xref: /petsc/src/vec/is/sf/interface/sf.c (revision eec179cf895b6fbcd6a0b58694319392b06e361b)
1af0996ceSBarry Smith #include <petsc/private/sfimpl.h> /*I "petscsf.h" I*/
2c4e6a40aSLawrence Mitchell #include <petsc/private/hashseti.h>
353dd6d7dSJunchao Zhang #include <petsc/private/viewerimpl.h>
4*eec179cfSJacob Faibussowitsch #include <petsc/private/hashmapi.h>
595fce210SBarry Smith 
67fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_CUDA)
77fd2d3dbSJunchao Zhang   #include <cuda_runtime.h>
87fd2d3dbSJunchao Zhang #endif
97fd2d3dbSJunchao Zhang 
107fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_HIP)
117fd2d3dbSJunchao Zhang   #include <hip/hip_runtime.h>
127fd2d3dbSJunchao Zhang #endif
137fd2d3dbSJunchao Zhang 
142abc8c78SJacob Faibussowitsch #if defined(PETSC_CLANG_STATIC_ANALYZER)
152abc8c78SJacob Faibussowitsch void PetscSFCheckGraphSet(PetscSF, int);
162abc8c78SJacob Faibussowitsch #else
1795fce210SBarry Smith   #if defined(PETSC_USE_DEBUG)
187a46b595SBarry Smith     #define PetscSFCheckGraphSet(sf, arg) PetscCheck((sf)->graphset, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetGraph() or PetscSFSetGraphWithPattern() on argument %d \"%s\" before %s()", (arg), #sf, PETSC_FUNCTION_NAME);
1995fce210SBarry Smith   #else
209371c9d4SSatish Balay     #define PetscSFCheckGraphSet(sf, arg) \
219371c9d4SSatish Balay       do { \
229371c9d4SSatish Balay       } while (0)
2395fce210SBarry Smith   #endif
242abc8c78SJacob Faibussowitsch #endif
2595fce210SBarry Smith 
264c8fdceaSLisandro Dalcin const char *const PetscSFDuplicateOptions[]     = {"CONFONLY", "RANKS", "GRAPH", "PetscSFDuplicateOption", "PETSCSF_DUPLICATE_", NULL};
271f40158dSVaclav Hapla const char *const PetscSFConcatenateRootModes[] = {"local", "shared", "global", "PetscSFConcatenateRootMode", "PETSCSF_CONCATENATE_ROOTMODE_", NULL};
2895fce210SBarry Smith 
298af6ec1cSBarry Smith /*@
3095fce210SBarry Smith    PetscSFCreate - create a star forest communication context
3195fce210SBarry Smith 
32d083f849SBarry Smith    Collective
3395fce210SBarry Smith 
344165533cSJose E. Roman    Input Parameter:
3595fce210SBarry Smith .  comm - communicator on which the star forest will operate
3695fce210SBarry Smith 
374165533cSJose E. Roman    Output Parameter:
3895fce210SBarry Smith .  sf - new star forest context
3995fce210SBarry Smith 
40dd5b3ca6SJunchao Zhang    Options Database Keys:
41cab54364SBarry Smith .  -sf_type type - value of type may be
42cab54364SBarry Smith .vb
43cab54364SBarry Smith     basic     -Use MPI persistent Isend/Irecv for communication (Default)
44cab54364SBarry Smith     window    -Use MPI-3 one-sided window for communication
45cab54364SBarry Smith     neighbor  -Use MPI-3 neighborhood collectives for communication
46cab54364SBarry Smith .ve
47dd5b3ca6SJunchao Zhang 
4895fce210SBarry Smith    Level: intermediate
4995fce210SBarry Smith 
50cab54364SBarry Smith    Note:
51cab54364SBarry Smith    When one knows the communication graph is one of the predefined graph, such as `MPI_Alltoall()`, `MPI_Allgatherv()`,
52cab54364SBarry Smith    `MPI_Gatherv()`, one can create a `PetscSF` and then set its graph with `PetscSFSetGraphWithPattern()`. These special
53cab54364SBarry Smith    `SF`s are optimized and they have better performance than general `SF`s.
54dd5b3ca6SJunchao Zhang 
55cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFSetGraphWithPattern()`, `PetscSFDestroy()`
5695fce210SBarry Smith @*/
57d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreate(MPI_Comm comm, PetscSF *sf)
58d71ae5a4SJacob Faibussowitsch {
5995fce210SBarry Smith   PetscSF b;
6095fce210SBarry Smith 
6195fce210SBarry Smith   PetscFunctionBegin;
6295fce210SBarry Smith   PetscValidPointer(sf, 2);
639566063dSJacob Faibussowitsch   PetscCall(PetscSFInitializePackage());
6495fce210SBarry Smith 
659566063dSJacob Faibussowitsch   PetscCall(PetscHeaderCreate(b, PETSCSF_CLASSID, "PetscSF", "Star Forest", "PetscSF", comm, PetscSFDestroy, PetscSFView));
6695fce210SBarry Smith 
6795fce210SBarry Smith   b->nroots    = -1;
6895fce210SBarry Smith   b->nleaves   = -1;
6929046d53SLisandro Dalcin   b->minleaf   = PETSC_MAX_INT;
7029046d53SLisandro Dalcin   b->maxleaf   = PETSC_MIN_INT;
7195fce210SBarry Smith   b->nranks    = -1;
7295fce210SBarry Smith   b->rankorder = PETSC_TRUE;
7395fce210SBarry Smith   b->ingroup   = MPI_GROUP_NULL;
7495fce210SBarry Smith   b->outgroup  = MPI_GROUP_NULL;
7595fce210SBarry Smith   b->graphset  = PETSC_FALSE;
7620c24465SJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
7720c24465SJunchao Zhang   b->use_gpu_aware_mpi    = use_gpu_aware_mpi;
7820c24465SJunchao Zhang   b->use_stream_aware_mpi = PETSC_FALSE;
7971438e86SJunchao Zhang   b->unknown_input_stream = PETSC_FALSE;
8027f636e8SJunchao Zhang   #if defined(PETSC_HAVE_KOKKOS) /* Prefer kokkos over cuda*/
8120c24465SJunchao Zhang   b->backend = PETSCSF_BACKEND_KOKKOS;
8227f636e8SJunchao Zhang   #elif defined(PETSC_HAVE_CUDA)
8327f636e8SJunchao Zhang   b->backend = PETSCSF_BACKEND_CUDA;
8459af0bd3SScott Kruger   #elif defined(PETSC_HAVE_HIP)
8559af0bd3SScott Kruger   b->backend = PETSCSF_BACKEND_HIP;
8620c24465SJunchao Zhang   #endif
8771438e86SJunchao Zhang 
8871438e86SJunchao Zhang   #if defined(PETSC_HAVE_NVSHMEM)
8971438e86SJunchao Zhang   b->use_nvshmem     = PETSC_FALSE; /* Default is not to try NVSHMEM */
9071438e86SJunchao Zhang   b->use_nvshmem_get = PETSC_FALSE; /* Default is to use nvshmem_put based protocol */
919566063dSJacob Faibussowitsch   PetscCall(PetscOptionsGetBool(NULL, NULL, "-use_nvshmem", &b->use_nvshmem, NULL));
929566063dSJacob Faibussowitsch   PetscCall(PetscOptionsGetBool(NULL, NULL, "-use_nvshmem_get", &b->use_nvshmem_get, NULL));
9371438e86SJunchao Zhang   #endif
9420c24465SJunchao Zhang #endif
9560c22052SBarry Smith   b->vscat.from_n = -1;
9660c22052SBarry Smith   b->vscat.to_n   = -1;
9760c22052SBarry Smith   b->vscat.unit   = MPIU_SCALAR;
9895fce210SBarry Smith   *sf             = b;
9995fce210SBarry Smith   PetscFunctionReturn(0);
10095fce210SBarry Smith }
10195fce210SBarry Smith 
10229046d53SLisandro Dalcin /*@
10395fce210SBarry Smith    PetscSFReset - Reset a star forest so that different sizes or neighbors can be used
10495fce210SBarry Smith 
10595fce210SBarry Smith    Collective
10695fce210SBarry Smith 
1074165533cSJose E. Roman    Input Parameter:
10895fce210SBarry Smith .  sf - star forest
10995fce210SBarry Smith 
11095fce210SBarry Smith    Level: advanced
11195fce210SBarry Smith 
112cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`, `PetscSFSetGraph()`, `PetscSFDestroy()`
11395fce210SBarry Smith @*/
114d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReset(PetscSF sf)
115d71ae5a4SJacob Faibussowitsch {
11695fce210SBarry Smith   PetscFunctionBegin;
11795fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
118dbbe0bcdSBarry Smith   PetscTryTypeMethod(sf, Reset);
11929046d53SLisandro Dalcin   sf->nroots   = -1;
12029046d53SLisandro Dalcin   sf->nleaves  = -1;
12129046d53SLisandro Dalcin   sf->minleaf  = PETSC_MAX_INT;
12229046d53SLisandro Dalcin   sf->maxleaf  = PETSC_MIN_INT;
12395fce210SBarry Smith   sf->mine     = NULL;
12495fce210SBarry Smith   sf->remote   = NULL;
12529046d53SLisandro Dalcin   sf->graphset = PETSC_FALSE;
1269566063dSJacob Faibussowitsch   PetscCall(PetscFree(sf->mine_alloc));
1279566063dSJacob Faibussowitsch   PetscCall(PetscFree(sf->remote_alloc));
12821c688dcSJed Brown   sf->nranks = -1;
1299566063dSJacob Faibussowitsch   PetscCall(PetscFree4(sf->ranks, sf->roffset, sf->rmine, sf->rremote));
13029046d53SLisandro Dalcin   sf->degreeknown = PETSC_FALSE;
1319566063dSJacob Faibussowitsch   PetscCall(PetscFree(sf->degree));
1329566063dSJacob Faibussowitsch   if (sf->ingroup != MPI_GROUP_NULL) PetscCallMPI(MPI_Group_free(&sf->ingroup));
1339566063dSJacob Faibussowitsch   if (sf->outgroup != MPI_GROUP_NULL) PetscCallMPI(MPI_Group_free(&sf->outgroup));
134013b3241SStefano Zampini   if (sf->multi) sf->multi->multi = NULL;
1359566063dSJacob Faibussowitsch   PetscCall(PetscSFDestroy(&sf->multi));
1369566063dSJacob Faibussowitsch   PetscCall(PetscLayoutDestroy(&sf->map));
13771438e86SJunchao Zhang 
13871438e86SJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
1399566063dSJacob Faibussowitsch   for (PetscInt i = 0; i < 2; i++) PetscCall(PetscSFFree(sf, PETSC_MEMTYPE_DEVICE, sf->rmine_d[i]));
14071438e86SJunchao Zhang #endif
14171438e86SJunchao Zhang 
14295fce210SBarry Smith   sf->setupcalled = PETSC_FALSE;
14395fce210SBarry Smith   PetscFunctionReturn(0);
14495fce210SBarry Smith }
14595fce210SBarry Smith 
14695fce210SBarry Smith /*@C
147cab54364SBarry Smith    PetscSFSetType - Set the `PetscSF` communication implementation
14895fce210SBarry Smith 
149cab54364SBarry Smith    Collective on sf
15095fce210SBarry Smith 
15195fce210SBarry Smith    Input Parameters:
152cab54364SBarry Smith +  sf - the `PetscSF` context
15395fce210SBarry Smith -  type - a known method
154cab54364SBarry Smith .vb
155cab54364SBarry Smith     PETSCSFWINDOW - MPI-2/3 one-sided
156cab54364SBarry Smith     PETSCSFBASIC - basic implementation using MPI-1 two-sided
157cab54364SBarry Smith .ve
15895fce210SBarry Smith 
15995fce210SBarry Smith    Options Database Key:
160cab54364SBarry Smith .  -sf_type <type> - Sets the method; for example basic or window use -help for a list of available methods (for instance, window, basic, neighbor)
161cab54364SBarry Smith 
162cab54364SBarry Smith   Level: intermediate
16395fce210SBarry Smith 
16495fce210SBarry Smith    Notes:
16595fce210SBarry Smith    See "include/petscsf.h" for available methods (for instance)
16695fce210SBarry Smith 
167db781477SPatrick Sanan .seealso: `PetscSFType`, `PetscSFCreate()`
16895fce210SBarry Smith @*/
169d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetType(PetscSF sf, PetscSFType type)
170d71ae5a4SJacob Faibussowitsch {
17195fce210SBarry Smith   PetscBool match;
1725f80ce2aSJacob Faibussowitsch   PetscErrorCode (*r)(PetscSF);
17395fce210SBarry Smith 
17495fce210SBarry Smith   PetscFunctionBegin;
17595fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
17695fce210SBarry Smith   PetscValidCharPointer(type, 2);
17795fce210SBarry Smith 
1789566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject)sf, type, &match));
17995fce210SBarry Smith   if (match) PetscFunctionReturn(0);
18095fce210SBarry Smith 
1819566063dSJacob Faibussowitsch   PetscCall(PetscFunctionListFind(PetscSFList, type, &r));
18228b400f6SJacob Faibussowitsch   PetscCheck(r, PETSC_COMM_SELF, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unable to find requested PetscSF type %s", type);
18329046d53SLisandro Dalcin   /* Destroy the previous PetscSF implementation context */
184dbbe0bcdSBarry Smith   PetscTryTypeMethod(sf, Destroy);
1859566063dSJacob Faibussowitsch   PetscCall(PetscMemzero(sf->ops, sizeof(*sf->ops)));
1869566063dSJacob Faibussowitsch   PetscCall(PetscObjectChangeTypeName((PetscObject)sf, type));
1879566063dSJacob Faibussowitsch   PetscCall((*r)(sf));
18895fce210SBarry Smith   PetscFunctionReturn(0);
18995fce210SBarry Smith }
19095fce210SBarry Smith 
19129046d53SLisandro Dalcin /*@C
192cab54364SBarry Smith   PetscSFGetType - Get the `PetscSF` communication implementation
19329046d53SLisandro Dalcin 
19429046d53SLisandro Dalcin   Not Collective
19529046d53SLisandro Dalcin 
19629046d53SLisandro Dalcin   Input Parameter:
197cab54364SBarry Smith . sf  - the `PetscSF` context
19829046d53SLisandro Dalcin 
19929046d53SLisandro Dalcin   Output Parameter:
200cab54364SBarry Smith . type - the `PetscSF` type name
20129046d53SLisandro Dalcin 
20229046d53SLisandro Dalcin   Level: intermediate
20329046d53SLisandro Dalcin 
204cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFSetType()`, `PetscSFCreate()`
20529046d53SLisandro Dalcin @*/
206d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetType(PetscSF sf, PetscSFType *type)
207d71ae5a4SJacob Faibussowitsch {
20829046d53SLisandro Dalcin   PetscFunctionBegin;
20929046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
21029046d53SLisandro Dalcin   PetscValidPointer(type, 2);
21129046d53SLisandro Dalcin   *type = ((PetscObject)sf)->type_name;
21229046d53SLisandro Dalcin   PetscFunctionReturn(0);
21329046d53SLisandro Dalcin }
21429046d53SLisandro Dalcin 
2151fb7b255SJunchao Zhang /*@C
21695fce210SBarry Smith    PetscSFDestroy - destroy star forest
21795fce210SBarry Smith 
21895fce210SBarry Smith    Collective
21995fce210SBarry Smith 
2204165533cSJose E. Roman    Input Parameter:
22195fce210SBarry Smith .  sf - address of star forest
22295fce210SBarry Smith 
22395fce210SBarry Smith    Level: intermediate
22495fce210SBarry Smith 
225cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFReset()`
22695fce210SBarry Smith @*/
227d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFDestroy(PetscSF *sf)
228d71ae5a4SJacob Faibussowitsch {
22995fce210SBarry Smith   PetscFunctionBegin;
23095fce210SBarry Smith   if (!*sf) PetscFunctionReturn(0);
23195fce210SBarry Smith   PetscValidHeaderSpecific((*sf), PETSCSF_CLASSID, 1);
2329371c9d4SSatish Balay   if (--((PetscObject)(*sf))->refct > 0) {
2339371c9d4SSatish Balay     *sf = NULL;
2349371c9d4SSatish Balay     PetscFunctionReturn(0);
2359371c9d4SSatish Balay   }
2369566063dSJacob Faibussowitsch   PetscCall(PetscSFReset(*sf));
237dbbe0bcdSBarry Smith   PetscTryTypeMethod((*sf), Destroy);
2389566063dSJacob Faibussowitsch   PetscCall(PetscSFDestroy(&(*sf)->vscat.lsf));
2399566063dSJacob Faibussowitsch   if ((*sf)->vscat.bs > 1) PetscCallMPI(MPI_Type_free(&(*sf)->vscat.unit));
2409566063dSJacob Faibussowitsch   PetscCall(PetscHeaderDestroy(sf));
24195fce210SBarry Smith   PetscFunctionReturn(0);
24295fce210SBarry Smith }
24395fce210SBarry Smith 
244d71ae5a4SJacob Faibussowitsch static PetscErrorCode PetscSFCheckGraphValid_Private(PetscSF sf)
245d71ae5a4SJacob Faibussowitsch {
246c4e6a40aSLawrence Mitchell   PetscInt           i, nleaves;
247c4e6a40aSLawrence Mitchell   PetscMPIInt        size;
248c4e6a40aSLawrence Mitchell   const PetscInt    *ilocal;
249c4e6a40aSLawrence Mitchell   const PetscSFNode *iremote;
250c4e6a40aSLawrence Mitchell 
251c4e6a40aSLawrence Mitchell   PetscFunctionBegin;
25276bd3646SJed Brown   if (!sf->graphset || !PetscDefined(USE_DEBUG)) PetscFunctionReturn(0);
2539566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sf, NULL, &nleaves, &ilocal, &iremote));
2549566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)sf), &size));
255c4e6a40aSLawrence Mitchell   for (i = 0; i < nleaves; i++) {
256c4e6a40aSLawrence Mitchell     const PetscInt rank   = iremote[i].rank;
257c4e6a40aSLawrence Mitchell     const PetscInt remote = iremote[i].index;
258c4e6a40aSLawrence Mitchell     const PetscInt leaf   = ilocal ? ilocal[i] : i;
259c9cc58a2SBarry Smith     PetscCheck(rank >= 0 && rank < size, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided rank (%" PetscInt_FMT ") for remote %" PetscInt_FMT " is invalid, should be in [0, %d)", rank, i, size);
26008401ef6SPierre Jolivet     PetscCheck(remote >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided index (%" PetscInt_FMT ") for remote %" PetscInt_FMT " is invalid, should be >= 0", remote, i);
26108401ef6SPierre Jolivet     PetscCheck(leaf >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided location (%" PetscInt_FMT ") for leaf %" PetscInt_FMT " is invalid, should be >= 0", leaf, i);
262c4e6a40aSLawrence Mitchell   }
263c4e6a40aSLawrence Mitchell   PetscFunctionReturn(0);
264c4e6a40aSLawrence Mitchell }
265c4e6a40aSLawrence Mitchell 
26695fce210SBarry Smith /*@
26795fce210SBarry Smith    PetscSFSetUp - set up communication structures
26895fce210SBarry Smith 
26995fce210SBarry Smith    Collective
27095fce210SBarry Smith 
2714165533cSJose E. Roman    Input Parameter:
27295fce210SBarry Smith .  sf - star forest communication object
27395fce210SBarry Smith 
27495fce210SBarry Smith    Level: beginner
27595fce210SBarry Smith 
276cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFSetFromOptions()`, `PetscSFSetType()`
27795fce210SBarry Smith @*/
278d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetUp(PetscSF sf)
279d71ae5a4SJacob Faibussowitsch {
28095fce210SBarry Smith   PetscFunctionBegin;
28129046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
28229046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf, 1);
28395fce210SBarry Smith   if (sf->setupcalled) PetscFunctionReturn(0);
2849566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_SetUp, sf, 0, 0, 0));
2859566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckGraphValid_Private(sf));
2869566063dSJacob Faibussowitsch   if (!((PetscObject)sf)->type_name) PetscCall(PetscSFSetType(sf, PETSCSFBASIC)); /* Zero all sf->ops */
287dbbe0bcdSBarry Smith   PetscTryTypeMethod(sf, SetUp);
28820c24465SJunchao Zhang #if defined(PETSC_HAVE_CUDA)
28920c24465SJunchao Zhang   if (sf->backend == PETSCSF_BACKEND_CUDA) {
29071438e86SJunchao Zhang     sf->ops->Malloc = PetscSFMalloc_CUDA;
29171438e86SJunchao Zhang     sf->ops->Free   = PetscSFFree_CUDA;
29220c24465SJunchao Zhang   }
29320c24465SJunchao Zhang #endif
29459af0bd3SScott Kruger #if defined(PETSC_HAVE_HIP)
29559af0bd3SScott Kruger   if (sf->backend == PETSCSF_BACKEND_HIP) {
29659af0bd3SScott Kruger     sf->ops->Malloc = PetscSFMalloc_HIP;
29759af0bd3SScott Kruger     sf->ops->Free   = PetscSFFree_HIP;
29859af0bd3SScott Kruger   }
29959af0bd3SScott Kruger #endif
30020c24465SJunchao Zhang 
30159af0bd3SScott Kruger #
30220c24465SJunchao Zhang #if defined(PETSC_HAVE_KOKKOS)
30320c24465SJunchao Zhang   if (sf->backend == PETSCSF_BACKEND_KOKKOS) {
30420c24465SJunchao Zhang     sf->ops->Malloc = PetscSFMalloc_Kokkos;
30520c24465SJunchao Zhang     sf->ops->Free   = PetscSFFree_Kokkos;
30620c24465SJunchao Zhang   }
30720c24465SJunchao Zhang #endif
3089566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_SetUp, sf, 0, 0, 0));
30995fce210SBarry Smith   sf->setupcalled = PETSC_TRUE;
31095fce210SBarry Smith   PetscFunctionReturn(0);
31195fce210SBarry Smith }
31295fce210SBarry Smith 
3138af6ec1cSBarry Smith /*@
314cab54364SBarry Smith    PetscSFSetFromOptions - set `PetscSF` options using the options database
31595fce210SBarry Smith 
31695fce210SBarry Smith    Logically Collective
31795fce210SBarry Smith 
3184165533cSJose E. Roman    Input Parameter:
31995fce210SBarry Smith .  sf - star forest
32095fce210SBarry Smith 
32195fce210SBarry Smith    Options Database Keys:
32260263706SJed Brown +  -sf_type               - implementation type, see PetscSFSetType()
32351ccb202SJunchao Zhang .  -sf_rank_order         - sort composite points for gathers and scatters in rank order, gathers are non-deterministic otherwise
324b85e67b7SJunchao Zhang .  -sf_use_default_stream - Assume callers of SF computed the input root/leafdata with the default cuda stream. SF will also
325c2a741eeSJunchao Zhang                             use the default stream to process data. Therefore, no stream synchronization is needed between SF and its caller (default: true).
326c06a8e02SRichard Tran Mills                             If true, this option only works with -use_gpu_aware_mpi 1.
32720c24465SJunchao Zhang .  -sf_use_stream_aware_mpi  - Assume the underlying MPI is cuda-stream aware and SF won't sync streams for send/recv buffers passed to MPI (default: false).
328c06a8e02SRichard Tran Mills                                If true, this option only works with -use_gpu_aware_mpi 1.
32995fce210SBarry Smith 
33059af0bd3SScott Kruger -  -sf_backend cuda | hip | kokkos -Select the device backend SF uses. Currently SF has these backends: cuda, hip and Kokkos.
33159af0bd3SScott Kruger                               On CUDA (HIP) devices, one can choose cuda (hip) or kokkos with the default being kokkos. On other devices,
33220c24465SJunchao Zhang                               the only available is kokkos.
