xref: /petsc/src/vec/is/sf/interface/sf.c (revision 08401ef684002a709c6d3db98a0c9f54a8bcf1ec)
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>
495fce210SBarry Smith #include <petscctable.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)
1895fce210SBarry Smith #  define PetscSFCheckGraphSet(sf,arg) do {                          \
1995fce210SBarry Smith     if (PetscUnlikely(!(sf)->graphset))                              \
2098921bdaSJacob Faibussowitsch       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must call PetscSFSetGraph() or PetscSFSetGraphWithPattern() on argument %d \"%s\" before %s()",(arg),#sf,PETSC_FUNCTION_NAME); \
2195fce210SBarry Smith   } while (0)
2295fce210SBarry Smith #else
2395fce210SBarry Smith #  define PetscSFCheckGraphSet(sf,arg) do {} while (0)
2495fce210SBarry Smith #endif
252abc8c78SJacob Faibussowitsch #endif
2695fce210SBarry Smith 
274c8fdceaSLisandro Dalcin const char *const PetscSFDuplicateOptions[] = {"CONFONLY","RANKS","GRAPH","PetscSFDuplicateOption","PETSCSF_DUPLICATE_",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:
41dd5b3ca6SJunchao Zhang +  -sf_type basic     -Use MPI persistent Isend/Irecv for communication (Default)
42dd5b3ca6SJunchao Zhang .  -sf_type window    -Use MPI-3 one-sided window for communication
43dd5b3ca6SJunchao Zhang -  -sf_type neighbor  -Use MPI-3 neighborhood collectives for communication
44dd5b3ca6SJunchao Zhang 
4595fce210SBarry Smith    Level: intermediate
4695fce210SBarry Smith 
47dd5b3ca6SJunchao Zhang    Notes:
48dd5b3ca6SJunchao Zhang    When one knows the communication graph is one of the predefined graph, such as MPI_Alltoall, MPI_Allgatherv,
49dd5b3ca6SJunchao Zhang    MPI_Gatherv, one can create a PetscSF and then set its graph with PetscSFSetGraphWithPattern(). These special
50dd5b3ca6SJunchao Zhang    SFs are optimized and they have better performance than general SFs.
51dd5b3ca6SJunchao Zhang 
52dd5b3ca6SJunchao Zhang .seealso: PetscSFSetGraph(), PetscSFSetGraphWithPattern(), PetscSFDestroy()
5395fce210SBarry Smith @*/
5495fce210SBarry Smith PetscErrorCode PetscSFCreate(MPI_Comm comm,PetscSF *sf)
5595fce210SBarry Smith {
5695fce210SBarry Smith   PetscSF        b;
5795fce210SBarry Smith 
5895fce210SBarry Smith   PetscFunctionBegin;
5995fce210SBarry Smith   PetscValidPointer(sf,2);
609566063dSJacob Faibussowitsch   PetscCall(PetscSFInitializePackage());
6195fce210SBarry Smith 
629566063dSJacob Faibussowitsch   PetscCall(PetscHeaderCreate(b,PETSCSF_CLASSID,"PetscSF","Star Forest","PetscSF",comm,PetscSFDestroy,PetscSFView));
6395fce210SBarry Smith 
6495fce210SBarry Smith   b->nroots    = -1;
6595fce210SBarry Smith   b->nleaves   = -1;
6629046d53SLisandro Dalcin   b->minleaf   = PETSC_MAX_INT;
6729046d53SLisandro Dalcin   b->maxleaf   = PETSC_MIN_INT;
6895fce210SBarry Smith   b->nranks    = -1;
6995fce210SBarry Smith   b->rankorder = PETSC_TRUE;
7095fce210SBarry Smith   b->ingroup   = MPI_GROUP_NULL;
7195fce210SBarry Smith   b->outgroup  = MPI_GROUP_NULL;
7295fce210SBarry Smith   b->graphset  = PETSC_FALSE;
7320c24465SJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
7420c24465SJunchao Zhang   b->use_gpu_aware_mpi    = use_gpu_aware_mpi;
7520c24465SJunchao Zhang   b->use_stream_aware_mpi = PETSC_FALSE;
7671438e86SJunchao Zhang   b->unknown_input_stream= PETSC_FALSE;
7727f636e8SJunchao Zhang   #if defined(PETSC_HAVE_KOKKOS) /* Prefer kokkos over cuda*/
7820c24465SJunchao Zhang     b->backend = PETSCSF_BACKEND_KOKKOS;
7927f636e8SJunchao Zhang   #elif defined(PETSC_HAVE_CUDA)
8027f636e8SJunchao Zhang     b->backend = PETSCSF_BACKEND_CUDA;
8159af0bd3SScott Kruger   #elif defined(PETSC_HAVE_HIP)
8259af0bd3SScott Kruger     b->backend = PETSCSF_BACKEND_HIP;
8320c24465SJunchao Zhang   #endif
8471438e86SJunchao Zhang 
8571438e86SJunchao Zhang   #if defined(PETSC_HAVE_NVSHMEM)
8671438e86SJunchao Zhang     b->use_nvshmem     = PETSC_FALSE; /* Default is not to try NVSHMEM */
8771438e86SJunchao Zhang     b->use_nvshmem_get = PETSC_FALSE; /* Default is to use nvshmem_put based protocol */
889566063dSJacob Faibussowitsch     PetscCall(PetscOptionsGetBool(NULL,NULL,"-use_nvshmem",&b->use_nvshmem,NULL));
899566063dSJacob Faibussowitsch     PetscCall(PetscOptionsGetBool(NULL,NULL,"-use_nvshmem_get",&b->use_nvshmem_get,NULL));
9071438e86SJunchao Zhang   #endif
9120c24465SJunchao Zhang #endif
9260c22052SBarry Smith   b->vscat.from_n = -1;
9360c22052SBarry Smith   b->vscat.to_n   = -1;
9460c22052SBarry Smith   b->vscat.unit   = MPIU_SCALAR;
9595fce210SBarry Smith  *sf = b;
9695fce210SBarry Smith   PetscFunctionReturn(0);
9795fce210SBarry Smith }
9895fce210SBarry Smith 
9929046d53SLisandro Dalcin /*@
10095fce210SBarry Smith    PetscSFReset - Reset a star forest so that different sizes or neighbors can be used
10195fce210SBarry Smith 
10295fce210SBarry Smith    Collective
10395fce210SBarry Smith 
1044165533cSJose E. Roman    Input Parameter:
10595fce210SBarry Smith .  sf - star forest
10695fce210SBarry Smith 
10795fce210SBarry Smith    Level: advanced
10895fce210SBarry Smith 
10995fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFSetGraph(), PetscSFDestroy()
11095fce210SBarry Smith @*/
11195fce210SBarry Smith PetscErrorCode PetscSFReset(PetscSF sf)
11295fce210SBarry Smith {
11395fce210SBarry Smith   PetscFunctionBegin;
11495fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
1159566063dSJacob Faibussowitsch   if (sf->ops->Reset) PetscCall((*sf->ops->Reset)(sf));
11629046d53SLisandro Dalcin   sf->nroots   = -1;
11729046d53SLisandro Dalcin   sf->nleaves  = -1;
11829046d53SLisandro Dalcin   sf->minleaf  = PETSC_MAX_INT;
11929046d53SLisandro Dalcin   sf->maxleaf  = PETSC_MIN_INT;
12095fce210SBarry Smith   sf->mine     = NULL;
12195fce210SBarry Smith   sf->remote   = NULL;
12229046d53SLisandro Dalcin   sf->graphset = PETSC_FALSE;
1239566063dSJacob Faibussowitsch   PetscCall(PetscFree(sf->mine_alloc));
1249566063dSJacob Faibussowitsch   PetscCall(PetscFree(sf->remote_alloc));
12521c688dcSJed Brown   sf->nranks = -1;
1269566063dSJacob Faibussowitsch   PetscCall(PetscFree4(sf->ranks,sf->roffset,sf->rmine,sf->rremote));
12729046d53SLisandro Dalcin   sf->degreeknown = PETSC_FALSE;
1289566063dSJacob Faibussowitsch   PetscCall(PetscFree(sf->degree));
1299566063dSJacob Faibussowitsch   if (sf->ingroup  != MPI_GROUP_NULL) PetscCallMPI(MPI_Group_free(&sf->ingroup));
1309566063dSJacob Faibussowitsch   if (sf->outgroup != MPI_GROUP_NULL) PetscCallMPI(MPI_Group_free(&sf->outgroup));
131013b3241SStefano Zampini   if (sf->multi) sf->multi->multi = NULL;
1329566063dSJacob Faibussowitsch   PetscCall(PetscSFDestroy(&sf->multi));
1339566063dSJacob Faibussowitsch   PetscCall(PetscLayoutDestroy(&sf->map));
13471438e86SJunchao Zhang 
13571438e86SJunchao Zhang  #if defined(PETSC_HAVE_DEVICE)
1369566063dSJacob Faibussowitsch   for (PetscInt i=0; i<2; i++) PetscCall(PetscSFFree(sf,PETSC_MEMTYPE_DEVICE,sf->rmine_d[i]));
13771438e86SJunchao Zhang  #endif
13871438e86SJunchao Zhang 
13995fce210SBarry Smith   sf->setupcalled = PETSC_FALSE;
14095fce210SBarry Smith   PetscFunctionReturn(0);
14195fce210SBarry Smith }
14295fce210SBarry Smith 
14395fce210SBarry Smith /*@C
14429046d53SLisandro Dalcin    PetscSFSetType - Set the PetscSF communication implementation
14595fce210SBarry Smith 
14695fce210SBarry Smith    Collective on PetscSF
14795fce210SBarry Smith 
14895fce210SBarry Smith    Input Parameters:
14995fce210SBarry Smith +  sf - the PetscSF context
15095fce210SBarry Smith -  type - a known method
15195fce210SBarry Smith 
15295fce210SBarry Smith    Options Database Key:
15395fce210SBarry Smith .  -sf_type <type> - Sets the method; use -help for a list
15470616304SStefano Zampini    of available methods (for instance, window, basic, neighbor)
15595fce210SBarry Smith 
15695fce210SBarry Smith    Notes:
15795fce210SBarry Smith    See "include/petscsf.h" for available methods (for instance)
15895fce210SBarry Smith +    PETSCSFWINDOW - MPI-2/3 one-sided
15995fce210SBarry Smith -    PETSCSFBASIC - basic implementation using MPI-1 two-sided
16095fce210SBarry Smith 
16195fce210SBarry Smith   Level: intermediate
16295fce210SBarry Smith 
16395fce210SBarry Smith .seealso: PetscSFType, PetscSFCreate()
16495fce210SBarry Smith @*/
16595fce210SBarry Smith PetscErrorCode PetscSFSetType(PetscSF sf,PetscSFType type)
16695fce210SBarry Smith {
16795fce210SBarry Smith   PetscBool      match;
1685f80ce2aSJacob Faibussowitsch   PetscErrorCode (*r)(PetscSF);
16995fce210SBarry Smith 
17095fce210SBarry Smith   PetscFunctionBegin;
17195fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
17295fce210SBarry Smith   PetscValidCharPointer(type,2);
17395fce210SBarry Smith 
1749566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject)sf,type,&match));
17595fce210SBarry Smith   if (match) PetscFunctionReturn(0);
17695fce210SBarry Smith 
1779566063dSJacob Faibussowitsch   PetscCall(PetscFunctionListFind(PetscSFList,type,&r));
17828b400f6SJacob Faibussowitsch   PetscCheck(r,PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unable to find requested PetscSF type %s",type);
17929046d53SLisandro Dalcin   /* Destroy the previous PetscSF implementation context */
1809566063dSJacob Faibussowitsch   if (sf->ops->Destroy) PetscCall((*(sf)->ops->Destroy)(sf));
1819566063dSJacob Faibussowitsch   PetscCall(PetscMemzero(sf->ops,sizeof(*sf->ops)));
1829566063dSJacob Faibussowitsch   PetscCall(PetscObjectChangeTypeName((PetscObject)sf,type));
1839566063dSJacob Faibussowitsch   PetscCall((*r)(sf));
18495fce210SBarry Smith   PetscFunctionReturn(0);
18595fce210SBarry Smith }
18695fce210SBarry Smith 
18729046d53SLisandro Dalcin /*@C
18829046d53SLisandro Dalcin   PetscSFGetType - Get the PetscSF communication implementation
18929046d53SLisandro Dalcin 
19029046d53SLisandro Dalcin   Not Collective
19129046d53SLisandro Dalcin 
19229046d53SLisandro Dalcin   Input Parameter:
19329046d53SLisandro Dalcin . sf  - the PetscSF context
19429046d53SLisandro Dalcin 
19529046d53SLisandro Dalcin   Output Parameter:
19629046d53SLisandro Dalcin . type - the PetscSF type name
19729046d53SLisandro Dalcin 
19829046d53SLisandro Dalcin   Level: intermediate
19929046d53SLisandro Dalcin 
20029046d53SLisandro Dalcin .seealso: PetscSFSetType(), PetscSFCreate()
20129046d53SLisandro Dalcin @*/
20229046d53SLisandro Dalcin PetscErrorCode PetscSFGetType(PetscSF sf, PetscSFType *type)
20329046d53SLisandro Dalcin {
20429046d53SLisandro Dalcin   PetscFunctionBegin;
20529046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID,1);
20629046d53SLisandro Dalcin   PetscValidPointer(type,2);
20729046d53SLisandro Dalcin   *type = ((PetscObject)sf)->type_name;
20829046d53SLisandro Dalcin   PetscFunctionReturn(0);
20929046d53SLisandro Dalcin }
21029046d53SLisandro Dalcin 
2111fb7b255SJunchao Zhang /*@C
21295fce210SBarry Smith    PetscSFDestroy - destroy star forest
21395fce210SBarry Smith 
21495fce210SBarry Smith    Collective
21595fce210SBarry Smith 
2164165533cSJose E. Roman    Input Parameter:
21795fce210SBarry Smith .  sf - address of star forest
21895fce210SBarry Smith 
21995fce210SBarry Smith    Level: intermediate
22095fce210SBarry Smith 
22195fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFReset()
22295fce210SBarry Smith @*/
22395fce210SBarry Smith PetscErrorCode PetscSFDestroy(PetscSF *sf)
22495fce210SBarry Smith {
22595fce210SBarry Smith   PetscFunctionBegin;
22695fce210SBarry Smith   if (!*sf) PetscFunctionReturn(0);
22795fce210SBarry Smith   PetscValidHeaderSpecific((*sf),PETSCSF_CLASSID,1);
22829046d53SLisandro Dalcin   if (--((PetscObject)(*sf))->refct > 0) {*sf = NULL; PetscFunctionReturn(0);}
2299566063dSJacob Faibussowitsch   PetscCall(PetscSFReset(*sf));
2309566063dSJacob Faibussowitsch   if ((*sf)->ops->Destroy) PetscCall((*(*sf)->ops->Destroy)(*sf));
2319566063dSJacob Faibussowitsch   PetscCall(PetscSFDestroy(&(*sf)->vscat.lsf));
2329566063dSJacob Faibussowitsch   if ((*sf)->vscat.bs > 1) PetscCallMPI(MPI_Type_free(&(*sf)->vscat.unit));
2339566063dSJacob Faibussowitsch   PetscCall(PetscHeaderDestroy(sf));
23495fce210SBarry Smith   PetscFunctionReturn(0);
23595fce210SBarry Smith }
23695fce210SBarry Smith 
237c4e6a40aSLawrence Mitchell static PetscErrorCode PetscSFCheckGraphValid_Private(PetscSF sf)
238c4e6a40aSLawrence Mitchell {
239c4e6a40aSLawrence Mitchell   PetscInt           i, nleaves;
240c4e6a40aSLawrence Mitchell   PetscMPIInt        size;
241c4e6a40aSLawrence Mitchell   const PetscInt    *ilocal;
242c4e6a40aSLawrence Mitchell   const PetscSFNode *iremote;
243c4e6a40aSLawrence Mitchell 
244c4e6a40aSLawrence Mitchell   PetscFunctionBegin;
24576bd3646SJed Brown   if (!sf->graphset || !PetscDefined(USE_DEBUG)) PetscFunctionReturn(0);
2469566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sf,NULL,&nleaves,&ilocal,&iremote));
2479566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)sf),&size));
248c4e6a40aSLawrence Mitchell   for (i = 0; i < nleaves; i++) {
249c4e6a40aSLawrence Mitchell     const PetscInt rank = iremote[i].rank;
250c4e6a40aSLawrence Mitchell     const PetscInt remote = iremote[i].index;
251c4e6a40aSLawrence Mitchell     const PetscInt leaf = ilocal ? ilocal[i] : i;
2522c71b3e2SJacob Faibussowitsch     PetscCheckFalse(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);
253*08401ef6SPierre 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);
254*08401ef6SPierre 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);
255c4e6a40aSLawrence Mitchell   }
256c4e6a40aSLawrence Mitchell   PetscFunctionReturn(0);
257c4e6a40aSLawrence Mitchell }
258c4e6a40aSLawrence Mitchell 
25995fce210SBarry Smith /*@
26095fce210SBarry Smith    PetscSFSetUp - set up communication structures
26195fce210SBarry Smith 
26295fce210SBarry Smith    Collective
26395fce210SBarry Smith 
2644165533cSJose E. Roman    Input Parameter:
26595fce210SBarry Smith .  sf - star forest communication object
26695fce210SBarry Smith 
26795fce210SBarry Smith    Level: beginner
26895fce210SBarry Smith 
26995fce210SBarry Smith .seealso: PetscSFSetFromOptions(), PetscSFSetType()
27095fce210SBarry Smith @*/
27195fce210SBarry Smith PetscErrorCode PetscSFSetUp(PetscSF sf)
27295fce210SBarry Smith {
27395fce210SBarry Smith   PetscFunctionBegin;
27429046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
27529046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
27695fce210SBarry Smith   if (sf->setupcalled) PetscFunctionReturn(0);
2779566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_SetUp,sf,0,0,0));
2789566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckGraphValid_Private(sf));
2799566063dSJacob Faibussowitsch   if (!((PetscObject)sf)->type_name) PetscCall(PetscSFSetType(sf,PETSCSFBASIC)); /* Zero all sf->ops */
2809566063dSJacob Faibussowitsch   if (sf->ops->SetUp) PetscCall((*sf->ops->SetUp)(sf));
28120c24465SJunchao Zhang #if defined(PETSC_HAVE_CUDA)
28220c24465SJunchao Zhang   if (sf->backend == PETSCSF_BACKEND_CUDA) {
28371438e86SJunchao Zhang     sf->ops->Malloc = PetscSFMalloc_CUDA;
28471438e86SJunchao Zhang     sf->ops->Free   = PetscSFFree_CUDA;
28520c24465SJunchao Zhang   }
28620c24465SJunchao Zhang #endif
28759af0bd3SScott Kruger #if defined(PETSC_HAVE_HIP)
28859af0bd3SScott Kruger   if (sf->backend == PETSCSF_BACKEND_HIP) {
28959af0bd3SScott Kruger     sf->ops->Malloc = PetscSFMalloc_HIP;
29059af0bd3SScott Kruger     sf->ops->Free   = PetscSFFree_HIP;
29159af0bd3SScott Kruger   }
29259af0bd3SScott Kruger #endif
29320c24465SJunchao Zhang 
29459af0bd3SScott Kruger #
29520c24465SJunchao Zhang #if defined(PETSC_HAVE_KOKKOS)
29620c24465SJunchao Zhang   if (sf->backend == PETSCSF_BACKEND_KOKKOS) {
29720c24465SJunchao Zhang     sf->ops->Malloc = PetscSFMalloc_Kokkos;
29820c24465SJunchao Zhang     sf->ops->Free   = PetscSFFree_Kokkos;
29920c24465SJunchao Zhang   }
30020c24465SJunchao Zhang #endif
3019566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_SetUp,sf,0,0,0));
30295fce210SBarry Smith   sf->setupcalled = PETSC_TRUE;
30395fce210SBarry Smith   PetscFunctionReturn(0);
30495fce210SBarry Smith }
30595fce210SBarry Smith 
3068af6ec1cSBarry Smith /*@
30795fce210SBarry Smith    PetscSFSetFromOptions - set PetscSF options using the options database
30895fce210SBarry Smith 
30995fce210SBarry Smith    Logically Collective
31095fce210SBarry Smith 
3114165533cSJose E. Roman    Input Parameter:
31295fce210SBarry Smith .  sf - star forest
31395fce210SBarry Smith 
31495fce210SBarry Smith    Options Database Keys:
31560263706SJed Brown +  -sf_type               - implementation type, see PetscSFSetType()
31651ccb202SJunchao Zhang .  -sf_rank_order         - sort composite points for gathers and scatters in rank order, gathers are non-deterministic otherwise
317b85e67b7SJunchao Zhang .  -sf_use_default_stream - Assume callers of SF computed the input root/leafdata with the default cuda stream. SF will also
318c2a741eeSJunchao Zhang                             use the default stream to process data. Therefore, no stream synchronization is needed between SF and its caller (default: true).
319c06a8e02SRichard Tran Mills                             If true, this option only works with -use_gpu_aware_mpi 1.
32020c24465SJunchao 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).
321c06a8e02SRichard Tran Mills                                If true, this option only works with -use_gpu_aware_mpi 1.
32295fce210SBarry Smith 
32359af0bd3SScott Kruger -  -sf_backend cuda | hip | kokkos -Select the device backend SF uses. Currently SF has these backends: cuda, hip and Kokkos.
32459af0bd3SScott Kruger                               On CUDA (HIP) devices, one can choose cuda (hip) or kokkos with the default being kokkos. On other devices,
32520c24465SJunchao Zhang                               the only available is kokkos.
