xref: /petsc/src/vec/is/sf/interface/sf.c (revision c06a8e02ff6b77f4014c18ceca68570eef2eedb9)
1af0996ceSBarry Smith #include <petsc/private/sfimpl.h> /*I "petscsf.h" I*/
2c4e6a40aSLawrence Mitchell #include <petsc/private/hashseti.h>
395fce210SBarry Smith #include <petscctable.h>
495fce210SBarry Smith 
595fce210SBarry Smith #if defined(PETSC_USE_DEBUG)
695fce210SBarry Smith #  define PetscSFCheckGraphSet(sf,arg) do {                          \
795fce210SBarry Smith     if (PetscUnlikely(!(sf)->graphset))                              \
8dd5b3ca6SJunchao Zhang       SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must call PetscSFSetGraph() or PetscSFSetGraphWithPattern() on argument %D \"%s\" before %s()",(arg),#sf,PETSC_FUNCTION_NAME); \
995fce210SBarry Smith   } while (0)
1095fce210SBarry Smith #else
1195fce210SBarry Smith #  define PetscSFCheckGraphSet(sf,arg) do {} while (0)
1295fce210SBarry Smith #endif
1395fce210SBarry Smith 
1495fce210SBarry Smith const char *const PetscSFDuplicateOptions[] = {"CONFONLY","RANKS","GRAPH","PetscSFDuplicateOption","PETSCSF_DUPLICATE_",0};
1595fce210SBarry Smith 
168af6ec1cSBarry Smith /*@
1795fce210SBarry Smith    PetscSFCreate - create a star forest communication context
1895fce210SBarry Smith 
19d083f849SBarry Smith    Collective
2095fce210SBarry Smith 
2195fce210SBarry Smith    Input Arguments:
2295fce210SBarry Smith .  comm - communicator on which the star forest will operate
2395fce210SBarry Smith 
2495fce210SBarry Smith    Output Arguments:
2595fce210SBarry Smith .  sf - new star forest context
2695fce210SBarry Smith 
27dd5b3ca6SJunchao Zhang    Options Database Keys:
28dd5b3ca6SJunchao Zhang +  -sf_type basic     -Use MPI persistent Isend/Irecv for communication (Default)
29dd5b3ca6SJunchao Zhang .  -sf_type window    -Use MPI-3 one-sided window for communication
30dd5b3ca6SJunchao Zhang -  -sf_type neighbor  -Use MPI-3 neighborhood collectives for communication
31dd5b3ca6SJunchao Zhang 
3295fce210SBarry Smith    Level: intermediate
3395fce210SBarry Smith 
34dd5b3ca6SJunchao Zhang    Notes:
35dd5b3ca6SJunchao Zhang    When one knows the communication graph is one of the predefined graph, such as MPI_Alltoall, MPI_Allgatherv,
36dd5b3ca6SJunchao Zhang    MPI_Gatherv, one can create a PetscSF and then set its graph with PetscSFSetGraphWithPattern(). These special
37dd5b3ca6SJunchao Zhang    SFs are optimized and they have better performance than general SFs.
38dd5b3ca6SJunchao Zhang 
39dd5b3ca6SJunchao Zhang .seealso: PetscSFSetGraph(), PetscSFSetGraphWithPattern(), PetscSFDestroy()
4095fce210SBarry Smith @*/
4195fce210SBarry Smith PetscErrorCode PetscSFCreate(MPI_Comm comm,PetscSF *sf)
4295fce210SBarry Smith {
4395fce210SBarry Smith   PetscErrorCode ierr;
4495fce210SBarry Smith   PetscSF        b;
4595fce210SBarry Smith 
4695fce210SBarry Smith   PetscFunctionBegin;
4795fce210SBarry Smith   PetscValidPointer(sf,2);
48607a6623SBarry Smith   ierr = PetscSFInitializePackage();CHKERRQ(ierr);
4995fce210SBarry Smith 
5073107ff1SLisandro Dalcin   ierr = PetscHeaderCreate(b,PETSCSF_CLASSID,"PetscSF","Star Forest","PetscSF",comm,PetscSFDestroy,PetscSFView);CHKERRQ(ierr);
5195fce210SBarry Smith 
5295fce210SBarry Smith   b->nroots    = -1;
5395fce210SBarry Smith   b->nleaves   = -1;
5429046d53SLisandro Dalcin   b->minleaf   = PETSC_MAX_INT;
5529046d53SLisandro Dalcin   b->maxleaf   = PETSC_MIN_INT;
5695fce210SBarry Smith   b->nranks    = -1;
5795fce210SBarry Smith   b->rankorder = PETSC_TRUE;
5895fce210SBarry Smith   b->ingroup   = MPI_GROUP_NULL;
5995fce210SBarry Smith   b->outgroup  = MPI_GROUP_NULL;
6095fce210SBarry Smith   b->graphset  = PETSC_FALSE;
6195fce210SBarry Smith 
6295fce210SBarry Smith   *sf = b;
6395fce210SBarry Smith   PetscFunctionReturn(0);
6495fce210SBarry Smith }
6595fce210SBarry Smith 
6629046d53SLisandro Dalcin /*@
6795fce210SBarry Smith    PetscSFReset - Reset a star forest so that different sizes or neighbors can be used
6895fce210SBarry Smith 
6995fce210SBarry Smith    Collective
7095fce210SBarry Smith 
7195fce210SBarry Smith    Input Arguments:
7295fce210SBarry Smith .  sf - star forest
7395fce210SBarry Smith 
7495fce210SBarry Smith    Level: advanced
7595fce210SBarry Smith 
7695fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFSetGraph(), PetscSFDestroy()
7795fce210SBarry Smith @*/
7895fce210SBarry Smith PetscErrorCode PetscSFReset(PetscSF sf)
7995fce210SBarry Smith {
8095fce210SBarry Smith   PetscErrorCode ierr;
8195fce210SBarry Smith 
8295fce210SBarry Smith   PetscFunctionBegin;
8395fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
8479715d56SJed Brown   if (sf->ops->Reset) {ierr = (*sf->ops->Reset)(sf);CHKERRQ(ierr);}
8529046d53SLisandro Dalcin   sf->nroots   = -1;
8629046d53SLisandro Dalcin   sf->nleaves  = -1;
8729046d53SLisandro Dalcin   sf->minleaf  = PETSC_MAX_INT;
8829046d53SLisandro Dalcin   sf->maxleaf  = PETSC_MIN_INT;
8995fce210SBarry Smith   sf->mine     = NULL;
9095fce210SBarry Smith   sf->remote   = NULL;
9129046d53SLisandro Dalcin   sf->graphset = PETSC_FALSE;
9229046d53SLisandro Dalcin   ierr = PetscFree(sf->mine_alloc);CHKERRQ(ierr);
9395fce210SBarry Smith   ierr = PetscFree(sf->remote_alloc);CHKERRQ(ierr);
9421c688dcSJed Brown   sf->nranks = -1;
9529046d53SLisandro Dalcin   ierr = PetscFree4(sf->ranks,sf->roffset,sf->rmine,sf->rremote);CHKERRQ(ierr);
96eb02082bSJunchao Zhang #if defined(PETSC_HAVE_CUDA)
97cd620004SJunchao Zhang   {
98cd620004SJunchao Zhang   PetscInt  i;
99cd620004SJunchao Zhang   for (i=0; i<2; i++) {if (sf->rmine_d[i]) {cudaError_t err = cudaFree(sf->rmine_d[i]);CHKERRCUDA(err);sf->rmine_d[i]=NULL;}}
100cd620004SJunchao Zhang   }
101eb02082bSJunchao Zhang #endif
10229046d53SLisandro Dalcin   sf->degreeknown = PETSC_FALSE;
10395fce210SBarry Smith   ierr = PetscFree(sf->degree);CHKERRQ(ierr);
10495fce210SBarry Smith   if (sf->ingroup  != MPI_GROUP_NULL) {ierr = MPI_Group_free(&sf->ingroup);CHKERRQ(ierr);}
10595fce210SBarry Smith   if (sf->outgroup != MPI_GROUP_NULL) {ierr = MPI_Group_free(&sf->outgroup);CHKERRQ(ierr);}
10695fce210SBarry Smith   ierr = PetscSFDestroy(&sf->multi);CHKERRQ(ierr);
107dd5b3ca6SJunchao Zhang   ierr = PetscLayoutDestroy(&sf->map);CHKERRQ(ierr);
10895fce210SBarry Smith   sf->setupcalled = PETSC_FALSE;
10995fce210SBarry Smith   PetscFunctionReturn(0);
11095fce210SBarry Smith }
11195fce210SBarry Smith 
11295fce210SBarry Smith /*@C
11329046d53SLisandro Dalcin    PetscSFSetType - Set the PetscSF communication implementation
11495fce210SBarry Smith 
11595fce210SBarry Smith    Collective on PetscSF
11695fce210SBarry Smith 
11795fce210SBarry Smith    Input Parameters:
11895fce210SBarry Smith +  sf - the PetscSF context
11995fce210SBarry Smith -  type - a known method
12095fce210SBarry Smith 
12195fce210SBarry Smith    Options Database Key:
12295fce210SBarry Smith .  -sf_type <type> - Sets the method; use -help for a list
12370616304SStefano Zampini    of available methods (for instance, window, basic, neighbor)
12495fce210SBarry Smith 
12595fce210SBarry Smith    Notes:
12695fce210SBarry Smith    See "include/petscsf.h" for available methods (for instance)
12795fce210SBarry Smith +    PETSCSFWINDOW - MPI-2/3 one-sided
12895fce210SBarry Smith -    PETSCSFBASIC - basic implementation using MPI-1 two-sided
12995fce210SBarry Smith 
13095fce210SBarry Smith   Level: intermediate
13195fce210SBarry Smith 
13295fce210SBarry Smith .seealso: PetscSFType, PetscSFCreate()
13395fce210SBarry Smith @*/
13495fce210SBarry Smith PetscErrorCode PetscSFSetType(PetscSF sf,PetscSFType type)
13595fce210SBarry Smith {
13695fce210SBarry Smith   PetscErrorCode ierr,(*r)(PetscSF);
13795fce210SBarry Smith   PetscBool      match;
13895fce210SBarry Smith 
13995fce210SBarry Smith   PetscFunctionBegin;
14095fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
14195fce210SBarry Smith   PetscValidCharPointer(type,2);
14295fce210SBarry Smith 
14395fce210SBarry Smith   ierr = PetscObjectTypeCompare((PetscObject)sf,type,&match);CHKERRQ(ierr);
14495fce210SBarry Smith   if (match) PetscFunctionReturn(0);
14595fce210SBarry Smith 
146adc40e5bSBarry Smith   ierr = PetscFunctionListFind(PetscSFList,type,&r);CHKERRQ(ierr);
14795fce210SBarry Smith   if (!r) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_UNKNOWN_TYPE,"Unable to find requested PetscSF type %s",type);
14829046d53SLisandro Dalcin   /* Destroy the previous PetscSF implementation context */
14929046d53SLisandro Dalcin   if (sf->ops->Destroy) {ierr = (*(sf)->ops->Destroy)(sf);CHKERRQ(ierr);}
15095fce210SBarry Smith   ierr = PetscMemzero(sf->ops,sizeof(*sf->ops));CHKERRQ(ierr);
15195fce210SBarry Smith   ierr = PetscObjectChangeTypeName((PetscObject)sf,type);CHKERRQ(ierr);
15295fce210SBarry Smith   ierr = (*r)(sf);CHKERRQ(ierr);
15395fce210SBarry Smith   PetscFunctionReturn(0);
15495fce210SBarry Smith }
15595fce210SBarry Smith 
15629046d53SLisandro Dalcin /*@C
15729046d53SLisandro Dalcin   PetscSFGetType - Get the PetscSF communication implementation
15829046d53SLisandro Dalcin 
15929046d53SLisandro Dalcin   Not Collective
16029046d53SLisandro Dalcin 
16129046d53SLisandro Dalcin   Input Parameter:
16229046d53SLisandro Dalcin . sf  - the PetscSF context
16329046d53SLisandro Dalcin 
16429046d53SLisandro Dalcin   Output Parameter:
16529046d53SLisandro Dalcin . type - the PetscSF type name
16629046d53SLisandro Dalcin 
16729046d53SLisandro Dalcin   Level: intermediate
16829046d53SLisandro Dalcin 
16929046d53SLisandro Dalcin .seealso: PetscSFSetType(), PetscSFCreate()
17029046d53SLisandro Dalcin @*/
17129046d53SLisandro Dalcin PetscErrorCode PetscSFGetType(PetscSF sf, PetscSFType *type)
17229046d53SLisandro Dalcin {
17329046d53SLisandro Dalcin   PetscFunctionBegin;
17429046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf, PETSCSF_CLASSID,1);
17529046d53SLisandro Dalcin   PetscValidPointer(type,2);
17629046d53SLisandro Dalcin   *type = ((PetscObject)sf)->type_name;
17729046d53SLisandro Dalcin   PetscFunctionReturn(0);
17829046d53SLisandro Dalcin }
17929046d53SLisandro Dalcin 
180d36d33e4SMatthew G. Knepley /*@
18195fce210SBarry Smith    PetscSFDestroy - destroy star forest
18295fce210SBarry Smith 
18395fce210SBarry Smith    Collective
18495fce210SBarry Smith 
18595fce210SBarry Smith    Input Arguments:
18695fce210SBarry Smith .  sf - address of star forest
18795fce210SBarry Smith 
18895fce210SBarry Smith    Level: intermediate
18995fce210SBarry Smith 
19095fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFReset()
19195fce210SBarry Smith @*/
19295fce210SBarry Smith PetscErrorCode PetscSFDestroy(PetscSF *sf)
19395fce210SBarry Smith {
19495fce210SBarry Smith   PetscErrorCode ierr;
19595fce210SBarry Smith 
19695fce210SBarry Smith   PetscFunctionBegin;
19795fce210SBarry Smith   if (!*sf) PetscFunctionReturn(0);
19895fce210SBarry Smith   PetscValidHeaderSpecific((*sf),PETSCSF_CLASSID,1);
19929046d53SLisandro Dalcin   if (--((PetscObject)(*sf))->refct > 0) {*sf = NULL; PetscFunctionReturn(0);}
20095fce210SBarry Smith   ierr = PetscSFReset(*sf);CHKERRQ(ierr);
20195fce210SBarry Smith   if ((*sf)->ops->Destroy) {ierr = (*(*sf)->ops->Destroy)(*sf);CHKERRQ(ierr);}
20295fce210SBarry Smith   ierr = PetscHeaderDestroy(sf);CHKERRQ(ierr);
20395fce210SBarry Smith   PetscFunctionReturn(0);
20495fce210SBarry Smith }
20595fce210SBarry Smith 
206c4e6a40aSLawrence Mitchell static PetscErrorCode PetscSFCheckGraphValid_Private(PetscSF sf)
207c4e6a40aSLawrence Mitchell {
208c4e6a40aSLawrence Mitchell #if defined(PETSC_USE_DEBUG)
209c4e6a40aSLawrence Mitchell   PetscInt           i, nleaves;
210c4e6a40aSLawrence Mitchell   PetscMPIInt        size;
211c4e6a40aSLawrence Mitchell   const PetscInt    *ilocal;
212c4e6a40aSLawrence Mitchell   const PetscSFNode *iremote;
213c4e6a40aSLawrence Mitchell   PetscErrorCode     ierr;
214c4e6a40aSLawrence Mitchell 
215c4e6a40aSLawrence Mitchell   PetscFunctionBegin;
216c4e6a40aSLawrence Mitchell   if (!sf->graphset) PetscFunctionReturn(0);
217c4e6a40aSLawrence Mitchell   ierr = PetscSFGetGraph(sf,NULL,&nleaves,&ilocal,&iremote);CHKERRQ(ierr);
218c4e6a40aSLawrence Mitchell   ierr = MPI_Comm_size(PetscObjectComm((PetscObject)sf),&size);CHKERRQ(ierr);
219c4e6a40aSLawrence Mitchell   for (i = 0; i < nleaves; i++) {
220c4e6a40aSLawrence Mitchell     const PetscInt rank = iremote[i].rank;
221c4e6a40aSLawrence Mitchell     const PetscInt remote = iremote[i].index;
222c4e6a40aSLawrence Mitchell     const PetscInt leaf = ilocal ? ilocal[i] : i;
223c4e6a40aSLawrence Mitchell     if (rank < 0 || rank >= size) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Provided rank (%D) for remote %D is invalid, should be in [0, %d)",rank,i,size);
224c4e6a40aSLawrence Mitchell     if (remote < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Provided index (%D) for remote %D is invalid, should be >= 0",remote,i);
225c4e6a40aSLawrence Mitchell     if (leaf < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Provided location (%D) for leaf %D is invalid, should be >= 0",leaf,i);
226c4e6a40aSLawrence Mitchell   }
227c4e6a40aSLawrence Mitchell   PetscFunctionReturn(0);
228c4e6a40aSLawrence Mitchell #else
229c4e6a40aSLawrence Mitchell   PetscFunctionBegin;
230c4e6a40aSLawrence Mitchell   PetscFunctionReturn(0);
231c4e6a40aSLawrence Mitchell #endif
232c4e6a40aSLawrence Mitchell }
233c4e6a40aSLawrence Mitchell 
23495fce210SBarry Smith /*@
23595fce210SBarry Smith    PetscSFSetUp - set up communication structures
23695fce210SBarry Smith 
23795fce210SBarry Smith    Collective
23895fce210SBarry Smith 
23995fce210SBarry Smith    Input Arguments:
24095fce210SBarry Smith .  sf - star forest communication object
24195fce210SBarry Smith 
24295fce210SBarry Smith    Level: beginner
24395fce210SBarry Smith 
24495fce210SBarry Smith .seealso: PetscSFSetFromOptions(), PetscSFSetType()
24595fce210SBarry Smith @*/
24695fce210SBarry Smith PetscErrorCode PetscSFSetUp(PetscSF sf)
24795fce210SBarry Smith {
24895fce210SBarry Smith   PetscErrorCode ierr;
24995fce210SBarry Smith 
25095fce210SBarry Smith   PetscFunctionBegin;
25129046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
25229046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
25395fce210SBarry Smith   if (sf->setupcalled) PetscFunctionReturn(0);
254c4e6a40aSLawrence Mitchell   ierr = PetscSFCheckGraphValid_Private(sf);CHKERRQ(ierr);
255c2a741eeSJunchao Zhang   sf->use_gpu_aware_mpi = use_gpu_aware_mpi;
25695fce210SBarry Smith   if (!((PetscObject)sf)->type_name) {ierr = PetscSFSetType(sf,PETSCSFBASIC);CHKERRQ(ierr);}
25729046d53SLisandro Dalcin   ierr = PetscLogEventBegin(PETSCSF_SetUp,sf,0,0,0);CHKERRQ(ierr);
25895fce210SBarry Smith   if (sf->ops->SetUp) {ierr = (*sf->ops->SetUp)(sf);CHKERRQ(ierr);}
25929046d53SLisandro Dalcin   ierr = PetscLogEventEnd(PETSCSF_SetUp,sf,0,0,0);CHKERRQ(ierr);
26095fce210SBarry Smith   sf->setupcalled = PETSC_TRUE;
26195fce210SBarry Smith   PetscFunctionReturn(0);
26295fce210SBarry Smith }
26395fce210SBarry Smith 
2648af6ec1cSBarry Smith /*@
26595fce210SBarry Smith    PetscSFSetFromOptions - set PetscSF options using the options database
26695fce210SBarry Smith 
26795fce210SBarry Smith    Logically Collective
26895fce210SBarry Smith 
26995fce210SBarry Smith    Input Arguments:
27095fce210SBarry Smith .  sf - star forest
27195fce210SBarry Smith 
27295fce210SBarry Smith    Options Database Keys:
27360263706SJed Brown +  -sf_type               - implementation type, see PetscSFSetType()
27451ccb202SJunchao Zhang .  -sf_rank_order         - sort composite points for gathers and scatters in rank order, gathers are non-deterministic otherwise
275b85e67b7SJunchao Zhang .  -sf_use_default_stream - Assume callers of SF computed the input root/leafdata with the default cuda stream. SF will also
276c2a741eeSJunchao Zhang                             use the default stream to process data. Therefore, no stream synchronization is needed between SF and its caller (default: true).