33320c24465SJunchao Zhang 
33495fce210SBarry Smith    Level: intermediate
335cab54364SBarry Smith 
336cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`, `PetscSFSetType()`
33795fce210SBarry Smith @*/
338d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetFromOptions(PetscSF sf)
339d71ae5a4SJacob Faibussowitsch {
34095fce210SBarry Smith   PetscSFType deft;
34195fce210SBarry Smith   char        type[256];
34295fce210SBarry Smith   PetscBool   flg;
34395fce210SBarry Smith 
34495fce210SBarry Smith   PetscFunctionBegin;
34595fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
346d0609cedSBarry Smith   PetscObjectOptionsBegin((PetscObject)sf);
34795fce210SBarry Smith   deft = ((PetscObject)sf)->type_name ? ((PetscObject)sf)->type_name : PETSCSFBASIC;
3489566063dSJacob Faibussowitsch   PetscCall(PetscOptionsFList("-sf_type", "PetscSF implementation type", "PetscSFSetType", PetscSFList, deft, type, sizeof(type), &flg));
3499566063dSJacob Faibussowitsch   PetscCall(PetscSFSetType(sf, flg ? type : deft));
3509566063dSJacob Faibussowitsch   PetscCall(PetscOptionsBool("-sf_rank_order", "sort composite points for gathers and scatters in rank order, gathers are non-deterministic otherwise", "PetscSFSetRankOrder", sf->rankorder, &sf->rankorder, NULL));
3517fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
35220c24465SJunchao Zhang   {
35320c24465SJunchao Zhang     char      backendstr[32] = {0};
35459af0bd3SScott Kruger     PetscBool isCuda = PETSC_FALSE, isHip = PETSC_FALSE, isKokkos = PETSC_FALSE, set;
35520c24465SJunchao Zhang     /* Change the defaults set in PetscSFCreate() with command line options */
356d5b43468SJose E. Roman     PetscCall(PetscOptionsBool("-sf_unknown_input_stream", "SF root/leafdata is computed on arbitrary streams unknown to SF", "PetscSFSetFromOptions", sf->unknown_input_stream, &sf->unknown_input_stream, NULL));
3579566063dSJacob Faibussowitsch     PetscCall(PetscOptionsBool("-sf_use_stream_aware_mpi", "Assume the underlying MPI is cuda-stream aware", "PetscSFSetFromOptions", sf->use_stream_aware_mpi, &sf->use_stream_aware_mpi, NULL));
3589566063dSJacob Faibussowitsch     PetscCall(PetscOptionsString("-sf_backend", "Select the device backend SF uses", "PetscSFSetFromOptions", NULL, backendstr, sizeof(backendstr), &set));
3599566063dSJacob Faibussowitsch     PetscCall(PetscStrcasecmp("cuda", backendstr, &isCuda));
3609566063dSJacob Faibussowitsch     PetscCall(PetscStrcasecmp("kokkos", backendstr, &isKokkos));
3619566063dSJacob Faibussowitsch     PetscCall(PetscStrcasecmp("hip", backendstr, &isHip));
36259af0bd3SScott Kruger   #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
36320c24465SJunchao Zhang     if (isCuda) sf->backend = PETSCSF_BACKEND_CUDA;
36420c24465SJunchao Zhang     else if (isKokkos) sf->backend = PETSCSF_BACKEND_KOKKOS;
36559af0bd3SScott Kruger     else if (isHip) sf->backend = PETSCSF_BACKEND_HIP;
36628b400f6SJacob Faibussowitsch     else PetscCheck(!set, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "-sf_backend %s is not supported. You may choose cuda, hip or kokkos (if installed)", backendstr);
36720c24465SJunchao Zhang   #elif defined(PETSC_HAVE_KOKKOS)
36808401ef6SPierre Jolivet     PetscCheck(!set || isKokkos, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "-sf_backend %s is not supported. You can only choose kokkos", backendstr);
36920c24465SJunchao Zhang   #endif
37020c24465SJunchao Zhang   }
371c2a741eeSJunchao Zhang #endif
372dbbe0bcdSBarry Smith   PetscTryTypeMethod(sf, SetFromOptions, PetscOptionsObject);
373d0609cedSBarry Smith   PetscOptionsEnd();
37495fce210SBarry Smith   PetscFunctionReturn(0);
37595fce210SBarry Smith }
37695fce210SBarry Smith 
37729046d53SLisandro Dalcin /*@
37895fce210SBarry Smith    PetscSFSetRankOrder - sort multi-points for gathers and scatters by rank order
37995fce210SBarry Smith 
38095fce210SBarry Smith    Logically Collective
38195fce210SBarry Smith 
3824165533cSJose E. Roman    Input Parameters:
38395fce210SBarry Smith +  sf - star forest
384cab54364SBarry Smith -  flg - `PETSC_TRUE` to sort, `PETSC_FALSE` to skip sorting (lower setup cost, but non-deterministic)
38595fce210SBarry Smith 
38695fce210SBarry Smith    Level: advanced
38795fce210SBarry Smith 
388cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFGatherBegin()`, `PetscSFScatterBegin()`
38995fce210SBarry Smith @*/
390d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetRankOrder(PetscSF sf, PetscBool flg)
391d71ae5a4SJacob Faibussowitsch {
39295fce210SBarry Smith   PetscFunctionBegin;
39395fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
39495fce210SBarry Smith   PetscValidLogicalCollectiveBool(sf, flg, 2);
39528b400f6SJacob Faibussowitsch   PetscCheck(!sf->multi, PetscObjectComm((PetscObject)sf), PETSC_ERR_ARG_WRONGSTATE, "Rank ordering must be set before first call to PetscSFGatherBegin() or PetscSFScatterBegin()");
39695fce210SBarry Smith   sf->rankorder = flg;
39795fce210SBarry Smith   PetscFunctionReturn(0);
39895fce210SBarry Smith }
39995fce210SBarry Smith 
4008dbb0df6SBarry Smith /*@C
40195fce210SBarry Smith    PetscSFSetGraph - Set a parallel star forest
40295fce210SBarry Smith 
40395fce210SBarry Smith    Collective
40495fce210SBarry Smith 
4054165533cSJose E. Roman    Input Parameters:
40695fce210SBarry Smith +  sf - star forest
40795fce210SBarry Smith .  nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves)
40895fce210SBarry Smith .  nleaves - number of leaf vertices on the current process, each of these references a root on any process
409c4e6a40aSLawrence Mitchell .  ilocal - locations of leaves in leafdata buffers, pass NULL for contiguous storage (locations must be >= 0, enforced
410c4e6a40aSLawrence Mitchell during setup in debug mode)
41195fce210SBarry Smith .  localmode - copy mode for ilocal
412c4e6a40aSLawrence Mitchell .  iremote - remote locations of root vertices for each leaf on the current process (locations must be >= 0, enforced
413c4e6a40aSLawrence Mitchell during setup in debug mode)
41495fce210SBarry Smith -  remotemode - copy mode for iremote
41595fce210SBarry Smith 
41695fce210SBarry Smith    Level: intermediate
41795fce210SBarry Smith 
41895452b02SPatrick Sanan    Notes:
419db2b9530SVaclav Hapla    Leaf indices in ilocal must be unique, otherwise an error occurs.
42038ab3f8aSBarry Smith 
421db2b9530SVaclav Hapla    Input arrays ilocal and iremote follow the PetscCopyMode semantics.
422cab54364SBarry Smith    In particular, if localmode/remotemode is `PETSC_OWN_POINTER` or `PETSC_USE_POINTER`,
423db2b9530SVaclav Hapla    PETSc might modify the respective array;
424cab54364SBarry Smith    if `PETSC_USE_POINTER`, the user must delete the array after PetscSFDestroy().
425cab54364SBarry Smith    Only if `PETSC_COPY_VALUES` is used, the respective array is guaranteed to stay intact and a const array can be passed (but a cast to non-const is needed).
426db2b9530SVaclav Hapla 
427cab54364SBarry Smith    Fortran Note:
428db2b9530SVaclav Hapla    In Fortran you must use PETSC_COPY_VALUES for localmode and remotemode.
429c4e6a40aSLawrence Mitchell 
430cab54364SBarry Smith    Developer Note:
431db2b9530SVaclav Hapla    We sort leaves to check for duplicates and contiguousness and to find minleaf/maxleaf.
432db2b9530SVaclav Hapla    This also allows to compare leaf sets of two SFs easily.
43372bf8598SVaclav Hapla 
434cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFGetGraph()`
43595fce210SBarry Smith @*/
436d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetGraph(PetscSF sf, PetscInt nroots, PetscInt nleaves, PetscInt *ilocal, PetscCopyMode localmode, PetscSFNode *iremote, PetscCopyMode remotemode)
437d71ae5a4SJacob Faibussowitsch {
438db2b9530SVaclav Hapla   PetscBool unique, contiguous;
43995fce210SBarry Smith 
44095fce210SBarry Smith   PetscFunctionBegin;
44195fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
44229046d53SLisandro Dalcin   if (nleaves > 0 && ilocal) PetscValidIntPointer(ilocal, 4);
44329046d53SLisandro Dalcin   if (nleaves > 0) PetscValidPointer(iremote, 6);
44408401ef6SPierre Jolivet   PetscCheck(nroots >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nroots %" PetscInt_FMT ", cannot be negative", nroots);
44508401ef6SPierre Jolivet   PetscCheck(nleaves >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nleaves %" PetscInt_FMT ", cannot be negative", nleaves);
4468da24d32SBarry Smith   /* enums may be handled as unsigned by some compilers, NVHPC for example, the int cast
4478da24d32SBarry Smith    * below is to prevent NVHPC from warning about meaningless comparison of unsigned with zero */
4488da24d32SBarry Smith   PetscCheck((int)localmode >= PETSC_COPY_VALUES && localmode <= PETSC_USE_POINTER, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Wrong localmode %d", localmode);
4498da24d32SBarry Smith   PetscCheck((int)remotemode >= PETSC_COPY_VALUES && remotemode <= PETSC_USE_POINTER, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Wrong remotemode %d", remotemode);
45029046d53SLisandro Dalcin 
4512a67d2daSStefano Zampini   if (sf->nroots >= 0) { /* Reset only if graph already set */
4529566063dSJacob Faibussowitsch     PetscCall(PetscSFReset(sf));
4532a67d2daSStefano Zampini   }
4542a67d2daSStefano Zampini 
4559566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_SetGraph, sf, 0, 0, 0));
45629046d53SLisandro Dalcin 
45795fce210SBarry Smith   sf->nroots  = nroots;
45895fce210SBarry Smith   sf->nleaves = nleaves;
45929046d53SLisandro Dalcin 
460db2b9530SVaclav Hapla   if (localmode == PETSC_COPY_VALUES && ilocal) {
461db2b9530SVaclav Hapla     PetscInt *tlocal = NULL;
462db2b9530SVaclav Hapla 
4639566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nleaves, &tlocal));
4649566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(tlocal, ilocal, nleaves));
465db2b9530SVaclav Hapla     ilocal = tlocal;
466db2b9530SVaclav Hapla   }
467db2b9530SVaclav Hapla   if (remotemode == PETSC_COPY_VALUES) {
468db2b9530SVaclav Hapla     PetscSFNode *tremote = NULL;
469db2b9530SVaclav Hapla 
4709566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nleaves, &tremote));
4719566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(tremote, iremote, nleaves));
472db2b9530SVaclav Hapla     iremote = tremote;
473db2b9530SVaclav Hapla   }
474db2b9530SVaclav Hapla 
47529046d53SLisandro Dalcin   if (nleaves && ilocal) {
476db2b9530SVaclav Hapla     PetscSFNode work;
477db2b9530SVaclav Hapla 
4789566063dSJacob Faibussowitsch     PetscCall(PetscSortIntWithDataArray(nleaves, ilocal, iremote, sizeof(PetscSFNode), &work));
4799566063dSJacob Faibussowitsch     PetscCall(PetscSortedCheckDupsInt(nleaves, ilocal, &unique));
480db2b9530SVaclav Hapla     unique = PetscNot(unique);
481db2b9530SVaclav Hapla     PetscCheck(sf->allow_multi_leaves || unique, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Input ilocal has duplicate entries which is not allowed for this PetscSF");
482db2b9530SVaclav Hapla     sf->minleaf = ilocal[0];
483db2b9530SVaclav Hapla     sf->maxleaf = ilocal[nleaves - 1];
484db2b9530SVaclav Hapla     contiguous  = (PetscBool)(unique && ilocal[0] == 0 && ilocal[nleaves - 1] == nleaves - 1);
48529046d53SLisandro Dalcin   } else {
48629046d53SLisandro Dalcin     sf->minleaf = 0;
48729046d53SLisandro Dalcin     sf->maxleaf = nleaves - 1;
488db2b9530SVaclav Hapla     unique      = PETSC_TRUE;
489db2b9530SVaclav Hapla     contiguous  = PETSC_TRUE;
49029046d53SLisandro Dalcin   }
49129046d53SLisandro Dalcin 
492db2b9530SVaclav Hapla   if (contiguous) {
493db2b9530SVaclav Hapla     if (localmode == PETSC_USE_POINTER) {
494db2b9530SVaclav Hapla       ilocal = NULL;
495db2b9530SVaclav Hapla     } else {
4969566063dSJacob Faibussowitsch       PetscCall(PetscFree(ilocal));
497db2b9530SVaclav Hapla     }
498db2b9530SVaclav Hapla   }
499db2b9530SVaclav Hapla   sf->mine = ilocal;
500db2b9530SVaclav Hapla   if (localmode == PETSC_USE_POINTER) {
50129046d53SLisandro Dalcin     sf->mine_alloc = NULL;
502db2b9530SVaclav Hapla   } else {
503db2b9530SVaclav Hapla     sf->mine_alloc = ilocal;
50495fce210SBarry Smith   }
505db2b9530SVaclav Hapla   sf->remote = iremote;
506db2b9530SVaclav Hapla   if (remotemode == PETSC_USE_POINTER) {
50729046d53SLisandro Dalcin     sf->remote_alloc = NULL;
508db2b9530SVaclav Hapla   } else {
509db2b9530SVaclav Hapla     sf->remote_alloc = iremote;
51095fce210SBarry Smith   }
5119566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_SetGraph, sf, 0, 0, 0));
51229046d53SLisandro Dalcin   sf->graphset = PETSC_TRUE;
51395fce210SBarry Smith   PetscFunctionReturn(0);
51495fce210SBarry Smith }
51595fce210SBarry Smith 
51629046d53SLisandro Dalcin /*@
517cab54364SBarry Smith   PetscSFSetGraphWithPattern - Sets the graph of a `PetscSF` with a specific pattern
518dd5b3ca6SJunchao Zhang 
519dd5b3ca6SJunchao Zhang   Collective
520dd5b3ca6SJunchao Zhang 
521dd5b3ca6SJunchao Zhang   Input Parameters:
522cab54364SBarry Smith + sf      - The `PetscSF`
523cab54364SBarry Smith . map     - Layout of roots over all processes (insignificant when pattern is `PETSCSF_PATTERN_ALLTOALL`)
524cab54364SBarry Smith - pattern - One of `PETSCSF_PATTERN_ALLGATHER`, `PETSCSF_PATTERN_GATHER`, `PETSCSF_PATTERN_ALLTOALL`
525cab54364SBarry Smith 
526cab54364SBarry Smith   Level: intermediate
527dd5b3ca6SJunchao Zhang 
528dd5b3ca6SJunchao Zhang   Notes:
529cab54364SBarry Smith   It is easier to explain `PetscSFPattern` using vectors. Suppose we have an MPI vector x and its layout is map.
530dd5b3ca6SJunchao Zhang   n and N are local and global sizes of x respectively.
531dd5b3ca6SJunchao Zhang 
532cab54364SBarry Smith   With `PETSCSF_PATTERN_ALLGATHER`, the routine creates a graph that if one does Bcast on it, it will copy x to
533dd5b3ca6SJunchao Zhang   sequential vectors y on all ranks.
534dd5b3ca6SJunchao Zhang 
535cab54364SBarry Smith   With `PETSCSF_PATTERN_GATHER`, the routine creates a graph that if one does Bcast on it, it will copy x to a
536dd5b3ca6SJunchao Zhang   sequential vector y on rank 0.
537dd5b3ca6SJunchao Zhang 
538dd5b3ca6SJunchao Zhang   In above cases, entries of x are roots and entries of y are leaves.
539dd5b3ca6SJunchao Zhang 
540cab54364SBarry Smith   With `PETSCSF_PATTERN_ALLTOALL`, map is insignificant. Suppose NP is size of sf's communicator. The routine
541dd5b3ca6SJunchao Zhang   creates a graph that every rank has NP leaves and NP roots. On rank i, its leaf j is connected to root i
542cab54364SBarry Smith   of rank j. Here 0 <=i,j<NP. It is a kind of `MPI_Alltoall()` with sendcount/recvcount being 1. Note that it does
543dd5b3ca6SJunchao Zhang   not mean one can not send multiple items. One just needs to create a new MPI datatype for the mulptiple data
544cab54364SBarry Smith   items with `MPI_Type_contiguous` and use that as the <unit> argument in SF routines.
545dd5b3ca6SJunchao Zhang 
546dd5b3ca6SJunchao Zhang   In this case, roots and leaves are symmetric.
547dd5b3ca6SJunchao Zhang 
548cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFGetGraph()`
549dd5b3ca6SJunchao Zhang  @*/
550d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetGraphWithPattern(PetscSF sf, PetscLayout map, PetscSFPattern pattern)
551d71ae5a4SJacob Faibussowitsch {
552dd5b3ca6SJunchao Zhang   MPI_Comm    comm;
553dd5b3ca6SJunchao Zhang   PetscInt    n, N, res[2];
554dd5b3ca6SJunchao Zhang   PetscMPIInt rank, size;
555dd5b3ca6SJunchao Zhang   PetscSFType type;
556dd5b3ca6SJunchao Zhang 
557dd5b3ca6SJunchao Zhang   PetscFunctionBegin;
5582abc8c78SJacob Faibussowitsch   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
5592abc8c78SJacob Faibussowitsch   if (pattern != PETSCSF_PATTERN_ALLTOALL) PetscValidPointer(map, 2);
5609566063dSJacob Faibussowitsch   PetscCall(PetscObjectGetComm((PetscObject)sf, &comm));
5612c71b3e2SJacob Faibussowitsch   PetscCheck(pattern >= PETSCSF_PATTERN_ALLGATHER && pattern <= PETSCSF_PATTERN_ALLTOALL, comm, PETSC_ERR_ARG_OUTOFRANGE, "Unsupported PetscSFPattern %d", pattern);
5629566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_rank(comm, &rank));
5639566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_size(comm, &size));
564dd5b3ca6SJunchao Zhang 
565dd5b3ca6SJunchao Zhang   if (pattern == PETSCSF_PATTERN_ALLTOALL) {
566dd5b3ca6SJunchao Zhang     type = PETSCSFALLTOALL;
5679566063dSJacob Faibussowitsch     PetscCall(PetscLayoutCreate(comm, &sf->map));
5689566063dSJacob Faibussowitsch     PetscCall(PetscLayoutSetLocalSize(sf->map, size));
5699566063dSJacob Faibussowitsch     PetscCall(PetscLayoutSetSize(sf->map, ((PetscInt)size) * size));
5709566063dSJacob Faibussowitsch     PetscCall(PetscLayoutSetUp(sf->map));
571dd5b3ca6SJunchao Zhang   } else {
5729566063dSJacob Faibussowitsch     PetscCall(PetscLayoutGetLocalSize(map, &n));
5739566063dSJacob Faibussowitsch     PetscCall(PetscLayoutGetSize(map, &N));
574dd5b3ca6SJunchao Zhang     res[0] = n;
575dd5b3ca6SJunchao Zhang     res[1] = -n;
576dd5b3ca6SJunchao Zhang     /* Check if n are same over all ranks so that we can optimize it */
5771c2dc1cbSBarry Smith     PetscCall(MPIU_Allreduce(MPI_IN_PLACE, res, 2, MPIU_INT, MPI_MAX, comm));
578dd5b3ca6SJunchao Zhang     if (res[0] == -res[1]) { /* same n */
579dd5b3ca6SJunchao Zhang       type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHER : PETSCSFGATHER;
580dd5b3ca6SJunchao Zhang     } else {
581dd5b3ca6SJunchao Zhang       type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHERV : PETSCSFGATHERV;
582dd5b3ca6SJunchao Zhang     }
5839566063dSJacob Faibussowitsch     PetscCall(PetscLayoutReference(map, &sf->map));
584dd5b3ca6SJunchao Zhang   }
5859566063dSJacob Faibussowitsch   PetscCall(PetscSFSetType(sf, type));
586dd5b3ca6SJunchao Zhang 
587dd5b3ca6SJunchao Zhang   sf->pattern = pattern;
588dd5b3ca6SJunchao Zhang   sf->mine    = NULL; /* Contiguous */
589dd5b3ca6SJunchao Zhang 
590dd5b3ca6SJunchao Zhang   /* Set nleaves, nroots here in case user calls PetscSFGetGraph, which is legal to call even before PetscSFSetUp is called.
591dd5b3ca6SJunchao Zhang      Also set other easy stuff.
592dd5b3ca6SJunchao Zhang    */
593dd5b3ca6SJunchao Zhang   if (pattern == PETSCSF_PATTERN_ALLGATHER) {
594dd5b3ca6SJunchao Zhang     sf->nleaves = N;
595dd5b3ca6SJunchao Zhang     sf->nroots  = n;
596dd5b3ca6SJunchao Zhang     sf->nranks  = size;
597dd5b3ca6SJunchao Zhang     sf->minleaf = 0;
598dd5b3ca6SJunchao Zhang     sf->maxleaf = N - 1;
599dd5b3ca6SJunchao Zhang   } else if (pattern == PETSCSF_PATTERN_GATHER) {
600dd5b3ca6SJunchao Zhang     sf->nleaves = rank ? 0 : N;
601dd5b3ca6SJunchao Zhang     sf->nroots  = n;
602dd5b3ca6SJunchao Zhang     sf->nranks  = rank ? 0 : size;
603dd5b3ca6SJunchao Zhang     sf->minleaf = 0;
604dd5b3ca6SJunchao Zhang     sf->maxleaf = rank ? -1 : N - 1;
605dd5b3ca6SJunchao Zhang   } else if (pattern == PETSCSF_PATTERN_ALLTOALL) {
606dd5b3ca6SJunchao Zhang     sf->nleaves = size;
607dd5b3ca6SJunchao Zhang     sf->nroots  = size;
608dd5b3ca6SJunchao Zhang     sf->nranks  = size;
609dd5b3ca6SJunchao Zhang     sf->minleaf = 0;
610dd5b3ca6SJunchao Zhang     sf->maxleaf = size - 1;
611dd5b3ca6SJunchao Zhang   }
612dd5b3ca6SJunchao Zhang   sf->ndranks  = 0; /* We do not need to separate out distinguished ranks for patterned graphs to improve communication performance */
613dd5b3ca6SJunchao Zhang   sf->graphset = PETSC_TRUE;
614dd5b3ca6SJunchao Zhang   PetscFunctionReturn(0);
615dd5b3ca6SJunchao Zhang }
616dd5b3ca6SJunchao Zhang 
617dd5b3ca6SJunchao Zhang /*@
618cab54364SBarry Smith    PetscSFCreateInverseSF - given a `PetscSF` in which all vertices have degree 1, creates the inverse map
61995fce210SBarry Smith 
62095fce210SBarry Smith    Collective
62195fce210SBarry Smith 
6224165533cSJose E. Roman    Input Parameter:
62395fce210SBarry Smith .  sf - star forest to invert
62495fce210SBarry Smith 
6254165533cSJose E. Roman    Output Parameter:
62695fce210SBarry Smith .  isf - inverse of sf
6274165533cSJose E. Roman 
62895fce210SBarry Smith    Level: advanced
62995fce210SBarry Smith 
63095fce210SBarry Smith    Notes:
63195fce210SBarry Smith    All roots must have degree 1.
63295fce210SBarry Smith 
63395fce210SBarry Smith    The local space may be a permutation, but cannot be sparse.