32620c24465SJunchao Zhang 
32795fce210SBarry Smith    Level: intermediate
32895fce210SBarry Smith @*/
32995fce210SBarry Smith PetscErrorCode PetscSFSetFromOptions(PetscSF sf)
33095fce210SBarry Smith {
33195fce210SBarry Smith   PetscSFType    deft;
33295fce210SBarry Smith   char           type[256];
33395fce210SBarry Smith   PetscErrorCode ierr;
33495fce210SBarry Smith   PetscBool      flg;
33595fce210SBarry Smith 
33695fce210SBarry Smith   PetscFunctionBegin;
33795fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
3389566063dSJacob Faibussowitsch   ierr = PetscObjectOptionsBegin((PetscObject)sf);PetscCall(ierr);
33995fce210SBarry Smith   deft = ((PetscObject)sf)->type_name ? ((PetscObject)sf)->type_name : PETSCSFBASIC;
3409566063dSJacob Faibussowitsch   PetscCall(PetscOptionsFList("-sf_type","PetscSF implementation type","PetscSFSetType",PetscSFList,deft,type,sizeof(type),&flg));
3419566063dSJacob Faibussowitsch   PetscCall(PetscSFSetType(sf,flg ? type : deft));
3429566063dSJacob 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));
3437fd2d3dbSJunchao Zhang  #if defined(PETSC_HAVE_DEVICE)
34420c24465SJunchao Zhang   {
34520c24465SJunchao Zhang     char        backendstr[32] = {0};
34659af0bd3SScott Kruger     PetscBool   isCuda = PETSC_FALSE,isHip = PETSC_FALSE,isKokkos = PETSC_FALSE,set;
34720c24465SJunchao Zhang     /* Change the defaults set in PetscSFCreate() with command line options */
3489566063dSJacob Faibussowitsch     PetscCall(PetscOptionsBool("-sf_unknown_input_stream","SF root/leafdata is computed on arbitary streams unknown to SF","PetscSFSetFromOptions",sf->unknown_input_stream,&sf->unknown_input_stream,NULL));
3499566063dSJacob 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));
3509566063dSJacob Faibussowitsch     PetscCall(PetscOptionsString("-sf_backend","Select the device backend SF uses","PetscSFSetFromOptions",NULL,backendstr,sizeof(backendstr),&set));
3519566063dSJacob Faibussowitsch     PetscCall(PetscStrcasecmp("cuda",backendstr,&isCuda));
3529566063dSJacob Faibussowitsch     PetscCall(PetscStrcasecmp("kokkos",backendstr,&isKokkos));
3539566063dSJacob Faibussowitsch     PetscCall(PetscStrcasecmp("hip",backendstr,&isHip));
35459af0bd3SScott Kruger   #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP)
35520c24465SJunchao Zhang     if (isCuda) sf->backend = PETSCSF_BACKEND_CUDA;
35620c24465SJunchao Zhang     else if (isKokkos) sf->backend = PETSCSF_BACKEND_KOKKOS;
35759af0bd3SScott Kruger     else if (isHip) sf->backend = PETSCSF_BACKEND_HIP;
35828b400f6SJacob 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);
35920c24465SJunchao Zhang   #elif defined(PETSC_HAVE_KOKKOS)
360*08401ef6SPierre Jolivet     PetscCheck(!set || isKokkos,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"-sf_backend %s is not supported. You can only choose kokkos", backendstr);
36120c24465SJunchao Zhang    #endif
36220c24465SJunchao Zhang   }
363c2a741eeSJunchao Zhang  #endif
3649566063dSJacob Faibussowitsch   if (sf->ops->SetFromOptions) PetscCall((*sf->ops->SetFromOptions)(PetscOptionsObject,sf));
3659566063dSJacob Faibussowitsch   ierr = PetscOptionsEnd();PetscCall(ierr);
36695fce210SBarry Smith   PetscFunctionReturn(0);
36795fce210SBarry Smith }
36895fce210SBarry Smith 
36929046d53SLisandro Dalcin /*@
37095fce210SBarry Smith    PetscSFSetRankOrder - sort multi-points for gathers and scatters by rank order
37195fce210SBarry Smith 
37295fce210SBarry Smith    Logically Collective
37395fce210SBarry Smith 
3744165533cSJose E. Roman    Input Parameters:
37595fce210SBarry Smith +  sf - star forest
37695fce210SBarry Smith -  flg - PETSC_TRUE to sort, PETSC_FALSE to skip sorting (lower setup cost, but non-deterministic)
37795fce210SBarry Smith 
37895fce210SBarry Smith    Level: advanced
37995fce210SBarry Smith 
38095fce210SBarry Smith .seealso: PetscSFGatherBegin(), PetscSFScatterBegin()
38195fce210SBarry Smith @*/
38295fce210SBarry Smith PetscErrorCode PetscSFSetRankOrder(PetscSF sf,PetscBool flg)
38395fce210SBarry Smith {
38495fce210SBarry Smith   PetscFunctionBegin;
38595fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
38695fce210SBarry Smith   PetscValidLogicalCollectiveBool(sf,flg,2);
38728b400f6SJacob Faibussowitsch   PetscCheck(!sf->multi,PetscObjectComm((PetscObject)sf),PETSC_ERR_ARG_WRONGSTATE,"Rank ordering must be set before first call to PetscSFGatherBegin() or PetscSFScatterBegin()");
38895fce210SBarry Smith   sf->rankorder = flg;
38995fce210SBarry Smith   PetscFunctionReturn(0);
39095fce210SBarry Smith }
39195fce210SBarry Smith 
3928dbb0df6SBarry Smith /*@C
39395fce210SBarry Smith    PetscSFSetGraph - Set a parallel star forest
39495fce210SBarry Smith 
39595fce210SBarry Smith    Collective
39695fce210SBarry Smith 
3974165533cSJose E. Roman    Input Parameters:
39895fce210SBarry Smith +  sf - star forest
39995fce210SBarry Smith .  nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves)
40095fce210SBarry Smith .  nleaves - number of leaf vertices on the current process, each of these references a root on any process
401c4e6a40aSLawrence Mitchell .  ilocal - locations of leaves in leafdata buffers, pass NULL for contiguous storage (locations must be >= 0, enforced
402c4e6a40aSLawrence Mitchell during setup in debug mode)
40395fce210SBarry Smith .  localmode - copy mode for ilocal
404c4e6a40aSLawrence Mitchell .  iremote - remote locations of root vertices for each leaf on the current process (locations must be >= 0, enforced
405c4e6a40aSLawrence Mitchell during setup in debug mode)
40695fce210SBarry Smith -  remotemode - copy mode for iremote
40795fce210SBarry Smith 
40895fce210SBarry Smith    Level: intermediate
40995fce210SBarry Smith 
41095452b02SPatrick Sanan    Notes:
411db2b9530SVaclav Hapla    Leaf indices in ilocal must be unique, otherwise an error occurs.
41238ab3f8aSBarry Smith 
413db2b9530SVaclav Hapla    Input arrays ilocal and iremote follow the PetscCopyMode semantics.
414db2b9530SVaclav Hapla    In particular, if localmode/remotemode is PETSC_OWN_POINTER or PETSC_USE_POINTER,
415db2b9530SVaclav Hapla    PETSc might modify the respective array;
416db2b9530SVaclav Hapla    if PETSC_USE_POINTER, the user must delete the array after PetscSFDestroy().
417e97d3de3SVaclav Hapla    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).
418db2b9530SVaclav Hapla 
419db2b9530SVaclav Hapla    Fortran Notes:
420db2b9530SVaclav Hapla    In Fortran you must use PETSC_COPY_VALUES for localmode and remotemode.
421c4e6a40aSLawrence Mitchell 
42272bf8598SVaclav Hapla    Developer Notes:
423db2b9530SVaclav Hapla    We sort leaves to check for duplicates and contiguousness and to find minleaf/maxleaf.
424db2b9530SVaclav Hapla    This also allows to compare leaf sets of two SFs easily.
42572bf8598SVaclav Hapla 
42695fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFView(), PetscSFGetGraph()
42795fce210SBarry Smith @*/
428db2b9530SVaclav Hapla PetscErrorCode PetscSFSetGraph(PetscSF sf,PetscInt nroots,PetscInt nleaves,PetscInt *ilocal,PetscCopyMode localmode,PetscSFNode *iremote,PetscCopyMode remotemode)
42995fce210SBarry Smith {
430db2b9530SVaclav Hapla   PetscBool       unique, contiguous;
43195fce210SBarry Smith 
43295fce210SBarry Smith   PetscFunctionBegin;
43395fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
43429046d53SLisandro Dalcin   if (nleaves > 0 && ilocal) PetscValidIntPointer(ilocal,4);
43529046d53SLisandro Dalcin   if (nleaves > 0) PetscValidPointer(iremote,6);
436*08401ef6SPierre Jolivet   PetscCheck(nroots  >= 0,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nroots %" PetscInt_FMT ", cannot be negative",nroots);
437*08401ef6SPierre Jolivet   PetscCheck(nleaves >= 0,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nleaves %" PetscInt_FMT ", cannot be negative",nleaves);
438db2b9530SVaclav Hapla   PetscCheck(localmode  >= PETSC_COPY_VALUES && localmode  <= PETSC_USE_POINTER,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Wrong localmode %d",localmode);
439db2b9530SVaclav Hapla   PetscCheck(remotemode >= PETSC_COPY_VALUES && remotemode <= PETSC_USE_POINTER,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Wrong remotemode %d",remotemode);
44029046d53SLisandro Dalcin 
4412a67d2daSStefano Zampini   if (sf->nroots >= 0) { /* Reset only if graph already set */
4429566063dSJacob Faibussowitsch     PetscCall(PetscSFReset(sf));
4432a67d2daSStefano Zampini   }
4442a67d2daSStefano Zampini 
4459566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_SetGraph,sf,0,0,0));
44629046d53SLisandro Dalcin 
44795fce210SBarry Smith   sf->nroots  = nroots;
44895fce210SBarry Smith   sf->nleaves = nleaves;
44929046d53SLisandro Dalcin 
450db2b9530SVaclav Hapla   if (localmode == PETSC_COPY_VALUES && ilocal) {
451db2b9530SVaclav Hapla     PetscInt *tlocal = NULL;
452db2b9530SVaclav Hapla 
4539566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nleaves,&tlocal));
4549566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(tlocal,ilocal,nleaves));
455db2b9530SVaclav Hapla     ilocal = tlocal;
456db2b9530SVaclav Hapla   }
457db2b9530SVaclav Hapla   if (remotemode == PETSC_COPY_VALUES) {
458db2b9530SVaclav Hapla     PetscSFNode *tremote = NULL;
459db2b9530SVaclav Hapla 
4609566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nleaves,&tremote));
4619566063dSJacob Faibussowitsch     PetscCall(PetscArraycpy(tremote,iremote,nleaves));
462db2b9530SVaclav Hapla     iremote = tremote;
463db2b9530SVaclav Hapla   }
464db2b9530SVaclav Hapla 
46529046d53SLisandro Dalcin   if (nleaves && ilocal) {
466db2b9530SVaclav Hapla     PetscSFNode   work;
467db2b9530SVaclav Hapla 
4689566063dSJacob Faibussowitsch     PetscCall(PetscSortIntWithDataArray(nleaves, ilocal, iremote, sizeof(PetscSFNode), &work));
4699566063dSJacob Faibussowitsch     PetscCall(PetscSortedCheckDupsInt(nleaves, ilocal, &unique));
470db2b9530SVaclav Hapla     unique = PetscNot(unique);
471db2b9530SVaclav 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");
472db2b9530SVaclav Hapla     sf->minleaf = ilocal[0];
473db2b9530SVaclav Hapla     sf->maxleaf = ilocal[nleaves-1];
474db2b9530SVaclav Hapla     contiguous = (PetscBool) (unique && ilocal[0] == 0 && ilocal[nleaves-1] == nleaves-1);
47529046d53SLisandro Dalcin   } else {
47629046d53SLisandro Dalcin     sf->minleaf = 0;
47729046d53SLisandro Dalcin     sf->maxleaf = nleaves - 1;
478db2b9530SVaclav Hapla     unique      = PETSC_TRUE;
479db2b9530SVaclav Hapla     contiguous  = PETSC_TRUE;
48029046d53SLisandro Dalcin   }
48129046d53SLisandro Dalcin 
482db2b9530SVaclav Hapla   if (contiguous) {
483db2b9530SVaclav Hapla     if (localmode == PETSC_USE_POINTER) {
484db2b9530SVaclav Hapla       ilocal = NULL;
485db2b9530SVaclav Hapla     } else {
4869566063dSJacob Faibussowitsch       PetscCall(PetscFree(ilocal));
487db2b9530SVaclav Hapla     }
488db2b9530SVaclav Hapla   }
489db2b9530SVaclav Hapla   sf->mine            = ilocal;
490db2b9530SVaclav Hapla   if (localmode == PETSC_USE_POINTER) {
49129046d53SLisandro Dalcin     sf->mine_alloc    = NULL;
492db2b9530SVaclav Hapla   } else {
493db2b9530SVaclav Hapla     sf->mine_alloc    = ilocal;
49495fce210SBarry Smith   }
495db2b9530SVaclav Hapla   sf->remote          = iremote;
496db2b9530SVaclav Hapla   if (remotemode == PETSC_USE_POINTER) {
49729046d53SLisandro Dalcin     sf->remote_alloc  = NULL;
498db2b9530SVaclav Hapla   } else {
499db2b9530SVaclav Hapla     sf->remote_alloc  = iremote;
50095fce210SBarry Smith   }
5019566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_SetGraph,sf,0,0,0));
50229046d53SLisandro Dalcin   sf->graphset = PETSC_TRUE;
50395fce210SBarry Smith   PetscFunctionReturn(0);
50495fce210SBarry Smith }
50595fce210SBarry Smith 
50629046d53SLisandro Dalcin /*@
507dd5b3ca6SJunchao Zhang   PetscSFSetGraphWithPattern - Sets the graph of an SF with a specific pattern
508dd5b3ca6SJunchao Zhang 
509dd5b3ca6SJunchao Zhang   Collective
510dd5b3ca6SJunchao Zhang 
511dd5b3ca6SJunchao Zhang   Input Parameters:
512dd5b3ca6SJunchao Zhang + sf      - The PetscSF
513dd5b3ca6SJunchao Zhang . map     - Layout of roots over all processes (insignificant when pattern is PETSCSF_PATTERN_ALLTOALL)
514dd5b3ca6SJunchao Zhang - pattern - One of PETSCSF_PATTERN_ALLGATHER, PETSCSF_PATTERN_GATHER, PETSCSF_PATTERN_ALLTOALL
515dd5b3ca6SJunchao Zhang 
516dd5b3ca6SJunchao Zhang   Notes:
517dd5b3ca6SJunchao Zhang   It is easier to explain PetscSFPattern using vectors. Suppose we have an MPI vector x and its layout is map.
518dd5b3ca6SJunchao Zhang   n and N are local and global sizes of x respectively.
519dd5b3ca6SJunchao Zhang 
520dd5b3ca6SJunchao Zhang   With PETSCSF_PATTERN_ALLGATHER, the routine creates a graph that if one does Bcast on it, it will copy x to
521dd5b3ca6SJunchao Zhang   sequential vectors y on all ranks.
522dd5b3ca6SJunchao Zhang 
523dd5b3ca6SJunchao Zhang   With PETSCSF_PATTERN_GATHER, the routine creates a graph that if one does Bcast on it, it will copy x to a
524dd5b3ca6SJunchao Zhang   sequential vector y on rank 0.
525dd5b3ca6SJunchao Zhang 
526dd5b3ca6SJunchao Zhang   In above cases, entries of x are roots and entries of y are leaves.
527dd5b3ca6SJunchao Zhang 
528dd5b3ca6SJunchao Zhang   With PETSCSF_PATTERN_ALLTOALL, map is insignificant. Suppose NP is size of sf's communicator. The routine
529dd5b3ca6SJunchao Zhang   creates a graph that every rank has NP leaves and NP roots. On rank i, its leaf j is connected to root i
530dd5b3ca6SJunchao Zhang   of rank j. Here 0 <=i,j<NP. It is a kind of MPI_Alltoall with sendcount/recvcount being 1. Note that it does
531dd5b3ca6SJunchao Zhang   not mean one can not send multiple items. One just needs to create a new MPI datatype for the mulptiple data
532dd5b3ca6SJunchao Zhang   items with MPI_Type_contiguous() and use that as the <unit> argument in SF routines.
533dd5b3ca6SJunchao Zhang 
534dd5b3ca6SJunchao Zhang   In this case, roots and leaves are symmetric.
535dd5b3ca6SJunchao Zhang 
536dd5b3ca6SJunchao Zhang   Level: intermediate
537dd5b3ca6SJunchao Zhang  @*/
538dd5b3ca6SJunchao Zhang PetscErrorCode PetscSFSetGraphWithPattern(PetscSF sf,PetscLayout map,PetscSFPattern pattern)
539dd5b3ca6SJunchao Zhang {
540dd5b3ca6SJunchao Zhang   MPI_Comm       comm;
541dd5b3ca6SJunchao Zhang   PetscInt       n,N,res[2];
542dd5b3ca6SJunchao Zhang   PetscMPIInt    rank,size;
543dd5b3ca6SJunchao Zhang   PetscSFType    type;
544dd5b3ca6SJunchao Zhang 
545dd5b3ca6SJunchao Zhang   PetscFunctionBegin;
5462abc8c78SJacob Faibussowitsch   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
5472abc8c78SJacob Faibussowitsch   if (pattern != PETSCSF_PATTERN_ALLTOALL) PetscValidPointer(map,2);
5489566063dSJacob Faibussowitsch   PetscCall(PetscObjectGetComm((PetscObject)sf, &comm));
5492c71b3e2SJacob Faibussowitsch   PetscCheck(pattern >= PETSCSF_PATTERN_ALLGATHER && pattern <= PETSCSF_PATTERN_ALLTOALL,comm,PETSC_ERR_ARG_OUTOFRANGE,"Unsupported PetscSFPattern %d",pattern);
5509566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_rank(comm,&rank));
5519566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_size(comm,&size));
552dd5b3ca6SJunchao Zhang 
553dd5b3ca6SJunchao Zhang   if (pattern == PETSCSF_PATTERN_ALLTOALL) {
554dd5b3ca6SJunchao Zhang     type = PETSCSFALLTOALL;
5559566063dSJacob Faibussowitsch     PetscCall(PetscLayoutCreate(comm,&sf->map));
5569566063dSJacob Faibussowitsch     PetscCall(PetscLayoutSetLocalSize(sf->map,size));
5579566063dSJacob Faibussowitsch     PetscCall(PetscLayoutSetSize(sf->map,((PetscInt)size)*size));
5589566063dSJacob Faibussowitsch     PetscCall(PetscLayoutSetUp(sf->map));
559dd5b3ca6SJunchao Zhang   } else {
5609566063dSJacob Faibussowitsch     PetscCall(PetscLayoutGetLocalSize(map,&n));
5619566063dSJacob Faibussowitsch     PetscCall(PetscLayoutGetSize(map,&N));
562dd5b3ca6SJunchao Zhang     res[0] = n;
563dd5b3ca6SJunchao Zhang     res[1] = -n;
564dd5b3ca6SJunchao Zhang     /* Check if n are same over all ranks so that we can optimize it */
5651c2dc1cbSBarry Smith     PetscCall(MPIU_Allreduce(MPI_IN_PLACE,res,2,MPIU_INT,MPI_MAX,comm));
566dd5b3ca6SJunchao Zhang     if (res[0] == -res[1]) { /* same n */
567dd5b3ca6SJunchao Zhang       type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHER  : PETSCSFGATHER;
568dd5b3ca6SJunchao Zhang     } else {
569dd5b3ca6SJunchao Zhang       type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHERV : PETSCSFGATHERV;
570dd5b3ca6SJunchao Zhang     }
5719566063dSJacob Faibussowitsch     PetscCall(PetscLayoutReference(map,&sf->map));
572dd5b3ca6SJunchao Zhang   }
5739566063dSJacob Faibussowitsch   PetscCall(PetscSFSetType(sf,type));
574dd5b3ca6SJunchao Zhang 
575dd5b3ca6SJunchao Zhang   sf->pattern = pattern;
576dd5b3ca6SJunchao Zhang   sf->mine    = NULL; /* Contiguous */
577dd5b3ca6SJunchao Zhang 
578dd5b3ca6SJunchao Zhang   /* Set nleaves, nroots here in case user calls PetscSFGetGraph, which is legal to call even before PetscSFSetUp is called.
579dd5b3ca6SJunchao Zhang      Also set other easy stuff.
580dd5b3ca6SJunchao Zhang    */
581dd5b3ca6SJunchao Zhang   if (pattern == PETSCSF_PATTERN_ALLGATHER) {
582dd5b3ca6SJunchao Zhang     sf->nleaves      = N;
583dd5b3ca6SJunchao Zhang     sf->nroots       = n;
584dd5b3ca6SJunchao Zhang     sf->nranks       = size;
585dd5b3ca6SJunchao Zhang     sf->minleaf      = 0;
586dd5b3ca6SJunchao Zhang     sf->maxleaf      = N - 1;
587dd5b3ca6SJunchao Zhang   } else if (pattern == PETSCSF_PATTERN_GATHER) {
588dd5b3ca6SJunchao Zhang     sf->nleaves      = rank ? 0 : N;
589dd5b3ca6SJunchao Zhang     sf->nroots       = n;
590dd5b3ca6SJunchao Zhang     sf->nranks       = rank ? 0 : size;
591dd5b3ca6SJunchao Zhang     sf->minleaf      = 0;
592dd5b3ca6SJunchao Zhang     sf->maxleaf      = rank ? -1 : N - 1;
593dd5b3ca6SJunchao Zhang   } else if (pattern == PETSCSF_PATTERN_ALLTOALL) {
594dd5b3ca6SJunchao Zhang     sf->nleaves      = size;
595dd5b3ca6SJunchao Zhang     sf->nroots       = size;
596dd5b3ca6SJunchao Zhang     sf->nranks       = size;
597dd5b3ca6SJunchao Zhang     sf->minleaf      = 0;
598dd5b3ca6SJunchao Zhang     sf->maxleaf      = size - 1;
599dd5b3ca6SJunchao Zhang   }
600dd5b3ca6SJunchao Zhang   sf->ndranks  = 0; /* We do not need to separate out distinguished ranks for patterned graphs to improve communication performance */
601dd5b3ca6SJunchao Zhang   sf->graphset = PETSC_TRUE;
602dd5b3ca6SJunchao Zhang   PetscFunctionReturn(0);
603dd5b3ca6SJunchao Zhang }
604dd5b3ca6SJunchao Zhang 
605dd5b3ca6SJunchao Zhang /*@
60695fce210SBarry Smith    PetscSFCreateInverseSF - given a PetscSF in which all vertices have degree 1, creates the inverse map
60795fce210SBarry Smith 
60895fce210SBarry Smith    Collective
60995fce210SBarry Smith 
6104165533cSJose E. Roman    Input Parameter:
61195fce210SBarry Smith .  sf - star forest to invert
61295fce210SBarry Smith 
6134165533cSJose E. Roman    Output Parameter:
61495fce210SBarry Smith .  isf - inverse of sf
6154165533cSJose E. Roman 
61695fce210SBarry Smith    Level: advanced
61795fce210SBarry Smith 
61895fce210SBarry Smith    Notes:
61995fce210SBarry Smith    All roots must have degree 1.