277*c06a8e02SRichard Tran Mills                             If true, this option only works with -use_gpu_aware_mpi 1.
278c2a741eeSJunchao 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).
279*c06a8e02SRichard Tran Mills                                If true, this option only works with -use_gpu_aware_mpi 1.
28095fce210SBarry Smith 
28195fce210SBarry Smith    Level: intermediate
28295fce210SBarry Smith @*/
28395fce210SBarry Smith PetscErrorCode PetscSFSetFromOptions(PetscSF sf)
28495fce210SBarry Smith {
28595fce210SBarry Smith   PetscSFType    deft;
28695fce210SBarry Smith   char           type[256];
28795fce210SBarry Smith   PetscErrorCode ierr;
28895fce210SBarry Smith   PetscBool      flg;
28995fce210SBarry Smith 
29095fce210SBarry Smith   PetscFunctionBegin;
29195fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
29295fce210SBarry Smith   ierr = PetscObjectOptionsBegin((PetscObject)sf);CHKERRQ(ierr);
29395fce210SBarry Smith   deft = ((PetscObject)sf)->type_name ? ((PetscObject)sf)->type_name : PETSCSFBASIC;
29429046d53SLisandro Dalcin   ierr = PetscOptionsFList("-sf_type","PetscSF implementation type","PetscSFSetType",PetscSFList,deft,type,sizeof(type),&flg);CHKERRQ(ierr);
29595fce210SBarry Smith   ierr = PetscSFSetType(sf,flg ? type : deft);CHKERRQ(ierr);
29695fce210SBarry Smith   ierr = 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);CHKERRQ(ierr);
297c2a741eeSJunchao Zhang 
298c2a741eeSJunchao Zhang #if defined(PETSC_HAVE_CUDA)
299c2a741eeSJunchao Zhang   sf->use_default_stream = PETSC_TRUE; /* The assumption is true for PETSc internal use of SF */
300cd620004SJunchao Zhang   ierr = PetscOptionsBool("-sf_use_default_stream","Assume SF's input and output root/leafdata is computed on the default stream","PetscSFSetFromOptions",sf->use_default_stream,&sf->use_default_stream,NULL);CHKERRQ(ierr);
301b85e67b7SJunchao Zhang   sf->use_stream_aware_mpi = PETSC_FALSE;
302b85e67b7SJunchao Zhang   ierr = 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);CHKERRQ(ierr);
303c2a741eeSJunchao Zhang #endif
304e55864a3SBarry Smith   if (sf->ops->SetFromOptions) {ierr = (*sf->ops->SetFromOptions)(PetscOptionsObject,sf);CHKERRQ(ierr);}
30595fce210SBarry Smith   ierr = PetscOptionsEnd();CHKERRQ(ierr);
30695fce210SBarry Smith   PetscFunctionReturn(0);
30795fce210SBarry Smith }
30895fce210SBarry Smith 
30929046d53SLisandro Dalcin /*@
31095fce210SBarry Smith    PetscSFSetRankOrder - sort multi-points for gathers and scatters by rank order
31195fce210SBarry Smith 
31295fce210SBarry Smith    Logically Collective
31395fce210SBarry Smith 
31495fce210SBarry Smith    Input Arguments:
31595fce210SBarry Smith +  sf - star forest
31695fce210SBarry Smith -  flg - PETSC_TRUE to sort, PETSC_FALSE to skip sorting (lower setup cost, but non-deterministic)
31795fce210SBarry Smith 
31895fce210SBarry Smith    Level: advanced
31995fce210SBarry Smith 
32095fce210SBarry Smith .seealso: PetscSFGatherBegin(), PetscSFScatterBegin()
32195fce210SBarry Smith @*/
32295fce210SBarry Smith PetscErrorCode PetscSFSetRankOrder(PetscSF sf,PetscBool flg)
32395fce210SBarry Smith {
32495fce210SBarry Smith   PetscFunctionBegin;
32595fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
32695fce210SBarry Smith   PetscValidLogicalCollectiveBool(sf,flg,2);
32795fce210SBarry Smith   if (sf->multi) SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_ARG_WRONGSTATE,"Rank ordering must be set before first call to PetscSFGatherBegin() or PetscSFScatterBegin()");
32895fce210SBarry Smith   sf->rankorder = flg;
32995fce210SBarry Smith   PetscFunctionReturn(0);
33095fce210SBarry Smith }
33195fce210SBarry Smith 
3328af6ec1cSBarry Smith /*@
33395fce210SBarry Smith    PetscSFSetGraph - Set a parallel star forest
33495fce210SBarry Smith 
33595fce210SBarry Smith    Collective
33695fce210SBarry Smith 
33795fce210SBarry Smith    Input Arguments:
33895fce210SBarry Smith +  sf - star forest
33995fce210SBarry Smith .  nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves)
34095fce210SBarry Smith .  nleaves - number of leaf vertices on the current process, each of these references a root on any process
341c4e6a40aSLawrence Mitchell .  ilocal - locations of leaves in leafdata buffers, pass NULL for contiguous storage (locations must be >= 0, enforced
342c4e6a40aSLawrence Mitchell during setup in debug mode)
34395fce210SBarry Smith .  localmode - copy mode for ilocal
344c4e6a40aSLawrence Mitchell .  iremote - remote locations of root vertices for each leaf on the current process (locations must be >= 0, enforced
345c4e6a40aSLawrence Mitchell during setup in debug mode)
34695fce210SBarry Smith -  remotemode - copy mode for iremote
34795fce210SBarry Smith 
34895fce210SBarry Smith    Level: intermediate
34995fce210SBarry Smith 
35095452b02SPatrick Sanan    Notes:
35195452b02SPatrick Sanan     In Fortran you must use PETSC_COPY_VALUES for localmode and remotemode
35238ab3f8aSBarry Smith 
3532ad1e87fSLisandro Dalcin    Developers Note: Local indices which are the identity permutation in the range [0,nleaves) are discarded as they
3542ad1e87fSLisandro Dalcin    encode contiguous storage. In such case, if localmode is PETSC_OWN_POINTER, the memory is deallocated as it is not
3552ad1e87fSLisandro Dalcin    needed
3562ad1e87fSLisandro Dalcin 
357c4e6a40aSLawrence Mitchell    Developers Note: This object does not necessarily encode a true star forest in the graph theoretic sense, since leaf
358c4e6a40aSLawrence Mitchell    indices are not required to be unique. Some functions, however, rely on unique leaf indices (checked in debug mode).
359c4e6a40aSLawrence Mitchell 
36095fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFView(), PetscSFGetGraph()
36195fce210SBarry Smith @*/
36295fce210SBarry Smith PetscErrorCode PetscSFSetGraph(PetscSF sf,PetscInt nroots,PetscInt nleaves,const PetscInt *ilocal,PetscCopyMode localmode,const PetscSFNode *iremote,PetscCopyMode remotemode)
36395fce210SBarry Smith {
36495fce210SBarry Smith   PetscErrorCode ierr;
36595fce210SBarry Smith 
36695fce210SBarry Smith   PetscFunctionBegin;
36795fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
36829046d53SLisandro Dalcin   if (nleaves > 0 && ilocal) PetscValidIntPointer(ilocal,4);
36929046d53SLisandro Dalcin   if (nleaves > 0) PetscValidPointer(iremote,6);
37029046d53SLisandro Dalcin   if (nroots  < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nroots %D, cannot be negative",nroots);
37195fce210SBarry Smith   if (nleaves < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"nleaves %D, cannot be negative",nleaves);
37229046d53SLisandro Dalcin 
3732a67d2daSStefano Zampini   if (sf->nroots >= 0) { /* Reset only if graph already set */
37495fce210SBarry Smith     ierr = PetscSFReset(sf);CHKERRQ(ierr);
3752a67d2daSStefano Zampini   }
3762a67d2daSStefano Zampini 
37729046d53SLisandro Dalcin   ierr = PetscLogEventBegin(PETSCSF_SetGraph,sf,0,0,0);CHKERRQ(ierr);
37829046d53SLisandro Dalcin 
37995fce210SBarry Smith   sf->nroots  = nroots;
38095fce210SBarry Smith   sf->nleaves = nleaves;
38129046d53SLisandro Dalcin 
38229046d53SLisandro Dalcin   if (nleaves && ilocal) {
38321c688dcSJed Brown     PetscInt i;
38429046d53SLisandro Dalcin     PetscInt minleaf = PETSC_MAX_INT;
38529046d53SLisandro Dalcin     PetscInt maxleaf = PETSC_MIN_INT;
3862ad1e87fSLisandro Dalcin     int      contiguous = 1;
38729046d53SLisandro Dalcin     for (i=0; i<nleaves; i++) {
38829046d53SLisandro Dalcin       minleaf = PetscMin(minleaf,ilocal[i]);
38929046d53SLisandro Dalcin       maxleaf = PetscMax(maxleaf,ilocal[i]);
3902ad1e87fSLisandro Dalcin       contiguous &= (ilocal[i] == i);
39129046d53SLisandro Dalcin     }
39229046d53SLisandro Dalcin     sf->minleaf = minleaf;
39329046d53SLisandro Dalcin     sf->maxleaf = maxleaf;
3942ad1e87fSLisandro Dalcin     if (contiguous) {
3952ad1e87fSLisandro Dalcin       if (localmode == PETSC_OWN_POINTER) {
3962ad1e87fSLisandro Dalcin         ierr = PetscFree(ilocal);CHKERRQ(ierr);
3972ad1e87fSLisandro Dalcin       }
3982ad1e87fSLisandro Dalcin       ilocal = NULL;
3992ad1e87fSLisandro Dalcin     }
40029046d53SLisandro Dalcin   } else {
40129046d53SLisandro Dalcin     sf->minleaf = 0;
40229046d53SLisandro Dalcin     sf->maxleaf = nleaves - 1;
40329046d53SLisandro Dalcin   }
40429046d53SLisandro Dalcin 
40529046d53SLisandro Dalcin   if (ilocal) {
40695fce210SBarry Smith     switch (localmode) {
40795fce210SBarry Smith     case PETSC_COPY_VALUES:
408785e854fSJed Brown       ierr = PetscMalloc1(nleaves,&sf->mine_alloc);CHKERRQ(ierr);
409580bdb30SBarry Smith       ierr = PetscArraycpy(sf->mine_alloc,ilocal,nleaves);CHKERRQ(ierr);
41095fce210SBarry Smith       sf->mine = sf->mine_alloc;
41195fce210SBarry Smith       break;
41295fce210SBarry Smith     case PETSC_OWN_POINTER:
41395fce210SBarry Smith       sf->mine_alloc = (PetscInt*)ilocal;
41495fce210SBarry Smith       sf->mine       = sf->mine_alloc;
41595fce210SBarry Smith       break;
41695fce210SBarry Smith     case PETSC_USE_POINTER:
41729046d53SLisandro Dalcin       sf->mine_alloc = NULL;
41895fce210SBarry Smith       sf->mine       = (PetscInt*)ilocal;
41995fce210SBarry Smith       break;
42095fce210SBarry Smith     default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_ARG_OUTOFRANGE,"Unknown localmode");
42195fce210SBarry Smith     }
42295fce210SBarry Smith   }
42329046d53SLisandro Dalcin 
42495fce210SBarry Smith   switch (remotemode) {
42595fce210SBarry Smith   case PETSC_COPY_VALUES:
426785e854fSJed Brown     ierr = PetscMalloc1(nleaves,&sf->remote_alloc);CHKERRQ(ierr);
427580bdb30SBarry Smith     ierr = PetscArraycpy(sf->remote_alloc,iremote,nleaves);CHKERRQ(ierr);
42895fce210SBarry Smith     sf->remote = sf->remote_alloc;
42995fce210SBarry Smith     break;
43095fce210SBarry Smith   case PETSC_OWN_POINTER:
43195fce210SBarry Smith     sf->remote_alloc = (PetscSFNode*)iremote;
43295fce210SBarry Smith     sf->remote       = sf->remote_alloc;
43395fce210SBarry Smith     break;
43495fce210SBarry Smith   case PETSC_USE_POINTER:
43529046d53SLisandro Dalcin     sf->remote_alloc = NULL;
43695fce210SBarry Smith     sf->remote       = (PetscSFNode*)iremote;
43795fce210SBarry Smith     break;
43895fce210SBarry Smith   default: SETERRQ(PetscObjectComm((PetscObject)sf),PETSC_ERR_ARG_OUTOFRANGE,"Unknown remotemode");
43995fce210SBarry Smith   }
44095fce210SBarry Smith 
441563ffbabSMatthew G. Knepley   ierr = PetscLogEventEnd(PETSCSF_SetGraph,sf,0,0,0);CHKERRQ(ierr);
44229046d53SLisandro Dalcin   sf->graphset = PETSC_TRUE;
44395fce210SBarry Smith   PetscFunctionReturn(0);
44495fce210SBarry Smith }
44595fce210SBarry Smith 
44629046d53SLisandro Dalcin /*@
447dd5b3ca6SJunchao Zhang   PetscSFSetGraphWithPattern - Sets the graph of an SF with a specific pattern
448dd5b3ca6SJunchao Zhang 
449dd5b3ca6SJunchao Zhang   Collective
450dd5b3ca6SJunchao Zhang 
451dd5b3ca6SJunchao Zhang   Input Parameters:
452dd5b3ca6SJunchao Zhang + sf      - The PetscSF
453dd5b3ca6SJunchao Zhang . map     - Layout of roots over all processes (insignificant when pattern is PETSCSF_PATTERN_ALLTOALL)
454dd5b3ca6SJunchao Zhang - pattern - One of PETSCSF_PATTERN_ALLGATHER, PETSCSF_PATTERN_GATHER, PETSCSF_PATTERN_ALLTOALL
455dd5b3ca6SJunchao Zhang 
456dd5b3ca6SJunchao Zhang   Notes:
457dd5b3ca6SJunchao Zhang   It is easier to explain PetscSFPattern using vectors. Suppose we have an MPI vector x and its layout is map.
458dd5b3ca6SJunchao Zhang   n and N are local and global sizes of x respectively.
459dd5b3ca6SJunchao Zhang 
460dd5b3ca6SJunchao Zhang   With PETSCSF_PATTERN_ALLGATHER, the routine creates a graph that if one does Bcast on it, it will copy x to
461dd5b3ca6SJunchao Zhang   sequential vectors y on all ranks.
462dd5b3ca6SJunchao Zhang 
463dd5b3ca6SJunchao Zhang   With PETSCSF_PATTERN_GATHER, the routine creates a graph that if one does Bcast on it, it will copy x to a
464dd5b3ca6SJunchao Zhang   sequential vector y on rank 0.
465dd5b3ca6SJunchao Zhang 
466dd5b3ca6SJunchao Zhang   In above cases, entries of x are roots and entries of y are leaves.
467dd5b3ca6SJunchao Zhang 
468dd5b3ca6SJunchao Zhang   With PETSCSF_PATTERN_ALLTOALL, map is insignificant. Suppose NP is size of sf's communicator. The routine
469dd5b3ca6SJunchao Zhang   creates a graph that every rank has NP leaves and NP roots. On rank i, its leaf j is connected to root i
470dd5b3ca6SJunchao 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
471dd5b3ca6SJunchao Zhang   not mean one can not send multiple items. One just needs to create a new MPI datatype for the mulptiple data
472dd5b3ca6SJunchao Zhang   items with MPI_Type_contiguous() and use that as the <unit> argument in SF routines.
473dd5b3ca6SJunchao Zhang 
474dd5b3ca6SJunchao Zhang   In this case, roots and leaves are symmetric.
475dd5b3ca6SJunchao Zhang 
476dd5b3ca6SJunchao Zhang   Level: intermediate
477dd5b3ca6SJunchao Zhang  @*/
478dd5b3ca6SJunchao Zhang PetscErrorCode PetscSFSetGraphWithPattern(PetscSF sf,PetscLayout map,PetscSFPattern pattern)
479dd5b3ca6SJunchao Zhang {
480dd5b3ca6SJunchao Zhang   MPI_Comm       comm;
481dd5b3ca6SJunchao Zhang   PetscInt       n,N,res[2];
482dd5b3ca6SJunchao Zhang   PetscMPIInt    rank,size;
483dd5b3ca6SJunchao Zhang   PetscSFType    type;
484dd5b3ca6SJunchao Zhang   PetscErrorCode ierr;
485dd5b3ca6SJunchao Zhang 
486dd5b3ca6SJunchao Zhang   PetscFunctionBegin;
487dd5b3ca6SJunchao Zhang   ierr = PetscObjectGetComm((PetscObject)sf, &comm);CHKERRQ(ierr);
488dd5b3ca6SJunchao Zhang   if (pattern < PETSCSF_PATTERN_ALLGATHER || pattern > PETSCSF_PATTERN_ALLTOALL) SETERRQ1(comm,PETSC_ERR_ARG_OUTOFRANGE,"Unsupported PetscSFPattern %D\n",pattern);
489dd5b3ca6SJunchao Zhang   ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr);
490dd5b3ca6SJunchao Zhang   ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr);
491dd5b3ca6SJunchao Zhang 
492dd5b3ca6SJunchao Zhang   if (pattern == PETSCSF_PATTERN_ALLTOALL) {
493dd5b3ca6SJunchao Zhang     type = PETSCSFALLTOALL;
494dd5b3ca6SJunchao Zhang     ierr = PetscLayoutCreate(comm,&sf->map);CHKERRQ(ierr);
495dd5b3ca6SJunchao Zhang     ierr = PetscLayoutSetLocalSize(sf->map,size);CHKERRQ(ierr);
496dd5b3ca6SJunchao Zhang     ierr = PetscLayoutSetSize(sf->map,((PetscInt)size)*size);CHKERRQ(ierr);
497dd5b3ca6SJunchao Zhang     ierr = PetscLayoutSetUp(sf->map);CHKERRQ(ierr);
498dd5b3ca6SJunchao Zhang   } else {
499dd5b3ca6SJunchao Zhang     ierr   = PetscLayoutGetLocalSize(map,&n);CHKERRQ(ierr);
500dd5b3ca6SJunchao Zhang     ierr   = PetscLayoutGetSize(map,&N);CHKERRQ(ierr);
501dd5b3ca6SJunchao Zhang     res[0] = n;
502dd5b3ca6SJunchao Zhang     res[1] = -n;
503dd5b3ca6SJunchao Zhang     /* Check if n are same over all ranks so that we can optimize it */
504dd5b3ca6SJunchao Zhang     ierr   = MPIU_Allreduce(MPI_IN_PLACE,res,2,MPIU_INT,MPI_MAX,comm);CHKERRQ(ierr);
505dd5b3ca6SJunchao Zhang     if (res[0] == -res[1]) { /* same n */
506dd5b3ca6SJunchao Zhang       type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHER  : PETSCSFGATHER;
507dd5b3ca6SJunchao Zhang     } else {
508dd5b3ca6SJunchao Zhang       type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHERV : PETSCSFGATHERV;
509dd5b3ca6SJunchao Zhang     }
510dd5b3ca6SJunchao Zhang     ierr = PetscLayoutReference(map,&sf->map);CHKERRQ(ierr);
511dd5b3ca6SJunchao Zhang   }
512dd5b3ca6SJunchao Zhang   ierr = PetscSFSetType(sf,type);CHKERRQ(ierr);
513dd5b3ca6SJunchao Zhang 
514dd5b3ca6SJunchao Zhang   sf->pattern = pattern;
515dd5b3ca6SJunchao Zhang   sf->mine    = NULL; /* Contiguous */
516dd5b3ca6SJunchao Zhang 
517dd5b3ca6SJunchao Zhang   /* Set nleaves, nroots here in case user calls PetscSFGetGraph, which is legal to call even before PetscSFSetUp is called.