63495fce210SBarry Smith 
635cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFSetGraph()`
63695fce210SBarry Smith @*/
637d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateInverseSF(PetscSF sf, PetscSF *isf)
638d71ae5a4SJacob Faibussowitsch {
63995fce210SBarry Smith   PetscMPIInt     rank;
64095fce210SBarry Smith   PetscInt        i, nroots, nleaves, maxlocal, count, *newilocal;
64195fce210SBarry Smith   const PetscInt *ilocal;
64295fce210SBarry Smith   PetscSFNode    *roots, *leaves;
64395fce210SBarry Smith 
64495fce210SBarry Smith   PetscFunctionBegin;
64529046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
64629046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf, 1);
64729046d53SLisandro Dalcin   PetscValidPointer(isf, 2);
64829046d53SLisandro Dalcin 
6499566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, NULL));
65029046d53SLisandro Dalcin   maxlocal = sf->maxleaf + 1; /* TODO: We should use PetscSFGetLeafRange() */
65129046d53SLisandro Dalcin 
6529566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank));
6539566063dSJacob Faibussowitsch   PetscCall(PetscMalloc2(nroots, &roots, maxlocal, &leaves));
654ae9aee6dSMatthew G. Knepley   for (i = 0; i < maxlocal; i++) {
65595fce210SBarry Smith     leaves[i].rank  = rank;
65695fce210SBarry Smith     leaves[i].index = i;
65795fce210SBarry Smith   }
65895fce210SBarry Smith   for (i = 0; i < nroots; i++) {
65995fce210SBarry Smith     roots[i].rank  = -1;
66095fce210SBarry Smith     roots[i].index = -1;
66195fce210SBarry Smith   }
6629566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceBegin(sf, MPIU_2INT, leaves, roots, MPI_REPLACE));
6639566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceEnd(sf, MPIU_2INT, leaves, roots, MPI_REPLACE));
66495fce210SBarry Smith 
66595fce210SBarry Smith   /* Check whether our leaves are sparse */
6669371c9d4SSatish Balay   for (i = 0, count = 0; i < nroots; i++)
6679371c9d4SSatish Balay     if (roots[i].rank >= 0) count++;
66895fce210SBarry Smith   if (count == nroots) newilocal = NULL;
6699371c9d4SSatish Balay   else { /* Index for sparse leaves and compact "roots" array (which is to become our leaves). */ PetscCall(PetscMalloc1(count, &newilocal));
67095fce210SBarry Smith     for (i = 0, count = 0; i < nroots; i++) {
67195fce210SBarry Smith       if (roots[i].rank >= 0) {
67295fce210SBarry Smith         newilocal[count]   = i;
67395fce210SBarry Smith         roots[count].rank  = roots[i].rank;
67495fce210SBarry Smith         roots[count].index = roots[i].index;
67595fce210SBarry Smith         count++;
67695fce210SBarry Smith       }
67795fce210SBarry Smith     }
67895fce210SBarry Smith   }
67995fce210SBarry Smith 
6809566063dSJacob Faibussowitsch   PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_CONFONLY, isf));
6819566063dSJacob Faibussowitsch   PetscCall(PetscSFSetGraph(*isf, maxlocal, count, newilocal, PETSC_OWN_POINTER, roots, PETSC_COPY_VALUES));
6829566063dSJacob Faibussowitsch   PetscCall(PetscFree2(roots, leaves));
68395fce210SBarry Smith   PetscFunctionReturn(0);
68495fce210SBarry Smith }
68595fce210SBarry Smith 
68695fce210SBarry Smith /*@
687cab54364SBarry Smith    PetscSFDuplicate - duplicate a `PetscSF`, optionally preserving rank connectivity and graph
68895fce210SBarry Smith 
68995fce210SBarry Smith    Collective
69095fce210SBarry Smith 
6914165533cSJose E. Roman    Input Parameters:
69295fce210SBarry Smith +  sf - communication object to duplicate
693cab54364SBarry Smith -  opt - `PETSCSF_DUPLICATE_CONFONLY`, `PETSCSF_DUPLICATE_RANKS`, or `PETSCSF_DUPLICATE_GRAPH` (see `PetscSFDuplicateOption`)
69495fce210SBarry Smith 
6954165533cSJose E. Roman    Output Parameter:
69695fce210SBarry Smith .  newsf - new communication object
69795fce210SBarry Smith 
69895fce210SBarry Smith    Level: beginner
69995fce210SBarry Smith 
700cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFSetType()`, `PetscSFSetGraph()`
70195fce210SBarry Smith @*/
702d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFDuplicate(PetscSF sf, PetscSFDuplicateOption opt, PetscSF *newsf)
703d71ae5a4SJacob Faibussowitsch {
70429046d53SLisandro Dalcin   PetscSFType  type;
70597929ea7SJunchao Zhang   MPI_Datatype dtype = MPIU_SCALAR;
70695fce210SBarry Smith 
70795fce210SBarry Smith   PetscFunctionBegin;
70829046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
70929046d53SLisandro Dalcin   PetscValidLogicalCollectiveEnum(sf, opt, 2);
71029046d53SLisandro Dalcin   PetscValidPointer(newsf, 3);
7119566063dSJacob Faibussowitsch   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sf), newsf));
7129566063dSJacob Faibussowitsch   PetscCall(PetscSFGetType(sf, &type));
7139566063dSJacob Faibussowitsch   if (type) PetscCall(PetscSFSetType(*newsf, type));
71454a24607SJunchao Zhang   (*newsf)->allow_multi_leaves = sf->allow_multi_leaves; /* Dup this flag ealier since PetscSFSetGraph() below checks on this flag */
71595fce210SBarry Smith   if (opt == PETSCSF_DUPLICATE_GRAPH) {
716dd5b3ca6SJunchao Zhang     PetscSFCheckGraphSet(sf, 1);
717dd5b3ca6SJunchao Zhang     if (sf->pattern == PETSCSF_PATTERN_GENERAL) {
71895fce210SBarry Smith       PetscInt           nroots, nleaves;
71995fce210SBarry Smith       const PetscInt    *ilocal;
72095fce210SBarry Smith       const PetscSFNode *iremote;
7219566063dSJacob Faibussowitsch       PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, &iremote));
7229566063dSJacob Faibussowitsch       PetscCall(PetscSFSetGraph(*newsf, nroots, nleaves, (PetscInt *)ilocal, PETSC_COPY_VALUES, (PetscSFNode *)iremote, PETSC_COPY_VALUES));
723dd5b3ca6SJunchao Zhang     } else {
7249566063dSJacob Faibussowitsch       PetscCall(PetscSFSetGraphWithPattern(*newsf, sf->map, sf->pattern));
725dd5b3ca6SJunchao Zhang     }
72695fce210SBarry Smith   }
72797929ea7SJunchao Zhang   /* Since oldtype is committed, so is newtype, according to MPI */
7289566063dSJacob Faibussowitsch   if (sf->vscat.bs > 1) PetscCallMPI(MPI_Type_dup(sf->vscat.unit, &dtype));
72997929ea7SJunchao Zhang   (*newsf)->vscat.bs     = sf->vscat.bs;
73097929ea7SJunchao Zhang   (*newsf)->vscat.unit   = dtype;
73197929ea7SJunchao Zhang   (*newsf)->vscat.to_n   = sf->vscat.to_n;
73297929ea7SJunchao Zhang   (*newsf)->vscat.from_n = sf->vscat.from_n;
73397929ea7SJunchao Zhang   /* Do not copy lsf. Build it on demand since it is rarely used */
73497929ea7SJunchao Zhang 
73520c24465SJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
73620c24465SJunchao Zhang   (*newsf)->backend              = sf->backend;
73771438e86SJunchao Zhang   (*newsf)->unknown_input_stream = sf->unknown_input_stream;
73820c24465SJunchao Zhang   (*newsf)->use_gpu_aware_mpi    = sf->use_gpu_aware_mpi;
73920c24465SJunchao Zhang   (*newsf)->use_stream_aware_mpi = sf->use_stream_aware_mpi;
74020c24465SJunchao Zhang #endif
741dbbe0bcdSBarry Smith   PetscTryTypeMethod(sf, Duplicate, opt, *newsf);
74220c24465SJunchao Zhang   /* Don't do PetscSFSetUp() since the new sf's graph might have not been set. */
74395fce210SBarry Smith   PetscFunctionReturn(0);
74495fce210SBarry Smith }
74595fce210SBarry Smith 
74695fce210SBarry Smith /*@C
74795fce210SBarry Smith    PetscSFGetGraph - Get the graph specifying a parallel star forest
74895fce210SBarry Smith 
74995fce210SBarry Smith    Not Collective
75095fce210SBarry Smith 
7514165533cSJose E. Roman    Input Parameter:
75295fce210SBarry Smith .  sf - star forest
75395fce210SBarry Smith 
7544165533cSJose E. Roman    Output Parameters:
75595fce210SBarry Smith +  nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves)
75695fce210SBarry Smith .  nleaves - number of leaf vertices on the current process, each of these references a root on any process
757bc6585dcSJunchao Zhang .  ilocal - locations of leaves in leafdata buffers (if returned value is NULL, it means leaves are in contiguous storage)
75895fce210SBarry Smith -  iremote - remote locations of root vertices for each leaf on the current process
75995fce210SBarry Smith 
760cab54364SBarry Smith    Level: intermediate
761cab54364SBarry Smith 
762373e0d91SLisandro Dalcin    Notes:
763373e0d91SLisandro Dalcin      We are not currently requiring that the graph is set, thus returning nroots=-1 if it has not been set yet
764373e0d91SLisandro Dalcin 
765cab54364SBarry Smith      The returned ilocal/iremote might contain values in different order than the input ones in `PetscSFSetGraph()`,
766db2b9530SVaclav Hapla      see its manpage for details.
767db2b9530SVaclav Hapla 
7688dbb0df6SBarry Smith    Fortran Notes:
7698dbb0df6SBarry Smith      The returned iremote array is a copy and must be deallocated after use. Consequently, if you
770cab54364SBarry Smith      want to update the graph, you must call `PetscSFSetGraph()` after modifying the iremote array.
7718dbb0df6SBarry Smith 
7728dbb0df6SBarry Smith      To check for a NULL ilocal use
7738dbb0df6SBarry Smith $      if (loc(ilocal) == loc(PETSC_NULL_INTEGER)) then
774ca797d7aSLawrence Mitchell 
775cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFSetGraph()`
77695fce210SBarry Smith @*/
777d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetGraph(PetscSF sf, PetscInt *nroots, PetscInt *nleaves, const PetscInt **ilocal, const PetscSFNode **iremote)
778d71ae5a4SJacob Faibussowitsch {
77995fce210SBarry Smith   PetscFunctionBegin;
78095fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
781b8dee149SJunchao Zhang   if (sf->ops->GetGraph) {
7829566063dSJacob Faibussowitsch     PetscCall((sf->ops->GetGraph)(sf, nroots, nleaves, ilocal, iremote));
783b8dee149SJunchao Zhang   } else {
78495fce210SBarry Smith     if (nroots) *nroots = sf->nroots;
78595fce210SBarry Smith     if (nleaves) *nleaves = sf->nleaves;
78695fce210SBarry Smith     if (ilocal) *ilocal = sf->mine;
78795fce210SBarry Smith     if (iremote) *iremote = sf->remote;
788b8dee149SJunchao Zhang   }
78995fce210SBarry Smith   PetscFunctionReturn(0);
79095fce210SBarry Smith }
79195fce210SBarry Smith 
79229046d53SLisandro Dalcin /*@
79395fce210SBarry Smith    PetscSFGetLeafRange - Get the active leaf ranges
79495fce210SBarry Smith 
79595fce210SBarry Smith    Not Collective
79695fce210SBarry Smith 
7974165533cSJose E. Roman    Input Parameter:
79895fce210SBarry Smith .  sf - star forest
79995fce210SBarry Smith 
8004165533cSJose E. Roman    Output Parameters:
801dd5b3ca6SJunchao Zhang +  minleaf - minimum active leaf on this process. Return 0 if there are no leaves.
802dd5b3ca6SJunchao Zhang -  maxleaf - maximum active leaf on this process. Return -1 if there are no leaves.
80395fce210SBarry Smith 
80495fce210SBarry Smith    Level: developer
80595fce210SBarry Smith 
806cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFSetGraph()`, `PetscSFGetGraph()`
80795fce210SBarry Smith @*/
808d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetLeafRange(PetscSF sf, PetscInt *minleaf, PetscInt *maxleaf)
809d71ae5a4SJacob Faibussowitsch {
81095fce210SBarry Smith   PetscFunctionBegin;
81195fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
81229046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf, 1);
81395fce210SBarry Smith   if (minleaf) *minleaf = sf->minleaf;
81495fce210SBarry Smith   if (maxleaf) *maxleaf = sf->maxleaf;
81595fce210SBarry Smith   PetscFunctionReturn(0);
81695fce210SBarry Smith }
81795fce210SBarry Smith 
81895fce210SBarry Smith /*@C
819cab54364SBarry Smith    PetscSFViewFromOptions - View a `PetscSF` based on arguments in the options database
820fe2efc57SMark 
821cab54364SBarry Smith    Collective on A
822fe2efc57SMark 
823fe2efc57SMark    Input Parameters:
824fe2efc57SMark +  A - the star forest
825cab54364SBarry Smith .  obj - Optional object that provides the prefix for the option names
826736c3998SJose E. Roman -  name - command line option
827fe2efc57SMark 
828fe2efc57SMark    Level: intermediate
829cab54364SBarry Smith 
830db781477SPatrick Sanan .seealso: `PetscSF`, `PetscSFView`, `PetscObjectViewFromOptions()`, `PetscSFCreate()`
831fe2efc57SMark @*/
832d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFViewFromOptions(PetscSF A, PetscObject obj, const char name[])
833d71ae5a4SJacob Faibussowitsch {
834fe2efc57SMark   PetscFunctionBegin;
835fe2efc57SMark   PetscValidHeaderSpecific(A, PETSCSF_CLASSID, 1);
8369566063dSJacob Faibussowitsch   PetscCall(PetscObjectViewFromOptions((PetscObject)A, obj, name));
837fe2efc57SMark   PetscFunctionReturn(0);
838fe2efc57SMark }
839fe2efc57SMark 
840fe2efc57SMark /*@C
84195fce210SBarry Smith    PetscSFView - view a star forest
84295fce210SBarry Smith 
84395fce210SBarry Smith    Collective
84495fce210SBarry Smith 
8454165533cSJose E. Roman    Input Parameters:
84695fce210SBarry Smith +  sf - star forest
847cab54364SBarry Smith -  viewer - viewer to display graph, for example `PETSC_VIEWER_STDOUT_WORLD`
84895fce210SBarry Smith 
84995fce210SBarry Smith    Level: beginner
85095fce210SBarry Smith 
851cab54364SBarry Smith .seealso: `PetscSF`, `PetscViewer`, `PetscSFCreate()`, `PetscSFSetGraph()`
85295fce210SBarry Smith @*/
853d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFView(PetscSF sf, PetscViewer viewer)
854d71ae5a4SJacob Faibussowitsch {
85595fce210SBarry Smith   PetscBool         iascii;
85695fce210SBarry Smith   PetscViewerFormat format;
85795fce210SBarry Smith 
85895fce210SBarry Smith   PetscFunctionBegin;
85995fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
8609566063dSJacob Faibussowitsch   if (!viewer) PetscCall(PetscViewerASCIIGetStdout(PetscObjectComm((PetscObject)sf), &viewer));
86195fce210SBarry Smith   PetscValidHeaderSpecific(viewer, PETSC_VIEWER_CLASSID, 2);
86295fce210SBarry Smith   PetscCheckSameComm(sf, 1, viewer, 2);
8639566063dSJacob Faibussowitsch   if (sf->graphset) PetscCall(PetscSFSetUp(sf));
8649566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii));
86553dd6d7dSJunchao Zhang   if (iascii && viewer->format != PETSC_VIEWER_ASCII_MATLAB) {
86695fce210SBarry Smith     PetscMPIInt rank;
86781bfa7aaSJed Brown     PetscInt    ii, i, j;
86895fce210SBarry Smith 
8699566063dSJacob Faibussowitsch     PetscCall(PetscObjectPrintClassNamePrefixType((PetscObject)sf, viewer));
8709566063dSJacob Faibussowitsch     PetscCall(PetscViewerASCIIPushTab(viewer));
871dd5b3ca6SJunchao Zhang     if (sf->pattern == PETSCSF_PATTERN_GENERAL) {
87280153354SVaclav Hapla       if (!sf->graphset) {
8739566063dSJacob Faibussowitsch         PetscCall(PetscViewerASCIIPrintf(viewer, "PetscSFSetGraph() has not been called yet\n"));
8749566063dSJacob Faibussowitsch         PetscCall(PetscViewerASCIIPopTab(viewer));
87580153354SVaclav Hapla         PetscFunctionReturn(0);
87680153354SVaclav Hapla       }
8779566063dSJacob Faibussowitsch       PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank));
8789566063dSJacob Faibussowitsch       PetscCall(PetscViewerASCIIPushSynchronized(viewer));
8799566063dSJacob Faibussowitsch       PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Number of roots=%" PetscInt_FMT ", leaves=%" PetscInt_FMT ", remote ranks=%" PetscInt_FMT "\n", rank, sf->nroots, sf->nleaves, sf->nranks));
88048a46eb9SPierre Jolivet       for (i = 0; i < sf->nleaves; i++) PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] %" PetscInt_FMT " <- (%" PetscInt_FMT ",%" PetscInt_FMT ")\n", rank, sf->mine ? sf->mine[i] : i, sf->remote[i].rank, sf->remote[i].index));
8819566063dSJacob Faibussowitsch       PetscCall(PetscViewerFlush(viewer));
8829566063dSJacob Faibussowitsch       PetscCall(PetscViewerGetFormat(viewer, &format));
88395fce210SBarry Smith       if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
88481bfa7aaSJed Brown         PetscMPIInt *tmpranks, *perm;
8859566063dSJacob Faibussowitsch         PetscCall(PetscMalloc2(sf->nranks, &tmpranks, sf->nranks, &perm));
8869566063dSJacob Faibussowitsch         PetscCall(PetscArraycpy(tmpranks, sf->ranks, sf->nranks));
88781bfa7aaSJed Brown         for (i = 0; i < sf->nranks; i++) perm[i] = i;
8889566063dSJacob Faibussowitsch         PetscCall(PetscSortMPIIntWithArray(sf->nranks, tmpranks, perm));
8899566063dSJacob Faibussowitsch         PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Roots referenced by my leaves, by rank\n", rank));
89081bfa7aaSJed Brown         for (ii = 0; ii < sf->nranks; ii++) {
89181bfa7aaSJed Brown           i = perm[ii];
8929566063dSJacob Faibussowitsch           PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] %d: %" PetscInt_FMT " edges\n", rank, sf->ranks[i], sf->roffset[i + 1] - sf->roffset[i]));
89348a46eb9SPierre Jolivet           for (j = sf->roffset[i]; j < sf->roffset[i + 1]; j++) PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d]    %" PetscInt_FMT " <- %" PetscInt_FMT "\n", rank, sf->rmine[j], sf->rremote[j]));
89495fce210SBarry Smith         }
8959566063dSJacob Faibussowitsch         PetscCall(PetscFree2(tmpranks, perm));
89695fce210SBarry Smith       }
8979566063dSJacob Faibussowitsch       PetscCall(PetscViewerFlush(viewer));
8989566063dSJacob Faibussowitsch       PetscCall(PetscViewerASCIIPopSynchronized(viewer));
899dd5b3ca6SJunchao Zhang     }
9009566063dSJacob Faibussowitsch     PetscCall(PetscViewerASCIIPopTab(viewer));
90195fce210SBarry Smith   }
902dbbe0bcdSBarry Smith   PetscTryTypeMethod(sf, View, viewer);
90395fce210SBarry Smith   PetscFunctionReturn(0);
90495fce210SBarry Smith }
90595fce210SBarry Smith 
90695fce210SBarry Smith /*@C
907dec1416fSJunchao Zhang    PetscSFGetRootRanks - Get root ranks and number of vertices referenced by leaves on this process
90895fce210SBarry Smith 
90995fce210SBarry Smith    Not Collective
91095fce210SBarry Smith 
9114165533cSJose E. Roman    Input Parameter:
91295fce210SBarry Smith .  sf - star forest
91395fce210SBarry Smith 
9144165533cSJose E. Roman    Output Parameters:
91595fce210SBarry Smith +  nranks - number of ranks referenced by local part
91695fce210SBarry Smith .  ranks - array of ranks
91795fce210SBarry Smith .  roffset - offset in rmine/rremote for each rank (length nranks+1)
91895fce210SBarry Smith .  rmine - concatenated array holding local indices referencing each remote rank
91995fce210SBarry Smith -  rremote - concatenated array holding remote indices referenced for each remote rank
92095fce210SBarry Smith 
92195fce210SBarry Smith    Level: developer
92295fce210SBarry Smith 
923cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetLeafRanks()`
92495fce210SBarry Smith @*/
925d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetRootRanks(PetscSF sf, PetscInt *nranks, const PetscMPIInt **ranks, const PetscInt **roffset, const PetscInt **rmine, const PetscInt **rremote)
926d71ae5a4SJacob Faibussowitsch {
92795fce210SBarry Smith   PetscFunctionBegin;
92895fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
92928b400f6SJacob Faibussowitsch   PetscCheck(sf->setupcalled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetUp() before obtaining ranks");
930dec1416fSJunchao Zhang   if (sf->ops->GetRootRanks) {
9319566063dSJacob Faibussowitsch     PetscCall((sf->ops->GetRootRanks)(sf, nranks, ranks, roffset, rmine, rremote));
932dec1416fSJunchao Zhang   } else {
933dec1416fSJunchao Zhang     /* The generic implementation */
93495fce210SBarry Smith     if (nranks) *nranks = sf->nranks;
93595fce210SBarry Smith     if (ranks) *ranks = sf->ranks;
93695fce210SBarry Smith     if (roffset) *roffset = sf->roffset;
93795fce210SBarry Smith     if (rmine) *rmine = sf->rmine;
93895fce210SBarry Smith     if (rremote) *rremote = sf->rremote;
939dec1416fSJunchao Zhang   }
94095fce210SBarry Smith   PetscFunctionReturn(0);
94195fce210SBarry Smith }
94295fce210SBarry Smith 
9438750ddebSJunchao Zhang /*@C
9448750ddebSJunchao Zhang    PetscSFGetLeafRanks - Get leaf ranks referencing roots on this process
9458750ddebSJunchao Zhang 
9468750ddebSJunchao Zhang    Not Collective
9478750ddebSJunchao Zhang 
9484165533cSJose E. Roman    Input Parameter:
9498750ddebSJunchao Zhang .  sf - star forest
9508750ddebSJunchao Zhang 
9514165533cSJose E. Roman    Output Parameters:
9528750ddebSJunchao Zhang +  niranks - number of leaf ranks referencing roots on this process
9538750ddebSJunchao Zhang .  iranks - array of ranks
9548750ddebSJunchao Zhang .  ioffset - offset in irootloc for each rank (length niranks+1)
9558750ddebSJunchao Zhang -  irootloc - concatenated array holding local indices of roots referenced by each leaf rank
9568750ddebSJunchao Zhang 
9578750ddebSJunchao Zhang    Level: developer
9588750ddebSJunchao Zhang 
959cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetRootRanks()`
9608750ddebSJunchao Zhang @*/
961d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetLeafRanks(PetscSF sf, PetscInt *niranks, const PetscMPIInt **iranks, const PetscInt **ioffset, const PetscInt **irootloc)
962d71ae5a4SJacob Faibussowitsch {
9638750ddebSJunchao Zhang   PetscFunctionBegin;
9648750ddebSJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
96528b400f6SJacob Faibussowitsch   PetscCheck(sf->setupcalled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetUp() before obtaining ranks");
9668750ddebSJunchao Zhang   if (sf->ops->GetLeafRanks) {
9679566063dSJacob Faibussowitsch     PetscCall((sf->ops->GetLeafRanks)(sf, niranks, iranks, ioffset, irootloc));
9688750ddebSJunchao Zhang   } else {
9698750ddebSJunchao Zhang     PetscSFType type;
9709566063dSJacob Faibussowitsch     PetscCall(PetscSFGetType(sf, &type));
97198921bdaSJacob Faibussowitsch     SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "PetscSFGetLeafRanks() is not supported on this StarForest type: %s", type);
9728750ddebSJunchao Zhang   }
9738750ddebSJunchao Zhang   PetscFunctionReturn(0);
9748750ddebSJunchao Zhang }
9758750ddebSJunchao Zhang 
976d71ae5a4SJacob Faibussowitsch static PetscBool InList(PetscMPIInt needle, PetscMPIInt n, const PetscMPIInt *list)
977d71ae5a4SJacob Faibussowitsch {
978b5a8e515SJed Brown   PetscInt i;
979b5a8e515SJed Brown   for (i = 0; i < n; i++) {
980b5a8e515SJed Brown     if (needle == list[i]) return PETSC_TRUE;
981b5a8e515SJed Brown   }
982b5a8e515SJed Brown   return PETSC_FALSE;
983b5a8e515SJed Brown }
984b5a8e515SJed Brown 
98595fce210SBarry Smith /*@C
986cab54364SBarry Smith    PetscSFSetUpRanks - Set up data structures associated with ranks; this is for internal use by `PetscSF` implementations.