62095fce210SBarry Smith 
62195fce210SBarry Smith    The local space may be a permutation, but cannot be sparse.
62295fce210SBarry Smith 
62395fce210SBarry Smith .seealso: PetscSFSetGraph()
62495fce210SBarry Smith @*/
62595fce210SBarry Smith PetscErrorCode PetscSFCreateInverseSF(PetscSF sf,PetscSF *isf)
62695fce210SBarry Smith {
62795fce210SBarry Smith   PetscMPIInt    rank;
62895fce210SBarry Smith   PetscInt       i,nroots,nleaves,maxlocal,count,*newilocal;
62995fce210SBarry Smith   const PetscInt *ilocal;
63095fce210SBarry Smith   PetscSFNode    *roots,*leaves;
63195fce210SBarry Smith 
63295fce210SBarry Smith   PetscFunctionBegin;
63329046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
63429046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
63529046d53SLisandro Dalcin   PetscValidPointer(isf,2);
63629046d53SLisandro Dalcin 
6379566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,NULL));
63829046d53SLisandro Dalcin   maxlocal = sf->maxleaf+1; /* TODO: We should use PetscSFGetLeafRange() */
63929046d53SLisandro Dalcin 
6409566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank));
6419566063dSJacob Faibussowitsch   PetscCall(PetscMalloc2(nroots,&roots,maxlocal,&leaves));
642ae9aee6dSMatthew G. Knepley   for (i=0; i<maxlocal; i++) {
64395fce210SBarry Smith     leaves[i].rank  = rank;
64495fce210SBarry Smith     leaves[i].index = i;
64595fce210SBarry Smith   }
64695fce210SBarry Smith   for (i=0; i <nroots; i++) {
64795fce210SBarry Smith     roots[i].rank  = -1;
64895fce210SBarry Smith     roots[i].index = -1;
64995fce210SBarry Smith   }
6509566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceBegin(sf,MPIU_2INT,leaves,roots,MPI_REPLACE));
6519566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceEnd(sf,MPIU_2INT,leaves,roots,MPI_REPLACE));
65295fce210SBarry Smith 
65395fce210SBarry Smith   /* Check whether our leaves are sparse */
65495fce210SBarry Smith   for (i=0,count=0; i<nroots; i++) if (roots[i].rank >= 0) count++;
65595fce210SBarry Smith   if (count == nroots) newilocal = NULL;
65695fce210SBarry Smith   else {                        /* Index for sparse leaves and compact "roots" array (which is to become our leaves). */
6579566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(count,&newilocal));
65895fce210SBarry Smith     for (i=0,count=0; i<nroots; i++) {
65995fce210SBarry Smith       if (roots[i].rank >= 0) {
66095fce210SBarry Smith         newilocal[count]   = i;
66195fce210SBarry Smith         roots[count].rank  = roots[i].rank;
66295fce210SBarry Smith         roots[count].index = roots[i].index;
66395fce210SBarry Smith         count++;
66495fce210SBarry Smith       }
66595fce210SBarry Smith     }
66695fce210SBarry Smith   }
66795fce210SBarry Smith 
6689566063dSJacob Faibussowitsch   PetscCall(PetscSFDuplicate(sf,PETSCSF_DUPLICATE_CONFONLY,isf));
6699566063dSJacob Faibussowitsch   PetscCall(PetscSFSetGraph(*isf,maxlocal,count,newilocal,PETSC_OWN_POINTER,roots,PETSC_COPY_VALUES));
6709566063dSJacob Faibussowitsch   PetscCall(PetscFree2(roots,leaves));
67195fce210SBarry Smith   PetscFunctionReturn(0);
67295fce210SBarry Smith }
67395fce210SBarry Smith 
67495fce210SBarry Smith /*@
67595fce210SBarry Smith    PetscSFDuplicate - duplicate a PetscSF, optionally preserving rank connectivity and graph
67695fce210SBarry Smith 
67795fce210SBarry Smith    Collective
67895fce210SBarry Smith 
6794165533cSJose E. Roman    Input Parameters:
68095fce210SBarry Smith +  sf - communication object to duplicate
68195fce210SBarry Smith -  opt - PETSCSF_DUPLICATE_CONFONLY, PETSCSF_DUPLICATE_RANKS, or PETSCSF_DUPLICATE_GRAPH (see PetscSFDuplicateOption)
68295fce210SBarry Smith 
6834165533cSJose E. Roman    Output Parameter:
68495fce210SBarry Smith .  newsf - new communication object
68595fce210SBarry Smith 
68695fce210SBarry Smith    Level: beginner
68795fce210SBarry Smith 
68895fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFSetType(), PetscSFSetGraph()
68995fce210SBarry Smith @*/
69095fce210SBarry Smith PetscErrorCode PetscSFDuplicate(PetscSF sf,PetscSFDuplicateOption opt,PetscSF *newsf)
69195fce210SBarry Smith {
69229046d53SLisandro Dalcin   PetscSFType    type;
69397929ea7SJunchao Zhang   MPI_Datatype   dtype=MPIU_SCALAR;
69495fce210SBarry Smith 
69595fce210SBarry Smith   PetscFunctionBegin;
69629046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
69729046d53SLisandro Dalcin   PetscValidLogicalCollectiveEnum(sf,opt,2);
69829046d53SLisandro Dalcin   PetscValidPointer(newsf,3);
6999566063dSJacob Faibussowitsch   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sf),newsf));
7009566063dSJacob Faibussowitsch   PetscCall(PetscSFGetType(sf,&type));
7019566063dSJacob Faibussowitsch   if (type) PetscCall(PetscSFSetType(*newsf,type));
70295fce210SBarry Smith   if (opt == PETSCSF_DUPLICATE_GRAPH) {
703dd5b3ca6SJunchao Zhang     PetscSFCheckGraphSet(sf,1);
704dd5b3ca6SJunchao Zhang     if (sf->pattern == PETSCSF_PATTERN_GENERAL) {
70595fce210SBarry Smith       PetscInt          nroots,nleaves;
70695fce210SBarry Smith       const PetscInt    *ilocal;
70795fce210SBarry Smith       const PetscSFNode *iremote;
7089566063dSJacob Faibussowitsch       PetscCall(PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,&iremote));
7099566063dSJacob Faibussowitsch       PetscCall(PetscSFSetGraph(*newsf,nroots,nleaves,(PetscInt*)ilocal,PETSC_COPY_VALUES,(PetscSFNode*)iremote,PETSC_COPY_VALUES));
710dd5b3ca6SJunchao Zhang     } else {
7119566063dSJacob Faibussowitsch       PetscCall(PetscSFSetGraphWithPattern(*newsf,sf->map,sf->pattern));
712dd5b3ca6SJunchao Zhang     }
71395fce210SBarry Smith   }
71497929ea7SJunchao Zhang   /* Since oldtype is committed, so is newtype, according to MPI */
7159566063dSJacob Faibussowitsch   if (sf->vscat.bs > 1) PetscCallMPI(MPI_Type_dup(sf->vscat.unit,&dtype));
71697929ea7SJunchao Zhang   (*newsf)->vscat.bs     = sf->vscat.bs;
71797929ea7SJunchao Zhang   (*newsf)->vscat.unit   = dtype;
71897929ea7SJunchao Zhang   (*newsf)->vscat.to_n   = sf->vscat.to_n;
71997929ea7SJunchao Zhang   (*newsf)->vscat.from_n = sf->vscat.from_n;
72097929ea7SJunchao Zhang   /* Do not copy lsf. Build it on demand since it is rarely used */
72197929ea7SJunchao Zhang 
72220c24465SJunchao Zhang #if defined(PETSC_HAVE_DEVICE)
72320c24465SJunchao Zhang   (*newsf)->backend              = sf->backend;
72471438e86SJunchao Zhang   (*newsf)->unknown_input_stream= sf->unknown_input_stream;
72520c24465SJunchao Zhang   (*newsf)->use_gpu_aware_mpi    = sf->use_gpu_aware_mpi;
72620c24465SJunchao Zhang   (*newsf)->use_stream_aware_mpi = sf->use_stream_aware_mpi;
72720c24465SJunchao Zhang #endif
7289566063dSJacob Faibussowitsch   if (sf->ops->Duplicate) PetscCall((*sf->ops->Duplicate)(sf,opt,*newsf));
72920c24465SJunchao Zhang   /* Don't do PetscSFSetUp() since the new sf's graph might have not been set. */
73095fce210SBarry Smith   PetscFunctionReturn(0);
73195fce210SBarry Smith }
73295fce210SBarry Smith 
73395fce210SBarry Smith /*@C
73495fce210SBarry Smith    PetscSFGetGraph - Get the graph specifying a parallel star forest
73595fce210SBarry Smith 
73695fce210SBarry Smith    Not Collective
73795fce210SBarry Smith 
7384165533cSJose E. Roman    Input Parameter:
73995fce210SBarry Smith .  sf - star forest
74095fce210SBarry Smith 
7414165533cSJose E. Roman    Output Parameters:
74295fce210SBarry Smith +  nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves)
74395fce210SBarry Smith .  nleaves - number of leaf vertices on the current process, each of these references a root on any process
744bc6585dcSJunchao Zhang .  ilocal - locations of leaves in leafdata buffers (if returned value is NULL, it means leaves are in contiguous storage)
74595fce210SBarry Smith -  iremote - remote locations of root vertices for each leaf on the current process
74695fce210SBarry Smith 
747373e0d91SLisandro Dalcin    Notes:
748373e0d91SLisandro Dalcin      We are not currently requiring that the graph is set, thus returning nroots=-1 if it has not been set yet
749373e0d91SLisandro Dalcin 
750db2b9530SVaclav Hapla      The returned ilocal/iremote might contain values in different order than the input ones in PetscSFSetGraph(),
751db2b9530SVaclav Hapla      see its manpage for details.
752db2b9530SVaclav Hapla 
7538dbb0df6SBarry Smith    Fortran Notes:
7548dbb0df6SBarry Smith      The returned iremote array is a copy and must be deallocated after use. Consequently, if you
7558dbb0df6SBarry Smith      want to update the graph, you must call PetscSFSetGraph() after modifying the iremote array.
7568dbb0df6SBarry Smith 
7578dbb0df6SBarry Smith      To check for a NULL ilocal use
7588dbb0df6SBarry Smith $      if (loc(ilocal) == loc(PETSC_NULL_INTEGER)) then
759ca797d7aSLawrence Mitchell 
76095fce210SBarry Smith    Level: intermediate
76195fce210SBarry Smith 
76295fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFView(), PetscSFSetGraph()
76395fce210SBarry Smith @*/
76495fce210SBarry Smith PetscErrorCode PetscSFGetGraph(PetscSF sf,PetscInt *nroots,PetscInt *nleaves,const PetscInt **ilocal,const PetscSFNode **iremote)
76595fce210SBarry Smith {
76695fce210SBarry Smith   PetscFunctionBegin;
76795fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
768b8dee149SJunchao Zhang   if (sf->ops->GetGraph) {
7699566063dSJacob Faibussowitsch     PetscCall((sf->ops->GetGraph)(sf,nroots,nleaves,ilocal,iremote));
770b8dee149SJunchao Zhang   } else {
77195fce210SBarry Smith     if (nroots) *nroots = sf->nroots;
77295fce210SBarry Smith     if (nleaves) *nleaves = sf->nleaves;
77395fce210SBarry Smith     if (ilocal) *ilocal = sf->mine;
77495fce210SBarry Smith     if (iremote) *iremote = sf->remote;
775b8dee149SJunchao Zhang   }
77695fce210SBarry Smith   PetscFunctionReturn(0);
77795fce210SBarry Smith }
77895fce210SBarry Smith 
77929046d53SLisandro Dalcin /*@
78095fce210SBarry Smith    PetscSFGetLeafRange - Get the active leaf ranges
78195fce210SBarry Smith 
78295fce210SBarry Smith    Not Collective
78395fce210SBarry Smith 
7844165533cSJose E. Roman    Input Parameter:
78595fce210SBarry Smith .  sf - star forest
78695fce210SBarry Smith 
7874165533cSJose E. Roman    Output Parameters:
788dd5b3ca6SJunchao Zhang +  minleaf - minimum active leaf on this process. Return 0 if there are no leaves.
789dd5b3ca6SJunchao Zhang -  maxleaf - maximum active leaf on this process. Return -1 if there are no leaves.
79095fce210SBarry Smith 
79195fce210SBarry Smith    Level: developer
79295fce210SBarry Smith 
79395fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFView(), PetscSFSetGraph(), PetscSFGetGraph()
79495fce210SBarry Smith @*/
79595fce210SBarry Smith PetscErrorCode PetscSFGetLeafRange(PetscSF sf,PetscInt *minleaf,PetscInt *maxleaf)
79695fce210SBarry Smith {
79795fce210SBarry Smith   PetscFunctionBegin;
79895fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
79929046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
80095fce210SBarry Smith   if (minleaf) *minleaf = sf->minleaf;
80195fce210SBarry Smith   if (maxleaf) *maxleaf = sf->maxleaf;
80295fce210SBarry Smith   PetscFunctionReturn(0);
80395fce210SBarry Smith }
80495fce210SBarry Smith 
80595fce210SBarry Smith /*@C
806fe2efc57SMark    PetscSFViewFromOptions - View from Options
807fe2efc57SMark 
808fe2efc57SMark    Collective on PetscSF
809fe2efc57SMark 
810fe2efc57SMark    Input Parameters:
811fe2efc57SMark +  A - the star forest
812736c3998SJose E. Roman .  obj - Optional object
813736c3998SJose E. Roman -  name - command line option
814fe2efc57SMark 
815fe2efc57SMark    Level: intermediate
816fe2efc57SMark .seealso:  PetscSF, PetscSFView, PetscObjectViewFromOptions(), PetscSFCreate()
817fe2efc57SMark @*/
818fe2efc57SMark PetscErrorCode  PetscSFViewFromOptions(PetscSF A,PetscObject obj,const char name[])
819fe2efc57SMark {
820fe2efc57SMark   PetscFunctionBegin;
821fe2efc57SMark   PetscValidHeaderSpecific(A,PETSCSF_CLASSID,1);
8229566063dSJacob Faibussowitsch   PetscCall(PetscObjectViewFromOptions((PetscObject)A,obj,name));
823fe2efc57SMark   PetscFunctionReturn(0);
824fe2efc57SMark }
825fe2efc57SMark 
826fe2efc57SMark /*@C
82795fce210SBarry Smith    PetscSFView - view a star forest
82895fce210SBarry Smith 
82995fce210SBarry Smith    Collective
83095fce210SBarry Smith 
8314165533cSJose E. Roman    Input Parameters:
83295fce210SBarry Smith +  sf - star forest
83395fce210SBarry Smith -  viewer - viewer to display graph, for example PETSC_VIEWER_STDOUT_WORLD
83495fce210SBarry Smith 
83595fce210SBarry Smith    Level: beginner
83695fce210SBarry Smith 
83795fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFSetGraph()
83895fce210SBarry Smith @*/
83995fce210SBarry Smith PetscErrorCode PetscSFView(PetscSF sf,PetscViewer viewer)
84095fce210SBarry Smith {
84195fce210SBarry Smith   PetscBool         iascii;
84295fce210SBarry Smith   PetscViewerFormat format;
84395fce210SBarry Smith 
84495fce210SBarry Smith   PetscFunctionBegin;
84595fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
8469566063dSJacob Faibussowitsch   if (!viewer) PetscCall(PetscViewerASCIIGetStdout(PetscObjectComm((PetscObject)sf),&viewer));
84795fce210SBarry Smith   PetscValidHeaderSpecific(viewer,PETSC_VIEWER_CLASSID,2);
84895fce210SBarry Smith   PetscCheckSameComm(sf,1,viewer,2);
8499566063dSJacob Faibussowitsch   if (sf->graphset) PetscCall(PetscSFSetUp(sf));
8509566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii));
85153dd6d7dSJunchao Zhang   if (iascii && viewer->format != PETSC_VIEWER_ASCII_MATLAB) {
85295fce210SBarry Smith     PetscMPIInt rank;
85381bfa7aaSJed Brown     PetscInt    ii,i,j;
85495fce210SBarry Smith 
8559566063dSJacob Faibussowitsch     PetscCall(PetscObjectPrintClassNamePrefixType((PetscObject)sf,viewer));
8569566063dSJacob Faibussowitsch     PetscCall(PetscViewerASCIIPushTab(viewer));
857dd5b3ca6SJunchao Zhang     if (sf->pattern == PETSCSF_PATTERN_GENERAL) {
85880153354SVaclav Hapla       if (!sf->graphset) {
8599566063dSJacob Faibussowitsch         PetscCall(PetscViewerASCIIPrintf(viewer,"PetscSFSetGraph() has not been called yet\n"));
8609566063dSJacob Faibussowitsch         PetscCall(PetscViewerASCIIPopTab(viewer));
86180153354SVaclav Hapla         PetscFunctionReturn(0);
86280153354SVaclav Hapla       }
8639566063dSJacob Faibussowitsch       PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank));
8649566063dSJacob Faibussowitsch       PetscCall(PetscViewerASCIIPushSynchronized(viewer));
8659566063dSJacob 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));
86695fce210SBarry Smith       for (i=0; i<sf->nleaves; i++) {
8679566063dSJacob Faibussowitsch         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));
86895fce210SBarry Smith       }
8699566063dSJacob Faibussowitsch       PetscCall(PetscViewerFlush(viewer));
8709566063dSJacob Faibussowitsch       PetscCall(PetscViewerGetFormat(viewer,&format));
87195fce210SBarry Smith       if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
87281bfa7aaSJed Brown         PetscMPIInt *tmpranks,*perm;
8739566063dSJacob Faibussowitsch         PetscCall(PetscMalloc2(sf->nranks,&tmpranks,sf->nranks,&perm));
8749566063dSJacob Faibussowitsch         PetscCall(PetscArraycpy(tmpranks,sf->ranks,sf->nranks));
87581bfa7aaSJed Brown         for (i=0; i<sf->nranks; i++) perm[i] = i;
8769566063dSJacob Faibussowitsch         PetscCall(PetscSortMPIIntWithArray(sf->nranks,tmpranks,perm));
8779566063dSJacob Faibussowitsch         PetscCall(PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Roots referenced by my leaves, by rank\n",rank));
87881bfa7aaSJed Brown         for (ii=0; ii<sf->nranks; ii++) {
87981bfa7aaSJed Brown           i = perm[ii];
8809566063dSJacob Faibussowitsch           PetscCall(PetscViewerASCIISynchronizedPrintf(viewer,"[%d] %d: %" PetscInt_FMT " edges\n",rank,sf->ranks[i],sf->roffset[i+1]-sf->roffset[i]));
88195fce210SBarry Smith           for (j=sf->roffset[i]; j<sf->roffset[i+1]; j++) {
8829566063dSJacob Faibussowitsch             PetscCall(PetscViewerASCIISynchronizedPrintf(viewer,"[%d]    %" PetscInt_FMT " <- %" PetscInt_FMT "\n",rank,sf->rmine[j],sf->rremote[j]));
88395fce210SBarry Smith           }
88495fce210SBarry Smith         }
8859566063dSJacob Faibussowitsch         PetscCall(PetscFree2(tmpranks,perm));
88695fce210SBarry Smith       }
8879566063dSJacob Faibussowitsch       PetscCall(PetscViewerFlush(viewer));
8889566063dSJacob Faibussowitsch       PetscCall(PetscViewerASCIIPopSynchronized(viewer));
889dd5b3ca6SJunchao Zhang     }
8909566063dSJacob Faibussowitsch     PetscCall(PetscViewerASCIIPopTab(viewer));
89195fce210SBarry Smith   }
8929566063dSJacob Faibussowitsch   if (sf->ops->View) PetscCall((*sf->ops->View)(sf,viewer));
89395fce210SBarry Smith   PetscFunctionReturn(0);
89495fce210SBarry Smith }
89595fce210SBarry Smith 
89695fce210SBarry Smith /*@C
897dec1416fSJunchao Zhang    PetscSFGetRootRanks - Get root ranks and number of vertices referenced by leaves on this process
89895fce210SBarry Smith 
89995fce210SBarry Smith    Not Collective
90095fce210SBarry Smith 
9014165533cSJose E. Roman    Input Parameter:
90295fce210SBarry Smith .  sf - star forest
90395fce210SBarry Smith 
9044165533cSJose E. Roman    Output Parameters:
90595fce210SBarry Smith +  nranks - number of ranks referenced by local part
90695fce210SBarry Smith .  ranks - array of ranks
90795fce210SBarry Smith .  roffset - offset in rmine/rremote for each rank (length nranks+1)
90895fce210SBarry Smith .  rmine - concatenated array holding local indices referencing each remote rank
90995fce210SBarry Smith -  rremote - concatenated array holding remote indices referenced for each remote rank
91095fce210SBarry Smith 
91195fce210SBarry Smith    Level: developer
91295fce210SBarry Smith 
913dec1416fSJunchao Zhang .seealso: PetscSFGetLeafRanks()
91495fce210SBarry Smith @*/
915dec1416fSJunchao Zhang PetscErrorCode PetscSFGetRootRanks(PetscSF sf,PetscInt *nranks,const PetscMPIInt **ranks,const PetscInt **roffset,const PetscInt **rmine,const PetscInt **rremote)
91695fce210SBarry Smith {
91795fce210SBarry Smith   PetscFunctionBegin;
91895fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
91928b400f6SJacob Faibussowitsch   PetscCheck(sf->setupcalled,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must call PetscSFSetUp() before obtaining ranks");
920dec1416fSJunchao Zhang   if (sf->ops->GetRootRanks) {
9219566063dSJacob Faibussowitsch     PetscCall((sf->ops->GetRootRanks)(sf,nranks,ranks,roffset,rmine,rremote));
922dec1416fSJunchao Zhang   } else {
923dec1416fSJunchao Zhang     /* The generic implementation */
92495fce210SBarry Smith     if (nranks)  *nranks  = sf->nranks;
92595fce210SBarry Smith     if (ranks)   *ranks   = sf->ranks;
92695fce210SBarry Smith     if (roffset) *roffset = sf->roffset;
92795fce210SBarry Smith     if (rmine)   *rmine   = sf->rmine;
92895fce210SBarry Smith     if (rremote) *rremote = sf->rremote;
929dec1416fSJunchao Zhang   }
93095fce210SBarry Smith   PetscFunctionReturn(0);
93195fce210SBarry Smith }
93295fce210SBarry Smith 
9338750ddebSJunchao Zhang /*@C
9348750ddebSJunchao Zhang    PetscSFGetLeafRanks - Get leaf ranks referencing roots on this process
9358750ddebSJunchao Zhang 
9368750ddebSJunchao Zhang    Not Collective
9378750ddebSJunchao Zhang 
9384165533cSJose E. Roman    Input Parameter:
9398750ddebSJunchao Zhang .  sf - star forest
9408750ddebSJunchao Zhang 
9414165533cSJose E. Roman    Output Parameters:
9428750ddebSJunchao Zhang +  niranks - number of leaf ranks referencing roots on this process
9438750ddebSJunchao Zhang .  iranks - array of ranks
9448750ddebSJunchao Zhang .  ioffset - offset in irootloc for each rank (length niranks+1)
9458750ddebSJunchao Zhang -  irootloc - concatenated array holding local indices of roots referenced by each leaf rank
9468750ddebSJunchao Zhang 
9478750ddebSJunchao Zhang    Level: developer
9488750ddebSJunchao Zhang 
949dec1416fSJunchao Zhang .seealso: PetscSFGetRootRanks()
9508750ddebSJunchao Zhang @*/
9518750ddebSJunchao Zhang PetscErrorCode PetscSFGetLeafRanks(PetscSF sf,PetscInt *niranks,const PetscMPIInt **iranks,const PetscInt **ioffset,const PetscInt **irootloc)
9528750ddebSJunchao Zhang {
9538750ddebSJunchao Zhang   PetscFunctionBegin;
9548750ddebSJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
95528b400f6SJacob Faibussowitsch   PetscCheck(sf->setupcalled,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must call PetscSFSetUp() before obtaining ranks");
9568750ddebSJunchao Zhang   if (sf->ops->GetLeafRanks) {
9579566063dSJacob Faibussowitsch     PetscCall((sf->ops->GetLeafRanks)(sf,niranks,iranks,ioffset,irootloc));
9588750ddebSJunchao Zhang   } else {
9598750ddebSJunchao Zhang     PetscSFType type;
9609566063dSJacob Faibussowitsch     PetscCall(PetscSFGetType(sf,&type));
96198921bdaSJacob Faibussowitsch     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"PetscSFGetLeafRanks() is not supported on this StarForest type: %s", type);
9628750ddebSJunchao Zhang   }
9638750ddebSJunchao Zhang   PetscFunctionReturn(0);
9648750ddebSJunchao Zhang }
9658750ddebSJunchao Zhang 
966b5a8e515SJed Brown static PetscBool InList(PetscMPIInt needle,PetscMPIInt n,const PetscMPIInt *list) {
967b5a8e515SJed Brown   PetscInt i;
968b5a8e515SJed Brown   for (i=0; i<n; i++) {
969b5a8e515SJed Brown     if (needle == list[i]) return PETSC_TRUE;
970b5a8e515SJed Brown   }
971b5a8e515SJed Brown   return PETSC_FALSE;
972b5a8e515SJed Brown }
973b5a8e515SJed Brown 
97495fce210SBarry Smith /*@C
97521c688dcSJed Brown    PetscSFSetUpRanks - Set up data structures associated with ranks; this is for internal use by PetscSF implementations.
