Computation or other communication can be performed between the begin and end calls.
     * This example sums the values, but other MPI_Ops can be used (e.g MPI_MAX, MPI_PROD). */
    PetscCall(PetscSFReduceBegin(sf, MPIU_INT, leafdata, rootdata, mop));
    PetscCall(PetscSFReduceEnd(sf, MPIU_INT, leafdata, rootdata, mop));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Reduce Leafdata\n"));
    PetscCall(PetscIntView(nleavesalloc, leafdata, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Reduce Rootdata\n"));
    PetscCall(PetscIntView(nrootsalloc, rootdata, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscFree2(rootdata, leafdata));
  }

  if (test_reduce && test_char) { /* Reduce with signed char */
    PetscInt     len;
    char         buf[256];
    signed char *rootdata, *leafdata;
    PetscCall(PetscMalloc2(nrootsalloc, &rootdata, nleavesalloc, &leafdata));
    /* Initialize rootdata buffer in which the result of the reduction will appear. */
    for (i = 0; i < nrootsalloc; i++) rootdata[i] = -1;
    for (i = 0; i < nroots; i++) rootdata[i * stride] = 10 * (rank + 1) + i;
    /* Set leaf values to reduce. */
    for (i = 0; i < nleavesalloc; i++) leafdata[i] = -1;
    for (i = 0; i < nleaves; i++) leafdata[i * stride] = 50 * (rank + 1) + 10 * i;
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Pre-Reduce Rootdata in type of signed char\n"));

    len = 0;
    PetscCall(PetscSNPrintf(buf, 256, "%4d:", rank));
    len += 5;
    for (i = 0; i < nrootsalloc; i++) {
      PetscCall(PetscSNPrintf(buf + len, 256 - len, "%5d", rootdata[i]));
      len += 5;
    }
    PetscCall(PetscSynchronizedPrintf(PETSC_COMM_WORLD, "%s\n", buf));
    PetscCall(PetscSynchronizedFlush(PETSC_COMM_WORLD, PETSC_STDOUT));

    /* Using MPI_CHAR should trigger an error since MPI standard does not support reduction on MPI_CHAR.
       Testing with -test_op max, one can see the sign does take effect in MPI_MAX.
     */
    PetscCall(PetscSFReduceBegin(sf, MPI_SIGNED_CHAR, leafdata, rootdata, mop));
    PetscCall(PetscSFReduceEnd(sf, MPI_SIGNED_CHAR, leafdata, rootdata, mop));

    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Reduce Leafdata in type of signed char\n"));
    len = 0;
    PetscCall(PetscSNPrintf(buf, 256, "%4d:", rank));
    len += 5;
    for (i = 0; i < nleavesalloc; i++) {
      PetscCall(PetscSNPrintf(buf + len, 256 - len, "%5d", leafdata[i]));
      len += 5;
    }
    PetscCall(PetscSynchronizedPrintf(PETSC_COMM_WORLD, "%s\n", buf));
    PetscCall(PetscSynchronizedFlush(PETSC_COMM_WORLD, PETSC_STDOUT));

    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Reduce Rootdata in type of signed char\n"));
    len = 0;
    PetscCall(PetscSNPrintf(buf, 256, "%4d:", rank));
    len += 5;
    for (i = 0; i < nrootsalloc; i++) {
      PetscCall(PetscSNPrintf(buf + len, 256 - len, "%5d", rootdata[i]));
      len += 5;
    }
    PetscCall(PetscSynchronizedPrintf(PETSC_COMM_WORLD, "%s\n", buf));
    PetscCall(PetscSynchronizedFlush(PETSC_COMM_WORLD, PETSC_STDOUT));

    PetscCall(PetscFree2(rootdata, leafdata));
  }

  if (test_reduce && test_char) { /* Reduce with unsigned char */
    PetscInt       len;
    char           buf[256];
    unsigned char *rootdata, *leafdata;
    PetscCall(PetscMalloc2(nrootsalloc, &rootdata, nleavesalloc, &leafdata));
    /* Initialize rootdata buffer in which the result of the reduction will appear. */
    for (i = 0; i < nrootsalloc; i++) rootdata[i] = 0;
    for (i = 0; i < nroots; i++) rootdata[i * stride] = 10 * (rank + 1) + i;
    /* Set leaf values to reduce. */
    for (i = 0; i < nleavesalloc; i++) leafdata[i] = 0;
    for (i = 0; i < nleaves; i++) leafdata[i * stride] = 50 * (rank + 1) + 10 * i;
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Pre-Reduce Rootdata in type of unsigned char\n"));