98721c688dcSJed Brown 
98821c688dcSJed Brown    Collective
98921c688dcSJed Brown 
9904165533cSJose E. Roman    Input Parameters:
991cab54364SBarry Smith +  sf - `PetscSF` to set up; `PetscSFSetGraph()` must have been called
992cab54364SBarry Smith -  dgroup - `MPI_Group` of ranks to be distinguished (e.g., for self or shared memory exchange)
99321c688dcSJed Brown 
99421c688dcSJed Brown    Level: developer
99521c688dcSJed Brown 
996cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetRootRanks()`
99721c688dcSJed Brown @*/
998d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetUpRanks(PetscSF sf, MPI_Group dgroup)
999d71ae5a4SJacob Faibussowitsch {
1000*eec179cfSJacob Faibussowitsch   PetscHMapI    table;
1001*eec179cfSJacob Faibussowitsch   PetscHashIter pos;
1002b5a8e515SJed Brown   PetscMPIInt   size, groupsize, *groupranks;
1003247e8311SStefano Zampini   PetscInt     *rcount, *ranks;
1004247e8311SStefano Zampini   PetscInt      i, irank = -1, orank = -1;
100521c688dcSJed Brown 
100621c688dcSJed Brown   PetscFunctionBegin;
100721c688dcSJed Brown   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
100829046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf, 1);
10099566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)sf), &size));
1010*eec179cfSJacob Faibussowitsch   PetscCall(PetscHMapICreateWithSize(10, &table));
101121c688dcSJed Brown   for (i = 0; i < sf->nleaves; i++) {
101221c688dcSJed Brown     /* Log 1-based rank */
1013*eec179cfSJacob Faibussowitsch     PetscCall(PetscHMapISetWithMode(table, sf->remote[i].rank + 1, 1, ADD_VALUES));
101421c688dcSJed Brown   }
1015*eec179cfSJacob Faibussowitsch   PetscCall(PetscHMapIGetSize(table, &sf->nranks));
10169566063dSJacob Faibussowitsch   PetscCall(PetscMalloc4(sf->nranks, &sf->ranks, sf->nranks + 1, &sf->roffset, sf->nleaves, &sf->rmine, sf->nleaves, &sf->rremote));
10179566063dSJacob Faibussowitsch   PetscCall(PetscMalloc2(sf->nranks, &rcount, sf->nranks, &ranks));
1018*eec179cfSJacob Faibussowitsch   PetscHashIterBegin(table, pos);
101921c688dcSJed Brown   for (i = 0; i < sf->nranks; i++) {
1020*eec179cfSJacob Faibussowitsch     PetscHashIterGetKey(table, pos, ranks[i]);
1021*eec179cfSJacob Faibussowitsch     PetscHashIterGetVal(table, pos, rcount[i]);
1022*eec179cfSJacob Faibussowitsch     PetscHashIterNext(table, pos);
102321c688dcSJed Brown     ranks[i]--; /* Convert back to 0-based */
102421c688dcSJed Brown   }
1025*eec179cfSJacob Faibussowitsch   PetscCall(PetscHMapIDestroy(&table));
1026b5a8e515SJed Brown 
1027b5a8e515SJed Brown   /* We expect that dgroup is reliably "small" while nranks could be large */
1028b5a8e515SJed Brown   {
10297fb8a5e4SKarl Rupp     MPI_Group    group = MPI_GROUP_NULL;
1030b5a8e515SJed Brown     PetscMPIInt *dgroupranks;
10319566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf), &group));
10329566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_size(dgroup, &groupsize));
10339566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(groupsize, &dgroupranks));
10349566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(groupsize, &groupranks));
1035b5a8e515SJed Brown     for (i = 0; i < groupsize; i++) dgroupranks[i] = i;
10369566063dSJacob Faibussowitsch     if (groupsize) PetscCallMPI(MPI_Group_translate_ranks(dgroup, groupsize, dgroupranks, group, groupranks));
10379566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_free(&group));
10389566063dSJacob Faibussowitsch     PetscCall(PetscFree(dgroupranks));
1039b5a8e515SJed Brown   }
1040b5a8e515SJed Brown 
1041b5a8e515SJed Brown   /* Partition ranks[] into distinguished (first sf->ndranks) followed by non-distinguished */
1042b5a8e515SJed Brown   for (sf->ndranks = 0, i = sf->nranks; sf->ndranks < i;) {
1043b5a8e515SJed Brown     for (i--; sf->ndranks < i; i--) { /* Scan i backward looking for distinguished rank */
1044b5a8e515SJed Brown       if (InList(ranks[i], groupsize, groupranks)) break;
1045b5a8e515SJed Brown     }
1046b5a8e515SJed Brown     for (; sf->ndranks <= i; sf->ndranks++) { /* Scan sf->ndranks forward looking for non-distinguished rank */
1047b5a8e515SJed Brown       if (!InList(ranks[sf->ndranks], groupsize, groupranks)) break;
1048b5a8e515SJed Brown     }
1049b5a8e515SJed Brown     if (sf->ndranks < i) { /* Swap ranks[sf->ndranks] with ranks[i] */
1050b5a8e515SJed Brown       PetscInt tmprank, tmpcount;
1051247e8311SStefano Zampini 
1052b5a8e515SJed Brown       tmprank             = ranks[i];
1053b5a8e515SJed Brown       tmpcount            = rcount[i];
1054b5a8e515SJed Brown       ranks[i]            = ranks[sf->ndranks];
1055b5a8e515SJed Brown       rcount[i]           = rcount[sf->ndranks];
1056b5a8e515SJed Brown       ranks[sf->ndranks]  = tmprank;
1057b5a8e515SJed Brown       rcount[sf->ndranks] = tmpcount;
1058b5a8e515SJed Brown       sf->ndranks++;
1059b5a8e515SJed Brown     }
1060b5a8e515SJed Brown   }
10619566063dSJacob Faibussowitsch   PetscCall(PetscFree(groupranks));
10629566063dSJacob Faibussowitsch   PetscCall(PetscSortIntWithArray(sf->ndranks, ranks, rcount));
10639566063dSJacob Faibussowitsch   PetscCall(PetscSortIntWithArray(sf->nranks - sf->ndranks, ranks + sf->ndranks, rcount + sf->ndranks));
106421c688dcSJed Brown   sf->roffset[0] = 0;
106521c688dcSJed Brown   for (i = 0; i < sf->nranks; i++) {
10669566063dSJacob Faibussowitsch     PetscCall(PetscMPIIntCast(ranks[i], sf->ranks + i));
106721c688dcSJed Brown     sf->roffset[i + 1] = sf->roffset[i] + rcount[i];
106821c688dcSJed Brown     rcount[i]          = 0;
106921c688dcSJed Brown   }
1070247e8311SStefano Zampini   for (i = 0, irank = -1, orank = -1; i < sf->nleaves; i++) {
1071247e8311SStefano Zampini     /* short circuit */
1072247e8311SStefano Zampini     if (orank != sf->remote[i].rank) {
107321c688dcSJed Brown       /* Search for index of iremote[i].rank in sf->ranks */
10749566063dSJacob Faibussowitsch       PetscCall(PetscFindMPIInt(sf->remote[i].rank, sf->ndranks, sf->ranks, &irank));
1075b5a8e515SJed Brown       if (irank < 0) {
10769566063dSJacob Faibussowitsch         PetscCall(PetscFindMPIInt(sf->remote[i].rank, sf->nranks - sf->ndranks, sf->ranks + sf->ndranks, &irank));
1077b5a8e515SJed Brown         if (irank >= 0) irank += sf->ndranks;
107821c688dcSJed Brown       }
1079247e8311SStefano Zampini       orank = sf->remote[i].rank;
1080247e8311SStefano Zampini     }
108108401ef6SPierre Jolivet     PetscCheck(irank >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Could not find rank %" PetscInt_FMT " in array", sf->remote[i].rank);
108221c688dcSJed Brown     sf->rmine[sf->roffset[irank] + rcount[irank]]   = sf->mine ? sf->mine[i] : i;
108321c688dcSJed Brown     sf->rremote[sf->roffset[irank] + rcount[irank]] = sf->remote[i].index;
108421c688dcSJed Brown     rcount[irank]++;
108521c688dcSJed Brown   }
10869566063dSJacob Faibussowitsch   PetscCall(PetscFree2(rcount, ranks));
108721c688dcSJed Brown   PetscFunctionReturn(0);
108821c688dcSJed Brown }
108921c688dcSJed Brown 
109021c688dcSJed Brown /*@C
109195fce210SBarry Smith    PetscSFGetGroups - gets incoming and outgoing process groups
109295fce210SBarry Smith 
109395fce210SBarry Smith    Collective
109495fce210SBarry Smith 
10954165533cSJose E. Roman    Input Parameter:
109695fce210SBarry Smith .  sf - star forest
109795fce210SBarry Smith 
10984165533cSJose E. Roman    Output Parameters:
109995fce210SBarry Smith +  incoming - group of origin processes for incoming edges (leaves that reference my roots)
110095fce210SBarry Smith -  outgoing - group of destination processes for outgoing edges (roots that I reference)
110195fce210SBarry Smith 
110295fce210SBarry Smith    Level: developer
110395fce210SBarry Smith 
1104cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetWindow()`, `PetscSFRestoreWindow()`
110595fce210SBarry Smith @*/
1106d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetGroups(PetscSF sf, MPI_Group *incoming, MPI_Group *outgoing)
1107d71ae5a4SJacob Faibussowitsch {
11087fb8a5e4SKarl Rupp   MPI_Group group = MPI_GROUP_NULL;
110995fce210SBarry Smith 
111095fce210SBarry Smith   PetscFunctionBegin;
111108401ef6SPierre Jolivet   PetscCheck(sf->nranks >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetUpRanks() before obtaining groups");
111295fce210SBarry Smith   if (sf->ingroup == MPI_GROUP_NULL) {
111395fce210SBarry Smith     PetscInt        i;
111495fce210SBarry Smith     const PetscInt *indegree;
111595fce210SBarry Smith     PetscMPIInt     rank, *outranks, *inranks;
111695fce210SBarry Smith     PetscSFNode    *remote;
111795fce210SBarry Smith     PetscSF         bgcount;
111895fce210SBarry Smith 
111995fce210SBarry Smith     /* Compute the number of incoming ranks */
11209566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(sf->nranks, &remote));
112195fce210SBarry Smith     for (i = 0; i < sf->nranks; i++) {
112295fce210SBarry Smith       remote[i].rank  = sf->ranks[i];
112395fce210SBarry Smith       remote[i].index = 0;
112495fce210SBarry Smith     }
11259566063dSJacob Faibussowitsch     PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_CONFONLY, &bgcount));
11269566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(bgcount, 1, sf->nranks, NULL, PETSC_COPY_VALUES, remote, PETSC_OWN_POINTER));
11279566063dSJacob Faibussowitsch     PetscCall(PetscSFComputeDegreeBegin(bgcount, &indegree));
11289566063dSJacob Faibussowitsch     PetscCall(PetscSFComputeDegreeEnd(bgcount, &indegree));
112995fce210SBarry Smith     /* Enumerate the incoming ranks */
11309566063dSJacob Faibussowitsch     PetscCall(PetscMalloc2(indegree[0], &inranks, sf->nranks, &outranks));
11319566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank));
113295fce210SBarry Smith     for (i = 0; i < sf->nranks; i++) outranks[i] = rank;
11339566063dSJacob Faibussowitsch     PetscCall(PetscSFGatherBegin(bgcount, MPI_INT, outranks, inranks));
11349566063dSJacob Faibussowitsch     PetscCall(PetscSFGatherEnd(bgcount, MPI_INT, outranks, inranks));
11359566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf), &group));
11369566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_incl(group, indegree[0], inranks, &sf->ingroup));
11379566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_free(&group));
11389566063dSJacob Faibussowitsch     PetscCall(PetscFree2(inranks, outranks));
11399566063dSJacob Faibussowitsch     PetscCall(PetscSFDestroy(&bgcount));
114095fce210SBarry Smith   }
114195fce210SBarry Smith   *incoming = sf->ingroup;
114295fce210SBarry Smith 
114395fce210SBarry Smith   if (sf->outgroup == MPI_GROUP_NULL) {
11449566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf), &group));
11459566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_incl(group, sf->nranks, sf->ranks, &sf->outgroup));
11469566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_free(&group));
114795fce210SBarry Smith   }
114895fce210SBarry Smith   *outgoing = sf->outgroup;
114995fce210SBarry Smith   PetscFunctionReturn(0);
115095fce210SBarry Smith }
115195fce210SBarry Smith 
115229046d53SLisandro Dalcin /*@
1153cab54364SBarry Smith    PetscSFGetMultiSF - gets the inner `PetscSF` implementing gathers and scatters
115495fce210SBarry Smith 
115595fce210SBarry Smith    Collective
115695fce210SBarry Smith 
11574165533cSJose E. Roman    Input Parameter:
115895fce210SBarry Smith .  sf - star forest that may contain roots with 0 or with more than 1 vertex
115995fce210SBarry Smith 
11604165533cSJose E. Roman    Output Parameter:
116195fce210SBarry Smith .  multi - star forest with split roots, such that each root has degree exactly 1
116295fce210SBarry Smith 
116395fce210SBarry Smith    Level: developer
116495fce210SBarry Smith 
1165cab54364SBarry Smith    Note:
1166cab54364SBarry Smith    In most cases, users should use `PetscSFGatherBegin()` and `PetscSFScatterBegin()` instead of manipulating multi
116795fce210SBarry Smith    directly. Since multi satisfies the stronger condition that each entry in the global space has exactly one incoming
116895fce210SBarry Smith    edge, it is a candidate for future optimization that might involve its removal.
116995fce210SBarry Smith 
1170cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFGatherBegin()`, `PetscSFScatterBegin()`, `PetscSFComputeMultiRootOriginalNumbering()`
117195fce210SBarry Smith @*/
1172d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetMultiSF(PetscSF sf, PetscSF *multi)
1173d71ae5a4SJacob Faibussowitsch {
117495fce210SBarry Smith   PetscFunctionBegin;
117595fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
117695fce210SBarry Smith   PetscValidPointer(multi, 2);
117795fce210SBarry Smith   if (sf->nroots < 0) { /* Graph has not been set yet; why do we need this? */
11789566063dSJacob Faibussowitsch     PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_RANKS, &sf->multi));
117995fce210SBarry Smith     *multi           = sf->multi;
1180013b3241SStefano Zampini     sf->multi->multi = sf->multi;
118195fce210SBarry Smith     PetscFunctionReturn(0);
118295fce210SBarry Smith   }
118395fce210SBarry Smith   if (!sf->multi) {
118495fce210SBarry Smith     const PetscInt *indegree;
11859837ea96SMatthew G. Knepley     PetscInt        i, *inoffset, *outones, *outoffset, maxlocal;
118695fce210SBarry Smith     PetscSFNode    *remote;
118729046d53SLisandro Dalcin     maxlocal = sf->maxleaf + 1; /* TODO: We should use PetscSFGetLeafRange() */
11889566063dSJacob Faibussowitsch     PetscCall(PetscSFComputeDegreeBegin(sf, &indegree));
11899566063dSJacob Faibussowitsch     PetscCall(PetscSFComputeDegreeEnd(sf, &indegree));
11909566063dSJacob Faibussowitsch     PetscCall(PetscMalloc3(sf->nroots + 1, &inoffset, maxlocal, &outones, maxlocal, &outoffset));
119195fce210SBarry Smith     inoffset[0] = 0;
119295fce210SBarry Smith     for (i = 0; i < sf->nroots; i++) inoffset[i + 1] = inoffset[i] + indegree[i];
11939837ea96SMatthew G. Knepley     for (i = 0; i < maxlocal; i++) outones[i] = 1;
11949566063dSJacob Faibussowitsch     PetscCall(PetscSFFetchAndOpBegin(sf, MPIU_INT, inoffset, outones, outoffset, MPI_SUM));
11959566063dSJacob Faibussowitsch     PetscCall(PetscSFFetchAndOpEnd(sf, MPIU_INT, inoffset, outones, outoffset, MPI_SUM));
119695fce210SBarry Smith     for (i = 0; i < sf->nroots; i++) inoffset[i] -= indegree[i]; /* Undo the increment */
119776bd3646SJed Brown     if (PetscDefined(USE_DEBUG)) {                               /* Check that the expected number of increments occurred */
1198ad540459SPierre Jolivet       for (i = 0; i < sf->nroots; i++) PetscCheck(inoffset[i] + indegree[i] == inoffset[i + 1], PETSC_COMM_SELF, PETSC_ERR_PLIB, "Incorrect result after PetscSFFetchAndOp");
119976bd3646SJed Brown     }
12009566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(sf->nleaves, &remote));
120195fce210SBarry Smith     for (i = 0; i < sf->nleaves; i++) {
120295fce210SBarry Smith       remote[i].rank  = sf->remote[i].rank;
120338e7336fSToby Isaac       remote[i].index = outoffset[sf->mine ? sf->mine[i] : i];
120495fce210SBarry Smith     }
12059566063dSJacob Faibussowitsch     PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_RANKS, &sf->multi));
1206013b3241SStefano Zampini     sf->multi->multi = sf->multi;
12079566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(sf->multi, inoffset[sf->nroots], sf->nleaves, sf->mine, PETSC_COPY_VALUES, remote, PETSC_OWN_POINTER));
120895fce210SBarry Smith     if (sf->rankorder) { /* Sort the ranks */
120995fce210SBarry Smith       PetscMPIInt  rank;
121095fce210SBarry Smith       PetscInt    *inranks, *newoffset, *outranks, *newoutoffset, *tmpoffset, maxdegree;
121195fce210SBarry Smith       PetscSFNode *newremote;
12129566063dSJacob Faibussowitsch       PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank));
121395fce210SBarry Smith       for (i = 0, maxdegree = 0; i < sf->nroots; i++) maxdegree = PetscMax(maxdegree, indegree[i]);
12149566063dSJacob Faibussowitsch       PetscCall(PetscMalloc5(sf->multi->nroots, &inranks, sf->multi->nroots, &newoffset, maxlocal, &outranks, maxlocal, &newoutoffset, maxdegree, &tmpoffset));
12159837ea96SMatthew G. Knepley       for (i = 0; i < maxlocal; i++) outranks[i] = rank;
12169566063dSJacob Faibussowitsch       PetscCall(PetscSFReduceBegin(sf->multi, MPIU_INT, outranks, inranks, MPI_REPLACE));
12179566063dSJacob Faibussowitsch       PetscCall(PetscSFReduceEnd(sf->multi, MPIU_INT, outranks, inranks, MPI_REPLACE));
121895fce210SBarry Smith       /* Sort the incoming ranks at each vertex, build the inverse map */
121995fce210SBarry Smith       for (i = 0; i < sf->nroots; i++) {
122095fce210SBarry Smith         PetscInt j;
122195fce210SBarry Smith         for (j = 0; j < indegree[i]; j++) tmpoffset[j] = j;
12229566063dSJacob Faibussowitsch         PetscCall(PetscSortIntWithArray(indegree[i], inranks + inoffset[i], tmpoffset));
122395fce210SBarry Smith         for (j = 0; j < indegree[i]; j++) newoffset[inoffset[i] + tmpoffset[j]] = inoffset[i] + j;
122495fce210SBarry Smith       }
12259566063dSJacob Faibussowitsch       PetscCall(PetscSFBcastBegin(sf->multi, MPIU_INT, newoffset, newoutoffset, MPI_REPLACE));
12269566063dSJacob Faibussowitsch       PetscCall(PetscSFBcastEnd(sf->multi, MPIU_INT, newoffset, newoutoffset, MPI_REPLACE));
12279566063dSJacob Faibussowitsch       PetscCall(PetscMalloc1(sf->nleaves, &newremote));
122895fce210SBarry Smith       for (i = 0; i < sf->nleaves; i++) {
122995fce210SBarry Smith         newremote[i].rank  = sf->remote[i].rank;
123001365b40SToby Isaac         newremote[i].index = newoutoffset[sf->mine ? sf->mine[i] : i];
123195fce210SBarry Smith       }
12329566063dSJacob Faibussowitsch       PetscCall(PetscSFSetGraph(sf->multi, inoffset[sf->nroots], sf->nleaves, sf->mine, PETSC_COPY_VALUES, newremote, PETSC_OWN_POINTER));
12339566063dSJacob Faibussowitsch       PetscCall(PetscFree5(inranks, newoffset, outranks, newoutoffset, tmpoffset));
123495fce210SBarry Smith     }
12359566063dSJacob Faibussowitsch     PetscCall(PetscFree3(inoffset, outones, outoffset));
123695fce210SBarry Smith   }
123795fce210SBarry Smith   *multi = sf->multi;
123895fce210SBarry Smith   PetscFunctionReturn(0);
123995fce210SBarry Smith }
124095fce210SBarry Smith 
124195fce210SBarry Smith /*@C
124272502a1fSJunchao Zhang    PetscSFCreateEmbeddedRootSF - removes edges from all but the selected roots, does not remap indices
124395fce210SBarry Smith 
124495fce210SBarry Smith    Collective
124595fce210SBarry Smith 
12464165533cSJose E. Roman    Input Parameters:
124795fce210SBarry Smith +  sf - original star forest
1248ba2a7774SJunchao Zhang .  nselected  - number of selected roots on this process
1249ba2a7774SJunchao Zhang -  selected   - indices of the selected roots on this process
125095fce210SBarry Smith 
12514165533cSJose E. Roman    Output Parameter:
1252cd620004SJunchao Zhang .  esf - new star forest
125395fce210SBarry Smith 
125495fce210SBarry Smith    Level: advanced
125595fce210SBarry Smith 
125695fce210SBarry Smith    Note:
1257cab54364SBarry Smith    To use the new `PetscSF`, it may be necessary to know the indices of the leaves that are still participating. This can
125895fce210SBarry Smith    be done by calling PetscSFGetGraph().