518dd5b3ca6SJunchao Zhang      Also set other easy stuff.
519dd5b3ca6SJunchao Zhang    */
520dd5b3ca6SJunchao Zhang   if (pattern == PETSCSF_PATTERN_ALLGATHER) {
521dd5b3ca6SJunchao Zhang     sf->nleaves      = N;
522dd5b3ca6SJunchao Zhang     sf->nroots       = n;
523dd5b3ca6SJunchao Zhang     sf->nranks       = size;
524dd5b3ca6SJunchao Zhang     sf->minleaf      = 0;
525dd5b3ca6SJunchao Zhang     sf->maxleaf      = N - 1;
526dd5b3ca6SJunchao Zhang   } else if (pattern == PETSCSF_PATTERN_GATHER) {
527dd5b3ca6SJunchao Zhang     sf->nleaves      = rank ? 0 : N;
528dd5b3ca6SJunchao Zhang     sf->nroots       = n;
529dd5b3ca6SJunchao Zhang     sf->nranks       = rank ? 0 : size;
530dd5b3ca6SJunchao Zhang     sf->minleaf      = 0;
531dd5b3ca6SJunchao Zhang     sf->maxleaf      = rank ? -1 : N - 1;
532dd5b3ca6SJunchao Zhang   } else if (pattern == PETSCSF_PATTERN_ALLTOALL) {
533dd5b3ca6SJunchao Zhang     sf->nleaves      = size;
534dd5b3ca6SJunchao Zhang     sf->nroots       = size;
535dd5b3ca6SJunchao Zhang     sf->nranks       = size;
536dd5b3ca6SJunchao Zhang     sf->minleaf      = 0;
537dd5b3ca6SJunchao Zhang     sf->maxleaf      = size - 1;
538dd5b3ca6SJunchao Zhang   }
539dd5b3ca6SJunchao Zhang   sf->ndranks  = 0; /* We do not need to separate out distinguished ranks for patterned graphs to improve communication performance */
540dd5b3ca6SJunchao Zhang   sf->graphset = PETSC_TRUE;
541dd5b3ca6SJunchao Zhang   PetscFunctionReturn(0);
542dd5b3ca6SJunchao Zhang }
543dd5b3ca6SJunchao Zhang 
544dd5b3ca6SJunchao Zhang /*@
54595fce210SBarry Smith    PetscSFCreateInverseSF - given a PetscSF in which all vertices have degree 1, creates the inverse map
54695fce210SBarry Smith 
54795fce210SBarry Smith    Collective
54895fce210SBarry Smith 
54995fce210SBarry Smith    Input Arguments:
550dd5b3ca6SJunchao Zhang 
55195fce210SBarry Smith .  sf - star forest to invert
55295fce210SBarry Smith 
55395fce210SBarry Smith    Output Arguments:
55495fce210SBarry Smith .  isf - inverse of sf
55595fce210SBarry Smith    Level: advanced
55695fce210SBarry Smith 
55795fce210SBarry Smith    Notes:
55895fce210SBarry Smith    All roots must have degree 1.
55995fce210SBarry Smith 
56095fce210SBarry Smith    The local space may be a permutation, but cannot be sparse.
56195fce210SBarry Smith 
56295fce210SBarry Smith .seealso: PetscSFSetGraph()
56395fce210SBarry Smith @*/
56495fce210SBarry Smith PetscErrorCode PetscSFCreateInverseSF(PetscSF sf,PetscSF *isf)
56595fce210SBarry Smith {
56695fce210SBarry Smith   PetscErrorCode ierr;
56795fce210SBarry Smith   PetscMPIInt    rank;
56895fce210SBarry Smith   PetscInt       i,nroots,nleaves,maxlocal,count,*newilocal;
56995fce210SBarry Smith   const PetscInt *ilocal;
57095fce210SBarry Smith   PetscSFNode    *roots,*leaves;
57195fce210SBarry Smith 
57295fce210SBarry Smith   PetscFunctionBegin;
57329046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
57429046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
57529046d53SLisandro Dalcin   PetscValidPointer(isf,2);
57629046d53SLisandro Dalcin 
57795fce210SBarry Smith   ierr = PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,NULL);CHKERRQ(ierr);
57829046d53SLisandro Dalcin   maxlocal = sf->maxleaf+1; /* TODO: We should use PetscSFGetLeafRange() */
57929046d53SLisandro Dalcin 
58029046d53SLisandro Dalcin   ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRQ(ierr);
581ae9aee6dSMatthew G. Knepley   ierr = PetscMalloc2(nroots,&roots,maxlocal,&leaves);CHKERRQ(ierr);
582ae9aee6dSMatthew G. Knepley   for (i=0; i<maxlocal; i++) {
58395fce210SBarry Smith     leaves[i].rank  = rank;
58495fce210SBarry Smith     leaves[i].index = i;
58595fce210SBarry Smith   }
58695fce210SBarry Smith   for (i=0; i <nroots; i++) {
58795fce210SBarry Smith     roots[i].rank  = -1;
58895fce210SBarry Smith     roots[i].index = -1;
58995fce210SBarry Smith   }
5908bfbc91cSJed Brown   ierr = PetscSFReduceBegin(sf,MPIU_2INT,leaves,roots,MPIU_REPLACE);CHKERRQ(ierr);
5918bfbc91cSJed Brown   ierr = PetscSFReduceEnd(sf,MPIU_2INT,leaves,roots,MPIU_REPLACE);CHKERRQ(ierr);
59295fce210SBarry Smith 
59395fce210SBarry Smith   /* Check whether our leaves are sparse */
59495fce210SBarry Smith   for (i=0,count=0; i<nroots; i++) if (roots[i].rank >= 0) count++;
59595fce210SBarry Smith   if (count == nroots) newilocal = NULL;
59695fce210SBarry Smith   else {                        /* Index for sparse leaves and compact "roots" array (which is to become our leaves). */
597785e854fSJed Brown     ierr = PetscMalloc1(count,&newilocal);CHKERRQ(ierr);
59895fce210SBarry Smith     for (i=0,count=0; i<nroots; i++) {
59995fce210SBarry Smith       if (roots[i].rank >= 0) {
60095fce210SBarry Smith         newilocal[count]   = i;
60195fce210SBarry Smith         roots[count].rank  = roots[i].rank;
60295fce210SBarry Smith         roots[count].index = roots[i].index;
60395fce210SBarry Smith         count++;
60495fce210SBarry Smith       }
60595fce210SBarry Smith     }
60695fce210SBarry Smith   }
60795fce210SBarry Smith 
60895fce210SBarry Smith   ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_CONFONLY,isf);CHKERRQ(ierr);
60995fce210SBarry Smith   ierr = PetscSFSetGraph(*isf,maxlocal,count,newilocal,PETSC_OWN_POINTER,roots,PETSC_COPY_VALUES);CHKERRQ(ierr);
61095fce210SBarry Smith   ierr = PetscFree2(roots,leaves);CHKERRQ(ierr);
61195fce210SBarry Smith   PetscFunctionReturn(0);
61295fce210SBarry Smith }
61395fce210SBarry Smith 
61495fce210SBarry Smith /*@
61595fce210SBarry Smith    PetscSFDuplicate - duplicate a PetscSF, optionally preserving rank connectivity and graph
61695fce210SBarry Smith 
61795fce210SBarry Smith    Collective
61895fce210SBarry Smith 
61995fce210SBarry Smith    Input Arguments:
62095fce210SBarry Smith +  sf - communication object to duplicate
62195fce210SBarry Smith -  opt - PETSCSF_DUPLICATE_CONFONLY, PETSCSF_DUPLICATE_RANKS, or PETSCSF_DUPLICATE_GRAPH (see PetscSFDuplicateOption)
62295fce210SBarry Smith 
62395fce210SBarry Smith    Output Arguments:
62495fce210SBarry Smith .  newsf - new communication object
62595fce210SBarry Smith 
62695fce210SBarry Smith    Level: beginner
62795fce210SBarry Smith 
62895fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFSetType(), PetscSFSetGraph()
62995fce210SBarry Smith @*/
63095fce210SBarry Smith PetscErrorCode PetscSFDuplicate(PetscSF sf,PetscSFDuplicateOption opt,PetscSF *newsf)
63195fce210SBarry Smith {
63229046d53SLisandro Dalcin   PetscSFType    type;
63395fce210SBarry Smith   PetscErrorCode ierr;
63495fce210SBarry Smith 
63595fce210SBarry Smith   PetscFunctionBegin;
63629046d53SLisandro Dalcin   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
63729046d53SLisandro Dalcin   PetscValidLogicalCollectiveEnum(sf,opt,2);
63829046d53SLisandro Dalcin   PetscValidPointer(newsf,3);
63995fce210SBarry Smith   ierr = PetscSFCreate(PetscObjectComm((PetscObject)sf),newsf);CHKERRQ(ierr);
64029046d53SLisandro Dalcin   ierr = PetscSFGetType(sf,&type);CHKERRQ(ierr);
64129046d53SLisandro Dalcin   if (type) {ierr = PetscSFSetType(*newsf,type);CHKERRQ(ierr);}
64295fce210SBarry Smith   if (opt == PETSCSF_DUPLICATE_GRAPH) {
643dd5b3ca6SJunchao Zhang     PetscSFCheckGraphSet(sf,1);
644dd5b3ca6SJunchao Zhang     if (sf->pattern == PETSCSF_PATTERN_GENERAL) {
64595fce210SBarry Smith       PetscInt          nroots,nleaves;
64695fce210SBarry Smith       const PetscInt    *ilocal;
64795fce210SBarry Smith       const PetscSFNode *iremote;
64895fce210SBarry Smith       ierr = PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,&iremote);CHKERRQ(ierr);
64995fce210SBarry Smith       ierr = PetscSFSetGraph(*newsf,nroots,nleaves,ilocal,PETSC_COPY_VALUES,iremote,PETSC_COPY_VALUES);CHKERRQ(ierr);
650dd5b3ca6SJunchao Zhang     } else {
651dd5b3ca6SJunchao Zhang       ierr = PetscSFSetGraphWithPattern(*newsf,sf->map,sf->pattern);CHKERRQ(ierr);
652dd5b3ca6SJunchao Zhang     }
65395fce210SBarry Smith   }
65429046d53SLisandro Dalcin   if (sf->ops->Duplicate) {ierr = (*sf->ops->Duplicate)(sf,opt,*newsf);CHKERRQ(ierr);}
65595fce210SBarry Smith   PetscFunctionReturn(0);
65695fce210SBarry Smith }
65795fce210SBarry Smith 
65895fce210SBarry Smith /*@C
65995fce210SBarry Smith    PetscSFGetGraph - Get the graph specifying a parallel star forest
66095fce210SBarry Smith 
66195fce210SBarry Smith    Not Collective
66295fce210SBarry Smith 
66395fce210SBarry Smith    Input Arguments:
66495fce210SBarry Smith .  sf - star forest
66595fce210SBarry Smith 
66695fce210SBarry Smith    Output Arguments:
66795fce210SBarry Smith +  nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves)
66895fce210SBarry Smith .  nleaves - number of leaf vertices on the current process, each of these references a root on any process
66995fce210SBarry Smith .  ilocal - locations of leaves in leafdata buffers
67095fce210SBarry Smith -  iremote - remote locations of root vertices for each leaf on the current process
67195fce210SBarry Smith 
672373e0d91SLisandro Dalcin    Notes:
673373e0d91SLisandro Dalcin    We are not currently requiring that the graph is set, thus returning nroots=-1 if it has not been set yet
674373e0d91SLisandro Dalcin 
675245d9833Sprj-    When called from Fortran, the returned iremote array is a copy and must be deallocated after use. Consequently, if you
676ca797d7aSLawrence Mitchell    want to update the graph, you must call PetscSFSetGraph after modifying the iremote array.
677ca797d7aSLawrence Mitchell 
67895fce210SBarry Smith    Level: intermediate
67995fce210SBarry Smith 
68095fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFView(), PetscSFSetGraph()
68195fce210SBarry Smith @*/
68295fce210SBarry Smith PetscErrorCode PetscSFGetGraph(PetscSF sf,PetscInt *nroots,PetscInt *nleaves,const PetscInt **ilocal,const PetscSFNode **iremote)
68395fce210SBarry Smith {
684b8dee149SJunchao Zhang   PetscErrorCode ierr;
68595fce210SBarry Smith 
68695fce210SBarry Smith   PetscFunctionBegin;
68795fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
688b8dee149SJunchao Zhang   if (sf->ops->GetGraph) {
689b8dee149SJunchao Zhang     ierr = (sf->ops->GetGraph)(sf,nroots,nleaves,ilocal,iremote);CHKERRQ(ierr);
690b8dee149SJunchao Zhang   } else {
69195fce210SBarry Smith     if (nroots) *nroots = sf->nroots;
69295fce210SBarry Smith     if (nleaves) *nleaves = sf->nleaves;
69395fce210SBarry Smith     if (ilocal) *ilocal = sf->mine;
69495fce210SBarry Smith     if (iremote) *iremote = sf->remote;
695b8dee149SJunchao Zhang   }
69695fce210SBarry Smith   PetscFunctionReturn(0);
69795fce210SBarry Smith }
69895fce210SBarry Smith 
69929046d53SLisandro Dalcin /*@
70095fce210SBarry Smith    PetscSFGetLeafRange - Get the active leaf ranges
70195fce210SBarry Smith 
70295fce210SBarry Smith    Not Collective
70395fce210SBarry Smith 
70495fce210SBarry Smith    Input Arguments:
70595fce210SBarry Smith .  sf - star forest
70695fce210SBarry Smith 
70795fce210SBarry Smith    Output Arguments:
708dd5b3ca6SJunchao Zhang +  minleaf - minimum active leaf on this process. Return 0 if there are no leaves.
709dd5b3ca6SJunchao Zhang -  maxleaf - maximum active leaf on this process. Return -1 if there are no leaves.