97621c688dcSJed Brown 
97721c688dcSJed Brown    Collective
97821c688dcSJed Brown 
9794165533cSJose E. Roman    Input Parameters:
980b5a8e515SJed Brown +  sf - PetscSF to set up; PetscSFSetGraph() must have been called
981b5a8e515SJed Brown -  dgroup - MPI_Group of ranks to be distinguished (e.g., for self or shared memory exchange)
98221c688dcSJed Brown 
98321c688dcSJed Brown    Level: developer
98421c688dcSJed Brown 
985dec1416fSJunchao Zhang .seealso: PetscSFGetRootRanks()
98621c688dcSJed Brown @*/
987b5a8e515SJed Brown PetscErrorCode PetscSFSetUpRanks(PetscSF sf,MPI_Group dgroup)
98821c688dcSJed Brown {
98921c688dcSJed Brown   PetscTable         table;
99021c688dcSJed Brown   PetscTablePosition pos;
991b5a8e515SJed Brown   PetscMPIInt        size,groupsize,*groupranks;
992247e8311SStefano Zampini   PetscInt           *rcount,*ranks;
993247e8311SStefano Zampini   PetscInt           i, irank = -1,orank = -1;
99421c688dcSJed Brown 
99521c688dcSJed Brown   PetscFunctionBegin;
99621c688dcSJed Brown   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
99729046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
9989566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)sf),&size));
9999566063dSJacob Faibussowitsch   PetscCall(PetscTableCreate(10,size,&table));
100021c688dcSJed Brown   for (i=0; i<sf->nleaves; i++) {
100121c688dcSJed Brown     /* Log 1-based rank */
10029566063dSJacob Faibussowitsch     PetscCall(PetscTableAdd(table,sf->remote[i].rank+1,1,ADD_VALUES));
100321c688dcSJed Brown   }
10049566063dSJacob Faibussowitsch   PetscCall(PetscTableGetCount(table,&sf->nranks));
10059566063dSJacob Faibussowitsch   PetscCall(PetscMalloc4(sf->nranks,&sf->ranks,sf->nranks+1,&sf->roffset,sf->nleaves,&sf->rmine,sf->nleaves,&sf->rremote));
10069566063dSJacob Faibussowitsch   PetscCall(PetscMalloc2(sf->nranks,&rcount,sf->nranks,&ranks));
10079566063dSJacob Faibussowitsch   PetscCall(PetscTableGetHeadPosition(table,&pos));
100821c688dcSJed Brown   for (i=0; i<sf->nranks; i++) {
10099566063dSJacob Faibussowitsch     PetscCall(PetscTableGetNext(table,&pos,&ranks[i],&rcount[i]));
101021c688dcSJed Brown     ranks[i]--;             /* Convert back to 0-based */
101121c688dcSJed Brown   }
10129566063dSJacob Faibussowitsch   PetscCall(PetscTableDestroy(&table));
1013b5a8e515SJed Brown 
1014b5a8e515SJed Brown   /* We expect that dgroup is reliably "small" while nranks could be large */
1015b5a8e515SJed Brown   {
10167fb8a5e4SKarl Rupp     MPI_Group group = MPI_GROUP_NULL;
1017b5a8e515SJed Brown     PetscMPIInt *dgroupranks;
10189566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf),&group));
10199566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_size(dgroup,&groupsize));
10209566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(groupsize,&dgroupranks));
10219566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(groupsize,&groupranks));
1022b5a8e515SJed Brown     for (i=0; i<groupsize; i++) dgroupranks[i] = i;
10239566063dSJacob Faibussowitsch     if (groupsize) PetscCallMPI(MPI_Group_translate_ranks(dgroup,groupsize,dgroupranks,group,groupranks));
10249566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_free(&group));
10259566063dSJacob Faibussowitsch     PetscCall(PetscFree(dgroupranks));
1026b5a8e515SJed Brown   }
1027b5a8e515SJed Brown 
1028b5a8e515SJed Brown   /* Partition ranks[] into distinguished (first sf->ndranks) followed by non-distinguished */
1029b5a8e515SJed Brown   for (sf->ndranks=0,i=sf->nranks; sf->ndranks<i;) {
1030b5a8e515SJed Brown     for (i--; sf->ndranks<i; i--) { /* Scan i backward looking for distinguished rank */
1031b5a8e515SJed Brown       if (InList(ranks[i],groupsize,groupranks)) break;
1032b5a8e515SJed Brown     }
1033b5a8e515SJed Brown     for (; sf->ndranks<=i; sf->ndranks++) { /* Scan sf->ndranks forward looking for non-distinguished rank */
1034b5a8e515SJed Brown       if (!InList(ranks[sf->ndranks],groupsize,groupranks)) break;
1035b5a8e515SJed Brown     }
1036b5a8e515SJed Brown     if (sf->ndranks < i) {                         /* Swap ranks[sf->ndranks] with ranks[i] */
1037b5a8e515SJed Brown       PetscInt tmprank,tmpcount;
1038247e8311SStefano Zampini 
1039b5a8e515SJed Brown       tmprank             = ranks[i];
1040b5a8e515SJed Brown       tmpcount            = rcount[i];
1041b5a8e515SJed Brown       ranks[i]            = ranks[sf->ndranks];
1042b5a8e515SJed Brown       rcount[i]           = rcount[sf->ndranks];
1043b5a8e515SJed Brown       ranks[sf->ndranks]  = tmprank;
1044b5a8e515SJed Brown       rcount[sf->ndranks] = tmpcount;
1045b5a8e515SJed Brown       sf->ndranks++;
1046b5a8e515SJed Brown     }
1047b5a8e515SJed Brown   }
10489566063dSJacob Faibussowitsch   PetscCall(PetscFree(groupranks));
10499566063dSJacob Faibussowitsch   PetscCall(PetscSortIntWithArray(sf->ndranks,ranks,rcount));
10509566063dSJacob Faibussowitsch   PetscCall(PetscSortIntWithArray(sf->nranks-sf->ndranks,ranks+sf->ndranks,rcount+sf->ndranks));
105121c688dcSJed Brown   sf->roffset[0] = 0;
105221c688dcSJed Brown   for (i=0; i<sf->nranks; i++) {
10539566063dSJacob Faibussowitsch     PetscCall(PetscMPIIntCast(ranks[i],sf->ranks+i));
105421c688dcSJed Brown     sf->roffset[i+1] = sf->roffset[i] + rcount[i];
105521c688dcSJed Brown     rcount[i]        = 0;
105621c688dcSJed Brown   }
1057247e8311SStefano Zampini   for (i=0, irank = -1, orank = -1; i<sf->nleaves; i++) {
1058247e8311SStefano Zampini     /* short circuit */
1059247e8311SStefano Zampini     if (orank != sf->remote[i].rank) {
106021c688dcSJed Brown       /* Search for index of iremote[i].rank in sf->ranks */
10619566063dSJacob Faibussowitsch       PetscCall(PetscFindMPIInt(sf->remote[i].rank,sf->ndranks,sf->ranks,&irank));
1062b5a8e515SJed Brown       if (irank < 0) {
10639566063dSJacob Faibussowitsch         PetscCall(PetscFindMPIInt(sf->remote[i].rank,sf->nranks-sf->ndranks,sf->ranks+sf->ndranks,&irank));
1064b5a8e515SJed Brown         if (irank >= 0) irank += sf->ndranks;
106521c688dcSJed Brown       }
1066247e8311SStefano Zampini       orank = sf->remote[i].rank;
1067247e8311SStefano Zampini     }
1068*08401ef6SPierre Jolivet     PetscCheck(irank >= 0,PETSC_COMM_SELF,PETSC_ERR_PLIB,"Could not find rank %" PetscInt_FMT " in array",sf->remote[i].rank);
106921c688dcSJed Brown     sf->rmine[sf->roffset[irank] + rcount[irank]]   = sf->mine ? sf->mine[i] : i;
107021c688dcSJed Brown     sf->rremote[sf->roffset[irank] + rcount[irank]] = sf->remote[i].index;
107121c688dcSJed Brown     rcount[irank]++;
107221c688dcSJed Brown   }
10739566063dSJacob Faibussowitsch   PetscCall(PetscFree2(rcount,ranks));
107421c688dcSJed Brown   PetscFunctionReturn(0);
107521c688dcSJed Brown }
107621c688dcSJed Brown 
107721c688dcSJed Brown /*@C
107895fce210SBarry Smith    PetscSFGetGroups - gets incoming and outgoing process groups
107995fce210SBarry Smith 
108095fce210SBarry Smith    Collective
108195fce210SBarry Smith 
10824165533cSJose E. Roman    Input Parameter:
108395fce210SBarry Smith .  sf - star forest
108495fce210SBarry Smith 
10854165533cSJose E. Roman    Output Parameters:
108695fce210SBarry Smith +  incoming - group of origin processes for incoming edges (leaves that reference my roots)
108795fce210SBarry Smith -  outgoing - group of destination processes for outgoing edges (roots that I reference)
108895fce210SBarry Smith 
108995fce210SBarry Smith    Level: developer
109095fce210SBarry Smith 
109195fce210SBarry Smith .seealso: PetscSFGetWindow(), PetscSFRestoreWindow()
109295fce210SBarry Smith @*/
109395fce210SBarry Smith PetscErrorCode PetscSFGetGroups(PetscSF sf,MPI_Group *incoming,MPI_Group *outgoing)
109495fce210SBarry Smith {
10957fb8a5e4SKarl Rupp   MPI_Group      group = MPI_GROUP_NULL;
109695fce210SBarry Smith 
109795fce210SBarry Smith   PetscFunctionBegin;
1098*08401ef6SPierre Jolivet   PetscCheck(sf->nranks >= 0,PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must call PetscSFSetUpRanks() before obtaining groups");
109995fce210SBarry Smith   if (sf->ingroup == MPI_GROUP_NULL) {
110095fce210SBarry Smith     PetscInt       i;
110195fce210SBarry Smith     const PetscInt *indegree;
110295fce210SBarry Smith     PetscMPIInt    rank,*outranks,*inranks;
110395fce210SBarry Smith     PetscSFNode    *remote;
110495fce210SBarry Smith     PetscSF        bgcount;
110595fce210SBarry Smith 
110695fce210SBarry Smith     /* Compute the number of incoming ranks */
11079566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(sf->nranks,&remote));
110895fce210SBarry Smith     for (i=0; i<sf->nranks; i++) {
110995fce210SBarry Smith       remote[i].rank  = sf->ranks[i];
111095fce210SBarry Smith       remote[i].index = 0;
111195fce210SBarry Smith     }
11129566063dSJacob Faibussowitsch     PetscCall(PetscSFDuplicate(sf,PETSCSF_DUPLICATE_CONFONLY,&bgcount));
11139566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(bgcount,1,sf->nranks,NULL,PETSC_COPY_VALUES,remote,PETSC_OWN_POINTER));
11149566063dSJacob Faibussowitsch     PetscCall(PetscSFComputeDegreeBegin(bgcount,&indegree));
11159566063dSJacob Faibussowitsch     PetscCall(PetscSFComputeDegreeEnd(bgcount,&indegree));
111695fce210SBarry Smith     /* Enumerate the incoming ranks */
11179566063dSJacob Faibussowitsch     PetscCall(PetscMalloc2(indegree[0],&inranks,sf->nranks,&outranks));
11189566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank));
111995fce210SBarry Smith     for (i=0; i<sf->nranks; i++) outranks[i] = rank;
11209566063dSJacob Faibussowitsch     PetscCall(PetscSFGatherBegin(bgcount,MPI_INT,outranks,inranks));
11219566063dSJacob Faibussowitsch     PetscCall(PetscSFGatherEnd(bgcount,MPI_INT,outranks,inranks));
11229566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf),&group));
11239566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_incl(group,indegree[0],inranks,&sf->ingroup));
11249566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_free(&group));
11259566063dSJacob Faibussowitsch     PetscCall(PetscFree2(inranks,outranks));
11269566063dSJacob Faibussowitsch     PetscCall(PetscSFDestroy(&bgcount));
112795fce210SBarry Smith   }
112895fce210SBarry Smith   *incoming = sf->ingroup;
112995fce210SBarry Smith 
113095fce210SBarry Smith   if (sf->outgroup == MPI_GROUP_NULL) {
11319566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf),&group));
11329566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_incl(group,sf->nranks,sf->ranks,&sf->outgroup));
11339566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Group_free(&group));
113495fce210SBarry Smith   }
113595fce210SBarry Smith   *outgoing = sf->outgroup;
113695fce210SBarry Smith   PetscFunctionReturn(0);
113795fce210SBarry Smith }
113895fce210SBarry Smith 
113929046d53SLisandro Dalcin /*@
11403b8d980fSPierre Jolivet    PetscSFGetMultiSF - gets the inner SF implementing gathers and scatters
114195fce210SBarry Smith 
114295fce210SBarry Smith    Collective
114395fce210SBarry Smith 
11444165533cSJose E. Roman    Input Parameter:
114595fce210SBarry Smith .  sf - star forest that may contain roots with 0 or with more than 1 vertex
114695fce210SBarry Smith 
11474165533cSJose E. Roman    Output Parameter:
114895fce210SBarry Smith .  multi - star forest with split roots, such that each root has degree exactly 1
114995fce210SBarry Smith 
115095fce210SBarry Smith    Level: developer
115195fce210SBarry Smith 
115295fce210SBarry Smith    Notes:
115395fce210SBarry Smith 
115495fce210SBarry Smith    In most cases, users should use PetscSFGatherBegin() and PetscSFScatterBegin() instead of manipulating multi
115595fce210SBarry Smith    directly. Since multi satisfies the stronger condition that each entry in the global space has exactly one incoming
115695fce210SBarry Smith    edge, it is a candidate for future optimization that might involve its removal.