125995fce210SBarry Smith 
1260cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFGetGraph()`
126195fce210SBarry Smith @*/
1262d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateEmbeddedRootSF(PetscSF sf, PetscInt nselected, const PetscInt *selected, PetscSF *esf)
1263d71ae5a4SJacob Faibussowitsch {
1264cd620004SJunchao Zhang   PetscInt           i, j, n, nroots, nleaves, esf_nleaves, *new_ilocal, minleaf, maxleaf, maxlocal;
1265cd620004SJunchao Zhang   const PetscInt    *ilocal;
1266cd620004SJunchao Zhang   signed char       *rootdata, *leafdata, *leafmem;
1267ba2a7774SJunchao Zhang   const PetscSFNode *iremote;
1268f659e5c7SJunchao Zhang   PetscSFNode       *new_iremote;
1269f659e5c7SJunchao Zhang   MPI_Comm           comm;
127095fce210SBarry Smith 
127195fce210SBarry Smith   PetscFunctionBegin;
127295fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
127329046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf, 1);
1274dadcf809SJacob Faibussowitsch   if (nselected) PetscValidIntPointer(selected, 3);
1275cd620004SJunchao Zhang   PetscValidPointer(esf, 4);
12760511a646SMatthew G. Knepley 
12779566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
12789566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_EmbedSF, sf, 0, 0, 0));
12799566063dSJacob Faibussowitsch   PetscCall(PetscObjectGetComm((PetscObject)sf, &comm));
12809566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, &iremote));
1281cd620004SJunchao Zhang 
128276bd3646SJed Brown   if (PetscDefined(USE_DEBUG)) { /* Error out if selected[] has dups or  out of range indices */
1283cd620004SJunchao Zhang     PetscBool dups;
12849566063dSJacob Faibussowitsch     PetscCall(PetscCheckDupsInt(nselected, selected, &dups));
128528b400f6SJacob Faibussowitsch     PetscCheck(!dups, comm, PETSC_ERR_ARG_WRONG, "selected[] has dups");
12869371c9d4SSatish Balay     for (i = 0; i < nselected; i++) PetscCheck(selected[i] >= 0 && selected[i] < nroots, comm, PETSC_ERR_ARG_OUTOFRANGE, "selected root indice %" PetscInt_FMT " is out of [0,%" PetscInt_FMT ")", selected[i], nroots);
1287cd620004SJunchao Zhang   }
1288f659e5c7SJunchao Zhang 
1289dbbe0bcdSBarry Smith   if (sf->ops->CreateEmbeddedRootSF) PetscUseTypeMethod(sf, CreateEmbeddedRootSF, nselected, selected, esf);
1290dbbe0bcdSBarry Smith   else {
1291cd620004SJunchao Zhang     /* A generic version of creating embedded sf */
12929566063dSJacob Faibussowitsch     PetscCall(PetscSFGetLeafRange(sf, &minleaf, &maxleaf));
1293cd620004SJunchao Zhang     maxlocal = maxleaf - minleaf + 1;
12949566063dSJacob Faibussowitsch     PetscCall(PetscCalloc2(nroots, &rootdata, maxlocal, &leafmem));
1295cd620004SJunchao Zhang     leafdata = leafmem - minleaf;
1296cd620004SJunchao Zhang     /* Tag selected roots and bcast to leaves */
1297cd620004SJunchao Zhang     for (i = 0; i < nselected; i++) rootdata[selected[i]] = 1;
12989566063dSJacob Faibussowitsch     PetscCall(PetscSFBcastBegin(sf, MPI_SIGNED_CHAR, rootdata, leafdata, MPI_REPLACE));
12999566063dSJacob Faibussowitsch     PetscCall(PetscSFBcastEnd(sf, MPI_SIGNED_CHAR, rootdata, leafdata, MPI_REPLACE));
1300ba2a7774SJunchao Zhang 
1301cd620004SJunchao Zhang     /* Build esf with leaves that are still connected */
1302cd620004SJunchao Zhang     esf_nleaves = 0;
1303cd620004SJunchao Zhang     for (i = 0; i < nleaves; i++) {
1304cd620004SJunchao Zhang       j = ilocal ? ilocal[i] : i;
1305cd620004SJunchao Zhang       /* esf_nleaves += leafdata[j] should work in theory, but failed with SFWindow bugs
1306cd620004SJunchao Zhang          with PetscSFBcast. See https://gitlab.com/petsc/petsc/issues/555
1307cd620004SJunchao Zhang       */
1308cd620004SJunchao Zhang       esf_nleaves += (leafdata[j] ? 1 : 0);
1309cd620004SJunchao Zhang     }
13109566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(esf_nleaves, &new_ilocal));
13119566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(esf_nleaves, &new_iremote));
1312cd620004SJunchao Zhang     for (i = n = 0; i < nleaves; i++) {
1313cd620004SJunchao Zhang       j = ilocal ? ilocal[i] : i;
1314cd620004SJunchao Zhang       if (leafdata[j]) {
1315cd620004SJunchao Zhang         new_ilocal[n]        = j;
1316cd620004SJunchao Zhang         new_iremote[n].rank  = iremote[i].rank;
1317cd620004SJunchao Zhang         new_iremote[n].index = iremote[i].index;
1318fc1ede2bSMatthew G. Knepley         ++n;
131995fce210SBarry Smith       }
132095fce210SBarry Smith     }
13219566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(comm, esf));
13229566063dSJacob Faibussowitsch     PetscCall(PetscSFSetFromOptions(*esf));
13239566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(*esf, nroots, esf_nleaves, new_ilocal, PETSC_OWN_POINTER, new_iremote, PETSC_OWN_POINTER));
13249566063dSJacob Faibussowitsch     PetscCall(PetscFree2(rootdata, leafmem));
1325f659e5c7SJunchao Zhang   }
13269566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_EmbedSF, sf, 0, 0, 0));
132795fce210SBarry Smith   PetscFunctionReturn(0);
132895fce210SBarry Smith }
132995fce210SBarry Smith 
13302f5fb4c2SMatthew G. Knepley /*@C
13312f5fb4c2SMatthew G. Knepley   PetscSFCreateEmbeddedLeafSF - removes edges from all but the selected leaves, does not remap indices
13322f5fb4c2SMatthew G. Knepley 
13332f5fb4c2SMatthew G. Knepley   Collective
13342f5fb4c2SMatthew G. Knepley 
13354165533cSJose E. Roman   Input Parameters:
13362f5fb4c2SMatthew G. Knepley + sf - original star forest
1337f659e5c7SJunchao Zhang . nselected  - number of selected leaves on this process
1338f659e5c7SJunchao Zhang - selected   - indices of the selected leaves on this process
13392f5fb4c2SMatthew G. Knepley 
13404165533cSJose E. Roman   Output Parameter:
13412f5fb4c2SMatthew G. Knepley .  newsf - new star forest
13422f5fb4c2SMatthew G. Knepley 
13432f5fb4c2SMatthew G. Knepley   Level: advanced
13442f5fb4c2SMatthew G. Knepley 
1345cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreateEmbeddedRootSF()`, `PetscSFSetGraph()`, `PetscSFGetGraph()`
13462f5fb4c2SMatthew G. Knepley @*/
1347d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateEmbeddedLeafSF(PetscSF sf, PetscInt nselected, const PetscInt *selected, PetscSF *newsf)
1348d71ae5a4SJacob Faibussowitsch {
1349f659e5c7SJunchao Zhang   const PetscSFNode *iremote;
1350f659e5c7SJunchao Zhang   PetscSFNode       *new_iremote;
1351f659e5c7SJunchao Zhang   const PetscInt    *ilocal;
1352f659e5c7SJunchao Zhang   PetscInt           i, nroots, *leaves, *new_ilocal;
1353f659e5c7SJunchao Zhang   MPI_Comm           comm;
13542f5fb4c2SMatthew G. Knepley 
13552f5fb4c2SMatthew G. Knepley   PetscFunctionBegin;
13562f5fb4c2SMatthew G. Knepley   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
135729046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf, 1);
1358dadcf809SJacob Faibussowitsch   if (nselected) PetscValidIntPointer(selected, 3);
13592f5fb4c2SMatthew G. Knepley   PetscValidPointer(newsf, 4);
13602f5fb4c2SMatthew G. Knepley 
1361f659e5c7SJunchao Zhang   /* Uniq selected[] and put results in leaves[] */
13629566063dSJacob Faibussowitsch   PetscCall(PetscObjectGetComm((PetscObject)sf, &comm));
13639566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(nselected, &leaves));
13649566063dSJacob Faibussowitsch   PetscCall(PetscArraycpy(leaves, selected, nselected));
13659566063dSJacob Faibussowitsch   PetscCall(PetscSortedRemoveDupsInt(&nselected, leaves));
136608401ef6SPierre Jolivet   PetscCheck(!nselected || !(leaves[0] < 0 || leaves[nselected - 1] >= sf->nleaves), comm, PETSC_ERR_ARG_OUTOFRANGE, "Min/Max leaf indices %" PetscInt_FMT "/%" PetscInt_FMT " are not in [0,%" PetscInt_FMT ")", leaves[0], leaves[nselected - 1], sf->nleaves);
1367f659e5c7SJunchao Zhang 
1368f659e5c7SJunchao Zhang   /* Optimize the routine only when sf is setup and hence we can reuse sf's communication pattern */
1369dbbe0bcdSBarry Smith   if (sf->setupcalled && sf->ops->CreateEmbeddedLeafSF) PetscUseTypeMethod(sf, CreateEmbeddedLeafSF, nselected, leaves, newsf);
1370dbbe0bcdSBarry Smith   else {
13719566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sf, &nroots, NULL, &ilocal, &iremote));
13729566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nselected, &new_ilocal));
13739566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nselected, &new_iremote));
1374f659e5c7SJunchao Zhang     for (i = 0; i < nselected; ++i) {
1375f659e5c7SJunchao Zhang       const PetscInt l     = leaves[i];
1376f659e5c7SJunchao Zhang       new_ilocal[i]        = ilocal ? ilocal[l] : l;
1377f659e5c7SJunchao Zhang       new_iremote[i].rank  = iremote[l].rank;
1378f659e5c7SJunchao Zhang       new_iremote[i].index = iremote[l].index;
13792f5fb4c2SMatthew G. Knepley     }
13809566063dSJacob Faibussowitsch     PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_CONFONLY, newsf));
13819566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(*newsf, nroots, nselected, new_ilocal, PETSC_OWN_POINTER, new_iremote, PETSC_OWN_POINTER));
1382f659e5c7SJunchao Zhang   }
13839566063dSJacob Faibussowitsch   PetscCall(PetscFree(leaves));
13842f5fb4c2SMatthew G. Knepley   PetscFunctionReturn(0);
13852f5fb4c2SMatthew G. Knepley }
13862f5fb4c2SMatthew G. Knepley 
138795fce210SBarry Smith /*@C
1388cab54364SBarry Smith    PetscSFBcastBegin - begin pointwise broadcast with root value being reduced to leaf value, to be concluded with call to `PetscSFBcastEnd()`
13893482bfa8SJunchao Zhang 
1390cab54364SBarry Smith    Collective on sf
13913482bfa8SJunchao Zhang 
13924165533cSJose E. Roman    Input Parameters:
13933482bfa8SJunchao Zhang +  sf - star forest on which to communicate
13943482bfa8SJunchao Zhang .  unit - data type associated with each node
13953482bfa8SJunchao Zhang .  rootdata - buffer to broadcast
13963482bfa8SJunchao Zhang -  op - operation to use for reduction
13973482bfa8SJunchao Zhang 
13984165533cSJose E. Roman    Output Parameter:
13993482bfa8SJunchao Zhang .  leafdata - buffer to be reduced with values from each leaf's respective root
14003482bfa8SJunchao Zhang 
14013482bfa8SJunchao Zhang    Level: intermediate
14023482bfa8SJunchao Zhang 
1403d0295fc0SJunchao Zhang    Notes:
1404d0295fc0SJunchao Zhang     When petsc is configured with device support, it will use its own mechanism to figure out whether the given data pointers
1405d0295fc0SJunchao Zhang     are host pointers or device pointers, which may incur a noticable cost. If you already knew the info, you should
1406cab54364SBarry Smith     use `PetscSFBcastWithMemTypeBegin()` instead.
1407cab54364SBarry Smith 
1408cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastEnd()`, `PetscSFBcastWithMemTypeBegin()`
14093482bfa8SJunchao Zhang @*/
1410d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastBegin(PetscSF sf, MPI_Datatype unit, const void *rootdata, void *leafdata, MPI_Op op)
1411d71ae5a4SJacob Faibussowitsch {
1412eb02082bSJunchao Zhang   PetscMemType rootmtype, leafmtype;
14133482bfa8SJunchao Zhang 
14143482bfa8SJunchao Zhang   PetscFunctionBegin;
14153482bfa8SJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
14169566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
14179566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin, sf, 0, 0, 0));
14189566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(rootdata, &rootmtype));
14199566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafdata, &leafmtype));
1420dbbe0bcdSBarry Smith   PetscUseTypeMethod(sf, BcastBegin, unit, rootmtype, rootdata, leafmtype, leafdata, op);
14219566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin, sf, 0, 0, 0));
14223482bfa8SJunchao Zhang   PetscFunctionReturn(0);
14233482bfa8SJunchao Zhang }
14243482bfa8SJunchao Zhang 
14253482bfa8SJunchao Zhang /*@C
1426cab54364SBarry Smith    PetscSFBcastWithMemTypeBegin - begin pointwise broadcast with root value being reduced to leaf value with explicit memory types, to be concluded with call to `PetscSFBcastEnd()`
1427d0295fc0SJunchao Zhang 
1428cab54364SBarry Smith    Collective on sf
1429d0295fc0SJunchao Zhang 
14304165533cSJose E. Roman    Input Parameters:
1431d0295fc0SJunchao Zhang +  sf - star forest on which to communicate
1432d0295fc0SJunchao Zhang .  unit - data type associated with each node
1433d0295fc0SJunchao Zhang .  rootmtype - memory type of rootdata
1434d0295fc0SJunchao Zhang .  rootdata - buffer to broadcast
1435d0295fc0SJunchao Zhang .  leafmtype - memory type of leafdata
1436d0295fc0SJunchao Zhang -  op - operation to use for reduction
1437d0295fc0SJunchao Zhang 
14384165533cSJose E. Roman    Output Parameter:
1439d0295fc0SJunchao Zhang .  leafdata - buffer to be reduced with values from each leaf's respective root
1440d0295fc0SJunchao Zhang 
1441d0295fc0SJunchao Zhang    Level: intermediate
1442d0295fc0SJunchao Zhang 
1443cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastEnd()`, `PetscSFBcastBegin()`
1444d0295fc0SJunchao Zhang @*/
1445d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastWithMemTypeBegin(PetscSF sf, MPI_Datatype unit, PetscMemType rootmtype, const void *rootdata, PetscMemType leafmtype, void *leafdata, MPI_Op op)
1446d71ae5a4SJacob Faibussowitsch {
1447d0295fc0SJunchao Zhang   PetscFunctionBegin;
1448d0295fc0SJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
14499566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
14509566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin, sf, 0, 0, 0));
1451dbbe0bcdSBarry Smith   PetscUseTypeMethod(sf, BcastBegin, unit, rootmtype, rootdata, leafmtype, leafdata, op);
14529566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin, sf, 0, 0, 0));
1453d0295fc0SJunchao Zhang   PetscFunctionReturn(0);
1454d0295fc0SJunchao Zhang }
1455d0295fc0SJunchao Zhang 
1456d0295fc0SJunchao Zhang /*@C
1457cab54364SBarry Smith    PetscSFBcastEnd - end a broadcast & reduce operation started with `PetscSFBcastBegin()`
14583482bfa8SJunchao Zhang 
14593482bfa8SJunchao Zhang    Collective
14603482bfa8SJunchao Zhang 
14614165533cSJose E. Roman    Input Parameters:
14623482bfa8SJunchao Zhang +  sf - star forest
14633482bfa8SJunchao Zhang .  unit - data type
14643482bfa8SJunchao Zhang .  rootdata - buffer to broadcast
14653482bfa8SJunchao Zhang -  op - operation to use for reduction
14663482bfa8SJunchao Zhang 
14674165533cSJose E. Roman    Output Parameter:
14683482bfa8SJunchao Zhang .  leafdata - buffer to be reduced with values from each leaf's respective root
14693482bfa8SJunchao Zhang 
14703482bfa8SJunchao Zhang    Level: intermediate
14713482bfa8SJunchao Zhang 
1472cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFReduceEnd()`
14733482bfa8SJunchao Zhang @*/
1474d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastEnd(PetscSF sf, MPI_Datatype unit, const void *rootdata, void *leafdata, MPI_Op op)
1475d71ae5a4SJacob Faibussowitsch {
14763482bfa8SJunchao Zhang   PetscFunctionBegin;
14773482bfa8SJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
14789566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastEnd, sf, 0, 0, 0));
1479dbbe0bcdSBarry Smith   PetscUseTypeMethod(sf, BcastEnd, unit, rootdata, leafdata, op);
14809566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastEnd, sf, 0, 0, 0));
14813482bfa8SJunchao Zhang   PetscFunctionReturn(0);
14823482bfa8SJunchao Zhang }
14833482bfa8SJunchao Zhang 
14843482bfa8SJunchao Zhang /*@C
1485cab54364SBarry Smith    PetscSFReduceBegin - begin reduction of leafdata into rootdata, to be completed with call to `PetscSFReduceEnd()`
148695fce210SBarry Smith 
148795fce210SBarry Smith    Collective
148895fce210SBarry Smith 
14894165533cSJose E. Roman    Input Parameters:
149095fce210SBarry Smith +  sf - star forest
149195fce210SBarry Smith .  unit - data type
149295fce210SBarry Smith .  leafdata - values to reduce
149395fce210SBarry Smith -  op - reduction operation
149495fce210SBarry Smith 
14954165533cSJose E. Roman    Output Parameter:
149695fce210SBarry Smith .  rootdata - result of reduction of values from all leaves of each root
149795fce210SBarry Smith 
149895fce210SBarry Smith    Level: intermediate
149995fce210SBarry Smith 
1500d0295fc0SJunchao Zhang    Notes:
1501d0295fc0SJunchao Zhang     When petsc is configured with device support, it will use its own mechanism to figure out whether the given data pointers
1502d0295fc0SJunchao Zhang     are host pointers or device pointers, which may incur a noticable cost. If you already knew the info, you should
1503cab54364SBarry Smith     use `PetscSFReduceWithMemTypeBegin()` instead.
1504d0295fc0SJunchao Zhang 
1505cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastBegin()`, `PetscSFReduceWithMemTypeBegin()`
150695fce210SBarry Smith @*/
1507d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReduceBegin(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *rootdata, MPI_Op op)
1508d71ae5a4SJacob Faibussowitsch {
1509eb02082bSJunchao Zhang   PetscMemType rootmtype, leafmtype;
151095fce210SBarry Smith 
151195fce210SBarry Smith   PetscFunctionBegin;
151295fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
15139566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
15149566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceBegin, sf, 0, 0, 0));
15159566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(rootdata, &rootmtype));
15169566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafdata, &leafmtype));
15179566063dSJacob Faibussowitsch   PetscCall((sf->ops->ReduceBegin)(sf, unit, leafmtype, leafdata, rootmtype, rootdata, op));
15189566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceBegin, sf, 0, 0, 0));
151995fce210SBarry Smith   PetscFunctionReturn(0);
152095fce210SBarry Smith }
152195fce210SBarry Smith 
152295fce210SBarry Smith /*@C
1523cab54364SBarry Smith    PetscSFReduceWithMemTypeBegin - begin reduction of leafdata into rootdata with explicit memory types, to be completed with call to `PetscSFReduceEnd()`
1524d0295fc0SJunchao Zhang 
1525d0295fc0SJunchao Zhang    Collective
1526d0295fc0SJunchao Zhang 
15274165533cSJose E. Roman    Input Parameters:
1528d0295fc0SJunchao Zhang +  sf - star forest
1529d0295fc0SJunchao Zhang .  unit - data type
1530d0295fc0SJunchao Zhang .  leafmtype - memory type of leafdata
1531d0295fc0SJunchao Zhang .  leafdata - values to reduce
1532d0295fc0SJunchao Zhang .  rootmtype - memory type of rootdata
1533d0295fc0SJunchao Zhang -  op - reduction operation
1534d0295fc0SJunchao Zhang 
15354165533cSJose E. Roman    Output Parameter:
1536d0295fc0SJunchao Zhang .  rootdata - result of reduction of values from all leaves of each root
1537d0295fc0SJunchao Zhang 
1538d0295fc0SJunchao Zhang    Level: intermediate
1539d0295fc0SJunchao Zhang 
1540cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastBegin()`, `PetscSFReduceBegin()`
1541d0295fc0SJunchao Zhang @*/
1542d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReduceWithMemTypeBegin(PetscSF sf, MPI_Datatype unit, PetscMemType leafmtype, const void *leafdata, PetscMemType rootmtype, void *rootdata, MPI_Op op)
1543d71ae5a4SJacob Faibussowitsch {
1544d0295fc0SJunchao Zhang   PetscFunctionBegin;
1545d0295fc0SJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
15469566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
15479566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceBegin, sf, 0, 0, 0));
15489566063dSJacob Faibussowitsch   PetscCall((sf->ops->ReduceBegin)(sf, unit, leafmtype, leafdata, rootmtype, rootdata, op));
15499566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceBegin, sf, 0, 0, 0));
1550d0295fc0SJunchao Zhang   PetscFunctionReturn(0);
1551d0295fc0SJunchao Zhang }
1552d0295fc0SJunchao Zhang 
1553d0295fc0SJunchao Zhang /*@C
1554cab54364SBarry Smith    PetscSFReduceEnd - end a reduction operation started with `PetscSFReduceBegin()`
155595fce210SBarry Smith 
155695fce210SBarry Smith    Collective
155795fce210SBarry Smith 
15584165533cSJose E. Roman    Input Parameters:
155995fce210SBarry Smith +  sf - star forest
156095fce210SBarry Smith .  unit - data type
156195fce210SBarry Smith .  leafdata - values to reduce
156295fce210SBarry Smith -  op - reduction operation
156395fce210SBarry Smith 
15644165533cSJose E. Roman    Output Parameter:
156595fce210SBarry Smith .  rootdata - result of reduction of values from all leaves of each root
156695fce210SBarry Smith 
156795fce210SBarry Smith    Level: intermediate
156895fce210SBarry Smith 
1569cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFBcastEnd()`
157095fce210SBarry Smith @*/
1571d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReduceEnd(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *rootdata, MPI_Op op)
1572d71ae5a4SJacob Faibussowitsch {
157395fce210SBarry Smith   PetscFunctionBegin;
157495fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
15759566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceEnd, sf, 0, 0, 0));
1576dbbe0bcdSBarry Smith   PetscUseTypeMethod(sf, ReduceEnd, unit, leafdata, rootdata, op);
15779566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceEnd, sf, 0, 0, 0));
157895fce210SBarry Smith   PetscFunctionReturn(0);
157995fce210SBarry Smith }
158095fce210SBarry Smith 
158195fce210SBarry Smith /*@C
1582cab54364SBarry Smith    PetscSFFetchAndOpBegin - begin operation that fetches values from root and updates atomically by applying operation using my leaf value,
1583cab54364SBarry Smith    to be completed with `PetscSFFetchAndOpEnd()`
1584a1729e3fSJunchao Zhang 
1585a1729e3fSJunchao Zhang    Collective
1586a1729e3fSJunchao Zhang 
15874165533cSJose E. Roman    Input Parameters:
1588a1729e3fSJunchao Zhang +  sf - star forest
1589a1729e3fSJunchao Zhang .  unit - data type
1590a1729e3fSJunchao Zhang .  leafdata - leaf values to use in reduction
1591a1729e3fSJunchao Zhang -  op - operation to use for reduction
1592a1729e3fSJunchao Zhang 
15934165533cSJose E. Roman    Output Parameters:
1594a1729e3fSJunchao Zhang +  rootdata - root values to be updated, input state is seen by first process to perform an update
1595a1729e3fSJunchao Zhang -  leafupdate - state at each leaf's respective root immediately prior to my atomic update
1596a1729e3fSJunchao Zhang 
1597a1729e3fSJunchao Zhang    Level: advanced
1598a1729e3fSJunchao Zhang 
1599a1729e3fSJunchao Zhang    Note:
1600a1729e3fSJunchao Zhang    The update is only atomic at the granularity provided by the hardware. Different roots referenced by the same process
1601a1729e3fSJunchao Zhang    might be updated in a different order. Furthermore, if a composite type is used for the unit datatype, atomicity is
1602a1729e3fSJunchao Zhang    not guaranteed across the whole vertex. Therefore, this function is mostly only used with primitive types such as
1603a1729e3fSJunchao Zhang    integers.