    len = 0;
    PetscCall(PetscSNPrintf(buf, 256, "%4d:", rank));
    len += 5;
    for (i = 0; i < nrootsalloc; i++) {
      PetscCall(PetscSNPrintf(buf + len, 256 - len, "%5u", rootdata[i]));
      len += 5;
    }
    PetscCall(PetscSynchronizedPrintf(PETSC_COMM_WORLD, "%s\n", buf));
    PetscCall(PetscSynchronizedFlush(PETSC_COMM_WORLD, PETSC_STDOUT));

    /* Using MPI_CHAR should trigger an error since MPI standard does not support reduction on MPI_CHAR.
       Testing with -test_op max, one can see the sign does take effect in MPI_MAX.
     */
    PetscCall(PetscSFReduceBegin(sf, MPI_UNSIGNED_CHAR, leafdata, rootdata, mop));
    PetscCall(PetscSFReduceEnd(sf, MPI_UNSIGNED_CHAR, leafdata, rootdata, mop));

    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Reduce Leafdata in type of unsigned char\n"));
    len = 0;
    PetscCall(PetscSNPrintf(buf, 256, "%4d:", rank));
    len += 5;
    for (i = 0; i < nleavesalloc; i++) {
      PetscCall(PetscSNPrintf(buf + len, 256 - len, "%5u", leafdata[i]));
      len += 5;
    }
    PetscCall(PetscSynchronizedPrintf(PETSC_COMM_WORLD, "%s\n", buf));
    PetscCall(PetscSynchronizedFlush(PETSC_COMM_WORLD, PETSC_STDOUT));

    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Reduce Rootdata in type of unsigned char\n"));
    len = 0;
    PetscCall(PetscSNPrintf(buf, 256, "%4d:", rank));
    len += 5;
    for (i = 0; i < nrootsalloc; i++) {
      PetscCall(PetscSNPrintf(buf + len, 256 - len, "%5u", rootdata[i]));
      len += 5;
    }
    PetscCall(PetscSynchronizedPrintf(PETSC_COMM_WORLD, "%s\n", buf));
    PetscCall(PetscSynchronizedFlush(PETSC_COMM_WORLD, PETSC_STDOUT));

    PetscCall(PetscFree2(rootdata, leafdata));
  }

  if (test_degree) {
    const PetscInt *degree;
    PetscCall(PetscSFComputeDegreeBegin(sf, &degree));
    PetscCall(PetscSFComputeDegreeEnd(sf, &degree));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Root degrees\n"));
    PetscCall(PetscIntView(nrootsalloc, degree, PETSC_VIEWER_STDOUT_WORLD));
  }

  if (test_fetchandop) {
    /* Cannot use text compare here because token ordering is not deterministic */
    PetscInt *leafdata, *leafupdate, *rootdata;
    PetscCall(PetscMalloc3(nleavesalloc, &leafdata, nleavesalloc, &leafupdate, nrootsalloc, &rootdata));
    for (i = 0; i < nleavesalloc; i++) leafdata[i] = -1;
    for (i = 0; i < nleaves; i++) leafdata[i * stride] = 1;
    for (i = 0; i < nrootsalloc; i++) rootdata[i] = -1;
    for (i = 0; i < nroots; i++) rootdata[i * stride] = 0;
    PetscCall(PetscSFFetchAndOpBegin(sf, MPIU_INT, rootdata, leafdata, leafupdate, mop));
    PetscCall(PetscSFFetchAndOpEnd(sf, MPIU_INT, rootdata, leafdata, leafupdate, mop));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Rootdata (sum of 1 from each leaf)\n"));
    PetscCall(PetscIntView(nrootsalloc, rootdata, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Leafupdate (value at roots prior to my atomic update)\n"));
    PetscCall(PetscIntView(nleavesalloc, leafupdate, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscFree3(leafdata, leafupdate, rootdata));
  }