71095fce210SBarry Smith 
71195fce210SBarry Smith    Level: developer
71295fce210SBarry Smith 
71395fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFView(), PetscSFSetGraph(), PetscSFGetGraph()
71495fce210SBarry Smith @*/
71595fce210SBarry Smith PetscErrorCode PetscSFGetLeafRange(PetscSF sf,PetscInt *minleaf,PetscInt *maxleaf)
71695fce210SBarry Smith {
71795fce210SBarry Smith 
71895fce210SBarry Smith   PetscFunctionBegin;
71995fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
72029046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
72195fce210SBarry Smith   if (minleaf) *minleaf = sf->minleaf;
72295fce210SBarry Smith   if (maxleaf) *maxleaf = sf->maxleaf;
72395fce210SBarry Smith   PetscFunctionReturn(0);
72495fce210SBarry Smith }
72595fce210SBarry Smith 
72695fce210SBarry Smith /*@C
727fe2efc57SMark    PetscSFViewFromOptions - View from Options
728fe2efc57SMark 
729fe2efc57SMark    Collective on PetscSF
730fe2efc57SMark 
731fe2efc57SMark    Input Parameters:
732fe2efc57SMark +  A - the star forest
733736c3998SJose E. Roman .  obj - Optional object
734736c3998SJose E. Roman -  name - command line option
735fe2efc57SMark 
736fe2efc57SMark    Level: intermediate
737fe2efc57SMark .seealso:  PetscSF, PetscSFView, PetscObjectViewFromOptions(), PetscSFCreate()
738fe2efc57SMark @*/
739fe2efc57SMark PetscErrorCode  PetscSFViewFromOptions(PetscSF A,PetscObject obj,const char name[])
740fe2efc57SMark {
741fe2efc57SMark   PetscErrorCode ierr;
742fe2efc57SMark 
743fe2efc57SMark   PetscFunctionBegin;
744fe2efc57SMark   PetscValidHeaderSpecific(A,PETSCSF_CLASSID,1);
745fe2efc57SMark   ierr = PetscObjectViewFromOptions((PetscObject)A,obj,name);CHKERRQ(ierr);
746fe2efc57SMark   PetscFunctionReturn(0);
747fe2efc57SMark }
748fe2efc57SMark 
749fe2efc57SMark /*@C
75095fce210SBarry Smith    PetscSFView - view a star forest
75195fce210SBarry Smith 
75295fce210SBarry Smith    Collective
75395fce210SBarry Smith 
75495fce210SBarry Smith    Input Arguments:
75595fce210SBarry Smith +  sf - star forest
75695fce210SBarry Smith -  viewer - viewer to display graph, for example PETSC_VIEWER_STDOUT_WORLD
75795fce210SBarry Smith 
75895fce210SBarry Smith    Level: beginner
75995fce210SBarry Smith 
76095fce210SBarry Smith .seealso: PetscSFCreate(), PetscSFSetGraph()
76195fce210SBarry Smith @*/
76295fce210SBarry Smith PetscErrorCode PetscSFView(PetscSF sf,PetscViewer viewer)
76395fce210SBarry Smith {
76495fce210SBarry Smith   PetscErrorCode    ierr;
76595fce210SBarry Smith   PetscBool         iascii;
76695fce210SBarry Smith   PetscViewerFormat format;
76795fce210SBarry Smith 
76895fce210SBarry Smith   PetscFunctionBegin;
76995fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
77095fce210SBarry Smith   if (!viewer) {ierr = PetscViewerASCIIGetStdout(PetscObjectComm((PetscObject)sf),&viewer);CHKERRQ(ierr);}
77195fce210SBarry Smith   PetscValidHeaderSpecific(viewer,PETSC_VIEWER_CLASSID,2);
77295fce210SBarry Smith   PetscCheckSameComm(sf,1,viewer,2);
77380153354SVaclav Hapla   if (sf->graphset) {ierr = PetscSFSetUp(sf);CHKERRQ(ierr);}
77495fce210SBarry Smith   ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr);
77595fce210SBarry Smith   if (iascii) {
77695fce210SBarry Smith     PetscMPIInt rank;
77781bfa7aaSJed Brown     PetscInt    ii,i,j;
77895fce210SBarry Smith 
779dae58748SBarry Smith     ierr = PetscObjectPrintClassNamePrefixType((PetscObject)sf,viewer);CHKERRQ(ierr);
78095fce210SBarry Smith     ierr = PetscViewerASCIIPushTab(viewer);CHKERRQ(ierr);
78195fce210SBarry Smith     if (sf->ops->View) {ierr = (*sf->ops->View)(sf,viewer);CHKERRQ(ierr);}
782dd5b3ca6SJunchao Zhang     if (sf->pattern == PETSCSF_PATTERN_GENERAL) {
78380153354SVaclav Hapla       if (!sf->graphset) {
78480153354SVaclav Hapla         ierr = PetscViewerASCIIPrintf(viewer,"PetscSFSetGraph() has not been called yet\n");CHKERRQ(ierr);
78580153354SVaclav Hapla         ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr);
78680153354SVaclav Hapla         PetscFunctionReturn(0);
78780153354SVaclav Hapla       }
78895fce210SBarry Smith       ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRQ(ierr);
7891575c14dSBarry Smith       ierr = PetscViewerASCIIPushSynchronized(viewer);CHKERRQ(ierr);
79095fce210SBarry Smith       ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Number of roots=%D, leaves=%D, remote ranks=%D\n",rank,sf->nroots,sf->nleaves,sf->nranks);CHKERRQ(ierr);
79195fce210SBarry Smith       for (i=0; i<sf->nleaves; i++) {
79295fce210SBarry Smith         ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] %D <- (%D,%D)\n",rank,sf->mine ? sf->mine[i] : i,sf->remote[i].rank,sf->remote[i].index);CHKERRQ(ierr);
79395fce210SBarry Smith       }
79495fce210SBarry Smith       ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
79595fce210SBarry Smith       ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr);
79695fce210SBarry Smith       if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
79781bfa7aaSJed Brown         PetscMPIInt *tmpranks,*perm;
79881bfa7aaSJed Brown         ierr = PetscMalloc2(sf->nranks,&tmpranks,sf->nranks,&perm);CHKERRQ(ierr);
799580bdb30SBarry Smith         ierr = PetscArraycpy(tmpranks,sf->ranks,sf->nranks);CHKERRQ(ierr);
80081bfa7aaSJed Brown         for (i=0; i<sf->nranks; i++) perm[i] = i;
80181bfa7aaSJed Brown         ierr = PetscSortMPIIntWithArray(sf->nranks,tmpranks,perm);CHKERRQ(ierr);
80295fce210SBarry Smith         ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Roots referenced by my leaves, by rank\n",rank);CHKERRQ(ierr);
80381bfa7aaSJed Brown         for (ii=0; ii<sf->nranks; ii++) {
80481bfa7aaSJed Brown           i = perm[ii];
8057904a332SBarry Smith           ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] %d: %D edges\n",rank,sf->ranks[i],sf->roffset[i+1]-sf->roffset[i]);CHKERRQ(ierr);
80695fce210SBarry Smith           for (j=sf->roffset[i]; j<sf->roffset[i+1]; j++) {
80795fce210SBarry Smith             ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d]    %D <- %D\n",rank,sf->rmine[j],sf->rremote[j]);CHKERRQ(ierr);
80895fce210SBarry Smith           }
80995fce210SBarry Smith         }
81081bfa7aaSJed Brown         ierr = PetscFree2(tmpranks,perm);CHKERRQ(ierr);
81195fce210SBarry Smith       }
81295fce210SBarry Smith       ierr = PetscViewerFlush(viewer);CHKERRQ(ierr);
8131575c14dSBarry Smith       ierr = PetscViewerASCIIPopSynchronized(viewer);CHKERRQ(ierr);
814dd5b3ca6SJunchao Zhang     }
81595fce210SBarry Smith     ierr = PetscViewerASCIIPopTab(viewer);CHKERRQ(ierr);
81695fce210SBarry Smith   }
81795fce210SBarry Smith   PetscFunctionReturn(0);
81895fce210SBarry Smith }
81995fce210SBarry Smith 
82095fce210SBarry Smith /*@C
821dec1416fSJunchao Zhang    PetscSFGetRootRanks - Get root ranks and number of vertices referenced by leaves on this process
82295fce210SBarry Smith 
82395fce210SBarry Smith    Not Collective
82495fce210SBarry Smith 
82595fce210SBarry Smith    Input Arguments:
82695fce210SBarry Smith .  sf - star forest
82795fce210SBarry Smith 
82895fce210SBarry Smith    Output Arguments:
82995fce210SBarry Smith +  nranks - number of ranks referenced by local part
83095fce210SBarry Smith .  ranks - array of ranks
83195fce210SBarry Smith .  roffset - offset in rmine/rremote for each rank (length nranks+1)
83295fce210SBarry Smith .  rmine - concatenated array holding local indices referencing each remote rank
83395fce210SBarry Smith -  rremote - concatenated array holding remote indices referenced for each remote rank
83495fce210SBarry Smith 
83595fce210SBarry Smith    Level: developer
83695fce210SBarry Smith 
837dec1416fSJunchao Zhang .seealso: PetscSFGetLeafRanks()
83895fce210SBarry Smith @*/
839dec1416fSJunchao Zhang PetscErrorCode PetscSFGetRootRanks(PetscSF sf,PetscInt *nranks,const PetscMPIInt **ranks,const PetscInt **roffset,const PetscInt **rmine,const PetscInt **rremote)
84095fce210SBarry Smith {
841dec1416fSJunchao Zhang   PetscErrorCode ierr;
84295fce210SBarry Smith 
84395fce210SBarry Smith   PetscFunctionBegin;
84495fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
84529046d53SLisandro Dalcin   if (!sf->setupcalled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must call PetscSFSetUp() before obtaining ranks");
846dec1416fSJunchao Zhang   if (sf->ops->GetRootRanks) {
847dec1416fSJunchao Zhang     ierr = (sf->ops->GetRootRanks)(sf,nranks,ranks,roffset,rmine,rremote);CHKERRQ(ierr);
848dec1416fSJunchao Zhang   } else {
849dec1416fSJunchao Zhang     /* The generic implementation */
85095fce210SBarry Smith     if (nranks)  *nranks  = sf->nranks;
85195fce210SBarry Smith     if (ranks)   *ranks   = sf->ranks;
85295fce210SBarry Smith     if (roffset) *roffset = sf->roffset;
85395fce210SBarry Smith     if (rmine)   *rmine   = sf->rmine;
85495fce210SBarry Smith     if (rremote) *rremote = sf->rremote;
855dec1416fSJunchao Zhang   }
85695fce210SBarry Smith   PetscFunctionReturn(0);
85795fce210SBarry Smith }
85895fce210SBarry Smith 
8598750ddebSJunchao Zhang /*@C
8608750ddebSJunchao Zhang    PetscSFGetLeafRanks - Get leaf ranks referencing roots on this process
8618750ddebSJunchao Zhang 
8628750ddebSJunchao Zhang    Not Collective
8638750ddebSJunchao Zhang 
8648750ddebSJunchao Zhang    Input Arguments:
8658750ddebSJunchao Zhang .  sf - star forest
8668750ddebSJunchao Zhang 
8678750ddebSJunchao Zhang    Output Arguments:
8688750ddebSJunchao Zhang +  niranks - number of leaf ranks referencing roots on this process
8698750ddebSJunchao Zhang .  iranks - array of ranks
8708750ddebSJunchao Zhang .  ioffset - offset in irootloc for each rank (length niranks+1)
8718750ddebSJunchao Zhang -  irootloc - concatenated array holding local indices of roots referenced by each leaf rank
8728750ddebSJunchao Zhang 
8738750ddebSJunchao Zhang    Level: developer
8748750ddebSJunchao Zhang 
875dec1416fSJunchao Zhang .seealso: PetscSFGetRootRanks()
8768750ddebSJunchao Zhang @*/
8778750ddebSJunchao Zhang PetscErrorCode PetscSFGetLeafRanks(PetscSF sf,PetscInt *niranks,const PetscMPIInt **iranks,const PetscInt **ioffset,const PetscInt **irootloc)
8788750ddebSJunchao Zhang {
8798750ddebSJunchao Zhang   PetscErrorCode ierr;
8808750ddebSJunchao Zhang 
8818750ddebSJunchao Zhang   PetscFunctionBegin;
8828750ddebSJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
8838750ddebSJunchao Zhang   if (!sf->setupcalled) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must call PetscSFSetUp() before obtaining ranks");
8848750ddebSJunchao Zhang   if (sf->ops->GetLeafRanks) {
8858750ddebSJunchao Zhang     ierr = (sf->ops->GetLeafRanks)(sf,niranks,iranks,ioffset,irootloc);CHKERRQ(ierr);
8868750ddebSJunchao Zhang   } else {
8878750ddebSJunchao Zhang     PetscSFType type;
8888750ddebSJunchao Zhang     ierr = PetscSFGetType(sf,&type);CHKERRQ(ierr);
8898750ddebSJunchao Zhang     SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"PetscSFGetLeafRanks() is not supported on this StarForest type: %s", type);
8908750ddebSJunchao Zhang   }
8918750ddebSJunchao Zhang   PetscFunctionReturn(0);
8928750ddebSJunchao Zhang }
8938750ddebSJunchao Zhang 
894b5a8e515SJed Brown static PetscBool InList(PetscMPIInt needle,PetscMPIInt n,const PetscMPIInt *list) {
895b5a8e515SJed Brown   PetscInt i;
896b5a8e515SJed Brown   for (i=0; i<n; i++) {
897b5a8e515SJed Brown     if (needle == list[i]) return PETSC_TRUE;
898b5a8e515SJed Brown   }
899b5a8e515SJed Brown   return PETSC_FALSE;
900b5a8e515SJed Brown }
901b5a8e515SJed Brown 
90295fce210SBarry Smith /*@C
90321c688dcSJed Brown    PetscSFSetUpRanks - Set up data structures associated with ranks; this is for internal use by PetscSF implementations.
90421c688dcSJed Brown 
90521c688dcSJed Brown    Collective
90621c688dcSJed Brown 
90721c688dcSJed Brown    Input Arguments:
908b5a8e515SJed Brown +  sf - PetscSF to set up; PetscSFSetGraph() must have been called
909b5a8e515SJed Brown -  dgroup - MPI_Group of ranks to be distinguished (e.g., for self or shared memory exchange)
91021c688dcSJed Brown 
91121c688dcSJed Brown    Level: developer
91221c688dcSJed Brown 
913dec1416fSJunchao Zhang .seealso: PetscSFGetRootRanks()
91421c688dcSJed Brown @*/
915b5a8e515SJed Brown PetscErrorCode PetscSFSetUpRanks(PetscSF sf,MPI_Group dgroup)
91621c688dcSJed Brown {
91721c688dcSJed Brown   PetscErrorCode     ierr;
91821c688dcSJed Brown   PetscTable         table;
91921c688dcSJed Brown   PetscTablePosition pos;
920b5a8e515SJed Brown   PetscMPIInt        size,groupsize,*groupranks;
921247e8311SStefano Zampini   PetscInt           *rcount,*ranks;
922247e8311SStefano Zampini   PetscInt           i, irank = -1,orank = -1;
92321c688dcSJed Brown 
92421c688dcSJed Brown   PetscFunctionBegin;
92521c688dcSJed Brown   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
92629046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
92721c688dcSJed Brown   ierr = MPI_Comm_size(PetscObjectComm((PetscObject)sf),&size);CHKERRQ(ierr);
92821c688dcSJed Brown   ierr = PetscTableCreate(10,size,&table);CHKERRQ(ierr);
92921c688dcSJed Brown   for (i=0; i<sf->nleaves; i++) {
93021c688dcSJed Brown     /* Log 1-based rank */
93121c688dcSJed Brown     ierr = PetscTableAdd(table,sf->remote[i].rank+1,1,ADD_VALUES);CHKERRQ(ierr);
93221c688dcSJed Brown   }
93321c688dcSJed Brown   ierr = PetscTableGetCount(table,&sf->nranks);CHKERRQ(ierr);
93421c688dcSJed Brown   ierr = PetscMalloc4(sf->nranks,&sf->ranks,sf->nranks+1,&sf->roffset,sf->nleaves,&sf->rmine,sf->nleaves,&sf->rremote);CHKERRQ(ierr);
93521c688dcSJed Brown   ierr = PetscMalloc2(sf->nranks,&rcount,sf->nranks,&ranks);CHKERRQ(ierr);
93621c688dcSJed Brown   ierr = PetscTableGetHeadPosition(table,&pos);CHKERRQ(ierr);
93721c688dcSJed Brown   for (i=0; i<sf->nranks; i++) {
93821c688dcSJed Brown     ierr = PetscTableGetNext(table,&pos,&ranks[i],&rcount[i]);CHKERRQ(ierr);
93921c688dcSJed Brown     ranks[i]--;             /* Convert back to 0-based */
94021c688dcSJed Brown   }
94121c688dcSJed Brown   ierr = PetscTableDestroy(&table);CHKERRQ(ierr);
942b5a8e515SJed Brown 
943b5a8e515SJed Brown   /* We expect that dgroup is reliably "small" while nranks could be large */
944b5a8e515SJed Brown   {
9457fb8a5e4SKarl Rupp     MPI_Group group = MPI_GROUP_NULL;
946b5a8e515SJed Brown     PetscMPIInt *dgroupranks;
947b5a8e515SJed Brown     ierr = MPI_Comm_group(PetscObjectComm((PetscObject)sf),&group);CHKERRQ(ierr);
948b5a8e515SJed Brown     ierr = MPI_Group_size(dgroup,&groupsize);CHKERRQ(ierr);
949b5a8e515SJed Brown     ierr = PetscMalloc1(groupsize,&dgroupranks);CHKERRQ(ierr);
950b5a8e515SJed Brown     ierr = PetscMalloc1(groupsize,&groupranks);CHKERRQ(ierr);
951b5a8e515SJed Brown     for (i=0; i<groupsize; i++) dgroupranks[i] = i;
952f3fc9a17SJed Brown     if (groupsize) {ierr = MPI_Group_translate_ranks(dgroup,groupsize,dgroupranks,group,groupranks);CHKERRQ(ierr);}
953b5a8e515SJed Brown     ierr = MPI_Group_free(&group);CHKERRQ(ierr);
954b5a8e515SJed Brown     ierr = PetscFree(dgroupranks);CHKERRQ(ierr);
955b5a8e515SJed Brown   }
956b5a8e515SJed Brown 
957b5a8e515SJed Brown   /* Partition ranks[] into distinguished (first sf->ndranks) followed by non-distinguished */
958b5a8e515SJed Brown   for (sf->ndranks=0,i=sf->nranks; sf->ndranks<i; ) {
959b5a8e515SJed Brown     for (i--; sf->ndranks<i; i--) { /* Scan i backward looking for distinguished rank */
960b5a8e515SJed Brown       if (InList(ranks[i],groupsize,groupranks)) break;
961b5a8e515SJed Brown     }
962b5a8e515SJed Brown     for ( ; sf->ndranks<=i; sf->ndranks++) { /* Scan sf->ndranks forward looking for non-distinguished rank */
963b5a8e515SJed Brown       if (!InList(ranks[sf->ndranks],groupsize,groupranks)) break;
964b5a8e515SJed Brown     }
965b5a8e515SJed Brown     if (sf->ndranks < i) {                         /* Swap ranks[sf->ndranks] with ranks[i] */
966b5a8e515SJed Brown       PetscInt tmprank,tmpcount;
967247e8311SStefano Zampini 
968b5a8e515SJed Brown       tmprank             = ranks[i];
969b5a8e515SJed Brown       tmpcount            = rcount[i];
970b5a8e515SJed Brown       ranks[i]            = ranks[sf->ndranks];
971b5a8e515SJed Brown       rcount[i]           = rcount[sf->ndranks];
972b5a8e515SJed Brown       ranks[sf->ndranks]  = tmprank;
973b5a8e515SJed Brown       rcount[sf->ndranks] = tmpcount;
974b5a8e515SJed Brown       sf->ndranks++;
975b5a8e515SJed Brown     }
976b5a8e515SJed Brown   }
977b5a8e515SJed Brown   ierr = PetscFree(groupranks);CHKERRQ(ierr);
978b5a8e515SJed Brown   ierr = PetscSortIntWithArray(sf->ndranks,ranks,rcount);CHKERRQ(ierr);
979b5a8e515SJed Brown   ierr = PetscSortIntWithArray(sf->nranks-sf->ndranks,ranks+sf->ndranks,rcount+sf->ndranks);CHKERRQ(ierr);
98021c688dcSJed Brown   sf->roffset[0] = 0;
98121c688dcSJed Brown   for (i=0; i<sf->nranks; i++) {
98221c688dcSJed Brown     ierr = PetscMPIIntCast(ranks[i],sf->ranks+i);CHKERRQ(ierr);
98321c688dcSJed Brown     sf->roffset[i+1] = sf->roffset[i] + rcount[i];
98421c688dcSJed Brown     rcount[i]        = 0;
98521c688dcSJed Brown   }
986247e8311SStefano Zampini   for (i=0, irank = -1, orank = -1; i<sf->nleaves; i++) {
987247e8311SStefano Zampini     /* short circuit */
988247e8311SStefano Zampini     if (orank != sf->remote[i].rank) {
98921c688dcSJed Brown       /* Search for index of iremote[i].rank in sf->ranks */
990b5a8e515SJed Brown       ierr = PetscFindMPIInt(sf->remote[i].rank,sf->ndranks,sf->ranks,&irank);CHKERRQ(ierr);
991b5a8e515SJed Brown       if (irank < 0) {
992b5a8e515SJed Brown         ierr = PetscFindMPIInt(sf->remote[i].rank,sf->nranks-sf->ndranks,sf->ranks+sf->ndranks,&irank);CHKERRQ(ierr);
993b5a8e515SJed Brown         if (irank >= 0) irank += sf->ndranks;
99421c688dcSJed Brown       }
995247e8311SStefano Zampini       orank = sf->remote[i].rank;
996247e8311SStefano Zampini     }
997b5a8e515SJed Brown     if (irank < 0) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Could not find rank %D in array",sf->remote[i].rank);
99821c688dcSJed Brown     sf->rmine[sf->roffset[irank] + rcount[irank]]   = sf->mine ? sf->mine[i] : i;
99921c688dcSJed Brown     sf->rremote[sf->roffset[irank] + rcount[irank]] = sf->remote[i].index;
100021c688dcSJed Brown     rcount[irank]++;
100121c688dcSJed Brown   }
100221c688dcSJed Brown   ierr = PetscFree2(rcount,ranks);CHKERRQ(ierr);
100321c688dcSJed Brown   PetscFunctionReturn(0);
100421c688dcSJed Brown }
100521c688dcSJed Brown 
100621c688dcSJed Brown /*@C
100795fce210SBarry Smith    PetscSFGetGroups - gets incoming and outgoing process groups
100895fce210SBarry Smith 
100995fce210SBarry Smith    Collective
101095fce210SBarry Smith 
101195fce210SBarry Smith    Input Argument:
101295fce210SBarry Smith .  sf - star forest
101395fce210SBarry Smith 
101495fce210SBarry Smith    Output Arguments:
101595fce210SBarry Smith +  incoming - group of origin processes for incoming edges (leaves that reference my roots)
101695fce210SBarry Smith -  outgoing - group of destination processes for outgoing edges (roots that I reference)
101795fce210SBarry Smith 
101895fce210SBarry Smith    Level: developer
101995fce210SBarry Smith 
102095fce210SBarry Smith .seealso: PetscSFGetWindow(), PetscSFRestoreWindow()
102195fce210SBarry Smith @*/
102295fce210SBarry Smith PetscErrorCode PetscSFGetGroups(PetscSF sf,MPI_Group *incoming,MPI_Group *outgoing)
102395fce210SBarry Smith {
102495fce210SBarry Smith   PetscErrorCode ierr;
10257fb8a5e4SKarl Rupp   MPI_Group      group = MPI_GROUP_NULL;
102695fce210SBarry Smith 
102795fce210SBarry Smith   PetscFunctionBegin;
102844ee17edSStefano Zampini   if (sf->nranks < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Must call PetscSFSetUpRanks() before obtaining groups");
102995fce210SBarry Smith   if (sf->ingroup == MPI_GROUP_NULL) {
103095fce210SBarry Smith     PetscInt       i;
103195fce210SBarry Smith     const PetscInt *indegree;
103295fce210SBarry Smith     PetscMPIInt    rank,*outranks,*inranks;
103395fce210SBarry Smith     PetscSFNode    *remote;
103495fce210SBarry Smith     PetscSF        bgcount;
103595fce210SBarry Smith 
103695fce210SBarry Smith     /* Compute the number of incoming ranks */
1037785e854fSJed Brown     ierr = PetscMalloc1(sf->nranks,&remote);CHKERRQ(ierr);
103895fce210SBarry Smith     for (i=0; i<sf->nranks; i++) {
103995fce210SBarry Smith       remote[i].rank  = sf->ranks[i];
104095fce210SBarry Smith       remote[i].index = 0;
104195fce210SBarry Smith     }
104295fce210SBarry Smith     ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_CONFONLY,&bgcount);CHKERRQ(ierr);
104395fce210SBarry Smith     ierr = PetscSFSetGraph(bgcount,1,sf->nranks,NULL,PETSC_COPY_VALUES,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
104495fce210SBarry Smith     ierr = PetscSFComputeDegreeBegin(bgcount,&indegree);CHKERRQ(ierr);
104595fce210SBarry Smith     ierr = PetscSFComputeDegreeEnd(bgcount,&indegree);CHKERRQ(ierr);
104695fce210SBarry Smith     /* Enumerate the incoming ranks */
1047dcca6d9dSJed Brown     ierr = PetscMalloc2(indegree[0],&inranks,sf->nranks,&outranks);CHKERRQ(ierr);
104895fce210SBarry Smith     ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRQ(ierr);
104995fce210SBarry Smith     for (i=0; i<sf->nranks; i++) outranks[i] = rank;
105095fce210SBarry Smith     ierr = PetscSFGatherBegin(bgcount,MPI_INT,outranks,inranks);CHKERRQ(ierr);
105195fce210SBarry Smith     ierr = PetscSFGatherEnd(bgcount,MPI_INT,outranks,inranks);CHKERRQ(ierr);
105295fce210SBarry Smith     ierr = MPI_Comm_group(PetscObjectComm((PetscObject)sf),&group);CHKERRQ(ierr);
105395fce210SBarry Smith     ierr = MPI_Group_incl(group,indegree[0],inranks,&sf->ingroup);CHKERRQ(ierr);
105495fce210SBarry Smith     ierr = MPI_Group_free(&group);CHKERRQ(ierr);
105595fce210SBarry Smith     ierr = PetscFree2(inranks,outranks);CHKERRQ(ierr);
105695fce210SBarry Smith     ierr = PetscSFDestroy(&bgcount);CHKERRQ(ierr);
105795fce210SBarry Smith   }
105895fce210SBarry Smith   *incoming = sf->ingroup;
105995fce210SBarry Smith 
106095fce210SBarry Smith   if (sf->outgroup == MPI_GROUP_NULL) {
106195fce210SBarry Smith     ierr = MPI_Comm_group(PetscObjectComm((PetscObject)sf),&group);CHKERRQ(ierr);
106295fce210SBarry Smith     ierr = MPI_Group_incl(group,sf->nranks,sf->ranks,&sf->outgroup);CHKERRQ(ierr);
106395fce210SBarry Smith     ierr = MPI_Group_free(&group);CHKERRQ(ierr);
106495fce210SBarry Smith   }
106595fce210SBarry Smith   *outgoing = sf->outgroup;
106695fce210SBarry Smith   PetscFunctionReturn(0);
106795fce210SBarry Smith }
106895fce210SBarry Smith 
106929046d53SLisandro Dalcin /*@
107095fce210SBarry Smith    PetscSFGetMultiSF - gets the inner SF implemeting gathers and scatters
107195fce210SBarry Smith 
107295fce210SBarry Smith    Collective
107395fce210SBarry Smith 
107495fce210SBarry Smith    Input Argument:
107595fce210SBarry Smith .  sf - star forest that may contain roots with 0 or with more than 1 vertex
107695fce210SBarry Smith 
107795fce210SBarry Smith    Output Arguments:
107895fce210SBarry Smith .  multi - star forest with split roots, such that each root has degree exactly 1
107995fce210SBarry Smith 
108095fce210SBarry Smith    Level: developer
108195fce210SBarry Smith 
108295fce210SBarry Smith    Notes:
108395fce210SBarry Smith 
108495fce210SBarry Smith    In most cases, users should use PetscSFGatherBegin() and PetscSFScatterBegin() instead of manipulating multi
108595fce210SBarry Smith    directly. Since multi satisfies the stronger condition that each entry in the global space has exactly one incoming
108695fce210SBarry Smith    edge, it is a candidate for future optimization that might involve its removal.