1604a1729e3fSJunchao Zhang 
1605cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFReduceBegin()`, `PetscSFSetGraph()`
1606a1729e3fSJunchao Zhang @*/
1607d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFFetchAndOpBegin(PetscSF sf, MPI_Datatype unit, void *rootdata, const void *leafdata, void *leafupdate, MPI_Op op)
1608d71ae5a4SJacob Faibussowitsch {
1609eb02082bSJunchao Zhang   PetscMemType rootmtype, leafmtype, leafupdatemtype;
1610a1729e3fSJunchao Zhang 
1611a1729e3fSJunchao Zhang   PetscFunctionBegin;
1612a1729e3fSJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
16139566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
16149566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0));
16159566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(rootdata, &rootmtype));
16169566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafdata, &leafmtype));
16179566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafupdate, &leafupdatemtype));
161808401ef6SPierre Jolivet   PetscCheck(leafmtype == leafupdatemtype, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for leafdata and leafupdate in different memory types");
1619dbbe0bcdSBarry Smith   PetscUseTypeMethod(sf, FetchAndOpBegin, unit, rootmtype, rootdata, leafmtype, leafdata, leafupdate, op);
16209566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0));
1621a1729e3fSJunchao Zhang   PetscFunctionReturn(0);
1622a1729e3fSJunchao Zhang }
1623a1729e3fSJunchao Zhang 
1624a1729e3fSJunchao Zhang /*@C
1625cab54364SBarry Smith    PetscSFFetchAndOpWithMemTypeBegin - begin operation with explicit memory types that fetches values from root and updates atomically by
1626cab54364SBarry Smith    applying operation using my leaf value, to be completed with `PetscSFFetchAndOpEnd()`
1627d3b3e55cSJunchao Zhang 
1628d3b3e55cSJunchao Zhang    Collective
1629d3b3e55cSJunchao Zhang 
1630d3b3e55cSJunchao Zhang    Input Parameters:
1631d3b3e55cSJunchao Zhang +  sf - star forest
1632d3b3e55cSJunchao Zhang .  unit - data type
1633d3b3e55cSJunchao Zhang .  rootmtype - memory type of rootdata
1634d3b3e55cSJunchao Zhang .  leafmtype - memory type of leafdata
1635d3b3e55cSJunchao Zhang .  leafdata - leaf values to use in reduction
1636d3b3e55cSJunchao Zhang .  leafupdatemtype - memory type of leafupdate
1637d3b3e55cSJunchao Zhang -  op - operation to use for reduction
1638d3b3e55cSJunchao Zhang 
1639d3b3e55cSJunchao Zhang    Output Parameters:
1640d3b3e55cSJunchao Zhang +  rootdata - root values to be updated, input state is seen by first process to perform an update
1641d3b3e55cSJunchao Zhang -  leafupdate - state at each leaf's respective root immediately prior to my atomic update
1642d3b3e55cSJunchao Zhang 
1643d3b3e55cSJunchao Zhang    Level: advanced
1644d3b3e55cSJunchao Zhang 
1645cab54364SBarry Smith    Note:
1646cab54364SBarry Smith    See `PetscSFFetchAndOpBegin()` for more details.
1647d3b3e55cSJunchao Zhang 
1648cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFFetchAndOpBegin()`, `PetscSFComputeDegreeBegin()`, `PetscSFReduceBegin()`, `PetscSFSetGraph()`
1649d3b3e55cSJunchao Zhang @*/
1650d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFFetchAndOpWithMemTypeBegin(PetscSF sf, MPI_Datatype unit, PetscMemType rootmtype, void *rootdata, PetscMemType leafmtype, const void *leafdata, PetscMemType leafupdatemtype, void *leafupdate, MPI_Op op)
1651d71ae5a4SJacob Faibussowitsch {
1652d3b3e55cSJunchao Zhang   PetscFunctionBegin;
1653d3b3e55cSJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
16549566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
16559566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0));
165608401ef6SPierre Jolivet   PetscCheck(leafmtype == leafupdatemtype, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for leafdata and leafupdate in different memory types");
1657dbbe0bcdSBarry Smith   PetscUseTypeMethod(sf, FetchAndOpBegin, unit, rootmtype, rootdata, leafmtype, leafdata, leafupdate, op);
16589566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0));
1659d3b3e55cSJunchao Zhang   PetscFunctionReturn(0);
1660d3b3e55cSJunchao Zhang }
1661d3b3e55cSJunchao Zhang 
1662d3b3e55cSJunchao Zhang /*@C
1663a1729e3fSJunchao Zhang    PetscSFFetchAndOpEnd - end operation started in matching call to PetscSFFetchAndOpBegin() to fetch values from roots and update atomically by applying operation using my leaf value
1664a1729e3fSJunchao Zhang 
1665a1729e3fSJunchao Zhang    Collective
1666a1729e3fSJunchao Zhang 
16674165533cSJose E. Roman    Input Parameters:
1668a1729e3fSJunchao Zhang +  sf - star forest
1669a1729e3fSJunchao Zhang .  unit - data type
1670a1729e3fSJunchao Zhang .  leafdata - leaf values to use in reduction
1671a1729e3fSJunchao Zhang -  op - operation to use for reduction
1672a1729e3fSJunchao Zhang 
16734165533cSJose E. Roman    Output Parameters:
1674a1729e3fSJunchao Zhang +  rootdata - root values to be updated, input state is seen by first process to perform an update
1675a1729e3fSJunchao Zhang -  leafupdate - state at each leaf's respective root immediately prior to my atomic update
1676a1729e3fSJunchao Zhang 
1677a1729e3fSJunchao Zhang    Level: advanced
1678a1729e3fSJunchao Zhang 
1679cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeEnd()`, `PetscSFReduceEnd()`, `PetscSFSetGraph()`
1680a1729e3fSJunchao Zhang @*/
1681d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFFetchAndOpEnd(PetscSF sf, MPI_Datatype unit, void *rootdata, const void *leafdata, void *leafupdate, MPI_Op op)
1682d71ae5a4SJacob Faibussowitsch {
1683a1729e3fSJunchao Zhang   PetscFunctionBegin;
1684a1729e3fSJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
16859566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpEnd, sf, 0, 0, 0));
1686dbbe0bcdSBarry Smith   PetscUseTypeMethod(sf, FetchAndOpEnd, unit, rootdata, leafdata, leafupdate, op);
16879566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpEnd, sf, 0, 0, 0));
1688a1729e3fSJunchao Zhang   PetscFunctionReturn(0);
1689a1729e3fSJunchao Zhang }
1690a1729e3fSJunchao Zhang 
1691a1729e3fSJunchao Zhang /*@C
1692cab54364SBarry Smith    PetscSFComputeDegreeBegin - begin computation of degree for each root vertex, to be completed with `PetscSFComputeDegreeEnd()`
169395fce210SBarry Smith 
169495fce210SBarry Smith    Collective
169595fce210SBarry Smith 
16964165533cSJose E. Roman    Input Parameter:
169795fce210SBarry Smith .  sf - star forest
169895fce210SBarry Smith 
16994165533cSJose E. Roman    Output Parameter:
170095fce210SBarry Smith .  degree - degree of each root vertex
170195fce210SBarry Smith 
170295fce210SBarry Smith    Level: advanced
170395fce210SBarry Smith 
1704cab54364SBarry Smith    Note:
1705cab54364SBarry Smith    The returned array is owned by `PetscSF` and automatically freed by `PetscSFDestroy()`. Hence no need to call `PetscFree()` on it.
1706ffe67aa5SVáclav Hapla 
1707cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGatherBegin()`, `PetscSFComputeDegreeEnd()`
170895fce210SBarry Smith @*/
1709d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComputeDegreeBegin(PetscSF sf, const PetscInt **degree)
1710d71ae5a4SJacob Faibussowitsch {
171195fce210SBarry Smith   PetscFunctionBegin;
171295fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
171395fce210SBarry Smith   PetscSFCheckGraphSet(sf, 1);
171495fce210SBarry Smith   PetscValidPointer(degree, 2);
1715803bd9e8SMatthew G. Knepley   if (!sf->degreeknown) {
17165b0d146aSStefano Zampini     PetscInt i, nroots = sf->nroots, maxlocal;
171728b400f6SJacob Faibussowitsch     PetscCheck(!sf->degree, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Calls to PetscSFComputeDegreeBegin() cannot be nested.");
17185b0d146aSStefano Zampini     maxlocal = sf->maxleaf - sf->minleaf + 1;
17199566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nroots, &sf->degree));
17209566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(PetscMax(maxlocal, 1), &sf->degreetmp)); /* allocate at least one entry, see check in PetscSFComputeDegreeEnd() */
172129046d53SLisandro Dalcin     for (i = 0; i < nroots; i++) sf->degree[i] = 0;
17229837ea96SMatthew G. Knepley     for (i = 0; i < maxlocal; i++) sf->degreetmp[i] = 1;
17239566063dSJacob Faibussowitsch     PetscCall(PetscSFReduceBegin(sf, MPIU_INT, sf->degreetmp - sf->minleaf, sf->degree, MPI_SUM));
172495fce210SBarry Smith   }
172595fce210SBarry Smith   *degree = NULL;
172695fce210SBarry Smith   PetscFunctionReturn(0);
172795fce210SBarry Smith }
172895fce210SBarry Smith 
172995fce210SBarry Smith /*@C
1730cab54364SBarry Smith    PetscSFComputeDegreeEnd - complete computation of degree for each root vertex, started with `PetscSFComputeDegreeBegin()`
173195fce210SBarry Smith 
173295fce210SBarry Smith    Collective
173395fce210SBarry Smith 
17344165533cSJose E. Roman    Input Parameter:
173595fce210SBarry Smith .  sf - star forest
173695fce210SBarry Smith 
17374165533cSJose E. Roman    Output Parameter:
173895fce210SBarry Smith .  degree - degree of each root vertex
173995fce210SBarry Smith 
174095fce210SBarry Smith    Level: developer
174195fce210SBarry Smith 
1742cab54364SBarry Smith    Note:
1743cab54364SBarry Smith    The returned array is owned by `PetscSF` and automatically freed by `PetscSFDestroy()`. Hence no need to call `PetscFree()` on it.
1744ffe67aa5SVáclav Hapla 
1745cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGatherBegin()`, `PetscSFComputeDegreeBegin()`
174695fce210SBarry Smith @*/
1747d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComputeDegreeEnd(PetscSF sf, const PetscInt **degree)
1748d71ae5a4SJacob Faibussowitsch {
174995fce210SBarry Smith   PetscFunctionBegin;
175095fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
175195fce210SBarry Smith   PetscSFCheckGraphSet(sf, 1);
175229046d53SLisandro Dalcin   PetscValidPointer(degree, 2);
175395fce210SBarry Smith   if (!sf->degreeknown) {
175428b400f6SJacob Faibussowitsch     PetscCheck(sf->degreetmp, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFComputeDegreeBegin() before PetscSFComputeDegreeEnd()");
17559566063dSJacob Faibussowitsch     PetscCall(PetscSFReduceEnd(sf, MPIU_INT, sf->degreetmp - sf->minleaf, sf->degree, MPI_SUM));
17569566063dSJacob Faibussowitsch     PetscCall(PetscFree(sf->degreetmp));
175795fce210SBarry Smith     sf->degreeknown = PETSC_TRUE;
175895fce210SBarry Smith   }
175995fce210SBarry Smith   *degree = sf->degree;
176095fce210SBarry Smith   PetscFunctionReturn(0);
176195fce210SBarry Smith }
176295fce210SBarry Smith 
1763673100f5SVaclav Hapla /*@C
1764cab54364SBarry Smith    PetscSFComputeMultiRootOriginalNumbering - Returns original numbering of multi-roots (roots of multi-SF returned by `PetscSFGetMultiSF()`).
176566dfcd1aSVaclav Hapla    Each multi-root is assigned index of the corresponding original root.
1766673100f5SVaclav Hapla 
1767673100f5SVaclav Hapla    Collective
1768673100f5SVaclav Hapla 
17694165533cSJose E. Roman    Input Parameters:
1770673100f5SVaclav Hapla +  sf - star forest
1771cab54364SBarry Smith -  degree - degree of each root vertex, computed with `PetscSFComputeDegreeBegin()`/`PetscSFComputeDegreeEnd()`
1772673100f5SVaclav Hapla 
17734165533cSJose E. Roman    Output Parameters:
177466dfcd1aSVaclav Hapla +  nMultiRoots - (optional) number of multi-roots (roots of multi-SF)
177566dfcd1aSVaclav Hapla -  multiRootsOrigNumbering - original indices of multi-roots; length of this array is nMultiRoots
1776673100f5SVaclav Hapla 
1777673100f5SVaclav Hapla    Level: developer
1778673100f5SVaclav Hapla 
1779cab54364SBarry Smith    Note:
1780cab54364SBarry Smith    The returned array multiRootsOrigNumbering is newly allocated and should be destroyed with `PetscFree()` when no longer needed.
1781ffe67aa5SVáclav Hapla 
1782cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFComputeDegreeEnd()`, `PetscSFGetMultiSF()`
1783673100f5SVaclav Hapla @*/
1784d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComputeMultiRootOriginalNumbering(PetscSF sf, const PetscInt degree[], PetscInt *nMultiRoots, PetscInt *multiRootsOrigNumbering[])
1785d71ae5a4SJacob Faibussowitsch {
1786673100f5SVaclav Hapla   PetscSF  msf;
1787673100f5SVaclav Hapla   PetscInt i, j, k, nroots, nmroots;
1788673100f5SVaclav Hapla 
1789673100f5SVaclav Hapla   PetscFunctionBegin;
1790673100f5SVaclav Hapla   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
17919566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sf, &nroots, NULL, NULL, NULL));
179229328920SVaclav Hapla   if (nroots) PetscValidIntPointer(degree, 2);
179366dfcd1aSVaclav Hapla   if (nMultiRoots) PetscValidIntPointer(nMultiRoots, 3);
179466dfcd1aSVaclav Hapla   PetscValidPointer(multiRootsOrigNumbering, 4);
17959566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf, &msf));
17969566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(msf, &nmroots, NULL, NULL, NULL));
17979566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(nmroots, multiRootsOrigNumbering));
1798673100f5SVaclav Hapla   for (i = 0, j = 0, k = 0; i < nroots; i++) {
1799673100f5SVaclav Hapla     if (!degree[i]) continue;
1800ad540459SPierre Jolivet     for (j = 0; j < degree[i]; j++, k++) (*multiRootsOrigNumbering)[k] = i;
1801673100f5SVaclav Hapla   }
180208401ef6SPierre Jolivet   PetscCheck(k == nmroots, PETSC_COMM_SELF, PETSC_ERR_PLIB, "sanity check fail");
180366dfcd1aSVaclav Hapla   if (nMultiRoots) *nMultiRoots = nmroots;
1804673100f5SVaclav Hapla   PetscFunctionReturn(0);
1805673100f5SVaclav Hapla }
1806673100f5SVaclav Hapla 
180795fce210SBarry Smith /*@C
1808cab54364SBarry Smith    PetscSFGatherBegin - begin pointwise gather of all leaves into multi-roots, to be completed with `PetscSFGatherEnd()`
180995fce210SBarry Smith 
181095fce210SBarry Smith    Collective
181195fce210SBarry Smith 
18124165533cSJose E. Roman    Input Parameters:
181395fce210SBarry Smith +  sf - star forest
181495fce210SBarry Smith .  unit - data type
181595fce210SBarry Smith -  leafdata - leaf data to gather to roots
181695fce210SBarry Smith 
18174165533cSJose E. Roman    Output Parameter:
181895fce210SBarry Smith .  multirootdata - root buffer to gather into, amount of space per root is equal to its degree
181995fce210SBarry Smith 
182095fce210SBarry Smith    Level: intermediate
182195fce210SBarry Smith 
1822cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFScatterBegin()`
182395fce210SBarry Smith @*/
1824d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGatherBegin(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *multirootdata)
1825d71ae5a4SJacob Faibussowitsch {
1826a5526d50SJunchao Zhang   PetscSF multi = NULL;
182795fce210SBarry Smith 
182895fce210SBarry Smith   PetscFunctionBegin;
182995fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
18309566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
18319566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf, &multi));
18329566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceBegin(multi, unit, leafdata, multirootdata, MPI_REPLACE));
183395fce210SBarry Smith   PetscFunctionReturn(0);
183495fce210SBarry Smith }
183595fce210SBarry Smith 
183695fce210SBarry Smith /*@C
1837cab54364SBarry Smith    PetscSFGatherEnd - ends pointwise gather operation that was started with `PetscSFGatherBegin()`
183895fce210SBarry Smith 
183995fce210SBarry Smith    Collective
184095fce210SBarry Smith 
18414165533cSJose E. Roman    Input Parameters:
184295fce210SBarry Smith +  sf - star forest
184395fce210SBarry Smith .  unit - data type
184495fce210SBarry Smith -  leafdata - leaf data to gather to roots
184595fce210SBarry Smith 
18464165533cSJose E. Roman    Output Parameter:
184795fce210SBarry Smith .  multirootdata - root buffer to gather into, amount of space per root is equal to its degree
184895fce210SBarry Smith 
184995fce210SBarry Smith    Level: intermediate
185095fce210SBarry Smith 
1851cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeEnd()`, `PetscSFScatterEnd()`
185295fce210SBarry Smith @*/
1853d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGatherEnd(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *multirootdata)
1854d71ae5a4SJacob Faibussowitsch {
1855a5526d50SJunchao Zhang   PetscSF multi = NULL;
185695fce210SBarry Smith 
185795fce210SBarry Smith   PetscFunctionBegin;
185895fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
18599566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf, &multi));
18609566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceEnd(multi, unit, leafdata, multirootdata, MPI_REPLACE));
186195fce210SBarry Smith   PetscFunctionReturn(0);
186295fce210SBarry Smith }
186395fce210SBarry Smith 
186495fce210SBarry Smith /*@C
1865cab54364SBarry Smith    PetscSFScatterBegin - begin pointwise scatter operation from multi-roots to leaves, to be completed with `PetscSFScatterEnd()`
186695fce210SBarry Smith 
186795fce210SBarry Smith    Collective
186895fce210SBarry Smith 
18694165533cSJose E. Roman    Input Parameters:
187095fce210SBarry Smith +  sf - star forest
187195fce210SBarry Smith .  unit - data type
187295fce210SBarry Smith -  multirootdata - root buffer to send to each leaf, one unit of data per leaf
187395fce210SBarry Smith 
18744165533cSJose E. Roman    Output Parameter:
187595fce210SBarry Smith .  leafdata - leaf data to be update with personal data from each respective root
187695fce210SBarry Smith 
187795fce210SBarry Smith    Level: intermediate
187895fce210SBarry Smith 
1879cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFScatterBegin()`
188095fce210SBarry Smith @*/
1881d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFScatterBegin(PetscSF sf, MPI_Datatype unit, const void *multirootdata, void *leafdata)
1882d71ae5a4SJacob Faibussowitsch {
1883a5526d50SJunchao Zhang   PetscSF multi = NULL;
188495fce210SBarry Smith 
188595fce210SBarry Smith   PetscFunctionBegin;
188695fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
18879566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
18889566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf, &multi));
18899566063dSJacob Faibussowitsch   PetscCall(PetscSFBcastBegin(multi, unit, multirootdata, leafdata, MPI_REPLACE));
189095fce210SBarry Smith   PetscFunctionReturn(0);
189195fce210SBarry Smith }
189295fce210SBarry Smith 
189395fce210SBarry Smith /*@C
1894cab54364SBarry Smith    PetscSFScatterEnd - ends pointwise scatter operation that was started with `PetscSFScatterBegin()`
189595fce210SBarry Smith 
189695fce210SBarry Smith    Collective
189795fce210SBarry Smith 
18984165533cSJose E. Roman    Input Parameters:
189995fce210SBarry Smith +  sf - star forest
190095fce210SBarry Smith .  unit - data type
190195fce210SBarry Smith -  multirootdata - root buffer to send to each leaf, one unit of data per leaf
190295fce210SBarry Smith 
19034165533cSJose E. Roman    Output Parameter:
190495fce210SBarry Smith .  leafdata - leaf data to be update with personal data from each respective root
190595fce210SBarry Smith 
190695fce210SBarry Smith    Level: intermediate
190795fce210SBarry Smith 
1908cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeEnd()`, `PetscSFScatterEnd()`
190995fce210SBarry Smith @*/
1910d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFScatterEnd(PetscSF sf, MPI_Datatype unit, const void *multirootdata, void *leafdata)
1911d71ae5a4SJacob Faibussowitsch {
1912a5526d50SJunchao Zhang   PetscSF multi = NULL;
191395fce210SBarry Smith 
191495fce210SBarry Smith   PetscFunctionBegin;
191595fce210SBarry Smith   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
19169566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf, &multi));
19179566063dSJacob Faibussowitsch   PetscCall(PetscSFBcastEnd(multi, unit, multirootdata, leafdata, MPI_REPLACE));
191895fce210SBarry Smith   PetscFunctionReturn(0);
191995fce210SBarry Smith }
1920a7b3aa13SAta Mesgarnejad 
1921d71ae5a4SJacob Faibussowitsch static PetscErrorCode PetscSFCheckLeavesUnique_Private(PetscSF sf)
1922d71ae5a4SJacob Faibussowitsch {
1923a072220fSLawrence Mitchell   PetscInt        i, n, nleaves;
1924a072220fSLawrence Mitchell   const PetscInt *ilocal = NULL;
1925a072220fSLawrence Mitchell   PetscHSetI      seen;
1926a072220fSLawrence Mitchell 
1927a072220fSLawrence Mitchell   PetscFunctionBegin;
1928b458e8f1SJose E. Roman   if (PetscDefined(USE_DEBUG)) {
19299566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sf, NULL, &nleaves, &ilocal, NULL));
19309566063dSJacob Faibussowitsch     PetscCall(PetscHSetICreate(&seen));
1931a072220fSLawrence Mitchell     for (i = 0; i < nleaves; i++) {
1932a072220fSLawrence Mitchell       const PetscInt leaf = ilocal ? ilocal[i] : i;
19339566063dSJacob Faibussowitsch       PetscCall(PetscHSetIAdd(seen, leaf));
1934a072220fSLawrence Mitchell     }
19359566063dSJacob Faibussowitsch     PetscCall(PetscHSetIGetSize(seen, &n));
193608401ef6SPierre Jolivet     PetscCheck(n == nleaves, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided leaves have repeated values: all leaves must be unique");
19379566063dSJacob Faibussowitsch     PetscCall(PetscHSetIDestroy(&seen));
1938b458e8f1SJose E. Roman   }
1939a072220fSLawrence Mitchell   PetscFunctionReturn(0);
1940a072220fSLawrence Mitchell }
194154729392SStefano Zampini 
1942a7b3aa13SAta Mesgarnejad /*@
1943cab54364SBarry Smith   PetscSFCompose - Compose a new `PetscSF` by putting the second `PetscSF` under the first one in a top (roots) down (leaves) view
1944a7b3aa13SAta Mesgarnejad 
1945a7b3aa13SAta Mesgarnejad   Input Parameters:
1946cab54364SBarry Smith + sfA - The first `PetscSF`
1947cab54364SBarry Smith - sfB - The second `PetscSF`
1948a7b3aa13SAta Mesgarnejad 
1949a7b3aa13SAta Mesgarnejad   Output Parameters:
1950cab54364SBarry Smith . sfBA - The composite `PetscSF`
1951a7b3aa13SAta Mesgarnejad 
1952a7b3aa13SAta Mesgarnejad   Level: developer
1953a7b3aa13SAta Mesgarnejad 
1954a072220fSLawrence Mitchell   Notes:
1955cab54364SBarry Smith   Currently, the two `PetscSF`s must be defined on congruent communicators and they must be true star
195654729392SStefano Zampini   forests, i.e. the same leaf is not connected with different roots.
195754729392SStefano Zampini 
195854729392SStefano Zampini   sfA's leaf space and sfB's root space might be partially overlapped. The composition builds
195954729392SStefano Zampini   a graph with sfA's roots and sfB's leaves only when there is a path between them. Unconnected
196054729392SStefano Zampini   nodes (roots or leaves) are not in sfBA. Doing a Bcast on the new SF is equivalent to doing a
196154729392SStefano Zampini   Bcast on sfA, then a Bcast on sfB, on connected nodes.