  if (test_gather) {
    const PetscInt *degree;
    PetscInt        inedges, *indata, *outdata;
    PetscCall(PetscSFComputeDegreeBegin(sf, &degree));
    PetscCall(PetscSFComputeDegreeEnd(sf, &degree));
    for (i = 0, inedges = 0; i < nrootsalloc; i++) inedges += degree[i];
    PetscCall(PetscMalloc2(inedges, &indata, nleavesalloc, &outdata));
    for (i = 0; i < nleavesalloc; i++) outdata[i] = -1;
    for (i = 0; i < nleaves; i++) outdata[i * stride] = 1000 * (rank + 1) + i;
    PetscCall(PetscSFGatherBegin(sf, MPIU_INT, outdata, indata));
    PetscCall(PetscSFGatherEnd(sf, MPIU_INT, outdata, indata));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Gathered data at multi-roots from leaves\n"));
    PetscCall(PetscIntView(inedges, indata, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscFree2(indata, outdata));
  }

  if (test_scatter) {
    const PetscInt *degree;
    PetscInt        j, count, inedges, *indata, *outdata;
    PetscCall(PetscSFComputeDegreeBegin(sf, &degree));
    PetscCall(PetscSFComputeDegreeEnd(sf, &degree));
    for (i = 0, inedges = 0; i < nrootsalloc; i++) inedges += degree[i];
    PetscCall(PetscMalloc2(inedges, &indata, nleavesalloc, &outdata));
    for (i = 0; i < nleavesalloc; i++) outdata[i] = -1;
    for (i = 0, count = 0; i < nrootsalloc; i++) {
      for (j = 0; j < degree[i]; j++) indata[count++] = 1000 * (rank + 1) + 100 * i + j;
    }
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Data at multi-roots, to scatter to leaves\n"));
    PetscCall(PetscIntView(inedges, indata, PETSC_VIEWER_STDOUT_WORLD));

    PetscCall(PetscSFScatterBegin(sf, MPIU_INT, indata, outdata));
    PetscCall(PetscSFScatterEnd(sf, MPIU_INT, indata, outdata));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Scattered data at leaves\n"));
    PetscCall(PetscIntView(nleavesalloc, outdata, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscFree2(indata, outdata));
  }

  if (test_embed) {
    const PetscInt nroots = 1 + (PetscInt)(rank == 0);
    PetscInt       selected[2];
    PetscSF        esf;

    selected[0] = stride;
    selected[1] = 2 * stride;
    PetscCall(PetscSFCreateEmbeddedRootSF(sf, nroots, selected, &esf));
    PetscCall(PetscSFSetUp(esf));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Embedded PetscSF\n"));
    PetscCall(PetscViewerPushFormat(PETSC_VIEWER_STDOUT_WORLD, PETSC_VIEWER_ASCII_INFO_DETAIL));
    PetscCall(PetscSFView(esf, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscViewerPopFormat(PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscSFDestroy(&esf));
  }

  if (test_invert) {
    const PetscInt *degree;
    PetscInt       *mRootsOrigNumbering;
    PetscInt        inedges;
    PetscSF         msf, imsf;

    PetscCall(PetscSFGetMultiSF(sf, &msf));
    PetscCall(PetscSFCreateInverseSF(msf, &imsf));
    PetscCall(PetscSFSetUp(msf));
    PetscCall(PetscSFSetUp(imsf));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Multi-SF\n"));
    PetscCall(PetscSFView(msf, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Multi-SF roots indices in original SF roots numbering\n"));
    PetscCall(PetscSFComputeDegreeBegin(sf, &degree));
    PetscCall(PetscSFComputeDegreeEnd(sf, &degree));
    PetscCall(PetscSFComputeMultiRootOriginalNumbering(sf, degree, &inedges, &mRootsOrigNumbering));
    PetscCall(PetscIntView(inedges, mRootsOrigNumbering, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Inverse of Multi-SF\n"));
    PetscCall(PetscSFView(imsf, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscViewerASCIIPrintf(PETSC_VIEWER_STDOUT_WORLD, "## Inverse of Multi-SF, original numbering\n"));
    PetscCall(PetscSFViewCustomLocals_Private(imsf, mRootsOrigNumbering, PETSC_VIEWER_STDOUT_WORLD));
    PetscCall(PetscSFDestroy(&imsf));
    PetscCall(PetscFree(mRootsOrigNumbering));
  }

  /* Clean storage for star forest. */
  PetscCall(PetscSFDestroy(&sf));
  PetscCall(PetscFinalize());
  return 0;
}