115795fce210SBarry Smith 
1158673100f5SVaclav Hapla .seealso: PetscSFSetGraph(), PetscSFGatherBegin(), PetscSFScatterBegin(), PetscSFComputeMultiRootOriginalNumbering()
115995fce210SBarry Smith @*/
116095fce210SBarry Smith PetscErrorCode PetscSFGetMultiSF(PetscSF sf,PetscSF *multi)
116195fce210SBarry Smith {
116295fce210SBarry Smith   PetscFunctionBegin;
116395fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
116495fce210SBarry Smith   PetscValidPointer(multi,2);
116595fce210SBarry Smith   if (sf->nroots < 0) {         /* Graph has not been set yet; why do we need this? */
11669566063dSJacob Faibussowitsch     PetscCall(PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&sf->multi));
116795fce210SBarry Smith     *multi = sf->multi;
1168013b3241SStefano Zampini     sf->multi->multi = sf->multi;
116995fce210SBarry Smith     PetscFunctionReturn(0);
117095fce210SBarry Smith   }
117195fce210SBarry Smith   if (!sf->multi) {
117295fce210SBarry Smith     const PetscInt *indegree;
11739837ea96SMatthew G. Knepley     PetscInt       i,*inoffset,*outones,*outoffset,maxlocal;
117495fce210SBarry Smith     PetscSFNode    *remote;
117529046d53SLisandro Dalcin     maxlocal = sf->maxleaf+1; /* TODO: We should use PetscSFGetLeafRange() */
11769566063dSJacob Faibussowitsch     PetscCall(PetscSFComputeDegreeBegin(sf,&indegree));
11779566063dSJacob Faibussowitsch     PetscCall(PetscSFComputeDegreeEnd(sf,&indegree));
11789566063dSJacob Faibussowitsch     PetscCall(PetscMalloc3(sf->nroots+1,&inoffset,maxlocal,&outones,maxlocal,&outoffset));
117995fce210SBarry Smith     inoffset[0] = 0;
118095fce210SBarry Smith     for (i=0; i<sf->nroots; i++) inoffset[i+1] = inoffset[i] + indegree[i];
11819837ea96SMatthew G. Knepley     for (i=0; i<maxlocal; i++) outones[i] = 1;
11829566063dSJacob Faibussowitsch     PetscCall(PetscSFFetchAndOpBegin(sf,MPIU_INT,inoffset,outones,outoffset,MPI_SUM));
11839566063dSJacob Faibussowitsch     PetscCall(PetscSFFetchAndOpEnd(sf,MPIU_INT,inoffset,outones,outoffset,MPI_SUM));
118495fce210SBarry Smith     for (i=0; i<sf->nroots; i++) inoffset[i] -= indegree[i]; /* Undo the increment */
118576bd3646SJed Brown     if (PetscDefined(USE_DEBUG)) { /* Check that the expected number of increments occurred */
118695fce210SBarry Smith       for (i=0; i<sf->nroots; i++) {
11872c71b3e2SJacob Faibussowitsch         PetscCheckFalse(inoffset[i] + indegree[i] != inoffset[i+1],PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect result after PetscSFFetchAndOp");
118895fce210SBarry Smith       }
118976bd3646SJed Brown     }
11909566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(sf->nleaves,&remote));
119195fce210SBarry Smith     for (i=0; i<sf->nleaves; i++) {
119295fce210SBarry Smith       remote[i].rank  = sf->remote[i].rank;
119338e7336fSToby Isaac       remote[i].index = outoffset[sf->mine ? sf->mine[i] : i];
119495fce210SBarry Smith     }
11959566063dSJacob Faibussowitsch     PetscCall(PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&sf->multi));
1196013b3241SStefano Zampini     sf->multi->multi = sf->multi;
11979566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(sf->multi,inoffset[sf->nroots],sf->nleaves,sf->mine,PETSC_COPY_VALUES,remote,PETSC_OWN_POINTER));
119895fce210SBarry Smith     if (sf->rankorder) {        /* Sort the ranks */
119995fce210SBarry Smith       PetscMPIInt rank;
120095fce210SBarry Smith       PetscInt    *inranks,*newoffset,*outranks,*newoutoffset,*tmpoffset,maxdegree;
120195fce210SBarry Smith       PetscSFNode *newremote;
12029566063dSJacob Faibussowitsch       PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank));
120395fce210SBarry Smith       for (i=0,maxdegree=0; i<sf->nroots; i++) maxdegree = PetscMax(maxdegree,indegree[i]);
12049566063dSJacob Faibussowitsch       PetscCall(PetscMalloc5(sf->multi->nroots,&inranks,sf->multi->nroots,&newoffset,maxlocal,&outranks,maxlocal,&newoutoffset,maxdegree,&tmpoffset));
12059837ea96SMatthew G. Knepley       for (i=0; i<maxlocal; i++) outranks[i] = rank;
12069566063dSJacob Faibussowitsch       PetscCall(PetscSFReduceBegin(sf->multi,MPIU_INT,outranks,inranks,MPI_REPLACE));
12079566063dSJacob Faibussowitsch       PetscCall(PetscSFReduceEnd(sf->multi,MPIU_INT,outranks,inranks,MPI_REPLACE));
120895fce210SBarry Smith       /* Sort the incoming ranks at each vertex, build the inverse map */
120995fce210SBarry Smith       for (i=0; i<sf->nroots; i++) {
121095fce210SBarry Smith         PetscInt j;
121195fce210SBarry Smith         for (j=0; j<indegree[i]; j++) tmpoffset[j] = j;
12129566063dSJacob Faibussowitsch         PetscCall(PetscSortIntWithArray(indegree[i],inranks+inoffset[i],tmpoffset));
121395fce210SBarry Smith         for (j=0; j<indegree[i]; j++) newoffset[inoffset[i] + tmpoffset[j]] = inoffset[i] + j;
121495fce210SBarry Smith       }
12159566063dSJacob Faibussowitsch       PetscCall(PetscSFBcastBegin(sf->multi,MPIU_INT,newoffset,newoutoffset,MPI_REPLACE));
12169566063dSJacob Faibussowitsch       PetscCall(PetscSFBcastEnd(sf->multi,MPIU_INT,newoffset,newoutoffset,MPI_REPLACE));
12179566063dSJacob Faibussowitsch       PetscCall(PetscMalloc1(sf->nleaves,&newremote));
121895fce210SBarry Smith       for (i=0; i<sf->nleaves; i++) {
121995fce210SBarry Smith         newremote[i].rank  = sf->remote[i].rank;
122001365b40SToby Isaac         newremote[i].index = newoutoffset[sf->mine ? sf->mine[i] : i];
122195fce210SBarry Smith       }
12229566063dSJacob Faibussowitsch       PetscCall(PetscSFSetGraph(sf->multi,inoffset[sf->nroots],sf->nleaves,sf->mine,PETSC_COPY_VALUES,newremote,PETSC_OWN_POINTER));
12239566063dSJacob Faibussowitsch       PetscCall(PetscFree5(inranks,newoffset,outranks,newoutoffset,tmpoffset));
122495fce210SBarry Smith     }
12259566063dSJacob Faibussowitsch     PetscCall(PetscFree3(inoffset,outones,outoffset));
122695fce210SBarry Smith   }
122795fce210SBarry Smith   *multi = sf->multi;
122895fce210SBarry Smith   PetscFunctionReturn(0);
122995fce210SBarry Smith }
123095fce210SBarry Smith 
123195fce210SBarry Smith /*@C
123272502a1fSJunchao Zhang    PetscSFCreateEmbeddedRootSF - removes edges from all but the selected roots, does not remap indices
123395fce210SBarry Smith 
123495fce210SBarry Smith    Collective
123595fce210SBarry Smith 
12364165533cSJose E. Roman    Input Parameters:
123795fce210SBarry Smith +  sf - original star forest
1238ba2a7774SJunchao Zhang .  nselected  - number of selected roots on this process
1239ba2a7774SJunchao Zhang -  selected   - indices of the selected roots on this process
124095fce210SBarry Smith 
12414165533cSJose E. Roman    Output Parameter:
1242cd620004SJunchao Zhang .  esf - new star forest
124395fce210SBarry Smith 
124495fce210SBarry Smith    Level: advanced
124595fce210SBarry Smith 
124695fce210SBarry Smith    Note:
124795fce210SBarry Smith    To use the new PetscSF, it may be necessary to know the indices of the leaves that are still participating. This can
124895fce210SBarry Smith    be done by calling PetscSFGetGraph().
124995fce210SBarry Smith 
125095fce210SBarry Smith .seealso: PetscSFSetGraph(), PetscSFGetGraph()
125195fce210SBarry Smith @*/
125272502a1fSJunchao Zhang PetscErrorCode PetscSFCreateEmbeddedRootSF(PetscSF sf,PetscInt nselected,const PetscInt *selected,PetscSF *esf)
125395fce210SBarry Smith {
1254cd620004SJunchao Zhang   PetscInt          i,j,n,nroots,nleaves,esf_nleaves,*new_ilocal,minleaf,maxleaf,maxlocal;
1255cd620004SJunchao Zhang   const PetscInt    *ilocal;
1256cd620004SJunchao Zhang   signed char       *rootdata,*leafdata,*leafmem;
1257ba2a7774SJunchao Zhang   const PetscSFNode *iremote;
1258f659e5c7SJunchao Zhang   PetscSFNode       *new_iremote;
1259f659e5c7SJunchao Zhang   MPI_Comm          comm;
126095fce210SBarry Smith 
126195fce210SBarry Smith   PetscFunctionBegin;
126295fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
126329046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
1264dadcf809SJacob Faibussowitsch   if (nselected) PetscValidIntPointer(selected,3);
1265cd620004SJunchao Zhang   PetscValidPointer(esf,4);
12660511a646SMatthew G. Knepley 
12679566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
12689566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_EmbedSF,sf,0,0,0));
12699566063dSJacob Faibussowitsch   PetscCall(PetscObjectGetComm((PetscObject)sf,&comm));
12709566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,&iremote));
1271cd620004SJunchao Zhang 
127276bd3646SJed Brown   if (PetscDefined(USE_DEBUG)) {  /* Error out if selected[] has dups or  out of range indices */
1273cd620004SJunchao Zhang     PetscBool dups;
12749566063dSJacob Faibussowitsch     PetscCall(PetscCheckDupsInt(nselected,selected,&dups));
127528b400f6SJacob Faibussowitsch     PetscCheck(!dups,comm,PETSC_ERR_ARG_WRONG,"selected[] has dups");
1276cd620004SJunchao Zhang     for (i=0; i<nselected; i++)
12772c71b3e2SJacob Faibussowitsch       PetscCheckFalse(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);
1278cd620004SJunchao Zhang   }
1279f659e5c7SJunchao Zhang 
128072502a1fSJunchao Zhang   if (sf->ops->CreateEmbeddedRootSF) {
12819566063dSJacob Faibussowitsch     PetscCall((*sf->ops->CreateEmbeddedRootSF)(sf,nselected,selected,esf));
1282f659e5c7SJunchao Zhang   } else {
1283cd620004SJunchao Zhang     /* A generic version of creating embedded sf */
12849566063dSJacob Faibussowitsch     PetscCall(PetscSFGetLeafRange(sf,&minleaf,&maxleaf));
1285cd620004SJunchao Zhang     maxlocal = maxleaf - minleaf + 1;
12869566063dSJacob Faibussowitsch     PetscCall(PetscCalloc2(nroots,&rootdata,maxlocal,&leafmem));
1287cd620004SJunchao Zhang     leafdata = leafmem - minleaf;
1288cd620004SJunchao Zhang     /* Tag selected roots and bcast to leaves */
1289cd620004SJunchao Zhang     for (i=0; i<nselected; i++) rootdata[selected[i]] = 1;
12909566063dSJacob Faibussowitsch     PetscCall(PetscSFBcastBegin(sf,MPI_SIGNED_CHAR,rootdata,leafdata,MPI_REPLACE));
12919566063dSJacob Faibussowitsch     PetscCall(PetscSFBcastEnd(sf,MPI_SIGNED_CHAR,rootdata,leafdata,MPI_REPLACE));
1292ba2a7774SJunchao Zhang 
1293cd620004SJunchao Zhang     /* Build esf with leaves that are still connected */
1294cd620004SJunchao Zhang     esf_nleaves = 0;
1295cd620004SJunchao Zhang     for (i=0; i<nleaves; i++) {
1296cd620004SJunchao Zhang       j = ilocal ? ilocal[i] : i;
1297cd620004SJunchao Zhang       /* esf_nleaves += leafdata[j] should work in theory, but failed with SFWindow bugs
1298cd620004SJunchao Zhang          with PetscSFBcast. See https://gitlab.com/petsc/petsc/issues/555
1299cd620004SJunchao Zhang       */
1300cd620004SJunchao Zhang       esf_nleaves += (leafdata[j] ? 1 : 0);
1301cd620004SJunchao Zhang     }
13029566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(esf_nleaves,&new_ilocal));
13039566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(esf_nleaves,&new_iremote));
1304cd620004SJunchao Zhang     for (i=n=0; i<nleaves; i++) {
1305cd620004SJunchao Zhang       j = ilocal ? ilocal[i] : i;
1306cd620004SJunchao Zhang       if (leafdata[j]) {
1307cd620004SJunchao Zhang         new_ilocal[n]        = j;
1308cd620004SJunchao Zhang         new_iremote[n].rank  = iremote[i].rank;
1309cd620004SJunchao Zhang         new_iremote[n].index = iremote[i].index;
1310fc1ede2bSMatthew G. Knepley         ++n;
131195fce210SBarry Smith       }
131295fce210SBarry Smith     }
13139566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(comm,esf));
13149566063dSJacob Faibussowitsch     PetscCall(PetscSFSetFromOptions(*esf));
13159566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(*esf,nroots,esf_nleaves,new_ilocal,PETSC_OWN_POINTER,new_iremote,PETSC_OWN_POINTER));
13169566063dSJacob Faibussowitsch     PetscCall(PetscFree2(rootdata,leafmem));
1317f659e5c7SJunchao Zhang   }
13189566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_EmbedSF,sf,0,0,0));
131995fce210SBarry Smith   PetscFunctionReturn(0);
132095fce210SBarry Smith }
132195fce210SBarry Smith 
13222f5fb4c2SMatthew G. Knepley /*@C
13232f5fb4c2SMatthew G. Knepley   PetscSFCreateEmbeddedLeafSF - removes edges from all but the selected leaves, does not remap indices
13242f5fb4c2SMatthew G. Knepley 
13252f5fb4c2SMatthew G. Knepley   Collective
13262f5fb4c2SMatthew G. Knepley 
13274165533cSJose E. Roman   Input Parameters:
13282f5fb4c2SMatthew G. Knepley + sf - original star forest
1329f659e5c7SJunchao Zhang . nselected  - number of selected leaves on this process
1330f659e5c7SJunchao Zhang - selected   - indices of the selected leaves on this process
13312f5fb4c2SMatthew G. Knepley 
13324165533cSJose E. Roman   Output Parameter:
13332f5fb4c2SMatthew G. Knepley .  newsf - new star forest
13342f5fb4c2SMatthew G. Knepley 
13352f5fb4c2SMatthew G. Knepley   Level: advanced
13362f5fb4c2SMatthew G. Knepley 
133772502a1fSJunchao Zhang .seealso: PetscSFCreateEmbeddedRootSF(), PetscSFSetGraph(), PetscSFGetGraph()
13382f5fb4c2SMatthew G. Knepley @*/
1339f659e5c7SJunchao Zhang PetscErrorCode PetscSFCreateEmbeddedLeafSF(PetscSF sf,PetscInt nselected,const PetscInt *selected,PetscSF *newsf)
13402f5fb4c2SMatthew G. Knepley {
1341f659e5c7SJunchao Zhang   const PetscSFNode *iremote;
1342f659e5c7SJunchao Zhang   PetscSFNode       *new_iremote;
1343f659e5c7SJunchao Zhang   const PetscInt    *ilocal;
1344f659e5c7SJunchao Zhang   PetscInt          i,nroots,*leaves,*new_ilocal;
1345f659e5c7SJunchao Zhang   MPI_Comm          comm;
13462f5fb4c2SMatthew G. Knepley 
13472f5fb4c2SMatthew G. Knepley   PetscFunctionBegin;
13482f5fb4c2SMatthew G. Knepley   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
134929046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
1350dadcf809SJacob Faibussowitsch   if (nselected) PetscValidIntPointer(selected,3);
13512f5fb4c2SMatthew G. Knepley   PetscValidPointer(newsf,4);
13522f5fb4c2SMatthew G. Knepley 
1353f659e5c7SJunchao Zhang   /* Uniq selected[] and put results in leaves[] */
13549566063dSJacob Faibussowitsch   PetscCall(PetscObjectGetComm((PetscObject)sf,&comm));
13559566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(nselected,&leaves));
13569566063dSJacob Faibussowitsch   PetscCall(PetscArraycpy(leaves,selected,nselected));
13579566063dSJacob Faibussowitsch   PetscCall(PetscSortedRemoveDupsInt(&nselected,leaves));
1358*08401ef6SPierre 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);
1359f659e5c7SJunchao Zhang 
1360f659e5c7SJunchao Zhang   /* Optimize the routine only when sf is setup and hence we can reuse sf's communication pattern */
1361f659e5c7SJunchao Zhang   if (sf->setupcalled && sf->ops->CreateEmbeddedLeafSF) {
13629566063dSJacob Faibussowitsch     PetscCall((*sf->ops->CreateEmbeddedLeafSF)(sf,nselected,leaves,newsf));
1363f659e5c7SJunchao Zhang   } else {
13649566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sf,&nroots,NULL,&ilocal,&iremote));
13659566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nselected,&new_ilocal));
13669566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nselected,&new_iremote));
1367f659e5c7SJunchao Zhang     for (i=0; i<nselected; ++i) {
1368f659e5c7SJunchao Zhang       const PetscInt l     = leaves[i];
1369f659e5c7SJunchao Zhang       new_ilocal[i]        = ilocal ? ilocal[l] : l;
1370f659e5c7SJunchao Zhang       new_iremote[i].rank  = iremote[l].rank;
1371f659e5c7SJunchao Zhang       new_iremote[i].index = iremote[l].index;
13722f5fb4c2SMatthew G. Knepley     }
13739566063dSJacob Faibussowitsch     PetscCall(PetscSFDuplicate(sf,PETSCSF_DUPLICATE_CONFONLY,newsf));
13749566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(*newsf,nroots,nselected,new_ilocal,PETSC_OWN_POINTER,new_iremote,PETSC_OWN_POINTER));
1375f659e5c7SJunchao Zhang   }
13769566063dSJacob Faibussowitsch   PetscCall(PetscFree(leaves));
13772f5fb4c2SMatthew G. Knepley   PetscFunctionReturn(0);
13782f5fb4c2SMatthew G. Knepley }
13792f5fb4c2SMatthew G. Knepley 
138095fce210SBarry Smith /*@C
1381ad227feaSJunchao Zhang    PetscSFBcastBegin - begin pointwise broadcast with root value being reduced to leaf value, to be concluded with call to PetscSFBcastEnd()
13823482bfa8SJunchao Zhang 
13833482bfa8SJunchao Zhang    Collective on PetscSF
13843482bfa8SJunchao Zhang 
13854165533cSJose E. Roman    Input Parameters:
13863482bfa8SJunchao Zhang +  sf - star forest on which to communicate
13873482bfa8SJunchao Zhang .  unit - data type associated with each node
13883482bfa8SJunchao Zhang .  rootdata - buffer to broadcast
13893482bfa8SJunchao Zhang -  op - operation to use for reduction
13903482bfa8SJunchao Zhang 
13914165533cSJose E. Roman    Output Parameter:
13923482bfa8SJunchao Zhang .  leafdata - buffer to be reduced with values from each leaf's respective root
13933482bfa8SJunchao Zhang 
13943482bfa8SJunchao Zhang    Level: intermediate
13953482bfa8SJunchao Zhang 
1396d0295fc0SJunchao Zhang    Notes:
1397d0295fc0SJunchao Zhang     When petsc is configured with device support, it will use its own mechanism to figure out whether the given data pointers
1398d0295fc0SJunchao Zhang     are host pointers or device pointers, which may incur a noticable cost. If you already knew the info, you should
1399ad227feaSJunchao Zhang     use PetscSFBcastWithMemTypeBegin() instead.
1400ad227feaSJunchao Zhang .seealso: PetscSFBcastEnd(), PetscSFBcastWithMemTypeBegin()
14013482bfa8SJunchao Zhang @*/
1402ad227feaSJunchao Zhang PetscErrorCode PetscSFBcastBegin(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata,MPI_Op op)
14033482bfa8SJunchao Zhang {
1404eb02082bSJunchao Zhang   PetscMemType   rootmtype,leafmtype;
14053482bfa8SJunchao Zhang 
14063482bfa8SJunchao Zhang   PetscFunctionBegin;
14073482bfa8SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
14089566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
14099566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin,sf,0,0,0));
14109566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(rootdata,&rootmtype));
14119566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafdata,&leafmtype));
14129566063dSJacob Faibussowitsch   PetscCall((*sf->ops->BcastBegin)(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op));
14139566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin,sf,0,0,0));
14143482bfa8SJunchao Zhang   PetscFunctionReturn(0);
14153482bfa8SJunchao Zhang }
14163482bfa8SJunchao Zhang 
14173482bfa8SJunchao Zhang /*@C
1418ad227feaSJunchao Zhang    PetscSFBcastWithMemTypeBegin - begin pointwise broadcast with root value being reduced to leaf value with explicit memory types, to be concluded with call to PetscSFBcastEnd()
1419d0295fc0SJunchao Zhang 
1420d0295fc0SJunchao Zhang    Collective on PetscSF
1421d0295fc0SJunchao Zhang 
14224165533cSJose E. Roman    Input Parameters:
1423d0295fc0SJunchao Zhang +  sf - star forest on which to communicate
1424d0295fc0SJunchao Zhang .  unit - data type associated with each node
1425d0295fc0SJunchao Zhang .  rootmtype - memory type of rootdata
1426d0295fc0SJunchao Zhang .  rootdata - buffer to broadcast
1427d0295fc0SJunchao Zhang .  leafmtype - memory type of leafdata
1428d0295fc0SJunchao Zhang -  op - operation to use for reduction
1429d0295fc0SJunchao Zhang 
14304165533cSJose E. Roman    Output Parameter:
1431d0295fc0SJunchao Zhang .  leafdata - buffer to be reduced with values from each leaf's respective root
1432d0295fc0SJunchao Zhang 
1433d0295fc0SJunchao Zhang    Level: intermediate
1434d0295fc0SJunchao Zhang 
1435ad227feaSJunchao Zhang .seealso: PetscSFBcastEnd(), PetscSFBcastBegin()
1436d0295fc0SJunchao Zhang @*/
1437ad227feaSJunchao Zhang PetscErrorCode PetscSFBcastWithMemTypeBegin(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata,MPI_Op op)
1438d0295fc0SJunchao Zhang {
1439d0295fc0SJunchao Zhang   PetscFunctionBegin;
1440d0295fc0SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
14419566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
14429566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin,sf,0,0,0));
14439566063dSJacob Faibussowitsch   PetscCall((*sf->ops->BcastBegin)(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op));
14449566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin,sf,0,0,0));
1445d0295fc0SJunchao Zhang   PetscFunctionReturn(0);
1446d0295fc0SJunchao Zhang }
1447d0295fc0SJunchao Zhang 
1448d0295fc0SJunchao Zhang /*@C
1449ad227feaSJunchao Zhang    PetscSFBcastEnd - end a broadcast & reduce operation started with PetscSFBcastBegin()
14503482bfa8SJunchao Zhang 
14513482bfa8SJunchao Zhang    Collective
14523482bfa8SJunchao Zhang 
14534165533cSJose E. Roman    Input Parameters:
14543482bfa8SJunchao Zhang +  sf - star forest
14553482bfa8SJunchao Zhang .  unit - data type
14563482bfa8SJunchao Zhang .  rootdata - buffer to broadcast
14573482bfa8SJunchao Zhang -  op - operation to use for reduction
14583482bfa8SJunchao Zhang 
14594165533cSJose E. Roman    Output Parameter:
14603482bfa8SJunchao Zhang .  leafdata - buffer to be reduced with values from each leaf's respective root
14613482bfa8SJunchao Zhang 
14623482bfa8SJunchao Zhang    Level: intermediate
14633482bfa8SJunchao Zhang 
14643482bfa8SJunchao Zhang .seealso: PetscSFSetGraph(), PetscSFReduceEnd()
14653482bfa8SJunchao Zhang @*/
1466ad227feaSJunchao Zhang PetscErrorCode PetscSFBcastEnd(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata,MPI_Op op)
14673482bfa8SJunchao Zhang {
14683482bfa8SJunchao Zhang   PetscFunctionBegin;
14693482bfa8SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
14709566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastEnd,sf,0,0,0));
14719566063dSJacob Faibussowitsch   PetscCall((*sf->ops->BcastEnd)(sf,unit,rootdata,leafdata,op));
14729566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastEnd,sf,0,0,0));
14733482bfa8SJunchao Zhang   PetscFunctionReturn(0);
14743482bfa8SJunchao Zhang }
14753482bfa8SJunchao Zhang 
14763482bfa8SJunchao Zhang /*@C
147795fce210SBarry Smith    PetscSFReduceBegin - begin reduction of leafdata into rootdata, to be completed with call to PetscSFReduceEnd()
147895fce210SBarry Smith 
147995fce210SBarry Smith    Collective
148095fce210SBarry Smith 
14814165533cSJose E. Roman    Input Parameters:
148295fce210SBarry Smith +  sf - star forest
148395fce210SBarry Smith .  unit - data type
148495fce210SBarry Smith .  leafdata - values to reduce
148595fce210SBarry Smith -  op - reduction operation
148695fce210SBarry Smith 
14874165533cSJose E. Roman    Output Parameter:
148895fce210SBarry Smith .  rootdata - result of reduction of values from all leaves of each root
148995fce210SBarry Smith 
149095fce210SBarry Smith    Level: intermediate
149195fce210SBarry Smith 
1492d0295fc0SJunchao Zhang    Notes:
1493d0295fc0SJunchao Zhang     When petsc is configured with device support, it will use its own mechanism to figure out whether the given data pointers
1494d0295fc0SJunchao Zhang     are host pointers or device pointers, which may incur a noticable cost. If you already knew the info, you should
1495d0295fc0SJunchao Zhang     use PetscSFReduceWithMemTypeBegin() instead.