1962a072220fSLawrence Mitchell 
1963db781477SPatrick Sanan .seealso: `PetscSF`, `PetscSFComposeInverse()`, `PetscSFGetGraph()`, `PetscSFSetGraph()`
1964a7b3aa13SAta Mesgarnejad @*/
1965d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCompose(PetscSF sfA, PetscSF sfB, PetscSF *sfBA)
1966d71ae5a4SJacob Faibussowitsch {
1967a7b3aa13SAta Mesgarnejad   const PetscSFNode *remotePointsA, *remotePointsB;
1968d41018fbSJunchao Zhang   PetscSFNode       *remotePointsBA = NULL, *reorderedRemotePointsA = NULL, *leafdataB;
196954729392SStefano Zampini   const PetscInt    *localPointsA, *localPointsB;
197054729392SStefano Zampini   PetscInt          *localPointsBA;
197154729392SStefano Zampini   PetscInt           i, numRootsA, numLeavesA, numRootsB, numLeavesB, minleaf, maxleaf, numLeavesBA;
197254729392SStefano Zampini   PetscBool          denseB;
1973a7b3aa13SAta Mesgarnejad 
1974a7b3aa13SAta Mesgarnejad   PetscFunctionBegin;
1975a7b3aa13SAta Mesgarnejad   PetscValidHeaderSpecific(sfA, PETSCSF_CLASSID, 1);
197629046d53SLisandro Dalcin   PetscSFCheckGraphSet(sfA, 1);
197729046d53SLisandro Dalcin   PetscValidHeaderSpecific(sfB, PETSCSF_CLASSID, 2);
197829046d53SLisandro Dalcin   PetscSFCheckGraphSet(sfB, 2);
197954729392SStefano Zampini   PetscCheckSameComm(sfA, 1, sfB, 2);
198029046d53SLisandro Dalcin   PetscValidPointer(sfBA, 3);
19819566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckLeavesUnique_Private(sfA));
19829566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckLeavesUnique_Private(sfB));
198354729392SStefano Zampini 
19849566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sfA, &numRootsA, &numLeavesA, &localPointsA, &remotePointsA));
19859566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sfB, &numRootsB, &numLeavesB, &localPointsB, &remotePointsB));
19860ea77edaSksagiyam   /* Make sure that PetscSFBcast{Begin, End}(sfB, ...) works with root data of size       */
19870ea77edaSksagiyam   /* numRootsB; otherwise, garbage will be broadcasted.                                   */
19880ea77edaSksagiyam   /* Example (comm size = 1):                                                             */
19890ea77edaSksagiyam   /* sfA: 0 <- (0, 0)                                                                     */
19900ea77edaSksagiyam   /* sfB: 100 <- (0, 0)                                                                   */
19910ea77edaSksagiyam   /*      101 <- (0, 1)                                                                   */
19920ea77edaSksagiyam   /* Here, we have remotePointsA = [(0, 0)], but for remotePointsA to be a valid tartget  */
19930ea77edaSksagiyam   /* of sfB, it has to be recasted as [(0, 0), (-1, -1)] so that points 100 and 101 would */
19940ea77edaSksagiyam   /* receive (0, 0) and (-1, -1), respectively, when PetscSFBcast(sfB, ...) is called on  */
19950ea77edaSksagiyam   /* remotePointsA; if not recasted, point 101 would receive a garbage value.             */
19969566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(numRootsB, &reorderedRemotePointsA));
199754729392SStefano Zampini   for (i = 0; i < numRootsB; i++) {
199854729392SStefano Zampini     reorderedRemotePointsA[i].rank  = -1;
199954729392SStefano Zampini     reorderedRemotePointsA[i].index = -1;
200054729392SStefano Zampini   }
200154729392SStefano Zampini   for (i = 0; i < numLeavesA; i++) {
20020ea77edaSksagiyam     PetscInt localp = localPointsA ? localPointsA[i] : i;
20030ea77edaSksagiyam 
20040ea77edaSksagiyam     if (localp >= numRootsB) continue;
20050ea77edaSksagiyam     reorderedRemotePointsA[localp] = remotePointsA[i];
200654729392SStefano Zampini   }
2007d41018fbSJunchao Zhang   remotePointsA = reorderedRemotePointsA;
20089566063dSJacob Faibussowitsch   PetscCall(PetscSFGetLeafRange(sfB, &minleaf, &maxleaf));
20099566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(maxleaf - minleaf + 1, &leafdataB));
20100ea77edaSksagiyam   for (i = 0; i < maxleaf - minleaf + 1; i++) {
20110ea77edaSksagiyam     leafdataB[i].rank  = -1;
20120ea77edaSksagiyam     leafdataB[i].index = -1;
20130ea77edaSksagiyam   }
20149566063dSJacob Faibussowitsch   PetscCall(PetscSFBcastBegin(sfB, MPIU_2INT, remotePointsA, leafdataB - minleaf, MPI_REPLACE));
20159566063dSJacob Faibussowitsch   PetscCall(PetscSFBcastEnd(sfB, MPIU_2INT, remotePointsA, leafdataB - minleaf, MPI_REPLACE));
20169566063dSJacob Faibussowitsch   PetscCall(PetscFree(reorderedRemotePointsA));
2017d41018fbSJunchao Zhang 
201854729392SStefano Zampini   denseB = (PetscBool)!localPointsB;
201954729392SStefano Zampini   for (i = 0, numLeavesBA = 0; i < numLeavesB; i++) {
202054729392SStefano Zampini     if (leafdataB[localPointsB ? localPointsB[i] - minleaf : i].rank == -1) denseB = PETSC_FALSE;
202154729392SStefano Zampini     else numLeavesBA++;
202254729392SStefano Zampini   }
202354729392SStefano Zampini   if (denseB) {
2024d41018fbSJunchao Zhang     localPointsBA  = NULL;
2025d41018fbSJunchao Zhang     remotePointsBA = leafdataB;
2026d41018fbSJunchao Zhang   } else {
20279566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(numLeavesBA, &localPointsBA));
20289566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(numLeavesBA, &remotePointsBA));
202954729392SStefano Zampini     for (i = 0, numLeavesBA = 0; i < numLeavesB; i++) {
203054729392SStefano Zampini       const PetscInt l = localPointsB ? localPointsB[i] : i;
203154729392SStefano Zampini 
203254729392SStefano Zampini       if (leafdataB[l - minleaf].rank == -1) continue;
203354729392SStefano Zampini       remotePointsBA[numLeavesBA] = leafdataB[l - minleaf];
203454729392SStefano Zampini       localPointsBA[numLeavesBA]  = l;
203554729392SStefano Zampini       numLeavesBA++;
203654729392SStefano Zampini     }
20379566063dSJacob Faibussowitsch     PetscCall(PetscFree(leafdataB));
2038d41018fbSJunchao Zhang   }
20399566063dSJacob Faibussowitsch   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sfA), sfBA));
20409566063dSJacob Faibussowitsch   PetscCall(PetscSFSetFromOptions(*sfBA));
20419566063dSJacob Faibussowitsch   PetscCall(PetscSFSetGraph(*sfBA, numRootsA, numLeavesBA, localPointsBA, PETSC_OWN_POINTER, remotePointsBA, PETSC_OWN_POINTER));
2042a7b3aa13SAta Mesgarnejad   PetscFunctionReturn(0);
2043a7b3aa13SAta Mesgarnejad }
20441c6ba672SJunchao Zhang 
204504c0ada0SJunchao Zhang /*@
2046cab54364SBarry Smith   PetscSFComposeInverse - Compose a new `PetscSF` by putting the inverse of the second `PetscSF` under the first one
204704c0ada0SJunchao Zhang 
204804c0ada0SJunchao Zhang   Input Parameters:
2049cab54364SBarry Smith + sfA - The first `PetscSF`
2050cab54364SBarry Smith - sfB - The second `PetscSF`
205104c0ada0SJunchao Zhang 
205204c0ada0SJunchao Zhang   Output Parameters:
2053cab54364SBarry Smith . sfBA - The composite `PetscSF`.
205404c0ada0SJunchao Zhang 
205504c0ada0SJunchao Zhang   Level: developer
205604c0ada0SJunchao Zhang 
205754729392SStefano Zampini   Notes:
205854729392SStefano Zampini   Currently, the two SFs must be defined on congruent communicators and they must be true star
205954729392SStefano Zampini   forests, i.e. the same leaf is not connected with different roots. Even more, all roots of the
206054729392SStefano Zampini   second SF must have a degree of 1, i.e., no roots have more than one leaf connected.
206154729392SStefano Zampini 
206254729392SStefano Zampini   sfA's leaf space and sfB's leaf space might be partially overlapped. The composition builds
206354729392SStefano Zampini   a graph with sfA's roots and sfB's roots only when there is a path between them. Unconnected
206454729392SStefano Zampini   roots are not in sfBA. Doing a Bcast on the new SF is equivalent to doing a Bcast on sfA, then
206554729392SStefano Zampini   a Reduce on sfB, on connected roots.
206654729392SStefano Zampini 
2067db781477SPatrick Sanan .seealso: `PetscSF`, `PetscSFCompose()`, `PetscSFGetGraph()`, `PetscSFSetGraph()`, `PetscSFCreateInverseSF()`
206804c0ada0SJunchao Zhang @*/
2069d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComposeInverse(PetscSF sfA, PetscSF sfB, PetscSF *sfBA)
2070d71ae5a4SJacob Faibussowitsch {
207104c0ada0SJunchao Zhang   const PetscSFNode *remotePointsA, *remotePointsB;
207204c0ada0SJunchao Zhang   PetscSFNode       *remotePointsBA;
207304c0ada0SJunchao Zhang   const PetscInt    *localPointsA, *localPointsB;
207454729392SStefano Zampini   PetscSFNode       *reorderedRemotePointsA = NULL;
207554729392SStefano Zampini   PetscInt           i, numRootsA, numLeavesA, numLeavesBA, numRootsB, numLeavesB, minleaf, maxleaf, *localPointsBA;
20765b0d146aSStefano Zampini   MPI_Op             op;
20775b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES)
20785b0d146aSStefano Zampini   PetscBool iswin;
20795b0d146aSStefano Zampini #endif
208004c0ada0SJunchao Zhang 
208104c0ada0SJunchao Zhang   PetscFunctionBegin;
208204c0ada0SJunchao Zhang   PetscValidHeaderSpecific(sfA, PETSCSF_CLASSID, 1);
208304c0ada0SJunchao Zhang   PetscSFCheckGraphSet(sfA, 1);
208404c0ada0SJunchao Zhang   PetscValidHeaderSpecific(sfB, PETSCSF_CLASSID, 2);
208504c0ada0SJunchao Zhang   PetscSFCheckGraphSet(sfB, 2);
208654729392SStefano Zampini   PetscCheckSameComm(sfA, 1, sfB, 2);
208704c0ada0SJunchao Zhang   PetscValidPointer(sfBA, 3);
20889566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckLeavesUnique_Private(sfA));
20899566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckLeavesUnique_Private(sfB));
209054729392SStefano Zampini 
20919566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sfA, &numRootsA, &numLeavesA, &localPointsA, &remotePointsA));
20929566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sfB, &numRootsB, &numLeavesB, &localPointsB, &remotePointsB));
20935b0d146aSStefano Zampini 
20945b0d146aSStefano Zampini   /* TODO: Check roots of sfB have degree of 1 */
20955b0d146aSStefano Zampini   /* Once we implement it, we can replace the MPI_MAXLOC
209683df288dSJunchao Zhang      with MPI_REPLACE. In that case, MPI_MAXLOC and MPI_REPLACE have the same effect.
20975b0d146aSStefano Zampini      We use MPI_MAXLOC only to have a deterministic output from this routine if
20985b0d146aSStefano Zampini      the root condition is not meet.
20995b0d146aSStefano Zampini    */
21005b0d146aSStefano Zampini   op = MPI_MAXLOC;
21015b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES)
21025b0d146aSStefano Zampini   /* we accept a non-deterministic output (if any) with PETSCSFWINDOW, since MPI_MAXLOC cannot operate on MPIU_2INT with MPI_Accumulate */
21039566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject)sfB, PETSCSFWINDOW, &iswin));
210483df288dSJunchao Zhang   if (iswin) op = MPI_REPLACE;
21055b0d146aSStefano Zampini #endif
21065b0d146aSStefano Zampini 
21079566063dSJacob Faibussowitsch   PetscCall(PetscSFGetLeafRange(sfB, &minleaf, &maxleaf));
21089566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(maxleaf - minleaf + 1, &reorderedRemotePointsA));
210954729392SStefano Zampini   for (i = 0; i < maxleaf - minleaf + 1; i++) {
211054729392SStefano Zampini     reorderedRemotePointsA[i].rank  = -1;
211154729392SStefano Zampini     reorderedRemotePointsA[i].index = -1;
211254729392SStefano Zampini   }
211354729392SStefano Zampini   if (localPointsA) {
211454729392SStefano Zampini     for (i = 0; i < numLeavesA; i++) {
211554729392SStefano Zampini       if (localPointsA[i] > maxleaf || localPointsA[i] < minleaf) continue;
211654729392SStefano Zampini       reorderedRemotePointsA[localPointsA[i] - minleaf] = remotePointsA[i];
211754729392SStefano Zampini     }
211854729392SStefano Zampini   } else {
211954729392SStefano Zampini     for (i = 0; i < numLeavesA; i++) {
212054729392SStefano Zampini       if (i > maxleaf || i < minleaf) continue;
212154729392SStefano Zampini       reorderedRemotePointsA[i - minleaf] = remotePointsA[i];
212254729392SStefano Zampini     }
212354729392SStefano Zampini   }
212454729392SStefano Zampini 
21259566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(numRootsB, &localPointsBA));
21269566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(numRootsB, &remotePointsBA));
212754729392SStefano Zampini   for (i = 0; i < numRootsB; i++) {
212854729392SStefano Zampini     remotePointsBA[i].rank  = -1;
212954729392SStefano Zampini     remotePointsBA[i].index = -1;
213054729392SStefano Zampini   }
213154729392SStefano Zampini 
21329566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceBegin(sfB, MPIU_2INT, reorderedRemotePointsA - minleaf, remotePointsBA, op));
21339566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceEnd(sfB, MPIU_2INT, reorderedRemotePointsA - minleaf, remotePointsBA, op));
21349566063dSJacob Faibussowitsch   PetscCall(PetscFree(reorderedRemotePointsA));
213554729392SStefano Zampini   for (i = 0, numLeavesBA = 0; i < numRootsB; i++) {
213654729392SStefano Zampini     if (remotePointsBA[i].rank == -1) continue;
213754729392SStefano Zampini     remotePointsBA[numLeavesBA].rank  = remotePointsBA[i].rank;
213854729392SStefano Zampini     remotePointsBA[numLeavesBA].index = remotePointsBA[i].index;
213954729392SStefano Zampini     localPointsBA[numLeavesBA]        = i;
214054729392SStefano Zampini     numLeavesBA++;
214154729392SStefano Zampini   }
21429566063dSJacob Faibussowitsch   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sfA), sfBA));
21439566063dSJacob Faibussowitsch   PetscCall(PetscSFSetFromOptions(*sfBA));
21449566063dSJacob Faibussowitsch   PetscCall(PetscSFSetGraph(*sfBA, numRootsA, numLeavesBA, localPointsBA, PETSC_OWN_POINTER, remotePointsBA, PETSC_OWN_POINTER));
214504c0ada0SJunchao Zhang   PetscFunctionReturn(0);
214604c0ada0SJunchao Zhang }
214704c0ada0SJunchao Zhang 
21481c6ba672SJunchao Zhang /*
2149cab54364SBarry Smith   PetscSFCreateLocalSF_Private - Creates a local `PetscSF` that only has intra-process edges of the global `PetscSF`
21501c6ba672SJunchao Zhang 
21511c6ba672SJunchao Zhang   Input Parameters:
2152cab54364SBarry Smith . sf - The global `PetscSF`
21531c6ba672SJunchao Zhang 
21541c6ba672SJunchao Zhang   Output Parameters:
2155cab54364SBarry Smith . out - The local `PetscSF`
2156cab54364SBarry Smith 
2157cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`
21581c6ba672SJunchao Zhang  */
2159d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateLocalSF_Private(PetscSF sf, PetscSF *out)
2160d71ae5a4SJacob Faibussowitsch {
21611c6ba672SJunchao Zhang   MPI_Comm           comm;
21621c6ba672SJunchao Zhang   PetscMPIInt        myrank;
21631c6ba672SJunchao Zhang   const PetscInt    *ilocal;
21641c6ba672SJunchao Zhang   const PetscSFNode *iremote;
21651c6ba672SJunchao Zhang   PetscInt           i, j, nroots, nleaves, lnleaves, *lilocal;
21661c6ba672SJunchao Zhang   PetscSFNode       *liremote;
21671c6ba672SJunchao Zhang   PetscSF            lsf;
21681c6ba672SJunchao Zhang 
21691c6ba672SJunchao Zhang   PetscFunctionBegin;
21701c6ba672SJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
2171dbbe0bcdSBarry Smith   if (sf->ops->CreateLocalSF) PetscUseTypeMethod(sf, CreateLocalSF, out);
2172dbbe0bcdSBarry Smith   else {
21731c6ba672SJunchao Zhang     /* Could use PetscSFCreateEmbeddedLeafSF, but since we know the comm is PETSC_COMM_SELF, we can make it fast */
21749566063dSJacob Faibussowitsch     PetscCall(PetscObjectGetComm((PetscObject)sf, &comm));
21759566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_rank(comm, &myrank));
21761c6ba672SJunchao Zhang 
21771c6ba672SJunchao Zhang     /* Find out local edges and build a local SF */
21789566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, &iremote));
21799371c9d4SSatish Balay     for (i = lnleaves = 0; i < nleaves; i++) {
21809371c9d4SSatish Balay       if (iremote[i].rank == (PetscInt)myrank) lnleaves++;
21819371c9d4SSatish Balay     }
21829566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(lnleaves, &lilocal));
21839566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(lnleaves, &liremote));
21841c6ba672SJunchao Zhang 
21851c6ba672SJunchao Zhang     for (i = j = 0; i < nleaves; i++) {
21861c6ba672SJunchao Zhang       if (iremote[i].rank == (PetscInt)myrank) {
21871c6ba672SJunchao Zhang         lilocal[j]        = ilocal ? ilocal[i] : i; /* ilocal=NULL for contiguous storage */
21881c6ba672SJunchao Zhang         liremote[j].rank  = 0;                      /* rank in PETSC_COMM_SELF */
21891c6ba672SJunchao Zhang         liremote[j].index = iremote[i].index;
21901c6ba672SJunchao Zhang         j++;
21911c6ba672SJunchao Zhang       }
21921c6ba672SJunchao Zhang     }
21939566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(PETSC_COMM_SELF, &lsf));
21949566063dSJacob Faibussowitsch     PetscCall(PetscSFSetFromOptions(lsf));
21959566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(lsf, nroots, lnleaves, lilocal, PETSC_OWN_POINTER, liremote, PETSC_OWN_POINTER));
21969566063dSJacob Faibussowitsch     PetscCall(PetscSFSetUp(lsf));
21971c6ba672SJunchao Zhang     *out = lsf;
21981c6ba672SJunchao Zhang   }
21991c6ba672SJunchao Zhang   PetscFunctionReturn(0);
22001c6ba672SJunchao Zhang }
2201dd5b3ca6SJunchao Zhang 
2202dd5b3ca6SJunchao Zhang /* Similar to PetscSFBcast, but only Bcast to leaves on rank 0 */
2203d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastToZero_Private(PetscSF sf, MPI_Datatype unit, const void *rootdata, void *leafdata)
2204d71ae5a4SJacob Faibussowitsch {
2205eb02082bSJunchao Zhang   PetscMemType rootmtype, leafmtype;
2206dd5b3ca6SJunchao Zhang 
2207dd5b3ca6SJunchao Zhang   PetscFunctionBegin;
2208dd5b3ca6SJunchao Zhang   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1);
22099566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
22109566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin, sf, 0, 0, 0));
22119566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(rootdata, &rootmtype));
22129566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafdata, &leafmtype));
2213dbbe0bcdSBarry Smith   PetscUseTypeMethod(sf, BcastToZero, unit, rootmtype, rootdata, leafmtype, leafdata);
22149566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin, sf, 0, 0, 0));
2215dd5b3ca6SJunchao Zhang   PetscFunctionReturn(0);
2216dd5b3ca6SJunchao Zhang }
2217dd5b3ca6SJunchao Zhang 
2218157edd7aSVaclav Hapla /*@
2219cab54364SBarry Smith   PetscSFConcatenate - concatenate multiple `PetscSF` into one
2220157edd7aSVaclav Hapla 
2221157edd7aSVaclav Hapla   Input Parameters:
2222157edd7aSVaclav Hapla + comm - the communicator
2223cab54364SBarry Smith . nsfs - the number of input `PetscSF`
2224cab54364SBarry Smith . sfs  - the array of input `PetscSF`
22251f40158dSVaclav Hapla . rootMode - the root mode specifying how roots are handled
2226cab54364SBarry Smith - leafOffsets - the array of local leaf offsets, one for each input `PetscSF`, or NULL for contiguous storage
2227157edd7aSVaclav Hapla 
2228157edd7aSVaclav Hapla   Output Parameters:
2229cab54364SBarry Smith . newsf - The resulting `PetscSF`
2230157edd7aSVaclav Hapla 
22311f40158dSVaclav Hapla   Level: advanced
2232157edd7aSVaclav Hapla 
2233157edd7aSVaclav Hapla   Notes:
2234157edd7aSVaclav Hapla   The communicator of all SFs in sfs must be comm.
2235157edd7aSVaclav Hapla 
22361f40158dSVaclav Hapla   Leaves are always concatenated locally, keeping them ordered by the input SF index and original local order.
2237157edd7aSVaclav Hapla   The offsets in leafOffsets are added to the original leaf indices.
2238157edd7aSVaclav Hapla   If all input SFs use contiguous leaf storage (ilocal = NULL), leafOffsets can be passed as NULL as well.
2239157edd7aSVaclav Hapla   In this case, NULL is also passed as ilocal to the resulting SF.
2240157edd7aSVaclav Hapla   If any input SF has non-null ilocal, leafOffsets is needed to distinguish leaves from different input SFs.
2241157edd7aSVaclav Hapla   In this case, user is responsible to provide correct offsets so that the resulting leaves are unique (otherwise an error occurs).
2242157edd7aSVaclav Hapla 
22431f40158dSVaclav Hapla   All root modes retain the essential connectivity condition:
22441f40158dSVaclav Hapla   If two leaves of the same input SF are connected (sharing the same root), they are also connected in the output SF.
22451f40158dSVaclav Hapla   Parameter rootMode controls how the input root spaces are combined.
22461f40158dSVaclav Hapla   For `PETSCSF_CONCATENATE_ROOTMODE_SHARED`, the root space is considered the same for each input SF (checked in debug mode) and is also the same in the output SF.
22471f40158dSVaclav Hapla   For `PETSCSF_CONCATENATE_ROOTMODE_LOCAL` and `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL`, the input root spaces are taken as separate and joined.
22481f40158dSVaclav Hapla   `PETSCSF_CONCATENATE_ROOTMODE_LOCAL` joins the root spaces locally;
22491f40158dSVaclav Hapla   roots of sfs[0], sfs[1], sfs[2], ... are joined on each rank separately, ordered by input SF and original local index, and renumbered contiguously.
22501f40158dSVaclav Hapla   `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL` joins the root spaces globally;
22511f40158dSVaclav Hapla   roots of sfs[0], sfs[1], sfs[2, ... are joined globally, ordered by input SF index and original global index, and renumbered contiguously;
22521f40158dSVaclav Hapla   the original root ranks are ignored.
22531f40158dSVaclav Hapla   For both `PETSCSF_CONCATENATE_ROOTMODE_LOCAL` and `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL`,
22541f40158dSVaclav Hapla   the output SF's root layout is such that the local number of roots is a sum of the input SF's local numbers of roots on each rank to keep the load balancing.