108795fce210SBarry Smith 
1088673100f5SVaclav Hapla .seealso: PetscSFSetGraph(), PetscSFGatherBegin(), PetscSFScatterBegin(), PetscSFComputeMultiRootOriginalNumbering()
108995fce210SBarry Smith @*/
109095fce210SBarry Smith PetscErrorCode PetscSFGetMultiSF(PetscSF sf,PetscSF *multi)
109195fce210SBarry Smith {
109295fce210SBarry Smith   PetscErrorCode ierr;
109395fce210SBarry Smith 
109495fce210SBarry Smith   PetscFunctionBegin;
109595fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
109695fce210SBarry Smith   PetscValidPointer(multi,2);
109795fce210SBarry Smith   if (sf->nroots < 0) {         /* Graph has not been set yet; why do we need this? */
109895fce210SBarry Smith     ierr   = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&sf->multi);CHKERRQ(ierr);
109995fce210SBarry Smith     *multi = sf->multi;
110095fce210SBarry Smith     PetscFunctionReturn(0);
110195fce210SBarry Smith   }
110295fce210SBarry Smith   if (!sf->multi) {
110395fce210SBarry Smith     const PetscInt *indegree;
11049837ea96SMatthew G. Knepley     PetscInt       i,*inoffset,*outones,*outoffset,maxlocal;
110595fce210SBarry Smith     PetscSFNode    *remote;
110629046d53SLisandro Dalcin     maxlocal = sf->maxleaf+1; /* TODO: We should use PetscSFGetLeafRange() */
110795fce210SBarry Smith     ierr = PetscSFComputeDegreeBegin(sf,&indegree);CHKERRQ(ierr);
110895fce210SBarry Smith     ierr = PetscSFComputeDegreeEnd(sf,&indegree);CHKERRQ(ierr);
11099837ea96SMatthew G. Knepley     ierr = PetscMalloc3(sf->nroots+1,&inoffset,maxlocal,&outones,maxlocal,&outoffset);CHKERRQ(ierr);
111095fce210SBarry Smith     inoffset[0] = 0;
111195fce210SBarry Smith     for (i=0; i<sf->nroots; i++) inoffset[i+1] = inoffset[i] + indegree[i];
11129837ea96SMatthew G. Knepley     for (i=0; i<maxlocal; i++) outones[i] = 1;
1113dbd2ff41SBarry Smith     ierr = PetscSFFetchAndOpBegin(sf,MPIU_INT,inoffset,outones,outoffset,MPI_SUM);CHKERRQ(ierr);
1114dbd2ff41SBarry Smith     ierr = PetscSFFetchAndOpEnd(sf,MPIU_INT,inoffset,outones,outoffset,MPI_SUM);CHKERRQ(ierr);
111595fce210SBarry Smith     for (i=0; i<sf->nroots; i++) inoffset[i] -= indegree[i]; /* Undo the increment */
111695fce210SBarry Smith #if defined(PETSC_USE_DEBUG)                                 /* Check that the expected number of increments occurred */
111795fce210SBarry Smith     for (i=0; i<sf->nroots; i++) {
111895fce210SBarry Smith       if (inoffset[i] + indegree[i] != inoffset[i+1]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Incorrect result after PetscSFFetchAndOp");
111995fce210SBarry Smith     }
112095fce210SBarry Smith #endif
1121785e854fSJed Brown     ierr = PetscMalloc1(sf->nleaves,&remote);CHKERRQ(ierr);
112295fce210SBarry Smith     for (i=0; i<sf->nleaves; i++) {
112395fce210SBarry Smith       remote[i].rank  = sf->remote[i].rank;
112438e7336fSToby Isaac       remote[i].index = outoffset[sf->mine ? sf->mine[i] : i];
112595fce210SBarry Smith     }
112695fce210SBarry Smith     ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_RANKS,&sf->multi);CHKERRQ(ierr);
112701365b40SToby Isaac     ierr = PetscSFSetGraph(sf->multi,inoffset[sf->nroots],sf->nleaves,sf->mine,PETSC_COPY_VALUES,remote,PETSC_OWN_POINTER);CHKERRQ(ierr);
112895fce210SBarry Smith     if (sf->rankorder) {        /* Sort the ranks */
112995fce210SBarry Smith       PetscMPIInt rank;
113095fce210SBarry Smith       PetscInt    *inranks,*newoffset,*outranks,*newoutoffset,*tmpoffset,maxdegree;
113195fce210SBarry Smith       PetscSFNode *newremote;
113295fce210SBarry Smith       ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRQ(ierr);
113395fce210SBarry Smith       for (i=0,maxdegree=0; i<sf->nroots; i++) maxdegree = PetscMax(maxdegree,indegree[i]);
11349837ea96SMatthew G. Knepley       ierr = PetscMalloc5(sf->multi->nroots,&inranks,sf->multi->nroots,&newoffset,maxlocal,&outranks,maxlocal,&newoutoffset,maxdegree,&tmpoffset);CHKERRQ(ierr);
11359837ea96SMatthew G. Knepley       for (i=0; i<maxlocal; i++) outranks[i] = rank;
11368bfbc91cSJed Brown       ierr = PetscSFReduceBegin(sf->multi,MPIU_INT,outranks,inranks,MPIU_REPLACE);CHKERRQ(ierr);
11378bfbc91cSJed Brown       ierr = PetscSFReduceEnd(sf->multi,MPIU_INT,outranks,inranks,MPIU_REPLACE);CHKERRQ(ierr);
113895fce210SBarry Smith       /* Sort the incoming ranks at each vertex, build the inverse map */
113995fce210SBarry Smith       for (i=0; i<sf->nroots; i++) {
114095fce210SBarry Smith         PetscInt j;
114195fce210SBarry Smith         for (j=0; j<indegree[i]; j++) tmpoffset[j] = j;
114295fce210SBarry Smith         ierr = PetscSortIntWithArray(indegree[i],inranks+inoffset[i],tmpoffset);CHKERRQ(ierr);
114395fce210SBarry Smith         for (j=0; j<indegree[i]; j++) newoffset[inoffset[i] + tmpoffset[j]] = inoffset[i] + j;
114495fce210SBarry Smith       }
114595fce210SBarry Smith       ierr = PetscSFBcastBegin(sf->multi,MPIU_INT,newoffset,newoutoffset);CHKERRQ(ierr);
114695fce210SBarry Smith       ierr = PetscSFBcastEnd(sf->multi,MPIU_INT,newoffset,newoutoffset);CHKERRQ(ierr);
1147785e854fSJed Brown       ierr = PetscMalloc1(sf->nleaves,&newremote);CHKERRQ(ierr);
114895fce210SBarry Smith       for (i=0; i<sf->nleaves; i++) {
114995fce210SBarry Smith         newremote[i].rank  = sf->remote[i].rank;
115001365b40SToby Isaac         newremote[i].index = newoutoffset[sf->mine ? sf->mine[i] : i];
115195fce210SBarry Smith       }
115201365b40SToby Isaac       ierr = PetscSFSetGraph(sf->multi,inoffset[sf->nroots],sf->nleaves,sf->mine,PETSC_COPY_VALUES,newremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
115395fce210SBarry Smith       ierr = PetscFree5(inranks,newoffset,outranks,newoutoffset,tmpoffset);CHKERRQ(ierr);
115495fce210SBarry Smith     }
115595fce210SBarry Smith     ierr = PetscFree3(inoffset,outones,outoffset);CHKERRQ(ierr);
115695fce210SBarry Smith   }
115795fce210SBarry Smith   *multi = sf->multi;
115895fce210SBarry Smith   PetscFunctionReturn(0);
115995fce210SBarry Smith }
116095fce210SBarry Smith 
116195fce210SBarry Smith /*@C
116295fce210SBarry Smith    PetscSFCreateEmbeddedSF - removes edges from all but the selected roots, does not remap indices
116395fce210SBarry Smith 
116495fce210SBarry Smith    Collective
116595fce210SBarry Smith 
116695fce210SBarry Smith    Input Arguments:
116795fce210SBarry Smith +  sf - original star forest
1168ba2a7774SJunchao Zhang .  nselected  - number of selected roots on this process
1169ba2a7774SJunchao Zhang -  selected   - indices of the selected roots on this process
117095fce210SBarry Smith 
117195fce210SBarry Smith    Output Arguments:
1172cd620004SJunchao Zhang .  esf - new star forest
117395fce210SBarry Smith 
117495fce210SBarry Smith    Level: advanced
117595fce210SBarry Smith 
117695fce210SBarry Smith    Note:
117795fce210SBarry Smith    To use the new PetscSF, it may be necessary to know the indices of the leaves that are still participating. This can
117895fce210SBarry Smith    be done by calling PetscSFGetGraph().
117995fce210SBarry Smith 
118095fce210SBarry Smith .seealso: PetscSFSetGraph(), PetscSFGetGraph()
118195fce210SBarry Smith @*/
1182cd620004SJunchao Zhang PetscErrorCode PetscSFCreateEmbeddedSF(PetscSF sf,PetscInt nselected,const PetscInt *selected,PetscSF *esf)
118395fce210SBarry Smith {
1184cd620004SJunchao Zhang   PetscInt          i,j,n,nroots,nleaves,esf_nleaves,*new_ilocal,minleaf,maxleaf,maxlocal;
1185cd620004SJunchao Zhang   const PetscInt    *ilocal;
1186cd620004SJunchao Zhang   signed char       *rootdata,*leafdata,*leafmem;
1187ba2a7774SJunchao Zhang   const PetscSFNode *iremote;
1188f659e5c7SJunchao Zhang   PetscSFNode       *new_iremote;
1189f659e5c7SJunchao Zhang   MPI_Comm          comm;
11900511a646SMatthew G. Knepley   PetscErrorCode    ierr;
119195fce210SBarry Smith 
119295fce210SBarry Smith   PetscFunctionBegin;
119395fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
119429046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
1195ba2a7774SJunchao Zhang   if (nselected) PetscValidPointer(selected,3);
1196cd620004SJunchao Zhang   PetscValidPointer(esf,4);
11970511a646SMatthew G. Knepley 
1198f659e5c7SJunchao Zhang   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
1199cd620004SJunchao Zhang 
1200140a1472SStefano Zampini   ierr = PetscLogEventBegin(PETSCSF_EmbedSF,sf,0,0,0);CHKERRQ(ierr);
1201f659e5c7SJunchao Zhang   ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr);
1202cd620004SJunchao Zhang   ierr = PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,&iremote);CHKERRQ(ierr);
1203cd620004SJunchao Zhang 
1204cd620004SJunchao Zhang #if defined(PETSC_USE_DEBUG)
1205cd620004SJunchao Zhang   /* Error out if selected[] has dups or  out of range indices */
1206cd620004SJunchao Zhang   {
1207cd620004SJunchao Zhang     PetscBool dups;
1208cd620004SJunchao Zhang     ierr = PetscCheckDupsInt(nselected,selected,&dups);CHKERRQ(ierr);
1209cd620004SJunchao Zhang     if (dups) SETERRQ(comm,PETSC_ERR_ARG_WRONG,"selected[] has dups");
1210cd620004SJunchao Zhang     for (i=0; i<nselected; i++)
1211cd620004SJunchao Zhang       if (selected[i] < 0 || selected[i] >= nroots) SETERRQ2(comm,PETSC_ERR_ARG_OUTOFRANGE,"selected root indice %D is out of [0,%D)",selected[i],nroots);
1212cd620004SJunchao Zhang   }
1213cd620004SJunchao Zhang #endif
1214f659e5c7SJunchao Zhang 
1215f659e5c7SJunchao Zhang   if (sf->ops->CreateEmbeddedSF) {
1216cd620004SJunchao Zhang     ierr = (*sf->ops->CreateEmbeddedSF)(sf,nselected,selected,esf);CHKERRQ(ierr);
1217f659e5c7SJunchao Zhang   } else {
1218cd620004SJunchao Zhang     /* A generic version of creating embedded sf */
1219cd620004SJunchao Zhang     ierr = PetscSFGetLeafRange(sf,&minleaf,&maxleaf);CHKERRQ(ierr);
1220cd620004SJunchao Zhang     maxlocal = maxleaf - minleaf + 1;
1221cd620004SJunchao Zhang     ierr = PetscCalloc2(nroots,&rootdata,maxlocal,&leafmem);CHKERRQ(ierr);
1222cd620004SJunchao Zhang     leafdata = leafmem - minleaf;
1223cd620004SJunchao Zhang     /* Tag selected roots and bcast to leaves */
1224cd620004SJunchao Zhang     for (i=0; i<nselected; i++) rootdata[selected[i]] = 1;
1225cd620004SJunchao Zhang     ierr = PetscSFBcastBegin(sf,MPI_SIGNED_CHAR,rootdata,leafdata);CHKERRQ(ierr);
1226cd620004SJunchao Zhang     ierr = PetscSFBcastEnd(sf,MPI_SIGNED_CHAR,rootdata,leafdata);CHKERRQ(ierr);
1227ba2a7774SJunchao Zhang 
1228cd620004SJunchao Zhang     /* Build esf with leaves that are still connected */
1229cd620004SJunchao Zhang     esf_nleaves = 0;
1230cd620004SJunchao Zhang     for (i=0; i<nleaves; i++) {
1231cd620004SJunchao Zhang       j = ilocal ? ilocal[i] : i;
1232cd620004SJunchao Zhang       /* esf_nleaves += leafdata[j] should work in theory, but failed with SFWindow bugs
1233cd620004SJunchao Zhang          with PetscSFBcast. See https://gitlab.com/petsc/petsc/issues/555
1234cd620004SJunchao Zhang       */
1235cd620004SJunchao Zhang       esf_nleaves += (leafdata[j] ? 1 : 0);
1236cd620004SJunchao Zhang     }
1237cd620004SJunchao Zhang     ierr = PetscMalloc1(esf_nleaves,&new_ilocal);CHKERRQ(ierr);
1238cd620004SJunchao Zhang     ierr = PetscMalloc1(esf_nleaves,&new_iremote);CHKERRQ(ierr);
1239cd620004SJunchao Zhang     for (i=n=0; i<nleaves; i++) {
1240cd620004SJunchao Zhang       j = ilocal ? ilocal[i] : i;
1241cd620004SJunchao Zhang       if (leafdata[j]) {
1242cd620004SJunchao Zhang         new_ilocal[n]        = j;
1243cd620004SJunchao Zhang         new_iremote[n].rank  = iremote[i].rank;
1244cd620004SJunchao Zhang         new_iremote[n].index = iremote[i].index;
1245fc1ede2bSMatthew G. Knepley         ++n;
124695fce210SBarry Smith       }
124795fce210SBarry Smith     }
1248cd620004SJunchao Zhang     ierr = PetscSFCreate(comm,esf);CHKERRQ(ierr);
1249cd620004SJunchao Zhang     ierr = PetscSFSetFromOptions(*esf);CHKERRQ(ierr);
1250cd620004SJunchao Zhang     ierr = PetscSFSetGraph(*esf,nroots,esf_nleaves,new_ilocal,PETSC_OWN_POINTER,new_iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
1251cd620004SJunchao Zhang     ierr = PetscFree2(rootdata,leafmem);CHKERRQ(ierr);
1252f659e5c7SJunchao Zhang   }
1253140a1472SStefano Zampini   ierr = PetscLogEventEnd(PETSCSF_EmbedSF,sf,0,0,0);CHKERRQ(ierr);
125495fce210SBarry Smith   PetscFunctionReturn(0);
125595fce210SBarry Smith }
125695fce210SBarry Smith 
12572f5fb4c2SMatthew G. Knepley /*@C
12582f5fb4c2SMatthew G. Knepley   PetscSFCreateEmbeddedLeafSF - removes edges from all but the selected leaves, does not remap indices
12592f5fb4c2SMatthew G. Knepley 
12602f5fb4c2SMatthew G. Knepley   Collective
12612f5fb4c2SMatthew G. Knepley 
12622f5fb4c2SMatthew G. Knepley   Input Arguments:
12632f5fb4c2SMatthew G. Knepley + sf - original star forest
1264f659e5c7SJunchao Zhang . nselected  - number of selected leaves on this process
1265f659e5c7SJunchao Zhang - selected   - indices of the selected leaves on this process
12662f5fb4c2SMatthew G. Knepley 
12672f5fb4c2SMatthew G. Knepley   Output Arguments:
12682f5fb4c2SMatthew G. Knepley .  newsf - new star forest
12692f5fb4c2SMatthew G. Knepley 
12702f5fb4c2SMatthew G. Knepley   Level: advanced
12712f5fb4c2SMatthew G. Knepley 
12722f5fb4c2SMatthew G. Knepley .seealso: PetscSFCreateEmbeddedSF(), PetscSFSetGraph(), PetscSFGetGraph()
12732f5fb4c2SMatthew G. Knepley @*/
1274f659e5c7SJunchao Zhang PetscErrorCode PetscSFCreateEmbeddedLeafSF(PetscSF sf,PetscInt nselected,const PetscInt *selected,PetscSF *newsf)
12752f5fb4c2SMatthew G. Knepley {
1276f659e5c7SJunchao Zhang   const PetscSFNode *iremote;
1277f659e5c7SJunchao Zhang   PetscSFNode       *new_iremote;
1278f659e5c7SJunchao Zhang   const PetscInt    *ilocal;
1279f659e5c7SJunchao Zhang   PetscInt          i,nroots,*leaves,*new_ilocal;
1280f659e5c7SJunchao Zhang   MPI_Comm          comm;
12812f5fb4c2SMatthew G. Knepley   PetscErrorCode    ierr;
12822f5fb4c2SMatthew G. Knepley 
12832f5fb4c2SMatthew G. Knepley   PetscFunctionBegin;
12842f5fb4c2SMatthew G. Knepley   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
128529046d53SLisandro Dalcin   PetscSFCheckGraphSet(sf,1);
1286f659e5c7SJunchao Zhang   if (nselected) PetscValidPointer(selected,3);
12872f5fb4c2SMatthew G. Knepley   PetscValidPointer(newsf,4);
12882f5fb4c2SMatthew G. Knepley 
1289f659e5c7SJunchao Zhang   /* Uniq selected[] and put results in leaves[] */
1290f659e5c7SJunchao Zhang   ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr);
1291f659e5c7SJunchao Zhang   ierr = PetscMalloc1(nselected,&leaves);CHKERRQ(ierr);
1292dd5b3ca6SJunchao Zhang   ierr = PetscArraycpy(leaves,selected,nselected);CHKERRQ(ierr);
1293f659e5c7SJunchao Zhang   ierr = PetscSortedRemoveDupsInt(&nselected,leaves);CHKERRQ(ierr);
1294f659e5c7SJunchao Zhang   if (nselected && (leaves[0] < 0 || leaves[nselected-1] >= sf->nleaves)) SETERRQ3(comm,PETSC_ERR_ARG_OUTOFRANGE,"Min/Max leaf indices %D/%D are not in [0,%D)",leaves[0],leaves[nselected-1],sf->nleaves);
1295f659e5c7SJunchao Zhang 
1296f659e5c7SJunchao Zhang   /* Optimize the routine only when sf is setup and hence we can reuse sf's communication pattern */