1496d0295fc0SJunchao Zhang 
1497d0295fc0SJunchao Zhang .seealso: PetscSFBcastBegin(), PetscSFReduceWithMemTypeBegin()
149895fce210SBarry Smith @*/
149995fce210SBarry Smith PetscErrorCode PetscSFReduceBegin(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *rootdata,MPI_Op op)
150095fce210SBarry Smith {
1501eb02082bSJunchao Zhang   PetscMemType   rootmtype,leafmtype;
150295fce210SBarry Smith 
150395fce210SBarry Smith   PetscFunctionBegin;
150495fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
15059566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
15069566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceBegin,sf,0,0,0));
15079566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(rootdata,&rootmtype));
15089566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafdata,&leafmtype));
15099566063dSJacob Faibussowitsch   PetscCall((sf->ops->ReduceBegin)(sf,unit,leafmtype,leafdata,rootmtype,rootdata,op));
15109566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceBegin,sf,0,0,0));
151195fce210SBarry Smith   PetscFunctionReturn(0);
151295fce210SBarry Smith }
151395fce210SBarry Smith 
151495fce210SBarry Smith /*@C
1515d0295fc0SJunchao Zhang    PetscSFReduceWithMemTypeBegin - begin reduction of leafdata into rootdata with explicit memory types, to be completed with call to PetscSFReduceEnd()
1516d0295fc0SJunchao Zhang 
1517d0295fc0SJunchao Zhang    Collective
1518d0295fc0SJunchao Zhang 
15194165533cSJose E. Roman    Input Parameters:
1520d0295fc0SJunchao Zhang +  sf - star forest
1521d0295fc0SJunchao Zhang .  unit - data type
1522d0295fc0SJunchao Zhang .  leafmtype - memory type of leafdata
1523d0295fc0SJunchao Zhang .  leafdata - values to reduce
1524d0295fc0SJunchao Zhang .  rootmtype - memory type of rootdata
1525d0295fc0SJunchao Zhang -  op - reduction operation
1526d0295fc0SJunchao Zhang 
15274165533cSJose E. Roman    Output Parameter:
1528d0295fc0SJunchao Zhang .  rootdata - result of reduction of values from all leaves of each root
1529d0295fc0SJunchao Zhang 
1530d0295fc0SJunchao Zhang    Level: intermediate
1531d0295fc0SJunchao Zhang 
1532d0295fc0SJunchao Zhang .seealso: PetscSFBcastBegin(), PetscSFReduceBegin()
1533d0295fc0SJunchao Zhang @*/
1534d0295fc0SJunchao Zhang PetscErrorCode PetscSFReduceWithMemTypeBegin(PetscSF sf,MPI_Datatype unit,PetscMemType leafmtype,const void *leafdata,PetscMemType rootmtype,void *rootdata,MPI_Op op)
1535d0295fc0SJunchao Zhang {
1536d0295fc0SJunchao Zhang   PetscFunctionBegin;
1537d0295fc0SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
15389566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
15399566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceBegin,sf,0,0,0));
15409566063dSJacob Faibussowitsch   PetscCall((sf->ops->ReduceBegin)(sf,unit,leafmtype,leafdata,rootmtype,rootdata,op));
15419566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceBegin,sf,0,0,0));
1542d0295fc0SJunchao Zhang   PetscFunctionReturn(0);
1543d0295fc0SJunchao Zhang }
1544d0295fc0SJunchao Zhang 
1545d0295fc0SJunchao Zhang /*@C
154695fce210SBarry Smith    PetscSFReduceEnd - end a reduction operation started with PetscSFReduceBegin()
154795fce210SBarry Smith 
154895fce210SBarry Smith    Collective
154995fce210SBarry Smith 
15504165533cSJose E. Roman    Input Parameters:
155195fce210SBarry Smith +  sf - star forest
155295fce210SBarry Smith .  unit - data type
155395fce210SBarry Smith .  leafdata - values to reduce
155495fce210SBarry Smith -  op - reduction operation
155595fce210SBarry Smith 
15564165533cSJose E. Roman    Output Parameter:
155795fce210SBarry Smith .  rootdata - result of reduction of values from all leaves of each root
155895fce210SBarry Smith 
155995fce210SBarry Smith    Level: intermediate
156095fce210SBarry Smith 
156195fce210SBarry Smith .seealso: PetscSFSetGraph(), PetscSFBcastEnd()
156295fce210SBarry Smith @*/
156395fce210SBarry Smith PetscErrorCode PetscSFReduceEnd(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *rootdata,MPI_Op op)
156495fce210SBarry Smith {
156595fce210SBarry Smith   PetscFunctionBegin;
156695fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
15679566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceEnd,sf,0,0,0));
15689566063dSJacob Faibussowitsch   PetscCall((*sf->ops->ReduceEnd)(sf,unit,leafdata,rootdata,op));
15699566063dSJacob Faibussowitsch   if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceEnd,sf,0,0,0));
157095fce210SBarry Smith   PetscFunctionReturn(0);
157195fce210SBarry Smith }
157295fce210SBarry Smith 
157395fce210SBarry Smith /*@C
1574a1729e3fSJunchao Zhang    PetscSFFetchAndOpBegin - begin operation that fetches values from root and updates atomically by applying operation using my leaf value, to be completed with PetscSFFetchAndOpEnd()
1575a1729e3fSJunchao Zhang 
1576a1729e3fSJunchao Zhang    Collective
1577a1729e3fSJunchao Zhang 
15784165533cSJose E. Roman    Input Parameters:
1579a1729e3fSJunchao Zhang +  sf - star forest
1580a1729e3fSJunchao Zhang .  unit - data type
1581a1729e3fSJunchao Zhang .  leafdata - leaf values to use in reduction
1582a1729e3fSJunchao Zhang -  op - operation to use for reduction
1583a1729e3fSJunchao Zhang 
15844165533cSJose E. Roman    Output Parameters:
1585a1729e3fSJunchao Zhang +  rootdata - root values to be updated, input state is seen by first process to perform an update
1586a1729e3fSJunchao Zhang -  leafupdate - state at each leaf's respective root immediately prior to my atomic update
1587a1729e3fSJunchao Zhang 
1588a1729e3fSJunchao Zhang    Level: advanced
1589a1729e3fSJunchao Zhang 
1590a1729e3fSJunchao Zhang    Note:
1591a1729e3fSJunchao Zhang    The update is only atomic at the granularity provided by the hardware. Different roots referenced by the same process
1592a1729e3fSJunchao Zhang    might be updated in a different order. Furthermore, if a composite type is used for the unit datatype, atomicity is
1593a1729e3fSJunchao Zhang    not guaranteed across the whole vertex. Therefore, this function is mostly only used with primitive types such as
1594a1729e3fSJunchao Zhang    integers.
1595a1729e3fSJunchao Zhang 
1596a1729e3fSJunchao Zhang .seealso: PetscSFComputeDegreeBegin(), PetscSFReduceBegin(), PetscSFSetGraph()
1597a1729e3fSJunchao Zhang @*/
1598a1729e3fSJunchao Zhang PetscErrorCode PetscSFFetchAndOpBegin(PetscSF sf,MPI_Datatype unit,void *rootdata,const void *leafdata,void *leafupdate,MPI_Op op)
1599a1729e3fSJunchao Zhang {
1600eb02082bSJunchao Zhang   PetscMemType   rootmtype,leafmtype,leafupdatemtype;
1601a1729e3fSJunchao Zhang 
1602a1729e3fSJunchao Zhang   PetscFunctionBegin;
1603a1729e3fSJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
16049566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
16059566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpBegin,sf,0,0,0));
16069566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(rootdata,&rootmtype));
16079566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafdata,&leafmtype));
16089566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafupdate,&leafupdatemtype));
1609*08401ef6SPierre Jolivet   PetscCheck(leafmtype == leafupdatemtype,PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for leafdata and leafupdate in different memory types");
16109566063dSJacob Faibussowitsch   PetscCall((*sf->ops->FetchAndOpBegin)(sf,unit,rootmtype,rootdata,leafmtype,leafdata,leafupdate,op));
16119566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpBegin,sf,0,0,0));
1612a1729e3fSJunchao Zhang   PetscFunctionReturn(0);
1613a1729e3fSJunchao Zhang }
1614a1729e3fSJunchao Zhang 
1615a1729e3fSJunchao Zhang /*@C
1616d3b3e55cSJunchao Zhang    PetscSFFetchAndOpWithMemTypeBegin - begin operation with explicit memory types that fetches values from root and updates atomically by applying operation using my leaf value, to be completed with PetscSFFetchAndOpEnd()
1617d3b3e55cSJunchao Zhang 
1618d3b3e55cSJunchao Zhang    Collective
1619d3b3e55cSJunchao Zhang 
1620d3b3e55cSJunchao Zhang    Input Parameters:
1621d3b3e55cSJunchao Zhang +  sf - star forest
1622d3b3e55cSJunchao Zhang .  unit - data type
1623d3b3e55cSJunchao Zhang .  rootmtype - memory type of rootdata
1624d3b3e55cSJunchao Zhang .  leafmtype - memory type of leafdata
1625d3b3e55cSJunchao Zhang .  leafdata - leaf values to use in reduction
1626d3b3e55cSJunchao Zhang .  leafupdatemtype - memory type of leafupdate
1627d3b3e55cSJunchao Zhang -  op - operation to use for reduction
1628d3b3e55cSJunchao Zhang 
1629d3b3e55cSJunchao Zhang    Output Parameters:
1630d3b3e55cSJunchao Zhang +  rootdata - root values to be updated, input state is seen by first process to perform an update
1631d3b3e55cSJunchao Zhang -  leafupdate - state at each leaf's respective root immediately prior to my atomic update
1632d3b3e55cSJunchao Zhang 
1633d3b3e55cSJunchao Zhang    Level: advanced
1634d3b3e55cSJunchao Zhang 
1635d3b3e55cSJunchao Zhang    Note: See PetscSFFetchAndOpBegin() for more details.
1636d3b3e55cSJunchao Zhang 
1637d3b3e55cSJunchao Zhang .seealso: PetscSFFetchAndOpBegin(),PetscSFComputeDegreeBegin(), PetscSFReduceBegin(), PetscSFSetGraph()
1638d3b3e55cSJunchao Zhang @*/
1639d3b3e55cSJunchao Zhang PetscErrorCode PetscSFFetchAndOpWithMemTypeBegin(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,void *rootdata,PetscMemType leafmtype,const void *leafdata,PetscMemType leafupdatemtype,void *leafupdate,MPI_Op op)
1640d3b3e55cSJunchao Zhang {
1641d3b3e55cSJunchao Zhang   PetscFunctionBegin;
1642d3b3e55cSJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
16439566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
16449566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpBegin,sf,0,0,0));
1645*08401ef6SPierre Jolivet   PetscCheck(leafmtype == leafupdatemtype,PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for leafdata and leafupdate in different memory types");
16469566063dSJacob Faibussowitsch   PetscCall((*sf->ops->FetchAndOpBegin)(sf,unit,rootmtype,rootdata,leafmtype,leafdata,leafupdate,op));
16479566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpBegin,sf,0,0,0));
1648d3b3e55cSJunchao Zhang   PetscFunctionReturn(0);
1649d3b3e55cSJunchao Zhang }
1650d3b3e55cSJunchao Zhang 
1651d3b3e55cSJunchao Zhang /*@C
1652a1729e3fSJunchao 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
1653a1729e3fSJunchao Zhang 
1654a1729e3fSJunchao Zhang    Collective
1655a1729e3fSJunchao Zhang 
16564165533cSJose E. Roman    Input Parameters:
1657a1729e3fSJunchao Zhang +  sf - star forest
1658a1729e3fSJunchao Zhang .  unit - data type
1659a1729e3fSJunchao Zhang .  leafdata - leaf values to use in reduction
1660a1729e3fSJunchao Zhang -  op - operation to use for reduction
1661a1729e3fSJunchao Zhang 
16624165533cSJose E. Roman    Output Parameters:
1663a1729e3fSJunchao Zhang +  rootdata - root values to be updated, input state is seen by first process to perform an update
1664a1729e3fSJunchao Zhang -  leafupdate - state at each leaf's respective root immediately prior to my atomic update
1665a1729e3fSJunchao Zhang 
1666a1729e3fSJunchao Zhang    Level: advanced
1667a1729e3fSJunchao Zhang 
1668a1729e3fSJunchao Zhang .seealso: PetscSFComputeDegreeEnd(), PetscSFReduceEnd(), PetscSFSetGraph()
1669a1729e3fSJunchao Zhang @*/
1670a1729e3fSJunchao Zhang PetscErrorCode PetscSFFetchAndOpEnd(PetscSF sf,MPI_Datatype unit,void *rootdata,const void *leafdata,void *leafupdate,MPI_Op op)
1671a1729e3fSJunchao Zhang {
1672a1729e3fSJunchao Zhang   PetscFunctionBegin;
1673a1729e3fSJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
16749566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpEnd,sf,0,0,0));
16759566063dSJacob Faibussowitsch   PetscCall((*sf->ops->FetchAndOpEnd)(sf,unit,rootdata,leafdata,leafupdate,op));
16769566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpEnd,sf,0,0,0));
1677a1729e3fSJunchao Zhang   PetscFunctionReturn(0);
1678a1729e3fSJunchao Zhang }
1679a1729e3fSJunchao Zhang 
1680a1729e3fSJunchao Zhang /*@C
168195fce210SBarry Smith    PetscSFComputeDegreeBegin - begin computation of degree for each root vertex, to be completed with PetscSFComputeDegreeEnd()
168295fce210SBarry Smith 
168395fce210SBarry Smith    Collective
168495fce210SBarry Smith 
16854165533cSJose E. Roman    Input Parameter:
168695fce210SBarry Smith .  sf - star forest
168795fce210SBarry Smith 
16884165533cSJose E. Roman    Output Parameter:
168995fce210SBarry Smith .  degree - degree of each root vertex
169095fce210SBarry Smith 
169195fce210SBarry Smith    Level: advanced
169295fce210SBarry Smith 
1693ffe67aa5SVáclav Hapla    Notes:
1694ffe67aa5SVáclav Hapla    The returned array is owned by PetscSF and automatically freed by PetscSFDestroy(). Hence no need to call PetscFree() on it.
1695ffe67aa5SVáclav Hapla 
169695fce210SBarry Smith .seealso: PetscSFGatherBegin()
169795fce210SBarry Smith @*/
169895fce210SBarry Smith PetscErrorCode PetscSFComputeDegreeBegin(PetscSF sf,const PetscInt **degree)
169995fce210SBarry Smith {
170095fce210SBarry Smith   PetscFunctionBegin;
170195fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
170295fce210SBarry Smith   PetscSFCheckGraphSet(sf,1);
170395fce210SBarry Smith   PetscValidPointer(degree,2);
1704803bd9e8SMatthew G. Knepley   if (!sf->degreeknown) {
17055b0d146aSStefano Zampini     PetscInt i, nroots = sf->nroots, maxlocal;
170628b400f6SJacob Faibussowitsch     PetscCheck(!sf->degree,PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Calls to PetscSFComputeDegreeBegin() cannot be nested.");
17075b0d146aSStefano Zampini     maxlocal = sf->maxleaf-sf->minleaf+1;
17089566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nroots,&sf->degree));
17099566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(PetscMax(maxlocal,1),&sf->degreetmp)); /* allocate at least one entry, see check in PetscSFComputeDegreeEnd() */
171029046d53SLisandro Dalcin     for (i=0; i<nroots; i++) sf->degree[i] = 0;
17119837ea96SMatthew G. Knepley     for (i=0; i<maxlocal; i++) sf->degreetmp[i] = 1;
17129566063dSJacob Faibussowitsch     PetscCall(PetscSFReduceBegin(sf,MPIU_INT,sf->degreetmp-sf->minleaf,sf->degree,MPI_SUM));
171395fce210SBarry Smith   }
171495fce210SBarry Smith   *degree = NULL;
171595fce210SBarry Smith   PetscFunctionReturn(0);
171695fce210SBarry Smith }
171795fce210SBarry Smith 
171895fce210SBarry Smith /*@C
171995fce210SBarry Smith    PetscSFComputeDegreeEnd - complete computation of degree for each root vertex, started with PetscSFComputeDegreeBegin()
172095fce210SBarry Smith 
172195fce210SBarry Smith    Collective
172295fce210SBarry Smith 
17234165533cSJose E. Roman    Input Parameter:
172495fce210SBarry Smith .  sf - star forest
172595fce210SBarry Smith 
17264165533cSJose E. Roman    Output Parameter:
172795fce210SBarry Smith .  degree - degree of each root vertex
172895fce210SBarry Smith 
172995fce210SBarry Smith    Level: developer
173095fce210SBarry Smith 
1731ffe67aa5SVáclav Hapla    Notes:
1732ffe67aa5SVáclav Hapla    The returned array is owned by PetscSF and automatically freed by PetscSFDestroy(). Hence no need to call PetscFree() on it.
1733ffe67aa5SVáclav Hapla 
173495fce210SBarry Smith .seealso:
173595fce210SBarry Smith @*/
173695fce210SBarry Smith PetscErrorCode PetscSFComputeDegreeEnd(PetscSF sf,const PetscInt **degree)
173795fce210SBarry Smith {
173895fce210SBarry Smith   PetscFunctionBegin;
173995fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
174095fce210SBarry Smith   PetscSFCheckGraphSet(sf,1);
174129046d53SLisandro Dalcin   PetscValidPointer(degree,2);
174295fce210SBarry Smith   if (!sf->degreeknown) {
174328b400f6SJacob Faibussowitsch     PetscCheck(sf->degreetmp,PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFComputeDegreeBegin() before PetscSFComputeDegreeEnd()");
17449566063dSJacob Faibussowitsch     PetscCall(PetscSFReduceEnd(sf,MPIU_INT,sf->degreetmp-sf->minleaf,sf->degree,MPI_SUM));
17459566063dSJacob Faibussowitsch     PetscCall(PetscFree(sf->degreetmp));
174695fce210SBarry Smith     sf->degreeknown = PETSC_TRUE;
174795fce210SBarry Smith   }
174895fce210SBarry Smith   *degree = sf->degree;
174995fce210SBarry Smith   PetscFunctionReturn(0);
175095fce210SBarry Smith }
175195fce210SBarry Smith 
1752673100f5SVaclav Hapla /*@C
175366dfcd1aSVaclav Hapla    PetscSFComputeMultiRootOriginalNumbering - Returns original numbering of multi-roots (roots of multi-SF returned by PetscSFGetMultiSF()).
175466dfcd1aSVaclav Hapla    Each multi-root is assigned index of the corresponding original root.
1755673100f5SVaclav Hapla 
1756673100f5SVaclav Hapla    Collective
1757673100f5SVaclav Hapla 
17584165533cSJose E. Roman    Input Parameters:
1759673100f5SVaclav Hapla +  sf - star forest
1760673100f5SVaclav Hapla -  degree - degree of each root vertex, computed with PetscSFComputeDegreeBegin()/PetscSFComputeDegreeEnd()
1761673100f5SVaclav Hapla 
17624165533cSJose E. Roman    Output Parameters:
176366dfcd1aSVaclav Hapla +  nMultiRoots - (optional) number of multi-roots (roots of multi-SF)
176466dfcd1aSVaclav Hapla -  multiRootsOrigNumbering - original indices of multi-roots; length of this array is nMultiRoots
1765673100f5SVaclav Hapla 
1766673100f5SVaclav Hapla    Level: developer
1767673100f5SVaclav Hapla 
1768ffe67aa5SVáclav Hapla    Notes:
1769ffe67aa5SVáclav Hapla    The returned array multiRootsOrigNumbering is newly allocated and should be destroyed with PetscFree() when no longer needed.