22551f40158dSVaclav Hapla   However, for `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL`, that means roots can move to different ranks.
22561f40158dSVaclav Hapla 
22571f40158dSVaclav Hapla    Example:
22581f40158dSVaclav Hapla    We can use src/vec/is/sf/tests/ex18.c to compare the root modes. By running
22591f40158dSVaclav Hapla $ make -C $PETSC_DIR/src/vec/is/sf/tests ex18
22601f40158dSVaclav Hapla $ for m in {local,global,shared}; do
22611f40158dSVaclav Hapla $   mpirun -n 2 $PETSC_DIR/src/vec/is/sf/tests/ex18 -nsfs 2 -n 2 -root_mode $m -sf_view
22621f40158dSVaclav Hapla $ done
22631f40158dSVaclav Hapla    we generate two identical SFs sf_0 and sf_1,
22641f40158dSVaclav Hapla $ PetscSF Object: sf_0 2 MPI processes
22651f40158dSVaclav Hapla $   type: basic
22661f40158dSVaclav Hapla $   rank #leaves #roots
22671f40158dSVaclav Hapla $   [ 0]       4      2
22681f40158dSVaclav Hapla $   [ 1]       4      2
22691f40158dSVaclav Hapla $   leaves      roots       roots in global numbering
22701f40158dSVaclav Hapla $   ( 0,  0) <- ( 0,  0)  =   0
22711f40158dSVaclav Hapla $   ( 0,  1) <- ( 0,  1)  =   1
22721f40158dSVaclav Hapla $   ( 0,  2) <- ( 1,  0)  =   2
22731f40158dSVaclav Hapla $   ( 0,  3) <- ( 1,  1)  =   3
22741f40158dSVaclav Hapla $   ( 1,  0) <- ( 0,  0)  =   0
22751f40158dSVaclav Hapla $   ( 1,  1) <- ( 0,  1)  =   1
22761f40158dSVaclav Hapla $   ( 1,  2) <- ( 1,  0)  =   2
22771f40158dSVaclav Hapla $   ( 1,  3) <- ( 1,  1)  =   3
22781f40158dSVaclav Hapla    and pass them to `PetscSFConcatenate()` along with different choices of rootMode, yielding different result_sf:
22791f40158dSVaclav Hapla $ rootMode = local:
22801f40158dSVaclav Hapla $ PetscSF Object: result_sf 2 MPI processes
22811f40158dSVaclav Hapla $   type: basic
22821f40158dSVaclav Hapla $   rank #leaves #roots
22831f40158dSVaclav Hapla $   [ 0]       8      4
22841f40158dSVaclav Hapla $   [ 1]       8      4
22851f40158dSVaclav Hapla $   leaves      roots       roots in global numbering
22861f40158dSVaclav Hapla $   ( 0,  0) <- ( 0,  0)  =   0
22871f40158dSVaclav Hapla $   ( 0,  1) <- ( 0,  1)  =   1
22881f40158dSVaclav Hapla $   ( 0,  2) <- ( 1,  0)  =   4
22891f40158dSVaclav Hapla $   ( 0,  3) <- ( 1,  1)  =   5
22901f40158dSVaclav Hapla $   ( 0,  4) <- ( 0,  2)  =   2
22911f40158dSVaclav Hapla $   ( 0,  5) <- ( 0,  3)  =   3
22921f40158dSVaclav Hapla $   ( 0,  6) <- ( 1,  2)  =   6
22931f40158dSVaclav Hapla $   ( 0,  7) <- ( 1,  3)  =   7
22941f40158dSVaclav Hapla $   ( 1,  0) <- ( 0,  0)  =   0
22951f40158dSVaclav Hapla $   ( 1,  1) <- ( 0,  1)  =   1
22961f40158dSVaclav Hapla $   ( 1,  2) <- ( 1,  0)  =   4
22971f40158dSVaclav Hapla $   ( 1,  3) <- ( 1,  1)  =   5
22981f40158dSVaclav Hapla $   ( 1,  4) <- ( 0,  2)  =   2
22991f40158dSVaclav Hapla $   ( 1,  5) <- ( 0,  3)  =   3
23001f40158dSVaclav Hapla $   ( 1,  6) <- ( 1,  2)  =   6
23011f40158dSVaclav Hapla $   ( 1,  7) <- ( 1,  3)  =   7
23021f40158dSVaclav Hapla $
23031f40158dSVaclav Hapla $ rootMode = global:
23041f40158dSVaclav Hapla $ PetscSF Object: result_sf 2 MPI processes
23051f40158dSVaclav Hapla $   type: basic
23061f40158dSVaclav Hapla $   rank #leaves #roots
23071f40158dSVaclav Hapla $   [ 0]       8      4
23081f40158dSVaclav Hapla $   [ 1]       8      4
23091f40158dSVaclav Hapla $   leaves      roots       roots in global numbering
23101f40158dSVaclav Hapla $   ( 0,  0) <- ( 0,  0)  =   0
23111f40158dSVaclav Hapla $   ( 0,  1) <- ( 0,  1)  =   1
23121f40158dSVaclav Hapla $   ( 0,  2) <- ( 0,  2)  =   2
23131f40158dSVaclav Hapla $   ( 0,  3) <- ( 0,  3)  =   3
23141f40158dSVaclav Hapla $   ( 0,  4) <- ( 1,  0)  =   4
23151f40158dSVaclav Hapla $   ( 0,  5) <- ( 1,  1)  =   5
23161f40158dSVaclav Hapla $   ( 0,  6) <- ( 1,  2)  =   6
23171f40158dSVaclav Hapla $   ( 0,  7) <- ( 1,  3)  =   7
23181f40158dSVaclav Hapla $   ( 1,  0) <- ( 0,  0)  =   0
23191f40158dSVaclav Hapla $   ( 1,  1) <- ( 0,  1)  =   1
23201f40158dSVaclav Hapla $   ( 1,  2) <- ( 0,  2)  =   2
23211f40158dSVaclav Hapla $   ( 1,  3) <- ( 0,  3)  =   3
23221f40158dSVaclav Hapla $   ( 1,  4) <- ( 1,  0)  =   4
23231f40158dSVaclav Hapla $   ( 1,  5) <- ( 1,  1)  =   5
23241f40158dSVaclav Hapla $   ( 1,  6) <- ( 1,  2)  =   6
23251f40158dSVaclav Hapla $   ( 1,  7) <- ( 1,  3)  =   7
23261f40158dSVaclav Hapla $
23271f40158dSVaclav Hapla $ rootMode = shared:
23281f40158dSVaclav Hapla $ PetscSF Object: result_sf 2 MPI processes
23291f40158dSVaclav Hapla $   type: basic
23301f40158dSVaclav Hapla $   rank #leaves #roots
23311f40158dSVaclav Hapla $   [ 0]       8      2
23321f40158dSVaclav Hapla $   [ 1]       8      2
23331f40158dSVaclav Hapla $   leaves      roots       roots in global numbering
23341f40158dSVaclav Hapla $   ( 0,  0) <- ( 0,  0)  =   0
23351f40158dSVaclav Hapla $   ( 0,  1) <- ( 0,  1)  =   1
23361f40158dSVaclav Hapla $   ( 0,  2) <- ( 1,  0)  =   2
23371f40158dSVaclav Hapla $   ( 0,  3) <- ( 1,  1)  =   3
23381f40158dSVaclav Hapla $   ( 0,  4) <- ( 0,  0)  =   0
23391f40158dSVaclav Hapla $   ( 0,  5) <- ( 0,  1)  =   1
23401f40158dSVaclav Hapla $   ( 0,  6) <- ( 1,  0)  =   2
23411f40158dSVaclav Hapla $   ( 0,  7) <- ( 1,  1)  =   3
23421f40158dSVaclav Hapla $   ( 1,  0) <- ( 0,  0)  =   0
23431f40158dSVaclav Hapla $   ( 1,  1) <- ( 0,  1)  =   1
23441f40158dSVaclav Hapla $   ( 1,  2) <- ( 1,  0)  =   2
23451f40158dSVaclav Hapla $   ( 1,  3) <- ( 1,  1)  =   3
23461f40158dSVaclav Hapla $   ( 1,  4) <- ( 0,  0)  =   0
23471f40158dSVaclav Hapla $   ( 1,  5) <- ( 0,  1)  =   1
23481f40158dSVaclav Hapla $   ( 1,  6) <- ( 1,  0)  =   2
23491f40158dSVaclav Hapla $   ( 1,  7) <- ( 1,  1)  =   3
23501f40158dSVaclav Hapla 
23511f40158dSVaclav Hapla .seealso: `PetscSF`, `PetscSFCompose()`, `PetscSFGetGraph()`, `PetscSFSetGraph()`, `PetscSFConcatenateRootMode`
2352157edd7aSVaclav Hapla @*/
23531f40158dSVaclav Hapla PetscErrorCode PetscSFConcatenate(MPI_Comm comm, PetscInt nsfs, PetscSF sfs[], PetscSFConcatenateRootMode rootMode, PetscInt leafOffsets[], PetscSF *newsf)
2354d71ae5a4SJacob Faibussowitsch {
2355157edd7aSVaclav Hapla   PetscInt     i, s, nLeaves, nRoots;
2356157edd7aSVaclav Hapla   PetscInt    *leafArrayOffsets;
2357157edd7aSVaclav Hapla   PetscInt    *ilocal_new;
2358157edd7aSVaclav Hapla   PetscSFNode *iremote_new;
2359157edd7aSVaclav Hapla   PetscBool    all_ilocal_null = PETSC_FALSE;
23601f40158dSVaclav Hapla   PetscLayout  glayout         = NULL;
23611f40158dSVaclav Hapla   PetscInt    *gremote         = NULL;
23621f40158dSVaclav Hapla   PetscMPIInt  rank, size;
2363157edd7aSVaclav Hapla 
2364157edd7aSVaclav Hapla   PetscFunctionBegin;
236512f479c1SVaclav Hapla   if (PetscDefined(USE_DEBUG)) {
2366157edd7aSVaclav Hapla     PetscSF dummy; /* just to have a PetscObject on comm for input validation */
2367157edd7aSVaclav Hapla 
23689566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(comm, &dummy));
2369157edd7aSVaclav Hapla     PetscValidLogicalCollectiveInt(dummy, nsfs, 2);
2370157edd7aSVaclav Hapla     PetscValidPointer(sfs, 3);
2371157edd7aSVaclav Hapla     for (i = 0; i < nsfs; i++) {
2372157edd7aSVaclav Hapla       PetscValidHeaderSpecific(sfs[i], PETSCSF_CLASSID, 3);
2373157edd7aSVaclav Hapla       PetscCheckSameComm(dummy, 1, sfs[i], 3);
2374157edd7aSVaclav Hapla     }
23751f40158dSVaclav Hapla     PetscValidLogicalCollectiveEnum(dummy, rootMode, 4);
2376157edd7aSVaclav Hapla     if (leafOffsets) PetscValidIntPointer(leafOffsets, 5);
2377157edd7aSVaclav Hapla     PetscValidPointer(newsf, 6);
23789566063dSJacob Faibussowitsch     PetscCall(PetscSFDestroy(&dummy));
2379157edd7aSVaclav Hapla   }
2380157edd7aSVaclav Hapla   if (!nsfs) {
23819566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(comm, newsf));
23829566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(*newsf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
2383157edd7aSVaclav Hapla     PetscFunctionReturn(0);
2384157edd7aSVaclav Hapla   }
23859566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_rank(comm, &rank));
23861f40158dSVaclav Hapla   PetscCallMPI(MPI_Comm_size(comm, &size));
2387157edd7aSVaclav Hapla 
23881f40158dSVaclav Hapla   /* Calculate leaf array offsets */
23899566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(nsfs + 1, &leafArrayOffsets));
2390157edd7aSVaclav Hapla   leafArrayOffsets[0] = 0;
2391157edd7aSVaclav Hapla   for (s = 0; s < nsfs; s++) {
2392157edd7aSVaclav Hapla     PetscInt nl;
2393157edd7aSVaclav Hapla 
23949566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sfs[s], NULL, &nl, NULL, NULL));
2395157edd7aSVaclav Hapla     leafArrayOffsets[s + 1] = leafArrayOffsets[s] + nl;
2396157edd7aSVaclav Hapla   }
2397157edd7aSVaclav Hapla   nLeaves = leafArrayOffsets[nsfs];
2398157edd7aSVaclav Hapla 
23991f40158dSVaclav Hapla   /* Calculate number of roots */
24001f40158dSVaclav Hapla   switch (rootMode) {
24011f40158dSVaclav Hapla   case PETSCSF_CONCATENATE_ROOTMODE_SHARED: {
24021f40158dSVaclav Hapla     PetscCall(PetscSFGetGraph(sfs[0], &nRoots, NULL, NULL, NULL));
24031f40158dSVaclav Hapla     if (PetscDefined(USE_DEBUG)) {
24041f40158dSVaclav Hapla       for (s = 1; s < nsfs; s++) {
24051f40158dSVaclav Hapla         PetscInt nr;
24061f40158dSVaclav Hapla 
24071f40158dSVaclav Hapla         PetscCall(PetscSFGetGraph(sfs[s], &nr, NULL, NULL, NULL));
24081f40158dSVaclav Hapla         PetscCheck(nr == nRoots, comm, PETSC_ERR_ARG_SIZ, "rootMode = %s but sfs[%" PetscInt_FMT "] has a different number of roots (%" PetscInt_FMT ") than sfs[0] (%" PetscInt_FMT ")", PetscSFConcatenateRootModes[rootMode], s, nr, nRoots);
24091f40158dSVaclav Hapla       }
24101f40158dSVaclav Hapla     }
24111f40158dSVaclav Hapla   } break;
24121f40158dSVaclav Hapla   case PETSCSF_CONCATENATE_ROOTMODE_GLOBAL: {
24131f40158dSVaclav Hapla     /* Calculate also global layout in this case */
24141f40158dSVaclav Hapla     PetscInt    *nls;
24151f40158dSVaclav Hapla     PetscLayout *lts;
24161f40158dSVaclav Hapla     PetscInt   **inds;
24171f40158dSVaclav Hapla     PetscInt     j;
24181f40158dSVaclav Hapla     PetscInt     rootOffset = 0;
24191f40158dSVaclav Hapla 
24201f40158dSVaclav Hapla     PetscCall(PetscCalloc3(nsfs, &lts, nsfs, &nls, nsfs, &inds));
24211f40158dSVaclav Hapla     PetscCall(PetscLayoutCreate(comm, &glayout));
24221f40158dSVaclav Hapla     glayout->bs = 1;
24231f40158dSVaclav Hapla     glayout->n  = 0;
24241f40158dSVaclav Hapla     glayout->N  = 0;
24251f40158dSVaclav Hapla     for (s = 0; s < nsfs; s++) {
24261f40158dSVaclav Hapla       PetscCall(PetscSFGetGraphLayout(sfs[s], &lts[s], &nls[s], NULL, &inds[s]));
24271f40158dSVaclav Hapla       glayout->n += lts[s]->n;
24281f40158dSVaclav Hapla       glayout->N += lts[s]->N;
24291f40158dSVaclav Hapla     }
24301f40158dSVaclav Hapla     PetscCall(PetscLayoutSetUp(glayout));
24311f40158dSVaclav Hapla     PetscCall(PetscMalloc1(nLeaves, &gremote));
24321f40158dSVaclav Hapla     for (s = 0, j = 0; s < nsfs; s++) {
24331f40158dSVaclav Hapla       for (i = 0; i < nls[s]; i++, j++) gremote[j] = inds[s][i] + rootOffset;
24341f40158dSVaclav Hapla       rootOffset += lts[s]->N;
24351f40158dSVaclav Hapla       PetscCall(PetscLayoutDestroy(&lts[s]));
24361f40158dSVaclav Hapla       PetscCall(PetscFree(inds[s]));
24371f40158dSVaclav Hapla     }
24381f40158dSVaclav Hapla     PetscCall(PetscFree3(lts, nls, inds));
24391f40158dSVaclav Hapla     nRoots = glayout->N;
24401f40158dSVaclav Hapla   } break;
24411f40158dSVaclav Hapla   case PETSCSF_CONCATENATE_ROOTMODE_LOCAL:
24421f40158dSVaclav Hapla     /* nRoots calculated later in this case */
24431f40158dSVaclav Hapla     break;
24441f40158dSVaclav Hapla   default:
24451f40158dSVaclav Hapla     SETERRQ(comm, PETSC_ERR_ARG_WRONG, "Invalid PetscSFConcatenateRootMode %d", rootMode);
24461f40158dSVaclav Hapla   }
24471f40158dSVaclav Hapla 
2448157edd7aSVaclav Hapla   if (!leafOffsets) {
2449157edd7aSVaclav Hapla     all_ilocal_null = PETSC_TRUE;
2450157edd7aSVaclav Hapla     for (s = 0; s < nsfs; s++) {
2451157edd7aSVaclav Hapla       const PetscInt *ilocal;
2452157edd7aSVaclav Hapla 
24539566063dSJacob Faibussowitsch       PetscCall(PetscSFGetGraph(sfs[s], NULL, NULL, &ilocal, NULL));
2454157edd7aSVaclav Hapla       if (ilocal) {
2455157edd7aSVaclav Hapla         all_ilocal_null = PETSC_FALSE;
2456157edd7aSVaclav Hapla         break;
2457157edd7aSVaclav Hapla       }
2458157edd7aSVaclav Hapla     }
2459157edd7aSVaclav Hapla     PetscCheck(all_ilocal_null, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "leafOffsets can be passed as NULL only if all SFs have ilocal = NULL");
2460157edd7aSVaclav Hapla   }
2461157edd7aSVaclav Hapla 
2462157edd7aSVaclav Hapla   /* Renumber and concatenate local leaves */
2463157edd7aSVaclav Hapla   ilocal_new = NULL;
2464157edd7aSVaclav Hapla   if (!all_ilocal_null) {
24659566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nLeaves, &ilocal_new));
2466157edd7aSVaclav Hapla     for (i = 0; i < nLeaves; i++) ilocal_new[i] = -1;
2467157edd7aSVaclav Hapla     for (s = 0; s < nsfs; s++) {
2468157edd7aSVaclav Hapla       const PetscInt *ilocal;
2469157edd7aSVaclav Hapla       PetscInt       *ilocal_l = &ilocal_new[leafArrayOffsets[s]];
2470157edd7aSVaclav Hapla       PetscInt        i, nleaves_l;
2471157edd7aSVaclav Hapla 
24729566063dSJacob Faibussowitsch       PetscCall(PetscSFGetGraph(sfs[s], NULL, &nleaves_l, &ilocal, NULL));
2473157edd7aSVaclav Hapla       for (i = 0; i < nleaves_l; i++) ilocal_l[i] = (ilocal ? ilocal[i] : i) + leafOffsets[s];
2474157edd7aSVaclav Hapla     }
2475157edd7aSVaclav Hapla   }
2476157edd7aSVaclav Hapla 
2477157edd7aSVaclav Hapla   /* Renumber and concatenate remote roots */
24781f40158dSVaclav Hapla   if (rootMode == PETSCSF_CONCATENATE_ROOTMODE_LOCAL || rootMode == PETSCSF_CONCATENATE_ROOTMODE_SHARED) {
24791f40158dSVaclav Hapla     PetscInt rootOffset = 0;
24801f40158dSVaclav Hapla 
24819566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nLeaves, &iremote_new));
2482157edd7aSVaclav Hapla     for (i = 0; i < nLeaves; i++) {
2483157edd7aSVaclav Hapla       iremote_new[i].rank  = -1;
2484157edd7aSVaclav Hapla       iremote_new[i].index = -1;
2485157edd7aSVaclav Hapla     }
2486157edd7aSVaclav Hapla     for (s = 0; s < nsfs; s++) {
2487157edd7aSVaclav Hapla       PetscInt           i, nl, nr;
2488157edd7aSVaclav Hapla       PetscSF            tmp_sf;
2489157edd7aSVaclav Hapla       const PetscSFNode *iremote;
2490157edd7aSVaclav Hapla       PetscSFNode       *tmp_rootdata;
2491157edd7aSVaclav Hapla       PetscSFNode       *tmp_leafdata = &iremote_new[leafArrayOffsets[s]];
2492157edd7aSVaclav Hapla 
24939566063dSJacob Faibussowitsch       PetscCall(PetscSFGetGraph(sfs[s], &nr, &nl, NULL, &iremote));
24949566063dSJacob Faibussowitsch       PetscCall(PetscSFCreate(comm, &tmp_sf));
2495157edd7aSVaclav Hapla       /* create helper SF with contiguous leaves */
24969566063dSJacob Faibussowitsch       PetscCall(PetscSFSetGraph(tmp_sf, nr, nl, NULL, PETSC_USE_POINTER, (PetscSFNode *)iremote, PETSC_COPY_VALUES));
24979566063dSJacob Faibussowitsch       PetscCall(PetscSFSetUp(tmp_sf));
24989566063dSJacob Faibussowitsch       PetscCall(PetscMalloc1(nr, &tmp_rootdata));
24991f40158dSVaclav Hapla       if (rootMode == PETSCSF_CONCATENATE_ROOTMODE_LOCAL) {
2500157edd7aSVaclav Hapla         for (i = 0; i < nr; i++) {
25011f40158dSVaclav Hapla           tmp_rootdata[i].index = i + rootOffset;
2502157edd7aSVaclav Hapla           tmp_rootdata[i].rank  = (PetscInt)rank;
2503157edd7aSVaclav Hapla         }
25041f40158dSVaclav Hapla         rootOffset += nr;
25051f40158dSVaclav Hapla       } else {
25061f40158dSVaclav Hapla         for (i = 0; i < nr; i++) {
25071f40158dSVaclav Hapla           tmp_rootdata[i].index = i;
25081f40158dSVaclav Hapla           tmp_rootdata[i].rank  = (PetscInt)rank;
25091f40158dSVaclav Hapla         }
25101f40158dSVaclav Hapla       }
25119566063dSJacob Faibussowitsch       PetscCall(PetscSFBcastBegin(tmp_sf, MPIU_2INT, tmp_rootdata, tmp_leafdata, MPI_REPLACE));
25129566063dSJacob Faibussowitsch       PetscCall(PetscSFBcastEnd(tmp_sf, MPIU_2INT, tmp_rootdata, tmp_leafdata, MPI_REPLACE));
25139566063dSJacob Faibussowitsch       PetscCall(PetscSFDestroy(&tmp_sf));
25149566063dSJacob Faibussowitsch       PetscCall(PetscFree(tmp_rootdata));
2515157edd7aSVaclav Hapla     }
25161f40158dSVaclav Hapla     if (rootMode == PETSCSF_CONCATENATE_ROOTMODE_LOCAL) { nRoots = rootOffset; } // else nRoots already calculated above
2517157edd7aSVaclav Hapla 
2518157edd7aSVaclav Hapla     /* Build the new SF */
25199566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(comm, newsf));
25209566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(*newsf, nRoots, nLeaves, ilocal_new, PETSC_OWN_POINTER, iremote_new, PETSC_OWN_POINTER));
25211f40158dSVaclav Hapla   } else {
25221f40158dSVaclav Hapla     /* Build the new SF */
25231f40158dSVaclav Hapla     PetscCall(PetscSFCreate(comm, newsf));
25241f40158dSVaclav Hapla     PetscCall(PetscSFSetGraphLayout(*newsf, glayout, nLeaves, ilocal_new, PETSC_OWN_POINTER, gremote));
25251f40158dSVaclav Hapla   }
25269566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(*newsf));
25271f40158dSVaclav Hapla   PetscCall(PetscSFViewFromOptions(*newsf, NULL, "-sf_concat_view"));
25281f40158dSVaclav Hapla   PetscCall(PetscLayoutDestroy(&glayout));
25291f40158dSVaclav Hapla   PetscCall(PetscFree(gremote));
25309566063dSJacob Faibussowitsch   PetscCall(PetscFree(leafArrayOffsets));
2531157edd7aSVaclav Hapla   PetscFunctionReturn(0);
2532157edd7aSVaclav Hapla }
2533