1297f659e5c7SJunchao Zhang   if (sf->setupcalled && sf->ops->CreateEmbeddedLeafSF) {
1298f659e5c7SJunchao Zhang     ierr = (*sf->ops->CreateEmbeddedLeafSF)(sf,nselected,leaves,newsf);CHKERRQ(ierr);
1299f659e5c7SJunchao Zhang   } else {
1300f659e5c7SJunchao Zhang     ierr = PetscSFGetGraph(sf,&nroots,NULL,&ilocal,&iremote);CHKERRQ(ierr);
1301f659e5c7SJunchao Zhang     ierr = PetscMalloc1(nselected,&new_ilocal);CHKERRQ(ierr);
1302f659e5c7SJunchao Zhang     ierr = PetscMalloc1(nselected,&new_iremote);CHKERRQ(ierr);
1303f659e5c7SJunchao Zhang     for (i=0; i<nselected; ++i) {
1304f659e5c7SJunchao Zhang       const PetscInt l     = leaves[i];
1305f659e5c7SJunchao Zhang       new_ilocal[i]        = ilocal ? ilocal[l] : l;
1306f659e5c7SJunchao Zhang       new_iremote[i].rank  = iremote[l].rank;
1307f659e5c7SJunchao Zhang       new_iremote[i].index = iremote[l].index;
13082f5fb4c2SMatthew G. Knepley     }
13094cc19a0cSStefano Zampini     ierr = PetscSFDuplicate(sf,PETSCSF_DUPLICATE_CONFONLY,newsf);CHKERRQ(ierr);
1310f659e5c7SJunchao Zhang     ierr = PetscSFSetGraph(*newsf,nroots,nselected,new_ilocal,PETSC_OWN_POINTER,new_iremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
1311f659e5c7SJunchao Zhang   }
1312f659e5c7SJunchao Zhang   ierr = PetscFree(leaves);CHKERRQ(ierr);
13132f5fb4c2SMatthew G. Knepley   PetscFunctionReturn(0);
13142f5fb4c2SMatthew G. Knepley }
13152f5fb4c2SMatthew G. Knepley 
131695fce210SBarry Smith /*@C
13173482bfa8SJunchao Zhang    PetscSFBcastAndOpBegin - begin pointwise broadcast with root value being reduced to leaf value, to be concluded with call to PetscSFBcastAndOpEnd()
13183482bfa8SJunchao Zhang 
13193482bfa8SJunchao Zhang    Collective on PetscSF
13203482bfa8SJunchao Zhang 
13213482bfa8SJunchao Zhang    Input Arguments:
13223482bfa8SJunchao Zhang +  sf - star forest on which to communicate
13233482bfa8SJunchao Zhang .  unit - data type associated with each node
13243482bfa8SJunchao Zhang .  rootdata - buffer to broadcast
13253482bfa8SJunchao Zhang -  op - operation to use for reduction
13263482bfa8SJunchao Zhang 
13273482bfa8SJunchao Zhang    Output Arguments:
13283482bfa8SJunchao Zhang .  leafdata - buffer to be reduced with values from each leaf's respective root
13293482bfa8SJunchao Zhang 
13303482bfa8SJunchao Zhang    Level: intermediate
13313482bfa8SJunchao Zhang 
13323482bfa8SJunchao Zhang .seealso: PetscSFBcastAndOpEnd(), PetscSFBcastBegin(), PetscSFBcastEnd()
13333482bfa8SJunchao Zhang @*/
13343482bfa8SJunchao Zhang PetscErrorCode PetscSFBcastAndOpBegin(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata,MPI_Op op)
13353482bfa8SJunchao Zhang {
13363482bfa8SJunchao Zhang   PetscErrorCode ierr;
1337eb02082bSJunchao Zhang   PetscMemType   rootmtype,leafmtype;
13383482bfa8SJunchao Zhang 
13393482bfa8SJunchao Zhang   PetscFunctionBegin;
13403482bfa8SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
13413482bfa8SJunchao Zhang   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
13423482bfa8SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_BcastAndOpBegin,sf,0,0,0);CHKERRQ(ierr);
1343eb02082bSJunchao Zhang   ierr = PetscGetMemType(rootdata,&rootmtype);CHKERRQ(ierr);
1344eb02082bSJunchao Zhang   ierr = PetscGetMemType(leafdata,&leafmtype);CHKERRQ(ierr);
1345eb02082bSJunchao Zhang   ierr = (*sf->ops->BcastAndOpBegin)(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op);CHKERRQ(ierr);
13463482bfa8SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_BcastAndOpBegin,sf,0,0,0);CHKERRQ(ierr);
13473482bfa8SJunchao Zhang   PetscFunctionReturn(0);
13483482bfa8SJunchao Zhang }
13493482bfa8SJunchao Zhang 
13503482bfa8SJunchao Zhang /*@C
13513482bfa8SJunchao Zhang    PetscSFBcastAndOpEnd - end a broadcast & reduce operation started with PetscSFBcastAndOpBegin()
13523482bfa8SJunchao Zhang 
13533482bfa8SJunchao Zhang    Collective
13543482bfa8SJunchao Zhang 
13553482bfa8SJunchao Zhang    Input Arguments:
13563482bfa8SJunchao Zhang +  sf - star forest
13573482bfa8SJunchao Zhang .  unit - data type
13583482bfa8SJunchao Zhang .  rootdata - buffer to broadcast
13593482bfa8SJunchao Zhang -  op - operation to use for reduction
13603482bfa8SJunchao Zhang 
13613482bfa8SJunchao Zhang    Output Arguments:
13623482bfa8SJunchao Zhang .  leafdata - buffer to be reduced with values from each leaf's respective root
13633482bfa8SJunchao Zhang 
13643482bfa8SJunchao Zhang    Level: intermediate
13653482bfa8SJunchao Zhang 
13663482bfa8SJunchao Zhang .seealso: PetscSFSetGraph(), PetscSFReduceEnd()
13673482bfa8SJunchao Zhang @*/
13683482bfa8SJunchao Zhang PetscErrorCode PetscSFBcastAndOpEnd(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata,MPI_Op op)
13693482bfa8SJunchao Zhang {
13703482bfa8SJunchao Zhang   PetscErrorCode ierr;
13713482bfa8SJunchao Zhang 
13723482bfa8SJunchao Zhang   PetscFunctionBegin;
13733482bfa8SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
13743482bfa8SJunchao Zhang   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
13753482bfa8SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_BcastAndOpEnd,sf,0,0,0);CHKERRQ(ierr);
137600816365SJunchao Zhang   ierr = (*sf->ops->BcastAndOpEnd)(sf,unit,rootdata,leafdata,op);CHKERRQ(ierr);
13773482bfa8SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_BcastAndOpEnd,sf,0,0,0);CHKERRQ(ierr);
13783482bfa8SJunchao Zhang   PetscFunctionReturn(0);
13793482bfa8SJunchao Zhang }
13803482bfa8SJunchao Zhang 
13813482bfa8SJunchao Zhang /*@C
138295fce210SBarry Smith    PetscSFReduceBegin - begin reduction of leafdata into rootdata, to be completed with call to PetscSFReduceEnd()
138395fce210SBarry Smith 
138495fce210SBarry Smith    Collective
138595fce210SBarry Smith 
138695fce210SBarry Smith    Input Arguments:
138795fce210SBarry Smith +  sf - star forest
138895fce210SBarry Smith .  unit - data type
138995fce210SBarry Smith .  leafdata - values to reduce
139095fce210SBarry Smith -  op - reduction operation
139195fce210SBarry Smith 
139295fce210SBarry Smith    Output Arguments:
139395fce210SBarry Smith .  rootdata - result of reduction of values from all leaves of each root
139495fce210SBarry Smith 
139595fce210SBarry Smith    Level: intermediate
139695fce210SBarry Smith 
139795fce210SBarry Smith .seealso: PetscSFBcastBegin()
139895fce210SBarry Smith @*/
139995fce210SBarry Smith PetscErrorCode PetscSFReduceBegin(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *rootdata,MPI_Op op)
140095fce210SBarry Smith {
140195fce210SBarry Smith   PetscErrorCode ierr;
1402eb02082bSJunchao Zhang   PetscMemType   rootmtype,leafmtype;
140395fce210SBarry Smith 
140495fce210SBarry Smith   PetscFunctionBegin;
140595fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
140695fce210SBarry Smith   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
140729046d53SLisandro Dalcin   ierr = PetscLogEventBegin(PETSCSF_ReduceBegin,sf,0,0,0);CHKERRQ(ierr);
1408eb02082bSJunchao Zhang   ierr = PetscGetMemType(rootdata,&rootmtype);CHKERRQ(ierr);
1409eb02082bSJunchao Zhang   ierr = PetscGetMemType(leafdata,&leafmtype);CHKERRQ(ierr);
1410eb02082bSJunchao Zhang   ierr = (sf->ops->ReduceBegin)(sf,unit,leafmtype,leafdata,rootmtype,rootdata,op);CHKERRQ(ierr);
1411563ffbabSMatthew G. Knepley   ierr = PetscLogEventEnd(PETSCSF_ReduceBegin,sf,0,0,0);CHKERRQ(ierr);
141295fce210SBarry Smith   PetscFunctionReturn(0);
141395fce210SBarry Smith }
141495fce210SBarry Smith 
141595fce210SBarry Smith /*@C
141695fce210SBarry Smith    PetscSFReduceEnd - end a reduction operation started with PetscSFReduceBegin()
141795fce210SBarry Smith 
141895fce210SBarry Smith    Collective
141995fce210SBarry Smith 
142095fce210SBarry Smith    Input Arguments:
142195fce210SBarry Smith +  sf - star forest
142295fce210SBarry Smith .  unit - data type
142395fce210SBarry Smith .  leafdata - values to reduce
142495fce210SBarry Smith -  op - reduction operation
142595fce210SBarry Smith 
142695fce210SBarry Smith    Output Arguments:
142795fce210SBarry Smith .  rootdata - result of reduction of values from all leaves of each root
142895fce210SBarry Smith 
142995fce210SBarry Smith    Level: intermediate
143095fce210SBarry Smith 
143195fce210SBarry Smith .seealso: PetscSFSetGraph(), PetscSFBcastEnd()
143295fce210SBarry Smith @*/
143395fce210SBarry Smith PetscErrorCode PetscSFReduceEnd(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *rootdata,MPI_Op op)
143495fce210SBarry Smith {
143595fce210SBarry Smith   PetscErrorCode ierr;
143695fce210SBarry Smith 
143795fce210SBarry Smith   PetscFunctionBegin;
143895fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
143995fce210SBarry Smith   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
144029046d53SLisandro Dalcin   ierr = PetscLogEventBegin(PETSCSF_ReduceEnd,sf,0,0,0);CHKERRQ(ierr);
144100816365SJunchao Zhang   ierr = (*sf->ops->ReduceEnd)(sf,unit,leafdata,rootdata,op);CHKERRQ(ierr);
1442563ffbabSMatthew G. Knepley   ierr = PetscLogEventEnd(PETSCSF_ReduceEnd,sf,0,0,0);CHKERRQ(ierr);
144395fce210SBarry Smith   PetscFunctionReturn(0);
144495fce210SBarry Smith }
144595fce210SBarry Smith 
144695fce210SBarry Smith /*@C
1447a1729e3fSJunchao Zhang    PetscSFFetchAndOpBegin - begin operation that fetches values from root and updates atomically by applying operation using my leaf value, to be completed with PetscSFFetchAndOpEnd()
1448a1729e3fSJunchao Zhang 
1449a1729e3fSJunchao Zhang    Collective
1450a1729e3fSJunchao Zhang 
1451a1729e3fSJunchao Zhang    Input Arguments:
1452a1729e3fSJunchao Zhang +  sf - star forest
1453a1729e3fSJunchao Zhang .  unit - data type
1454a1729e3fSJunchao Zhang .  leafdata - leaf values to use in reduction
1455a1729e3fSJunchao Zhang -  op - operation to use for reduction
1456a1729e3fSJunchao Zhang 
1457a1729e3fSJunchao Zhang    Output Arguments:
1458a1729e3fSJunchao Zhang +  rootdata - root values to be updated, input state is seen by first process to perform an update
1459a1729e3fSJunchao Zhang -  leafupdate - state at each leaf's respective root immediately prior to my atomic update
1460a1729e3fSJunchao Zhang 
1461a1729e3fSJunchao Zhang    Level: advanced
1462a1729e3fSJunchao Zhang 
1463a1729e3fSJunchao Zhang    Note:
1464a1729e3fSJunchao Zhang    The update is only atomic at the granularity provided by the hardware. Different roots referenced by the same process
1465a1729e3fSJunchao Zhang    might be updated in a different order. Furthermore, if a composite type is used for the unit datatype, atomicity is
1466a1729e3fSJunchao Zhang    not guaranteed across the whole vertex. Therefore, this function is mostly only used with primitive types such as
1467a1729e3fSJunchao Zhang    integers.
1468a1729e3fSJunchao Zhang 
1469a1729e3fSJunchao Zhang .seealso: PetscSFComputeDegreeBegin(), PetscSFReduceBegin(), PetscSFSetGraph()
1470a1729e3fSJunchao Zhang @*/
1471a1729e3fSJunchao Zhang PetscErrorCode PetscSFFetchAndOpBegin(PetscSF sf,MPI_Datatype unit,void *rootdata,const void *leafdata,void *leafupdate,MPI_Op op)
1472a1729e3fSJunchao Zhang {
1473a1729e3fSJunchao Zhang   PetscErrorCode ierr;
1474eb02082bSJunchao Zhang   PetscMemType   rootmtype,leafmtype,leafupdatemtype;
1475a1729e3fSJunchao Zhang 
1476a1729e3fSJunchao Zhang   PetscFunctionBegin;
1477a1729e3fSJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
1478a1729e3fSJunchao Zhang   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
1479a1729e3fSJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_FetchAndOpBegin,sf,0,0,0);CHKERRQ(ierr);
1480eb02082bSJunchao Zhang   ierr = PetscGetMemType(rootdata,&rootmtype);CHKERRQ(ierr);
1481eb02082bSJunchao Zhang   ierr = PetscGetMemType(leafdata,&leafmtype);CHKERRQ(ierr);
1482eb02082bSJunchao Zhang   ierr = PetscGetMemType(leafupdate,&leafupdatemtype);CHKERRQ(ierr);
1483eb02082bSJunchao Zhang   if (leafmtype != leafupdatemtype) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support for leafdata and leafupdate in different memory types");
1484eb02082bSJunchao Zhang   ierr = (*sf->ops->FetchAndOpBegin)(sf,unit,rootmtype,rootdata,leafmtype,leafdata,leafupdate,op);CHKERRQ(ierr);
1485a1729e3fSJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_FetchAndOpBegin,sf,0,0,0);CHKERRQ(ierr);
1486a1729e3fSJunchao Zhang   PetscFunctionReturn(0);
1487a1729e3fSJunchao Zhang }
1488a1729e3fSJunchao Zhang 
1489a1729e3fSJunchao Zhang /*@C
1490a1729e3fSJunchao 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
1491a1729e3fSJunchao Zhang 
1492a1729e3fSJunchao Zhang    Collective
1493a1729e3fSJunchao Zhang 
1494a1729e3fSJunchao Zhang    Input Arguments:
1495a1729e3fSJunchao Zhang +  sf - star forest
1496a1729e3fSJunchao Zhang .  unit - data type
1497a1729e3fSJunchao Zhang .  leafdata - leaf values to use in reduction
1498a1729e3fSJunchao Zhang -  op - operation to use for reduction
1499a1729e3fSJunchao Zhang 
1500a1729e3fSJunchao Zhang    Output Arguments:
1501a1729e3fSJunchao Zhang +  rootdata - root values to be updated, input state is seen by first process to perform an update
1502a1729e3fSJunchao Zhang -  leafupdate - state at each leaf's respective root immediately prior to my atomic update
1503a1729e3fSJunchao Zhang 
1504a1729e3fSJunchao Zhang    Level: advanced
1505a1729e3fSJunchao Zhang 
1506a1729e3fSJunchao Zhang .seealso: PetscSFComputeDegreeEnd(), PetscSFReduceEnd(), PetscSFSetGraph()
1507a1729e3fSJunchao Zhang @*/
1508a1729e3fSJunchao Zhang PetscErrorCode PetscSFFetchAndOpEnd(PetscSF sf,MPI_Datatype unit,void *rootdata,const void *leafdata,void *leafupdate,MPI_Op op)
1509a1729e3fSJunchao Zhang {
1510a1729e3fSJunchao Zhang   PetscErrorCode ierr;
1511a1729e3fSJunchao Zhang 
1512a1729e3fSJunchao Zhang   PetscFunctionBegin;
1513a1729e3fSJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
1514a1729e3fSJunchao Zhang   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
1515a1729e3fSJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_FetchAndOpEnd,sf,0,0,0);CHKERRQ(ierr);
151600816365SJunchao Zhang   ierr = (*sf->ops->FetchAndOpEnd)(sf,unit,rootdata,leafdata,leafupdate,op);CHKERRQ(ierr);
1517a1729e3fSJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_FetchAndOpEnd,sf,0,0,0);CHKERRQ(ierr);
1518a1729e3fSJunchao Zhang   PetscFunctionReturn(0);
1519a1729e3fSJunchao Zhang }
1520a1729e3fSJunchao Zhang 
1521a1729e3fSJunchao Zhang /*@C
152295fce210SBarry Smith    PetscSFComputeDegreeBegin - begin computation of degree for each root vertex, to be completed with PetscSFComputeDegreeEnd()
152395fce210SBarry Smith 
152495fce210SBarry Smith    Collective
152595fce210SBarry Smith 
152695fce210SBarry Smith    Input Arguments:
152795fce210SBarry Smith .  sf - star forest
152895fce210SBarry Smith 
152995fce210SBarry Smith    Output Arguments:
153095fce210SBarry Smith .  degree - degree of each root vertex
153195fce210SBarry Smith 
153295fce210SBarry Smith    Level: advanced
153395fce210SBarry Smith 
1534ffe67aa5SVáclav Hapla    Notes:
1535ffe67aa5SVáclav Hapla    The returned array is owned by PetscSF and automatically freed by PetscSFDestroy(). Hence no need to call PetscFree() on it.