1770ffe67aa5SVáclav Hapla 
1771673100f5SVaclav Hapla .seealso: PetscSFComputeDegreeBegin(), PetscSFComputeDegreeEnd(), PetscSFGetMultiSF()
1772673100f5SVaclav Hapla @*/
177366dfcd1aSVaclav Hapla PetscErrorCode PetscSFComputeMultiRootOriginalNumbering(PetscSF sf, const PetscInt degree[], PetscInt *nMultiRoots, PetscInt *multiRootsOrigNumbering[])
1774673100f5SVaclav Hapla {
1775673100f5SVaclav Hapla   PetscSF             msf;
1776673100f5SVaclav Hapla   PetscInt            i, j, k, nroots, nmroots;
1777673100f5SVaclav Hapla 
1778673100f5SVaclav Hapla   PetscFunctionBegin;
1779673100f5SVaclav Hapla   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
17809566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sf, &nroots, NULL, NULL, NULL));
178129328920SVaclav Hapla   if (nroots) PetscValidIntPointer(degree,2);
178266dfcd1aSVaclav Hapla   if (nMultiRoots) PetscValidIntPointer(nMultiRoots,3);
178366dfcd1aSVaclav Hapla   PetscValidPointer(multiRootsOrigNumbering,4);
17849566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf,&msf));
17859566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(msf, &nmroots, NULL, NULL, NULL));
17869566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(nmroots, multiRootsOrigNumbering));
1787673100f5SVaclav Hapla   for (i=0,j=0,k=0; i<nroots; i++) {
1788673100f5SVaclav Hapla     if (!degree[i]) continue;
1789673100f5SVaclav Hapla     for (j=0; j<degree[i]; j++,k++) {
179066dfcd1aSVaclav Hapla       (*multiRootsOrigNumbering)[k] = i;
1791673100f5SVaclav Hapla     }
1792673100f5SVaclav Hapla   }
1793*08401ef6SPierre Jolivet   PetscCheck(k == nmroots,PETSC_COMM_SELF,PETSC_ERR_PLIB,"sanity check fail");
179466dfcd1aSVaclav Hapla   if (nMultiRoots) *nMultiRoots = nmroots;
1795673100f5SVaclav Hapla   PetscFunctionReturn(0);
1796673100f5SVaclav Hapla }
1797673100f5SVaclav Hapla 
179895fce210SBarry Smith /*@C
179995fce210SBarry Smith    PetscSFGatherBegin - begin pointwise gather of all leaves into multi-roots, to be completed with PetscSFGatherEnd()
180095fce210SBarry Smith 
180195fce210SBarry Smith    Collective
180295fce210SBarry Smith 
18034165533cSJose E. Roman    Input Parameters:
180495fce210SBarry Smith +  sf - star forest
180595fce210SBarry Smith .  unit - data type
180695fce210SBarry Smith -  leafdata - leaf data to gather to roots
180795fce210SBarry Smith 
18084165533cSJose E. Roman    Output Parameter:
180995fce210SBarry Smith .  multirootdata - root buffer to gather into, amount of space per root is equal to its degree
181095fce210SBarry Smith 
181195fce210SBarry Smith    Level: intermediate
181295fce210SBarry Smith 
181395fce210SBarry Smith .seealso: PetscSFComputeDegreeBegin(), PetscSFScatterBegin()
181495fce210SBarry Smith @*/
181595fce210SBarry Smith PetscErrorCode PetscSFGatherBegin(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *multirootdata)
181695fce210SBarry Smith {
1817a5526d50SJunchao Zhang   PetscSF        multi = NULL;
181895fce210SBarry Smith 
181995fce210SBarry Smith   PetscFunctionBegin;
182095fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
18219566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
18229566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf,&multi));
18239566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceBegin(multi,unit,leafdata,multirootdata,MPI_REPLACE));
182495fce210SBarry Smith   PetscFunctionReturn(0);
182595fce210SBarry Smith }
182695fce210SBarry Smith 
182795fce210SBarry Smith /*@C
182895fce210SBarry Smith    PetscSFGatherEnd - ends pointwise gather operation that was started with PetscSFGatherBegin()
182995fce210SBarry Smith 
183095fce210SBarry Smith    Collective
183195fce210SBarry Smith 
18324165533cSJose E. Roman    Input Parameters:
183395fce210SBarry Smith +  sf - star forest
183495fce210SBarry Smith .  unit - data type
183595fce210SBarry Smith -  leafdata - leaf data to gather to roots
183695fce210SBarry Smith 
18374165533cSJose E. Roman    Output Parameter:
183895fce210SBarry Smith .  multirootdata - root buffer to gather into, amount of space per root is equal to its degree
183995fce210SBarry Smith 
184095fce210SBarry Smith    Level: intermediate
184195fce210SBarry Smith 
184295fce210SBarry Smith .seealso: PetscSFComputeDegreeEnd(), PetscSFScatterEnd()
184395fce210SBarry Smith @*/
184495fce210SBarry Smith PetscErrorCode PetscSFGatherEnd(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *multirootdata)
184595fce210SBarry Smith {
1846a5526d50SJunchao Zhang   PetscSF        multi = NULL;
184795fce210SBarry Smith 
184895fce210SBarry Smith   PetscFunctionBegin;
184995fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
18509566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf,&multi));
18519566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceEnd(multi,unit,leafdata,multirootdata,MPI_REPLACE));
185295fce210SBarry Smith   PetscFunctionReturn(0);
185395fce210SBarry Smith }
185495fce210SBarry Smith 
185595fce210SBarry Smith /*@C
185695fce210SBarry Smith    PetscSFScatterBegin - begin pointwise scatter operation from multi-roots to leaves, to be completed with PetscSFScatterEnd()
185795fce210SBarry Smith 
185895fce210SBarry Smith    Collective
185995fce210SBarry Smith 
18604165533cSJose E. Roman    Input Parameters:
186195fce210SBarry Smith +  sf - star forest
186295fce210SBarry Smith .  unit - data type
186395fce210SBarry Smith -  multirootdata - root buffer to send to each leaf, one unit of data per leaf
186495fce210SBarry Smith 
18654165533cSJose E. Roman    Output Parameter:
186695fce210SBarry Smith .  leafdata - leaf data to be update with personal data from each respective root
186795fce210SBarry Smith 
186895fce210SBarry Smith    Level: intermediate
186995fce210SBarry Smith 
187095fce210SBarry Smith .seealso: PetscSFComputeDegreeBegin(), PetscSFScatterBegin()
187195fce210SBarry Smith @*/
187295fce210SBarry Smith PetscErrorCode PetscSFScatterBegin(PetscSF sf,MPI_Datatype unit,const void *multirootdata,void *leafdata)
187395fce210SBarry Smith {
1874a5526d50SJunchao Zhang   PetscSF        multi = NULL;
187595fce210SBarry Smith 
187695fce210SBarry Smith   PetscFunctionBegin;
187795fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
18789566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
18799566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf,&multi));
18809566063dSJacob Faibussowitsch   PetscCall(PetscSFBcastBegin(multi,unit,multirootdata,leafdata,MPI_REPLACE));
188195fce210SBarry Smith   PetscFunctionReturn(0);
188295fce210SBarry Smith }
188395fce210SBarry Smith 
188495fce210SBarry Smith /*@C
188595fce210SBarry Smith    PetscSFScatterEnd - ends pointwise scatter operation that was started with PetscSFScatterBegin()
188695fce210SBarry Smith 
188795fce210SBarry Smith    Collective
188895fce210SBarry Smith 
18894165533cSJose E. Roman    Input Parameters:
189095fce210SBarry Smith +  sf - star forest
189195fce210SBarry Smith .  unit - data type
189295fce210SBarry Smith -  multirootdata - root buffer to send to each leaf, one unit of data per leaf
189395fce210SBarry Smith 
18944165533cSJose E. Roman    Output Parameter:
189595fce210SBarry Smith .  leafdata - leaf data to be update with personal data from each respective root
189695fce210SBarry Smith 
189795fce210SBarry Smith    Level: intermediate
189895fce210SBarry Smith 
189995fce210SBarry Smith .seealso: PetscSFComputeDegreeEnd(), PetscSFScatterEnd()
190095fce210SBarry Smith @*/
190195fce210SBarry Smith PetscErrorCode PetscSFScatterEnd(PetscSF sf,MPI_Datatype unit,const void *multirootdata,void *leafdata)
190295fce210SBarry Smith {
1903a5526d50SJunchao Zhang   PetscSF        multi = NULL;
190495fce210SBarry Smith 
190595fce210SBarry Smith   PetscFunctionBegin;
190695fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
19079566063dSJacob Faibussowitsch   PetscCall(PetscSFGetMultiSF(sf,&multi));
19089566063dSJacob Faibussowitsch   PetscCall(PetscSFBcastEnd(multi,unit,multirootdata,leafdata,MPI_REPLACE));
190995fce210SBarry Smith   PetscFunctionReturn(0);
191095fce210SBarry Smith }
1911a7b3aa13SAta Mesgarnejad 
1912a072220fSLawrence Mitchell static PetscErrorCode PetscSFCheckLeavesUnique_Private(PetscSF sf)
1913a072220fSLawrence Mitchell {
1914a072220fSLawrence Mitchell   PetscInt        i, n, nleaves;
1915a072220fSLawrence Mitchell   const PetscInt *ilocal = NULL;
1916a072220fSLawrence Mitchell   PetscHSetI      seen;
1917a072220fSLawrence Mitchell 
1918a072220fSLawrence Mitchell   PetscFunctionBegin;
1919b458e8f1SJose E. Roman   if (PetscDefined(USE_DEBUG)) {
19209566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sf,NULL,&nleaves,&ilocal,NULL));
19219566063dSJacob Faibussowitsch     PetscCall(PetscHSetICreate(&seen));
1922a072220fSLawrence Mitchell     for (i = 0; i < nleaves; i++) {
1923a072220fSLawrence Mitchell       const PetscInt leaf = ilocal ? ilocal[i] : i;
19249566063dSJacob Faibussowitsch       PetscCall(PetscHSetIAdd(seen,leaf));
1925a072220fSLawrence Mitchell     }
19269566063dSJacob Faibussowitsch     PetscCall(PetscHSetIGetSize(seen,&n));
1927*08401ef6SPierre Jolivet     PetscCheck(n == nleaves,PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Provided leaves have repeated values: all leaves must be unique");
19289566063dSJacob Faibussowitsch     PetscCall(PetscHSetIDestroy(&seen));
1929b458e8f1SJose E. Roman   }
1930a072220fSLawrence Mitchell   PetscFunctionReturn(0);
1931a072220fSLawrence Mitchell }
193254729392SStefano Zampini 
1933a7b3aa13SAta Mesgarnejad /*@
193404c0ada0SJunchao Zhang   PetscSFCompose - Compose a new PetscSF by putting the second SF under the first one in a top (roots) down (leaves) view
1935a7b3aa13SAta Mesgarnejad 
1936a7b3aa13SAta Mesgarnejad   Input Parameters:
193754729392SStefano Zampini + sfA - The first PetscSF
193854729392SStefano Zampini - sfB - The second PetscSF
1939a7b3aa13SAta Mesgarnejad 
1940a7b3aa13SAta Mesgarnejad   Output Parameters:
194154729392SStefano Zampini . sfBA - The composite SF
1942a7b3aa13SAta Mesgarnejad 
1943a7b3aa13SAta Mesgarnejad   Level: developer
1944a7b3aa13SAta Mesgarnejad 
1945a072220fSLawrence Mitchell   Notes:
194654729392SStefano Zampini   Currently, the two SFs must be defined on congruent communicators and they must be true star
194754729392SStefano Zampini   forests, i.e. the same leaf is not connected with different roots.
194854729392SStefano Zampini 
194954729392SStefano Zampini   sfA's leaf space and sfB's root space might be partially overlapped. The composition builds
195054729392SStefano Zampini   a graph with sfA's roots and sfB's leaves only when there is a path between them. Unconnected
195154729392SStefano Zampini   nodes (roots or leaves) are not in sfBA. Doing a Bcast on the new SF is equivalent to doing a
195254729392SStefano Zampini   Bcast on sfA, then a Bcast on sfB, on connected nodes.
1953a072220fSLawrence Mitchell 
195404c0ada0SJunchao Zhang .seealso: PetscSF, PetscSFComposeInverse(), PetscSFGetGraph(), PetscSFSetGraph()
1955a7b3aa13SAta Mesgarnejad @*/
1956a7b3aa13SAta Mesgarnejad PetscErrorCode PetscSFCompose(PetscSF sfA,PetscSF sfB,PetscSF *sfBA)
1957a7b3aa13SAta Mesgarnejad {
1958a7b3aa13SAta Mesgarnejad   const PetscSFNode *remotePointsA,*remotePointsB;
1959d41018fbSJunchao Zhang   PetscSFNode       *remotePointsBA=NULL,*reorderedRemotePointsA = NULL,*leafdataB;
196054729392SStefano Zampini   const PetscInt    *localPointsA,*localPointsB;
196154729392SStefano Zampini   PetscInt          *localPointsBA;
196254729392SStefano Zampini   PetscInt          i,numRootsA,numLeavesA,numRootsB,numLeavesB,minleaf,maxleaf,numLeavesBA;
196354729392SStefano Zampini   PetscBool         denseB;
1964a7b3aa13SAta Mesgarnejad 
1965a7b3aa13SAta Mesgarnejad   PetscFunctionBegin;
1966a7b3aa13SAta Mesgarnejad   PetscValidHeaderSpecific(sfA,PETSCSF_CLASSID,1);
196729046d53SLisandro Dalcin   PetscSFCheckGraphSet(sfA,1);
196829046d53SLisandro Dalcin   PetscValidHeaderSpecific(sfB,PETSCSF_CLASSID,2);
196929046d53SLisandro Dalcin   PetscSFCheckGraphSet(sfB,2);
197054729392SStefano Zampini   PetscCheckSameComm(sfA,1,sfB,2);
197129046d53SLisandro Dalcin   PetscValidPointer(sfBA,3);
19729566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckLeavesUnique_Private(sfA));
19739566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckLeavesUnique_Private(sfB));
197454729392SStefano Zampini 
19759566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sfA,&numRootsA,&numLeavesA,&localPointsA,&remotePointsA));
19769566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sfB,&numRootsB,&numLeavesB,&localPointsB,&remotePointsB));
1977d41018fbSJunchao Zhang   if (localPointsA) {
19789566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(numRootsB,&reorderedRemotePointsA));
197954729392SStefano Zampini     for (i=0; i<numRootsB; i++) {
198054729392SStefano Zampini       reorderedRemotePointsA[i].rank = -1;
198154729392SStefano Zampini       reorderedRemotePointsA[i].index = -1;
198254729392SStefano Zampini     }
198354729392SStefano Zampini     for (i=0; i<numLeavesA; i++) {
198454729392SStefano Zampini       if (localPointsA[i] >= numRootsB) continue;
198554729392SStefano Zampini       reorderedRemotePointsA[localPointsA[i]] = remotePointsA[i];
198654729392SStefano Zampini     }
1987d41018fbSJunchao Zhang     remotePointsA = reorderedRemotePointsA;
1988d41018fbSJunchao Zhang   }
19899566063dSJacob Faibussowitsch   PetscCall(PetscSFGetLeafRange(sfB,&minleaf,&maxleaf));
19909566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(maxleaf-minleaf+1,&leafdataB));
19919566063dSJacob Faibussowitsch   PetscCall(PetscSFBcastBegin(sfB,MPIU_2INT,remotePointsA,leafdataB-minleaf,MPI_REPLACE));
19929566063dSJacob Faibussowitsch   PetscCall(PetscSFBcastEnd(sfB,MPIU_2INT,remotePointsA,leafdataB-minleaf,MPI_REPLACE));
19939566063dSJacob Faibussowitsch   PetscCall(PetscFree(reorderedRemotePointsA));
1994d41018fbSJunchao Zhang 
199554729392SStefano Zampini   denseB = (PetscBool)!localPointsB;
199654729392SStefano Zampini   for (i=0,numLeavesBA=0; i<numLeavesB; i++) {
199754729392SStefano Zampini     if (leafdataB[localPointsB ? localPointsB[i]-minleaf : i].rank == -1) denseB = PETSC_FALSE;
199854729392SStefano Zampini     else numLeavesBA++;
199954729392SStefano Zampini   }
200054729392SStefano Zampini   if (denseB) {
2001d41018fbSJunchao Zhang     localPointsBA  = NULL;
2002d41018fbSJunchao Zhang     remotePointsBA = leafdataB;
2003d41018fbSJunchao Zhang   } else {
20049566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(numLeavesBA,&localPointsBA));
20059566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(numLeavesBA,&remotePointsBA));
200654729392SStefano Zampini     for (i=0,numLeavesBA=0; i<numLeavesB; i++) {
200754729392SStefano Zampini       const PetscInt l = localPointsB ? localPointsB[i] : i;
200854729392SStefano Zampini 
200954729392SStefano Zampini       if (leafdataB[l-minleaf].rank == -1) continue;
201054729392SStefano Zampini       remotePointsBA[numLeavesBA] = leafdataB[l-minleaf];
201154729392SStefano Zampini       localPointsBA[numLeavesBA] = l;
201254729392SStefano Zampini       numLeavesBA++;
201354729392SStefano Zampini     }
20149566063dSJacob Faibussowitsch     PetscCall(PetscFree(leafdataB));
2015d41018fbSJunchao Zhang   }
20169566063dSJacob Faibussowitsch   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sfA),sfBA));
20179566063dSJacob Faibussowitsch   PetscCall(PetscSFSetFromOptions(*sfBA));
20189566063dSJacob Faibussowitsch   PetscCall(PetscSFSetGraph(*sfBA,numRootsA,numLeavesBA,localPointsBA,PETSC_OWN_POINTER,remotePointsBA,PETSC_OWN_POINTER));
2019a7b3aa13SAta Mesgarnejad   PetscFunctionReturn(0);
2020a7b3aa13SAta Mesgarnejad }
20211c6ba672SJunchao Zhang 
202204c0ada0SJunchao Zhang /*@
202354729392SStefano Zampini   PetscSFComposeInverse - Compose a new PetscSF by putting the inverse of the second SF under the first one
202404c0ada0SJunchao Zhang 
202504c0ada0SJunchao Zhang   Input Parameters:
202654729392SStefano Zampini + sfA - The first PetscSF
202754729392SStefano Zampini - sfB - The second PetscSF
202804c0ada0SJunchao Zhang 
202904c0ada0SJunchao Zhang   Output Parameters:
203054729392SStefano Zampini . sfBA - The composite SF.
203104c0ada0SJunchao Zhang 
203204c0ada0SJunchao Zhang   Level: developer
203304c0ada0SJunchao Zhang 
203454729392SStefano Zampini   Notes:
203554729392SStefano Zampini   Currently, the two SFs must be defined on congruent communicators and they must be true star
203654729392SStefano Zampini   forests, i.e. the same leaf is not connected with different roots. Even more, all roots of the
203754729392SStefano Zampini   second SF must have a degree of 1, i.e., no roots have more than one leaf connected.
203854729392SStefano Zampini 
203954729392SStefano Zampini   sfA's leaf space and sfB's leaf space might be partially overlapped. The composition builds
204054729392SStefano Zampini   a graph with sfA's roots and sfB's roots only when there is a path between them. Unconnected
204154729392SStefano Zampini   roots are not in sfBA. Doing a Bcast on the new SF is equivalent to doing a Bcast on sfA, then
204254729392SStefano Zampini   a Reduce on sfB, on connected roots.
204354729392SStefano Zampini 
204454729392SStefano Zampini .seealso: PetscSF, PetscSFCompose(), PetscSFGetGraph(), PetscSFSetGraph(), PetscSFCreateInverseSF()
204504c0ada0SJunchao Zhang @*/
204604c0ada0SJunchao Zhang PetscErrorCode PetscSFComposeInverse(PetscSF sfA,PetscSF sfB,PetscSF *sfBA)
204704c0ada0SJunchao Zhang {
204804c0ada0SJunchao Zhang   const PetscSFNode *remotePointsA,*remotePointsB;
204904c0ada0SJunchao Zhang   PetscSFNode       *remotePointsBA;
205004c0ada0SJunchao Zhang   const PetscInt    *localPointsA,*localPointsB;
205154729392SStefano Zampini   PetscSFNode       *reorderedRemotePointsA = NULL;
205254729392SStefano Zampini   PetscInt          i,numRootsA,numLeavesA,numLeavesBA,numRootsB,numLeavesB,minleaf,maxleaf,*localPointsBA;
20535b0d146aSStefano Zampini   MPI_Op            op;
20545b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES)
20555b0d146aSStefano Zampini   PetscBool         iswin;
20565b0d146aSStefano Zampini #endif
205704c0ada0SJunchao Zhang 
205804c0ada0SJunchao Zhang   PetscFunctionBegin;
205904c0ada0SJunchao Zhang   PetscValidHeaderSpecific(sfA,PETSCSF_CLASSID,1);
206004c0ada0SJunchao Zhang   PetscSFCheckGraphSet(sfA,1);
206104c0ada0SJunchao Zhang   PetscValidHeaderSpecific(sfB,PETSCSF_CLASSID,2);
206204c0ada0SJunchao Zhang   PetscSFCheckGraphSet(sfB,2);
206354729392SStefano Zampini   PetscCheckSameComm(sfA,1,sfB,2);
206404c0ada0SJunchao Zhang   PetscValidPointer(sfBA,3);
20659566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckLeavesUnique_Private(sfA));
20669566063dSJacob Faibussowitsch   PetscCall(PetscSFCheckLeavesUnique_Private(sfB));
206754729392SStefano Zampini 
20689566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sfA, &numRootsA, &numLeavesA, &localPointsA, &remotePointsA));
20699566063dSJacob Faibussowitsch   PetscCall(PetscSFGetGraph(sfB, &numRootsB, &numLeavesB, &localPointsB, &remotePointsB));
20705b0d146aSStefano Zampini 
20715b0d146aSStefano Zampini   /* TODO: Check roots of sfB have degree of 1 */
20725b0d146aSStefano Zampini   /* Once we implement it, we can replace the MPI_MAXLOC
207383df288dSJunchao Zhang      with MPI_REPLACE. In that case, MPI_MAXLOC and MPI_REPLACE have the same effect.
20745b0d146aSStefano Zampini      We use MPI_MAXLOC only to have a deterministic output from this routine if
20755b0d146aSStefano Zampini      the root condition is not meet.