/*TEST

   test:
      nsize: 4
      filter: grep -v "type" | grep -v "sort"
      args: -test_bcast -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor {{create dynamic allocate}}
      requires: defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   test:
      suffix: 2
      nsize: 4
      filter: grep -v "type" | grep -v "sort"
      args: -test_reduce -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor {{create dynamic allocate}}
      requires: defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   test:
      suffix: 2_basic
      nsize: 4
      args: -test_reduce -sf_type basic

   test:
      suffix: 3
      nsize: 4
      filter: grep -v "type" | grep -v "sort"
      args: -test_degree -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor {{create dynamic allocate}}
      requires: defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   test:
      suffix: 3_basic
      nsize: 4
      args: -test_degree -sf_type basic

   test:
      suffix: 4
      nsize: 4
      filter: grep -v "type" | grep -v "sort"
      args: -test_gather -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor {{create dynamic allocate}}
      requires: defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   test:
      suffix: 4_basic
      nsize: 4
      args: -test_gather -sf_type basic

   test:
      suffix: 4_stride
      nsize: 4
      args: -test_gather -sf_type basic -stride 2

   test:
      suffix: 5
      nsize: 4
      filter: grep -v "type" | grep -v "sort"
      args: -test_scatter -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor {{create dynamic allocate}}
      requires: defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   test:
      suffix: 5_basic
      nsize: 4
      args: -test_scatter -sf_type basic

   test:
      suffix: 5_stride
      nsize: 4
      args: -test_scatter -sf_type basic -stride 2

   test:
      suffix: 6
      nsize: 4
      filter: grep -v "type" | grep -v "sort"
      # No -sf_window_flavor dynamic due to bug https://gitlab.com/petsc/petsc/issues/555
      args: -test_embed -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor {{create allocate}}
      requires: defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   test:
      suffix: 6_basic
      nsize: 4
      args: -test_embed -sf_type basic

   test:
      suffix: 7
      nsize: 4
      filter: grep -v "type" | grep -v "sort"
      args: -test_invert -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor {{create dynamic allocate}}
      requires: defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   test:
      suffix: 7_basic
      nsize: 4
      args: -test_invert -sf_type basic

   test:
      suffix: basic
      nsize: 4
      args: -test_bcast -sf_type basic
      output_file: output/ex1_1_basic.out

   test:
      suffix: bcastop_basic
      nsize: 4
      args: -test_bcastop -sf_type basic
      output_file: output/ex1_bcastop_basic.out

   test:
      suffix: 8
      nsize: 3
      filter: grep -v "type" | grep -v "sort"
      args: -test_bcast -test_sf_distribute -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor {{create dynamic allocate}}
      requires: defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   test:
      suffix: 8_basic
      nsize: 3
      args: -test_bcast -test_sf_distribute -sf_type basic

   test:
      suffix: 9_char
      nsize: 4
      args: -sf_type basic -test_bcast -test_reduce -test_op max -test_char

   # Here we do not test -sf_window_flavor dynamic since it is designed for repeated SFs with few different rootdata pointers
   test:
      suffix: 10
      filter: grep -v "type" | grep -v "sort"
      nsize: 4
      args: -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor {{create allocate}} -test_all -test_bcastop 0 -test_fetchandop 0
      requires: defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   # The nightly test suite with MPICH uses ch3:sock, which is broken when winsize == 0 in some of the processes
   test:
      suffix: 10_shared
      output_file: output/ex1_10.out
      filter: grep -v "type" | grep -v "sort"
      nsize: 4
      args: -sf_type window -sf_window_sync {{fence active lock}} -sf_window_flavor shared -test_all -test_bcastop 0 -test_fetchandop 0
      requires: defined(PETSC_HAVE_MPI_PROCESS_SHARED_MEMORY) !defined(PETSC_HAVE_MPICH_NUMVERSION) defined(PETSC_HAVE_MPI_ONE_SIDED) defined(PETSC_HAVE_MPI_FEATURE_DYNAMIC_WINDOW)

   test:
      suffix: 10_basic
      nsize: 4
      args: -sf_type basic -test_all -test_bcastop 0 -test_fetchandop 0

   test:
      suffix: 10_basic_vector
      nsize: 4
      args: -sf_type basic -test_all -test_bcastop 0 -test_fetchandop 0 -test_vector

TEST*/