1536ffe67aa5SVáclav Hapla 
153795fce210SBarry Smith .seealso: PetscSFGatherBegin()
153895fce210SBarry Smith @*/
153995fce210SBarry Smith PetscErrorCode PetscSFComputeDegreeBegin(PetscSF sf,const PetscInt **degree)
154095fce210SBarry Smith {
154195fce210SBarry Smith   PetscErrorCode ierr;
154295fce210SBarry Smith 
154395fce210SBarry Smith   PetscFunctionBegin;
154495fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
154595fce210SBarry Smith   PetscSFCheckGraphSet(sf,1);
154695fce210SBarry Smith   PetscValidPointer(degree,2);
1547803bd9e8SMatthew G. Knepley   if (!sf->degreeknown) {
15485b0d146aSStefano Zampini     PetscInt i, nroots = sf->nroots, maxlocal;
1549803bd9e8SMatthew G. Knepley     if (sf->degree) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Calls to PetscSFComputeDegreeBegin() cannot be nested.");
15505b0d146aSStefano Zampini     maxlocal = sf->maxleaf-sf->minleaf+1;
155129046d53SLisandro Dalcin     ierr = PetscMalloc1(nroots,&sf->degree);CHKERRQ(ierr);
155229046d53SLisandro Dalcin     ierr = PetscMalloc1(PetscMax(maxlocal,1),&sf->degreetmp);CHKERRQ(ierr); /* allocate at least one entry, see check in PetscSFComputeDegreeEnd() */
155329046d53SLisandro Dalcin     for (i=0; i<nroots; i++) sf->degree[i] = 0;
15549837ea96SMatthew G. Knepley     for (i=0; i<maxlocal; i++) sf->degreetmp[i] = 1;
15555b0d146aSStefano Zampini     ierr = PetscSFReduceBegin(sf,MPIU_INT,sf->degreetmp-sf->minleaf,sf->degree,MPI_SUM);CHKERRQ(ierr);
155695fce210SBarry Smith   }
155795fce210SBarry Smith   *degree = NULL;
155895fce210SBarry Smith   PetscFunctionReturn(0);
155995fce210SBarry Smith }
156095fce210SBarry Smith 
156195fce210SBarry Smith /*@C
156295fce210SBarry Smith    PetscSFComputeDegreeEnd - complete computation of degree for each root vertex, started with PetscSFComputeDegreeBegin()
156395fce210SBarry Smith 
156495fce210SBarry Smith    Collective
156595fce210SBarry Smith 
156695fce210SBarry Smith    Input Arguments:
156795fce210SBarry Smith .  sf - star forest
156895fce210SBarry Smith 
156995fce210SBarry Smith    Output Arguments:
157095fce210SBarry Smith .  degree - degree of each root vertex
157195fce210SBarry Smith 
157295fce210SBarry Smith    Level: developer
157395fce210SBarry Smith 
1574ffe67aa5SVáclav Hapla    Notes:
1575ffe67aa5SVáclav Hapla    The returned array is owned by PetscSF and automatically freed by PetscSFDestroy(). Hence no need to call PetscFree() on it.
1576ffe67aa5SVáclav Hapla 
157795fce210SBarry Smith .seealso:
157895fce210SBarry Smith @*/
157995fce210SBarry Smith PetscErrorCode PetscSFComputeDegreeEnd(PetscSF sf,const PetscInt **degree)
158095fce210SBarry Smith {
158195fce210SBarry Smith   PetscErrorCode ierr;
158295fce210SBarry Smith 
158395fce210SBarry Smith   PetscFunctionBegin;
158495fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
158595fce210SBarry Smith   PetscSFCheckGraphSet(sf,1);
158629046d53SLisandro Dalcin   PetscValidPointer(degree,2);
158795fce210SBarry Smith   if (!sf->degreeknown) {
158829046d53SLisandro Dalcin     if (!sf->degreetmp) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFComputeDegreeBegin() before PetscSFComputeDegreeEnd()");
15895b0d146aSStefano Zampini     ierr = PetscSFReduceEnd(sf,MPIU_INT,sf->degreetmp-sf->minleaf,sf->degree,MPI_SUM);CHKERRQ(ierr);
159095fce210SBarry Smith     ierr = PetscFree(sf->degreetmp);CHKERRQ(ierr);
159195fce210SBarry Smith     sf->degreeknown = PETSC_TRUE;
159295fce210SBarry Smith   }
159395fce210SBarry Smith   *degree = sf->degree;
159495fce210SBarry Smith   PetscFunctionReturn(0);
159595fce210SBarry Smith }
159695fce210SBarry Smith 
1597673100f5SVaclav Hapla 
1598673100f5SVaclav Hapla /*@C
159966dfcd1aSVaclav Hapla    PetscSFComputeMultiRootOriginalNumbering - Returns original numbering of multi-roots (roots of multi-SF returned by PetscSFGetMultiSF()).
160066dfcd1aSVaclav Hapla    Each multi-root is assigned index of the corresponding original root.
1601673100f5SVaclav Hapla 
1602673100f5SVaclav Hapla    Collective
1603673100f5SVaclav Hapla 
1604673100f5SVaclav Hapla    Input Arguments:
1605673100f5SVaclav Hapla +  sf - star forest
1606673100f5SVaclav Hapla -  degree - degree of each root vertex, computed with PetscSFComputeDegreeBegin()/PetscSFComputeDegreeEnd()
1607673100f5SVaclav Hapla 
1608673100f5SVaclav Hapla    Output Arguments:
160966dfcd1aSVaclav Hapla +  nMultiRoots - (optional) number of multi-roots (roots of multi-SF)
161066dfcd1aSVaclav Hapla -  multiRootsOrigNumbering - original indices of multi-roots; length of this array is nMultiRoots
1611673100f5SVaclav Hapla 
1612673100f5SVaclav Hapla    Level: developer
1613673100f5SVaclav Hapla 
1614ffe67aa5SVáclav Hapla    Notes:
1615ffe67aa5SVáclav Hapla    The returned array multiRootsOrigNumbering is newly allocated and should be destroyed with PetscFree() when no longer needed.
1616ffe67aa5SVáclav Hapla 
1617673100f5SVaclav Hapla .seealso: PetscSFComputeDegreeBegin(), PetscSFComputeDegreeEnd(), PetscSFGetMultiSF()
1618673100f5SVaclav Hapla @*/
161966dfcd1aSVaclav Hapla PetscErrorCode PetscSFComputeMultiRootOriginalNumbering(PetscSF sf, const PetscInt degree[], PetscInt *nMultiRoots, PetscInt *multiRootsOrigNumbering[])
1620673100f5SVaclav Hapla {
1621673100f5SVaclav Hapla   PetscSF             msf;
1622673100f5SVaclav Hapla   PetscInt            i, j, k, nroots, nmroots;
1623673100f5SVaclav Hapla   PetscErrorCode      ierr;
1624673100f5SVaclav Hapla 
1625673100f5SVaclav Hapla   PetscFunctionBegin;
1626673100f5SVaclav Hapla   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
1627673100f5SVaclav Hapla   ierr = PetscSFGetGraph(sf, &nroots, NULL, NULL, NULL);CHKERRQ(ierr);
162829328920SVaclav Hapla   if (nroots) PetscValidIntPointer(degree,2);
162966dfcd1aSVaclav Hapla   if (nMultiRoots) PetscValidIntPointer(nMultiRoots,3);
163066dfcd1aSVaclav Hapla   PetscValidPointer(multiRootsOrigNumbering,4);
1631673100f5SVaclav Hapla   ierr = PetscSFGetMultiSF(sf,&msf);CHKERRQ(ierr);
1632673100f5SVaclav Hapla   ierr = PetscSFGetGraph(msf, &nmroots, NULL, NULL, NULL);CHKERRQ(ierr);
163366dfcd1aSVaclav Hapla   ierr = PetscMalloc1(nmroots, multiRootsOrigNumbering);CHKERRQ(ierr);
1634673100f5SVaclav Hapla   for (i=0,j=0,k=0; i<nroots; i++) {
1635673100f5SVaclav Hapla     if (!degree[i]) continue;
1636673100f5SVaclav Hapla     for (j=0; j<degree[i]; j++,k++) {
163766dfcd1aSVaclav Hapla       (*multiRootsOrigNumbering)[k] = i;
1638673100f5SVaclav Hapla     }
1639673100f5SVaclav Hapla   }
1640673100f5SVaclav Hapla #if defined(PETSC_USE_DEBUG)
1641673100f5SVaclav Hapla   if (PetscUnlikely(k != nmroots)) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"sanity check fail");
1642673100f5SVaclav Hapla #endif
164366dfcd1aSVaclav Hapla   if (nMultiRoots) *nMultiRoots = nmroots;
1644673100f5SVaclav Hapla   PetscFunctionReturn(0);
1645673100f5SVaclav Hapla }
1646673100f5SVaclav Hapla 
164795fce210SBarry Smith /*@C
164895fce210SBarry Smith    PetscSFGatherBegin - begin pointwise gather of all leaves into multi-roots, to be completed with PetscSFGatherEnd()
164995fce210SBarry Smith 
165095fce210SBarry Smith    Collective
165195fce210SBarry Smith 
165295fce210SBarry Smith    Input Arguments:
165395fce210SBarry Smith +  sf - star forest
165495fce210SBarry Smith .  unit - data type
165595fce210SBarry Smith -  leafdata - leaf data to gather to roots
165695fce210SBarry Smith 
165795fce210SBarry Smith    Output Argument:
165895fce210SBarry Smith .  multirootdata - root buffer to gather into, amount of space per root is equal to its degree
165995fce210SBarry Smith 
166095fce210SBarry Smith    Level: intermediate
166195fce210SBarry Smith 
166295fce210SBarry Smith .seealso: PetscSFComputeDegreeBegin(), PetscSFScatterBegin()
166395fce210SBarry Smith @*/
166495fce210SBarry Smith PetscErrorCode PetscSFGatherBegin(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *multirootdata)
166595fce210SBarry Smith {
166695fce210SBarry Smith   PetscErrorCode ierr;
1667a5526d50SJunchao Zhang   PetscSF        multi = NULL;
166895fce210SBarry Smith 
166995fce210SBarry Smith   PetscFunctionBegin;
167095fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
167129046d53SLisandro Dalcin   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
167295fce210SBarry Smith   ierr = PetscSFGetMultiSF(sf,&multi);CHKERRQ(ierr);
16738bfbc91cSJed Brown   ierr = PetscSFReduceBegin(multi,unit,leafdata,multirootdata,MPIU_REPLACE);CHKERRQ(ierr);
167495fce210SBarry Smith   PetscFunctionReturn(0);
167595fce210SBarry Smith }
167695fce210SBarry Smith 
167795fce210SBarry Smith /*@C
167895fce210SBarry Smith    PetscSFGatherEnd - ends pointwise gather operation that was started with PetscSFGatherBegin()
167995fce210SBarry Smith 
168095fce210SBarry Smith    Collective
168195fce210SBarry Smith 
168295fce210SBarry Smith    Input Arguments:
168395fce210SBarry Smith +  sf - star forest
168495fce210SBarry Smith .  unit - data type
168595fce210SBarry Smith -  leafdata - leaf data to gather to roots
168695fce210SBarry Smith 
168795fce210SBarry Smith    Output Argument:
168895fce210SBarry Smith .  multirootdata - root buffer to gather into, amount of space per root is equal to its degree
168995fce210SBarry Smith 
169095fce210SBarry Smith    Level: intermediate
169195fce210SBarry Smith 
169295fce210SBarry Smith .seealso: PetscSFComputeDegreeEnd(), PetscSFScatterEnd()
169395fce210SBarry Smith @*/
169495fce210SBarry Smith PetscErrorCode PetscSFGatherEnd(PetscSF sf,MPI_Datatype unit,const void *leafdata,void *multirootdata)
169595fce210SBarry Smith {
169695fce210SBarry Smith   PetscErrorCode ierr;
1697a5526d50SJunchao Zhang   PetscSF        multi = NULL;
169895fce210SBarry Smith 
169995fce210SBarry Smith   PetscFunctionBegin;
170095fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
170195fce210SBarry Smith   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
170295fce210SBarry Smith   ierr = PetscSFGetMultiSF(sf,&multi);CHKERRQ(ierr);
17038bfbc91cSJed Brown   ierr = PetscSFReduceEnd(multi,unit,leafdata,multirootdata,MPIU_REPLACE);CHKERRQ(ierr);
170495fce210SBarry Smith   PetscFunctionReturn(0);
170595fce210SBarry Smith }
170695fce210SBarry Smith 
170795fce210SBarry Smith /*@C
170895fce210SBarry Smith    PetscSFScatterBegin - begin pointwise scatter operation from multi-roots to leaves, to be completed with PetscSFScatterEnd()
170995fce210SBarry Smith 
171095fce210SBarry Smith    Collective
171195fce210SBarry Smith 
171295fce210SBarry Smith    Input Arguments:
171395fce210SBarry Smith +  sf - star forest
171495fce210SBarry Smith .  unit - data type
171595fce210SBarry Smith -  multirootdata - root buffer to send to each leaf, one unit of data per leaf
171695fce210SBarry Smith 
171795fce210SBarry Smith    Output Argument:
171895fce210SBarry Smith .  leafdata - leaf data to be update with personal data from each respective root
171995fce210SBarry Smith 
172095fce210SBarry Smith    Level: intermediate
172195fce210SBarry Smith 
172295fce210SBarry Smith .seealso: PetscSFComputeDegreeBegin(), PetscSFScatterBegin()
172395fce210SBarry Smith @*/
172495fce210SBarry Smith PetscErrorCode PetscSFScatterBegin(PetscSF sf,MPI_Datatype unit,const void *multirootdata,void *leafdata)
172595fce210SBarry Smith {
172695fce210SBarry Smith   PetscErrorCode ierr;
1727a5526d50SJunchao Zhang   PetscSF        multi = NULL;
172895fce210SBarry Smith 
172995fce210SBarry Smith   PetscFunctionBegin;
173095fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
173195fce210SBarry Smith   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
173295fce210SBarry Smith   ierr = PetscSFGetMultiSF(sf,&multi);CHKERRQ(ierr);
173395fce210SBarry Smith   ierr = PetscSFBcastBegin(multi,unit,multirootdata,leafdata);CHKERRQ(ierr);
173495fce210SBarry Smith   PetscFunctionReturn(0);
173595fce210SBarry Smith }
173695fce210SBarry Smith 
173795fce210SBarry Smith /*@C
173895fce210SBarry Smith    PetscSFScatterEnd - ends pointwise scatter operation that was started with PetscSFScatterBegin()
173995fce210SBarry Smith 
174095fce210SBarry Smith    Collective
174195fce210SBarry Smith 
174295fce210SBarry Smith    Input Arguments:
174395fce210SBarry Smith +  sf - star forest
174495fce210SBarry Smith .  unit - data type
174595fce210SBarry Smith -  multirootdata - root buffer to send to each leaf, one unit of data per leaf
174695fce210SBarry Smith 
174795fce210SBarry Smith    Output Argument:
174895fce210SBarry Smith .  leafdata - leaf data to be update with personal data from each respective root
174995fce210SBarry Smith 
175095fce210SBarry Smith    Level: intermediate
175195fce210SBarry Smith 
175295fce210SBarry Smith .seealso: PetscSFComputeDegreeEnd(), PetscSFScatterEnd()
175395fce210SBarry Smith @*/
175495fce210SBarry Smith PetscErrorCode PetscSFScatterEnd(PetscSF sf,MPI_Datatype unit,const void *multirootdata,void *leafdata)
175595fce210SBarry Smith {
175695fce210SBarry Smith   PetscErrorCode ierr;
1757a5526d50SJunchao Zhang   PetscSF        multi = NULL;
175895fce210SBarry Smith 
175995fce210SBarry Smith   PetscFunctionBegin;
176095fce210SBarry Smith   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
176195fce210SBarry Smith   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
176295fce210SBarry Smith   ierr = PetscSFGetMultiSF(sf,&multi);CHKERRQ(ierr);
176395fce210SBarry Smith   ierr = PetscSFBcastEnd(multi,unit,multirootdata,leafdata);CHKERRQ(ierr);
176495fce210SBarry Smith   PetscFunctionReturn(0);
176595fce210SBarry Smith }
1766a7b3aa13SAta Mesgarnejad 
1767a072220fSLawrence Mitchell static PetscErrorCode PetscSFCheckLeavesUnique_Private(PetscSF sf)
1768a072220fSLawrence Mitchell {
1769a072220fSLawrence Mitchell #if defined(PETSC_USE_DEBUG)
1770a072220fSLawrence Mitchell   PetscInt        i, n, nleaves;
1771a072220fSLawrence Mitchell   const PetscInt *ilocal = NULL;
1772a072220fSLawrence Mitchell   PetscHSetI      seen;
1773a072220fSLawrence Mitchell   PetscErrorCode  ierr;
1774a072220fSLawrence Mitchell 
1775a072220fSLawrence Mitchell   PetscFunctionBegin;
1776a072220fSLawrence Mitchell   ierr = PetscSFGetGraph(sf,NULL,&nleaves,&ilocal,NULL);CHKERRQ(ierr);
1777a072220fSLawrence Mitchell   ierr = PetscHSetICreate(&seen);CHKERRQ(ierr);
1778a072220fSLawrence Mitchell   for (i = 0; i < nleaves; i++) {
1779a072220fSLawrence Mitchell     const PetscInt leaf = ilocal ? ilocal[i] : i;
1780a072220fSLawrence Mitchell     ierr = PetscHSetIAdd(seen,leaf);CHKERRQ(ierr);
1781a072220fSLawrence Mitchell   }
1782a072220fSLawrence Mitchell   ierr = PetscHSetIGetSize(seen,&n);CHKERRQ(ierr);
1783a072220fSLawrence Mitchell   if (n != nleaves) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Provided leaves have repeated values: all leaves must be unique");
1784a072220fSLawrence Mitchell   ierr = PetscHSetIDestroy(&seen);CHKERRQ(ierr);
1785a072220fSLawrence Mitchell   PetscFunctionReturn(0);
1786a072220fSLawrence Mitchell #else
1787a072220fSLawrence Mitchell   PetscFunctionBegin;
1788a072220fSLawrence Mitchell   PetscFunctionReturn(0);
1789a072220fSLawrence Mitchell #endif
1790a072220fSLawrence Mitchell }
179154729392SStefano Zampini 
1792a7b3aa13SAta Mesgarnejad /*@
179304c0ada0SJunchao Zhang   PetscSFCompose - Compose a new PetscSF by putting the second SF under the first one in a top (roots) down (leaves) view
1794a7b3aa13SAta Mesgarnejad 
1795a7b3aa13SAta Mesgarnejad   Input Parameters:
179654729392SStefano Zampini + sfA - The first PetscSF
179754729392SStefano Zampini - sfB - The second PetscSF
1798a7b3aa13SAta Mesgarnejad 
1799a7b3aa13SAta Mesgarnejad   Output Parameters:
180054729392SStefano Zampini . sfBA - The composite SF
1801a7b3aa13SAta Mesgarnejad 
1802a7b3aa13SAta Mesgarnejad   Level: developer
1803a7b3aa13SAta Mesgarnejad 
1804a072220fSLawrence Mitchell   Notes:
180554729392SStefano Zampini   Currently, the two SFs must be defined on congruent communicators and they must be true star
180654729392SStefano Zampini   forests, i.e. the same leaf is not connected with different roots.