20765b0d146aSStefano Zampini    */
20775b0d146aSStefano Zampini   op = MPI_MAXLOC;
20785b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES)
20795b0d146aSStefano Zampini   /* we accept a non-deterministic output (if any) with PETSCSFWINDOW, since MPI_MAXLOC cannot operate on MPIU_2INT with MPI_Accumulate */
20809566063dSJacob Faibussowitsch   PetscCall(PetscObjectTypeCompare((PetscObject)sfB,PETSCSFWINDOW,&iswin));
208183df288dSJunchao Zhang   if (iswin) op = MPI_REPLACE;
20825b0d146aSStefano Zampini #endif
20835b0d146aSStefano Zampini 
20849566063dSJacob Faibussowitsch   PetscCall(PetscSFGetLeafRange(sfB, &minleaf, &maxleaf));
20859566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(maxleaf - minleaf + 1,&reorderedRemotePointsA));
208654729392SStefano Zampini   for (i=0; i<maxleaf - minleaf + 1; i++) {
208754729392SStefano Zampini     reorderedRemotePointsA[i].rank = -1;
208854729392SStefano Zampini     reorderedRemotePointsA[i].index = -1;
208954729392SStefano Zampini   }
209054729392SStefano Zampini   if (localPointsA) {
209154729392SStefano Zampini     for (i=0; i<numLeavesA; i++) {
209254729392SStefano Zampini       if (localPointsA[i] > maxleaf || localPointsA[i] < minleaf) continue;
209354729392SStefano Zampini       reorderedRemotePointsA[localPointsA[i] - minleaf] = remotePointsA[i];
209454729392SStefano Zampini     }
209554729392SStefano Zampini   } else {
209654729392SStefano Zampini     for (i=0; i<numLeavesA; i++) {
209754729392SStefano Zampini       if (i > maxleaf || i < minleaf) continue;
209854729392SStefano Zampini       reorderedRemotePointsA[i - minleaf] = remotePointsA[i];
209954729392SStefano Zampini     }
210054729392SStefano Zampini   }
210154729392SStefano Zampini 
21029566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(numRootsB,&localPointsBA));
21039566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(numRootsB,&remotePointsBA));
210454729392SStefano Zampini   for (i=0; i<numRootsB; i++) {
210554729392SStefano Zampini     remotePointsBA[i].rank = -1;
210654729392SStefano Zampini     remotePointsBA[i].index = -1;
210754729392SStefano Zampini   }
210854729392SStefano Zampini 
21099566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceBegin(sfB,MPIU_2INT,reorderedRemotePointsA-minleaf,remotePointsBA,op));
21109566063dSJacob Faibussowitsch   PetscCall(PetscSFReduceEnd(sfB,MPIU_2INT,reorderedRemotePointsA-minleaf,remotePointsBA,op));
21119566063dSJacob Faibussowitsch   PetscCall(PetscFree(reorderedRemotePointsA));
211254729392SStefano Zampini   for (i=0,numLeavesBA=0; i<numRootsB; i++) {
211354729392SStefano Zampini     if (remotePointsBA[i].rank == -1) continue;
211454729392SStefano Zampini     remotePointsBA[numLeavesBA].rank = remotePointsBA[i].rank;
211554729392SStefano Zampini     remotePointsBA[numLeavesBA].index = remotePointsBA[i].index;
211654729392SStefano Zampini     localPointsBA[numLeavesBA] = i;
211754729392SStefano Zampini     numLeavesBA++;
211854729392SStefano Zampini   }
21199566063dSJacob Faibussowitsch   PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sfA),sfBA));
21209566063dSJacob Faibussowitsch   PetscCall(PetscSFSetFromOptions(*sfBA));
21219566063dSJacob Faibussowitsch   PetscCall(PetscSFSetGraph(*sfBA,numRootsA,numLeavesBA,localPointsBA,PETSC_OWN_POINTER,remotePointsBA,PETSC_OWN_POINTER));
212204c0ada0SJunchao Zhang   PetscFunctionReturn(0);
212304c0ada0SJunchao Zhang }
212404c0ada0SJunchao Zhang 
21251c6ba672SJunchao Zhang /*
21261c6ba672SJunchao Zhang   PetscSFCreateLocalSF_Private - Creates a local PetscSF that only has intra-process edges of the global PetscSF
21271c6ba672SJunchao Zhang 
21281c6ba672SJunchao Zhang   Input Parameters:
21291c6ba672SJunchao Zhang . sf - The global PetscSF
21301c6ba672SJunchao Zhang 
21311c6ba672SJunchao Zhang   Output Parameters:
21321c6ba672SJunchao Zhang . out - The local PetscSF
21331c6ba672SJunchao Zhang  */
21341c6ba672SJunchao Zhang PetscErrorCode PetscSFCreateLocalSF_Private(PetscSF sf,PetscSF *out)
21351c6ba672SJunchao Zhang {
21361c6ba672SJunchao Zhang   MPI_Comm           comm;
21371c6ba672SJunchao Zhang   PetscMPIInt        myrank;
21381c6ba672SJunchao Zhang   const PetscInt     *ilocal;
21391c6ba672SJunchao Zhang   const PetscSFNode  *iremote;
21401c6ba672SJunchao Zhang   PetscInt           i,j,nroots,nleaves,lnleaves,*lilocal;
21411c6ba672SJunchao Zhang   PetscSFNode        *liremote;
21421c6ba672SJunchao Zhang   PetscSF            lsf;
21431c6ba672SJunchao Zhang 
21441c6ba672SJunchao Zhang   PetscFunctionBegin;
21451c6ba672SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
21461c6ba672SJunchao Zhang   if (sf->ops->CreateLocalSF) {
21479566063dSJacob Faibussowitsch     PetscCall((*sf->ops->CreateLocalSF)(sf,out));
21481c6ba672SJunchao Zhang   } else {
21491c6ba672SJunchao Zhang     /* Could use PetscSFCreateEmbeddedLeafSF, but since we know the comm is PETSC_COMM_SELF, we can make it fast */
21509566063dSJacob Faibussowitsch     PetscCall(PetscObjectGetComm((PetscObject)sf,&comm));
21519566063dSJacob Faibussowitsch     PetscCallMPI(MPI_Comm_rank(comm,&myrank));
21521c6ba672SJunchao Zhang 
21531c6ba672SJunchao Zhang     /* Find out local edges and build a local SF */
21549566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,&iremote));
21551c6ba672SJunchao Zhang     for (i=lnleaves=0; i<nleaves; i++) {if (iremote[i].rank == (PetscInt)myrank) lnleaves++;}
21569566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(lnleaves,&lilocal));
21579566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(lnleaves,&liremote));
21581c6ba672SJunchao Zhang 
21591c6ba672SJunchao Zhang     for (i=j=0; i<nleaves; i++) {
21601c6ba672SJunchao Zhang       if (iremote[i].rank == (PetscInt)myrank) {
21611c6ba672SJunchao Zhang         lilocal[j]        = ilocal? ilocal[i] : i; /* ilocal=NULL for contiguous storage */
21621c6ba672SJunchao Zhang         liremote[j].rank  = 0; /* rank in PETSC_COMM_SELF */
21631c6ba672SJunchao Zhang         liremote[j].index = iremote[i].index;
21641c6ba672SJunchao Zhang         j++;
21651c6ba672SJunchao Zhang       }
21661c6ba672SJunchao Zhang     }
21679566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(PETSC_COMM_SELF,&lsf));
21689566063dSJacob Faibussowitsch     PetscCall(PetscSFSetFromOptions(lsf));
21699566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(lsf,nroots,lnleaves,lilocal,PETSC_OWN_POINTER,liremote,PETSC_OWN_POINTER));
21709566063dSJacob Faibussowitsch     PetscCall(PetscSFSetUp(lsf));
21711c6ba672SJunchao Zhang     *out = lsf;
21721c6ba672SJunchao Zhang   }
21731c6ba672SJunchao Zhang   PetscFunctionReturn(0);
21741c6ba672SJunchao Zhang }
2175dd5b3ca6SJunchao Zhang 
2176dd5b3ca6SJunchao Zhang /* Similar to PetscSFBcast, but only Bcast to leaves on rank 0 */
2177dd5b3ca6SJunchao Zhang PetscErrorCode PetscSFBcastToZero_Private(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata)
2178dd5b3ca6SJunchao Zhang {
2179eb02082bSJunchao Zhang   PetscMemType       rootmtype,leafmtype;
2180dd5b3ca6SJunchao Zhang 
2181dd5b3ca6SJunchao Zhang   PetscFunctionBegin;
2182dd5b3ca6SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
21839566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(sf));
21849566063dSJacob Faibussowitsch   PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin,sf,0,0,0));
21859566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(rootdata,&rootmtype));
21869566063dSJacob Faibussowitsch   PetscCall(PetscGetMemType(leafdata,&leafmtype));
2187dd5b3ca6SJunchao Zhang   if (sf->ops->BcastToZero) {
21889566063dSJacob Faibussowitsch     PetscCall((*sf->ops->BcastToZero)(sf,unit,rootmtype,rootdata,leafmtype,leafdata));
2189dd5b3ca6SJunchao Zhang   } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"PetscSFBcastToZero_Private is not supported on this SF type");
21909566063dSJacob Faibussowitsch   PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin,sf,0,0,0));
2191dd5b3ca6SJunchao Zhang   PetscFunctionReturn(0);
2192dd5b3ca6SJunchao Zhang }
2193dd5b3ca6SJunchao Zhang 
2194157edd7aSVaclav Hapla /*@
2195157edd7aSVaclav Hapla   PetscSFConcatenate - concatenate multiple SFs into one
2196157edd7aSVaclav Hapla 
2197157edd7aSVaclav Hapla   Input Parameters:
2198157edd7aSVaclav Hapla + comm - the communicator
2199157edd7aSVaclav Hapla . nsfs - the number of input PetscSF
2200157edd7aSVaclav Hapla . sfs  - the array of input PetscSF
2201157edd7aSVaclav Hapla . shareRoots - the flag whether roots of input PetscSFs are taken as shared (PETSC_TRUE), or separate and concatenated (PETSC_FALSE)
2202157edd7aSVaclav Hapla - leafOffsets - the array of local leaf offsets, one for each input PetscSF, or NULL for contiguous storage
2203157edd7aSVaclav Hapla 
2204157edd7aSVaclav Hapla   Output Parameters:
2205157edd7aSVaclav Hapla . newsf - The resulting PetscSF
2206157edd7aSVaclav Hapla 
2207157edd7aSVaclav Hapla   Level: developer
2208157edd7aSVaclav Hapla 
2209157edd7aSVaclav Hapla   Notes:
2210157edd7aSVaclav Hapla   The communicator of all SFs in sfs must be comm.
2211157edd7aSVaclav Hapla 
2212157edd7aSVaclav Hapla   The offsets in leafOffsets are added to the original leaf indices.
2213157edd7aSVaclav Hapla 
2214157edd7aSVaclav Hapla   If all input SFs use contiguous leaf storage (ilocal = NULL), leafOffsets can be passed as NULL as well.
2215157edd7aSVaclav Hapla   In this case, NULL is also passed as ilocal to the resulting SF.
2216157edd7aSVaclav Hapla 
2217157edd7aSVaclav Hapla   If any input SF has non-null ilocal, leafOffsets is needed to distinguish leaves from different input SFs.
2218157edd7aSVaclav Hapla   In this case, user is responsible to provide correct offsets so that the resulting leaves are unique (otherwise an error occurs).
2219157edd7aSVaclav Hapla 
2220157edd7aSVaclav Hapla .seealso: PetscSF, PetscSFCompose(), PetscSFGetGraph(), PetscSFSetGraph()
2221157edd7aSVaclav Hapla @*/
2222157edd7aSVaclav Hapla PetscErrorCode PetscSFConcatenate(MPI_Comm comm, PetscInt nsfs, PetscSF sfs[], PetscBool shareRoots, PetscInt leafOffsets[], PetscSF *newsf)
2223157edd7aSVaclav Hapla {
2224157edd7aSVaclav Hapla   PetscInt            i, s, nLeaves, nRoots;
2225157edd7aSVaclav Hapla   PetscInt           *leafArrayOffsets;
2226157edd7aSVaclav Hapla   PetscInt           *ilocal_new;
2227157edd7aSVaclav Hapla   PetscSFNode        *iremote_new;
2228157edd7aSVaclav Hapla   PetscInt           *rootOffsets;
2229157edd7aSVaclav Hapla   PetscBool           all_ilocal_null = PETSC_FALSE;
2230157edd7aSVaclav Hapla   PetscMPIInt         rank;
2231157edd7aSVaclav Hapla 
2232157edd7aSVaclav Hapla   PetscFunctionBegin;
2233157edd7aSVaclav Hapla   {
2234157edd7aSVaclav Hapla     PetscSF dummy; /* just to have a PetscObject on comm for input validation */
2235157edd7aSVaclav Hapla 
22369566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(comm,&dummy));
2237157edd7aSVaclav Hapla     PetscValidLogicalCollectiveInt(dummy,nsfs,2);
2238157edd7aSVaclav Hapla     PetscValidPointer(sfs,3);
2239157edd7aSVaclav Hapla     for (i=0; i<nsfs; i++) {
2240157edd7aSVaclav Hapla       PetscValidHeaderSpecific(sfs[i],PETSCSF_CLASSID,3);
2241157edd7aSVaclav Hapla       PetscCheckSameComm(dummy,1,sfs[i],3);
2242157edd7aSVaclav Hapla     }
2243157edd7aSVaclav Hapla     PetscValidLogicalCollectiveBool(dummy,shareRoots,4);
2244157edd7aSVaclav Hapla     if (leafOffsets) PetscValidIntPointer(leafOffsets,5);
2245157edd7aSVaclav Hapla     PetscValidPointer(newsf,6);
22469566063dSJacob Faibussowitsch     PetscCall(PetscSFDestroy(&dummy));
2247157edd7aSVaclav Hapla   }
2248157edd7aSVaclav Hapla   if (!nsfs) {
22499566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(comm, newsf));
22509566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(*newsf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER));
2251157edd7aSVaclav Hapla     PetscFunctionReturn(0);
2252157edd7aSVaclav Hapla   }
22539566063dSJacob Faibussowitsch   PetscCallMPI(MPI_Comm_rank(comm, &rank));
2254157edd7aSVaclav Hapla 
22559566063dSJacob Faibussowitsch   PetscCall(PetscCalloc1(nsfs+1, &rootOffsets));
2256157edd7aSVaclav Hapla   if (shareRoots) {
22579566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sfs[0], &nRoots, NULL, NULL, NULL));
2258157edd7aSVaclav Hapla     if (PetscDefined(USE_DEBUG)) {
2259157edd7aSVaclav Hapla       for (s=1; s<nsfs; s++) {
2260157edd7aSVaclav Hapla         PetscInt nr;
2261157edd7aSVaclav Hapla 
22629566063dSJacob Faibussowitsch         PetscCall(PetscSFGetGraph(sfs[s], &nr, NULL, NULL, NULL));
2263157edd7aSVaclav Hapla         PetscCheck(nr == nRoots, comm, PETSC_ERR_ARG_SIZ, "shareRoots = PETSC_TRUE but sfs[%" PetscInt_FMT "] has a different number of roots (%" PetscInt_FMT ") than sfs[0] (%" PetscInt_FMT ")", s, nr, nRoots);
2264157edd7aSVaclav Hapla       }
2265157edd7aSVaclav Hapla     }
2266157edd7aSVaclav Hapla   } else {
2267157edd7aSVaclav Hapla     for (s=0; s<nsfs; s++) {
2268157edd7aSVaclav Hapla       PetscInt nr;
2269157edd7aSVaclav Hapla 
22709566063dSJacob Faibussowitsch       PetscCall(PetscSFGetGraph(sfs[s], &nr, NULL, NULL, NULL));
2271157edd7aSVaclav Hapla       rootOffsets[s+1] = rootOffsets[s] + nr;
2272157edd7aSVaclav Hapla     }
2273157edd7aSVaclav Hapla     nRoots = rootOffsets[nsfs];
2274157edd7aSVaclav Hapla   }
2275157edd7aSVaclav Hapla 
2276157edd7aSVaclav Hapla   /* Calculate leaf array offsets and automatic root offsets */
22779566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(nsfs+1,&leafArrayOffsets));
2278157edd7aSVaclav Hapla   leafArrayOffsets[0] = 0;
2279157edd7aSVaclav Hapla   for (s=0; s<nsfs; s++) {
2280157edd7aSVaclav Hapla     PetscInt        nl;
2281157edd7aSVaclav Hapla 
22829566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sfs[s], NULL, &nl, NULL, NULL));
2283157edd7aSVaclav Hapla     leafArrayOffsets[s+1] = leafArrayOffsets[s] + nl;
2284157edd7aSVaclav Hapla   }
2285157edd7aSVaclav Hapla   nLeaves = leafArrayOffsets[nsfs];
2286157edd7aSVaclav Hapla 
2287157edd7aSVaclav Hapla   if (!leafOffsets) {
2288157edd7aSVaclav Hapla     all_ilocal_null = PETSC_TRUE;
2289157edd7aSVaclav Hapla     for (s=0; s<nsfs; s++) {
2290157edd7aSVaclav Hapla       const PetscInt *ilocal;
2291157edd7aSVaclav Hapla 
22929566063dSJacob Faibussowitsch       PetscCall(PetscSFGetGraph(sfs[s], NULL, NULL, &ilocal, NULL));
2293157edd7aSVaclav Hapla       if (ilocal) {
2294157edd7aSVaclav Hapla         all_ilocal_null = PETSC_FALSE;
2295157edd7aSVaclav Hapla         break;
2296157edd7aSVaclav Hapla       }
2297157edd7aSVaclav Hapla     }
2298157edd7aSVaclav 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");
2299157edd7aSVaclav Hapla   }
2300157edd7aSVaclav Hapla 
2301157edd7aSVaclav Hapla   /* Renumber and concatenate local leaves */
2302157edd7aSVaclav Hapla   ilocal_new = NULL;
2303157edd7aSVaclav Hapla   if (!all_ilocal_null) {
23049566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nLeaves, &ilocal_new));
2305157edd7aSVaclav Hapla     for (i = 0; i<nLeaves; i++) ilocal_new[i] = -1;
2306157edd7aSVaclav Hapla     for (s = 0; s<nsfs; s++) {
2307157edd7aSVaclav Hapla       const PetscInt   *ilocal;
2308157edd7aSVaclav Hapla       PetscInt         *ilocal_l = &ilocal_new[leafArrayOffsets[s]];
2309157edd7aSVaclav Hapla       PetscInt          i, nleaves_l;
2310157edd7aSVaclav Hapla 
23119566063dSJacob Faibussowitsch       PetscCall(PetscSFGetGraph(sfs[s], NULL, &nleaves_l, &ilocal, NULL));
2312157edd7aSVaclav Hapla       for (i=0; i<nleaves_l; i++) ilocal_l[i] = (ilocal ? ilocal[i] : i) + leafOffsets[s];
2313157edd7aSVaclav Hapla     }
2314157edd7aSVaclav Hapla   }
2315157edd7aSVaclav Hapla 
2316157edd7aSVaclav Hapla   /* Renumber and concatenate remote roots */
23179566063dSJacob Faibussowitsch   PetscCall(PetscMalloc1(nLeaves, &iremote_new));
2318157edd7aSVaclav Hapla   for (i = 0; i < nLeaves; i++) {
2319157edd7aSVaclav Hapla     iremote_new[i].rank   = -1;
2320157edd7aSVaclav Hapla     iremote_new[i].index  = -1;
2321157edd7aSVaclav Hapla   }
2322157edd7aSVaclav Hapla   for (s = 0; s<nsfs; s++) {
2323157edd7aSVaclav Hapla     PetscInt            i, nl, nr;
2324157edd7aSVaclav Hapla     PetscSF             tmp_sf;
2325157edd7aSVaclav Hapla     const PetscSFNode  *iremote;
2326157edd7aSVaclav Hapla     PetscSFNode        *tmp_rootdata;
2327157edd7aSVaclav Hapla     PetscSFNode        *tmp_leafdata = &iremote_new[leafArrayOffsets[s]];
2328157edd7aSVaclav Hapla 
23299566063dSJacob Faibussowitsch     PetscCall(PetscSFGetGraph(sfs[s], &nr, &nl, NULL, &iremote));
23309566063dSJacob Faibussowitsch     PetscCall(PetscSFCreate(comm, &tmp_sf));
2331157edd7aSVaclav Hapla     /* create helper SF with contiguous leaves */
23329566063dSJacob Faibussowitsch     PetscCall(PetscSFSetGraph(tmp_sf, nr, nl, NULL, PETSC_USE_POINTER, (PetscSFNode*) iremote, PETSC_COPY_VALUES));
23339566063dSJacob Faibussowitsch     PetscCall(PetscSFSetUp(tmp_sf));
23349566063dSJacob Faibussowitsch     PetscCall(PetscMalloc1(nr, &tmp_rootdata));
2335157edd7aSVaclav Hapla     for (i = 0; i < nr; i++) {
2336157edd7aSVaclav Hapla       tmp_rootdata[i].index = i + rootOffsets[s];
2337157edd7aSVaclav Hapla       tmp_rootdata[i].rank  = (PetscInt) rank;
2338157edd7aSVaclav Hapla     }
23399566063dSJacob Faibussowitsch     PetscCall(PetscSFBcastBegin(tmp_sf, MPIU_2INT, tmp_rootdata, tmp_leafdata, MPI_REPLACE));
23409566063dSJacob Faibussowitsch     PetscCall(PetscSFBcastEnd(  tmp_sf, MPIU_2INT, tmp_rootdata, tmp_leafdata, MPI_REPLACE));
23419566063dSJacob Faibussowitsch     PetscCall(PetscSFDestroy(&tmp_sf));
23429566063dSJacob Faibussowitsch     PetscCall(PetscFree(tmp_rootdata));
2343157edd7aSVaclav Hapla   }
2344157edd7aSVaclav Hapla 
2345157edd7aSVaclav Hapla   /* Build the new SF */
23469566063dSJacob Faibussowitsch   PetscCall(PetscSFCreate(comm, newsf));
23479566063dSJacob Faibussowitsch   PetscCall(PetscSFSetGraph(*newsf, nRoots, nLeaves, ilocal_new, PETSC_OWN_POINTER, iremote_new, PETSC_OWN_POINTER));
23489566063dSJacob Faibussowitsch   PetscCall(PetscSFSetUp(*newsf));
23499566063dSJacob Faibussowitsch   PetscCall(PetscFree(rootOffsets));
23509566063dSJacob Faibussowitsch   PetscCall(PetscFree(leafArrayOffsets));
2351157edd7aSVaclav Hapla   PetscFunctionReturn(0);
2352157edd7aSVaclav Hapla }
2353