180754729392SStefano Zampini 
180854729392SStefano Zampini   sfA's leaf space and sfB's root space might be partially overlapped. The composition builds
180954729392SStefano Zampini   a graph with sfA's roots and sfB's leaves only when there is a path between them. Unconnected
181054729392SStefano Zampini   nodes (roots or leaves) are not in sfBA. Doing a Bcast on the new SF is equivalent to doing a
181154729392SStefano Zampini   Bcast on sfA, then a Bcast on sfB, on connected nodes.
1812a072220fSLawrence Mitchell 
181304c0ada0SJunchao Zhang .seealso: PetscSF, PetscSFComposeInverse(), PetscSFGetGraph(), PetscSFSetGraph()
1814a7b3aa13SAta Mesgarnejad @*/
1815a7b3aa13SAta Mesgarnejad PetscErrorCode PetscSFCompose(PetscSF sfA,PetscSF sfB,PetscSF *sfBA)
1816a7b3aa13SAta Mesgarnejad {
181704c0ada0SJunchao Zhang   PetscErrorCode    ierr;
1818a7b3aa13SAta Mesgarnejad   const PetscSFNode *remotePointsA,*remotePointsB;
1819d41018fbSJunchao Zhang   PetscSFNode       *remotePointsBA=NULL,*reorderedRemotePointsA = NULL,*leafdataB;
182054729392SStefano Zampini   const PetscInt    *localPointsA,*localPointsB;
182154729392SStefano Zampini   PetscInt          *localPointsBA;
182254729392SStefano Zampini   PetscInt          i,numRootsA,numLeavesA,numRootsB,numLeavesB,minleaf,maxleaf,numLeavesBA;
182354729392SStefano Zampini   PetscBool         denseB;
1824a7b3aa13SAta Mesgarnejad 
1825a7b3aa13SAta Mesgarnejad   PetscFunctionBegin;
1826a7b3aa13SAta Mesgarnejad   PetscValidHeaderSpecific(sfA,PETSCSF_CLASSID,1);
182729046d53SLisandro Dalcin   PetscSFCheckGraphSet(sfA,1);
182829046d53SLisandro Dalcin   PetscValidHeaderSpecific(sfB,PETSCSF_CLASSID,2);
182929046d53SLisandro Dalcin   PetscSFCheckGraphSet(sfB,2);
183054729392SStefano Zampini   PetscCheckSameComm(sfA,1,sfB,2);
183129046d53SLisandro Dalcin   PetscValidPointer(sfBA,3);
183254729392SStefano Zampini   ierr = PetscSFCheckLeavesUnique_Private(sfA);CHKERRQ(ierr);
183354729392SStefano Zampini   ierr = PetscSFCheckLeavesUnique_Private(sfB);CHKERRQ(ierr);
183454729392SStefano Zampini 
1835a7b3aa13SAta Mesgarnejad   ierr = PetscSFGetGraph(sfA,&numRootsA,&numLeavesA,&localPointsA,&remotePointsA);CHKERRQ(ierr);
1836a7b3aa13SAta Mesgarnejad   ierr = PetscSFGetGraph(sfB,&numRootsB,&numLeavesB,&localPointsB,&remotePointsB);CHKERRQ(ierr);
1837d41018fbSJunchao Zhang   if (localPointsA) {
183854729392SStefano Zampini     ierr = PetscMalloc1(numRootsB,&reorderedRemotePointsA);CHKERRQ(ierr);
183954729392SStefano Zampini     for (i=0; i<numRootsB; i++) {
184054729392SStefano Zampini       reorderedRemotePointsA[i].rank = -1;
184154729392SStefano Zampini       reorderedRemotePointsA[i].index = -1;
184254729392SStefano Zampini     }
184354729392SStefano Zampini     for (i=0; i<numLeavesA; i++) {
184454729392SStefano Zampini       if (localPointsA[i] >= numRootsB) continue;
184554729392SStefano Zampini       reorderedRemotePointsA[localPointsA[i]] = remotePointsA[i];
184654729392SStefano Zampini     }
1847d41018fbSJunchao Zhang     remotePointsA = reorderedRemotePointsA;
1848d41018fbSJunchao Zhang   }
1849d41018fbSJunchao Zhang   ierr = PetscSFGetLeafRange(sfB,&minleaf,&maxleaf);CHKERRQ(ierr);
1850d41018fbSJunchao Zhang   ierr = PetscMalloc1(maxleaf-minleaf+1,&leafdataB);CHKERRQ(ierr);
1851d41018fbSJunchao Zhang   ierr = PetscSFBcastBegin(sfB,MPIU_2INT,remotePointsA,leafdataB-minleaf);CHKERRQ(ierr);
1852d41018fbSJunchao Zhang   ierr = PetscSFBcastEnd(sfB,MPIU_2INT,remotePointsA,leafdataB-minleaf);CHKERRQ(ierr);
1853d41018fbSJunchao Zhang   ierr = PetscFree(reorderedRemotePointsA);CHKERRQ(ierr);
1854d41018fbSJunchao Zhang 
185554729392SStefano Zampini   denseB = (PetscBool)!localPointsB;
185654729392SStefano Zampini   for (i=0,numLeavesBA=0; i<numLeavesB; i++) {
185754729392SStefano Zampini     if (leafdataB[localPointsB ? localPointsB[i]-minleaf : i].rank == -1) denseB = PETSC_FALSE;
185854729392SStefano Zampini     else numLeavesBA++;
185954729392SStefano Zampini   }
186054729392SStefano Zampini   if (denseB) {
1861d41018fbSJunchao Zhang     localPointsBA  = NULL;
1862d41018fbSJunchao Zhang     remotePointsBA = leafdataB;
1863d41018fbSJunchao Zhang   } else {
186454729392SStefano Zampini     ierr = PetscMalloc1(numLeavesBA,&localPointsBA);CHKERRQ(ierr);
186554729392SStefano Zampini     ierr = PetscMalloc1(numLeavesBA,&remotePointsBA);CHKERRQ(ierr);
186654729392SStefano Zampini     for (i=0,numLeavesBA=0; i<numLeavesB; i++) {
186754729392SStefano Zampini       const PetscInt l = localPointsB ? localPointsB[i] : i;
186854729392SStefano Zampini 
186954729392SStefano Zampini       if (leafdataB[l-minleaf].rank == -1) continue;
187054729392SStefano Zampini       remotePointsBA[numLeavesBA] = leafdataB[l-minleaf];
187154729392SStefano Zampini       localPointsBA[numLeavesBA] = l;
187254729392SStefano Zampini       numLeavesBA++;
187354729392SStefano Zampini     }
1874d41018fbSJunchao Zhang     ierr = PetscFree(leafdataB);CHKERRQ(ierr);
1875d41018fbSJunchao Zhang   }
187654729392SStefano Zampini   ierr = PetscSFCreate(PetscObjectComm((PetscObject)sfA),sfBA);CHKERRQ(ierr);
187754729392SStefano Zampini   ierr = PetscSFSetGraph(*sfBA,numRootsA,numLeavesBA,localPointsBA,PETSC_OWN_POINTER,remotePointsBA,PETSC_OWN_POINTER);CHKERRQ(ierr);
1878a7b3aa13SAta Mesgarnejad   PetscFunctionReturn(0);
1879a7b3aa13SAta Mesgarnejad }
18801c6ba672SJunchao Zhang 
188104c0ada0SJunchao Zhang /*@
188254729392SStefano Zampini   PetscSFComposeInverse - Compose a new PetscSF by putting the inverse of the second SF under the first one
188304c0ada0SJunchao Zhang 
188404c0ada0SJunchao Zhang   Input Parameters:
188554729392SStefano Zampini + sfA - The first PetscSF
188654729392SStefano Zampini - sfB - The second PetscSF
188704c0ada0SJunchao Zhang 
188804c0ada0SJunchao Zhang   Output Parameters:
188954729392SStefano Zampini . sfBA - The composite SF.
189004c0ada0SJunchao Zhang 
189104c0ada0SJunchao Zhang   Level: developer
189204c0ada0SJunchao Zhang 
189354729392SStefano Zampini   Notes:
189454729392SStefano Zampini   Currently, the two SFs must be defined on congruent communicators and they must be true star
189554729392SStefano Zampini   forests, i.e. the same leaf is not connected with different roots. Even more, all roots of the
189654729392SStefano Zampini   second SF must have a degree of 1, i.e., no roots have more than one leaf connected.
189754729392SStefano Zampini 
189854729392SStefano Zampini   sfA's leaf space and sfB's leaf space might be partially overlapped. The composition builds
189954729392SStefano Zampini   a graph with sfA's roots and sfB's roots only when there is a path between them. Unconnected
190054729392SStefano Zampini   roots are not in sfBA. Doing a Bcast on the new SF is equivalent to doing a Bcast on sfA, then
190154729392SStefano Zampini   a Reduce on sfB, on connected roots.
190254729392SStefano Zampini 
190354729392SStefano Zampini .seealso: PetscSF, PetscSFCompose(), PetscSFGetGraph(), PetscSFSetGraph(), PetscSFCreateInverseSF()
190404c0ada0SJunchao Zhang @*/
190504c0ada0SJunchao Zhang PetscErrorCode PetscSFComposeInverse(PetscSF sfA,PetscSF sfB,PetscSF *sfBA)
190604c0ada0SJunchao Zhang {
190704c0ada0SJunchao Zhang   PetscErrorCode    ierr;
190804c0ada0SJunchao Zhang   const PetscSFNode *remotePointsA,*remotePointsB;
190904c0ada0SJunchao Zhang   PetscSFNode       *remotePointsBA;
191004c0ada0SJunchao Zhang   const PetscInt    *localPointsA,*localPointsB;
191154729392SStefano Zampini   PetscSFNode       *reorderedRemotePointsA = NULL;
191254729392SStefano Zampini   PetscInt          i,numRootsA,numLeavesA,numLeavesBA,numRootsB,numLeavesB,minleaf,maxleaf,*localPointsBA;
19135b0d146aSStefano Zampini   MPI_Op            op;
19145b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES)
19155b0d146aSStefano Zampini   PetscBool         iswin;
19165b0d146aSStefano Zampini #endif
191704c0ada0SJunchao Zhang 
191804c0ada0SJunchao Zhang   PetscFunctionBegin;
191904c0ada0SJunchao Zhang   PetscValidHeaderSpecific(sfA,PETSCSF_CLASSID,1);
192004c0ada0SJunchao Zhang   PetscSFCheckGraphSet(sfA,1);
192104c0ada0SJunchao Zhang   PetscValidHeaderSpecific(sfB,PETSCSF_CLASSID,2);
192204c0ada0SJunchao Zhang   PetscSFCheckGraphSet(sfB,2);
192354729392SStefano Zampini   PetscCheckSameComm(sfA,1,sfB,2);
192404c0ada0SJunchao Zhang   PetscValidPointer(sfBA,3);
192554729392SStefano Zampini   ierr = PetscSFCheckLeavesUnique_Private(sfA);CHKERRQ(ierr);
192654729392SStefano Zampini   ierr = PetscSFCheckLeavesUnique_Private(sfB);CHKERRQ(ierr);
192754729392SStefano Zampini 
192804c0ada0SJunchao Zhang   ierr = PetscSFGetGraph(sfA, &numRootsA, &numLeavesA, &localPointsA, &remotePointsA);CHKERRQ(ierr);
192904c0ada0SJunchao Zhang   ierr = PetscSFGetGraph(sfB, &numRootsB, &numLeavesB, &localPointsB, &remotePointsB);CHKERRQ(ierr);
19305b0d146aSStefano Zampini 
19315b0d146aSStefano Zampini   /* TODO: Check roots of sfB have degree of 1 */
19325b0d146aSStefano Zampini   /* Once we implement it, we can replace the MPI_MAXLOC
19335b0d146aSStefano Zampini      with MPIU_REPLACE. In that case, MPI_MAXLOC and MPIU_REPLACE have the same effect.
19345b0d146aSStefano Zampini      We use MPI_MAXLOC only to have a deterministic output from this routine if
19355b0d146aSStefano Zampini      the root condition is not meet.
19365b0d146aSStefano Zampini    */
19375b0d146aSStefano Zampini   op = MPI_MAXLOC;
19385b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES)
19395b0d146aSStefano Zampini   /* we accept a non-deterministic output (if any) with PETSCSFWINDOW, since MPI_MAXLOC cannot operate on MPIU_2INT with MPI_Accumulate */
19405b0d146aSStefano Zampini   ierr = PetscObjectTypeCompare((PetscObject)sfB,PETSCSFWINDOW,&iswin);CHKERRQ(ierr);
19415b0d146aSStefano Zampini   if (iswin) op = MPIU_REPLACE;
19425b0d146aSStefano Zampini #endif
19435b0d146aSStefano Zampini 
194454729392SStefano Zampini   ierr = PetscSFGetLeafRange(sfB, &minleaf, &maxleaf);CHKERRQ(ierr);
194554729392SStefano Zampini   ierr = PetscMalloc1(maxleaf - minleaf + 1,&reorderedRemotePointsA);CHKERRQ(ierr);
194654729392SStefano Zampini   for (i=0; i<maxleaf - minleaf + 1; i++) {
194754729392SStefano Zampini     reorderedRemotePointsA[i].rank = -1;
194854729392SStefano Zampini     reorderedRemotePointsA[i].index = -1;
194954729392SStefano Zampini   }
195054729392SStefano Zampini   if (localPointsA) {
195154729392SStefano Zampini     for (i=0; i<numLeavesA; i++) {
195254729392SStefano Zampini       if (localPointsA[i] > maxleaf || localPointsA[i] < minleaf) continue;
195354729392SStefano Zampini       reorderedRemotePointsA[localPointsA[i] - minleaf] = remotePointsA[i];
195454729392SStefano Zampini     }
195554729392SStefano Zampini   } else {
195654729392SStefano Zampini     for (i=0; i<numLeavesA; i++) {
195754729392SStefano Zampini       if (i > maxleaf || i < minleaf) continue;
195854729392SStefano Zampini       reorderedRemotePointsA[i - minleaf] = remotePointsA[i];
195954729392SStefano Zampini     }
196054729392SStefano Zampini   }
196154729392SStefano Zampini 
196254729392SStefano Zampini   ierr = PetscMalloc1(numRootsB,&localPointsBA);CHKERRQ(ierr);
196304c0ada0SJunchao Zhang   ierr = PetscMalloc1(numRootsB,&remotePointsBA);CHKERRQ(ierr);
196454729392SStefano Zampini   for (i=0; i<numRootsB; i++) {
196554729392SStefano Zampini     remotePointsBA[i].rank = -1;
196654729392SStefano Zampini     remotePointsBA[i].index = -1;
196754729392SStefano Zampini   }
196854729392SStefano Zampini 
19695b0d146aSStefano Zampini   ierr = PetscSFReduceBegin(sfB,MPIU_2INT,reorderedRemotePointsA-minleaf,remotePointsBA,op);CHKERRQ(ierr);
19705b0d146aSStefano Zampini   ierr = PetscSFReduceEnd(sfB,MPIU_2INT,reorderedRemotePointsA-minleaf,remotePointsBA,op);CHKERRQ(ierr);
197154729392SStefano Zampini   ierr = PetscFree(reorderedRemotePointsA);CHKERRQ(ierr);
197254729392SStefano Zampini   for (i=0,numLeavesBA=0; i<numRootsB; i++) {
197354729392SStefano Zampini     if (remotePointsBA[i].rank == -1) continue;
197454729392SStefano Zampini     remotePointsBA[numLeavesBA].rank = remotePointsBA[i].rank;
197554729392SStefano Zampini     remotePointsBA[numLeavesBA].index = remotePointsBA[i].index;
197654729392SStefano Zampini     localPointsBA[numLeavesBA] = i;
197754729392SStefano Zampini     numLeavesBA++;
197854729392SStefano Zampini   }
197954729392SStefano Zampini   ierr = PetscSFCreate(PetscObjectComm((PetscObject)sfA),sfBA);CHKERRQ(ierr);
198054729392SStefano Zampini   ierr = PetscSFSetGraph(*sfBA,numRootsA,numLeavesBA,localPointsBA,PETSC_OWN_POINTER,remotePointsBA,PETSC_OWN_POINTER);CHKERRQ(ierr);
198104c0ada0SJunchao Zhang   PetscFunctionReturn(0);
198204c0ada0SJunchao Zhang }
198304c0ada0SJunchao Zhang 
19841c6ba672SJunchao Zhang /*
19851c6ba672SJunchao Zhang   PetscSFCreateLocalSF_Private - Creates a local PetscSF that only has intra-process edges of the global PetscSF
19861c6ba672SJunchao Zhang 
19871c6ba672SJunchao Zhang   Input Parameters:
19881c6ba672SJunchao Zhang . sf - The global PetscSF
19891c6ba672SJunchao Zhang 
19901c6ba672SJunchao Zhang   Output Parameters:
19911c6ba672SJunchao Zhang . out - The local PetscSF
19921c6ba672SJunchao Zhang  */
19931c6ba672SJunchao Zhang PetscErrorCode PetscSFCreateLocalSF_Private(PetscSF sf,PetscSF *out)
19941c6ba672SJunchao Zhang {
19951c6ba672SJunchao Zhang   MPI_Comm           comm;
19961c6ba672SJunchao Zhang   PetscMPIInt        myrank;
19971c6ba672SJunchao Zhang   const PetscInt     *ilocal;
19981c6ba672SJunchao Zhang   const PetscSFNode  *iremote;
19991c6ba672SJunchao Zhang   PetscInt           i,j,nroots,nleaves,lnleaves,*lilocal;
20001c6ba672SJunchao Zhang   PetscSFNode        *liremote;
20011c6ba672SJunchao Zhang   PetscSF            lsf;
20021c6ba672SJunchao Zhang   PetscErrorCode     ierr;
20031c6ba672SJunchao Zhang 
20041c6ba672SJunchao Zhang   PetscFunctionBegin;
20051c6ba672SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
20061c6ba672SJunchao Zhang   if (sf->ops->CreateLocalSF) {
20071c6ba672SJunchao Zhang     ierr = (*sf->ops->CreateLocalSF)(sf,out);CHKERRQ(ierr);
20081c6ba672SJunchao Zhang   } else {
20091c6ba672SJunchao Zhang     /* Could use PetscSFCreateEmbeddedLeafSF, but since we know the comm is PETSC_COMM_SELF, we can make it fast */
20101c6ba672SJunchao Zhang     ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr);
20111c6ba672SJunchao Zhang     ierr = MPI_Comm_rank(comm,&myrank);CHKERRQ(ierr);
20121c6ba672SJunchao Zhang 
20131c6ba672SJunchao Zhang     /* Find out local edges and build a local SF */
20141c6ba672SJunchao Zhang     ierr = PetscSFGetGraph(sf,&nroots,&nleaves,&ilocal,&iremote);CHKERRQ(ierr);
20151c6ba672SJunchao Zhang     for (i=lnleaves=0; i<nleaves; i++) {if (iremote[i].rank == (PetscInt)myrank) lnleaves++;}
20161c6ba672SJunchao Zhang     ierr = PetscMalloc1(lnleaves,&lilocal);CHKERRQ(ierr);
20171c6ba672SJunchao Zhang     ierr = PetscMalloc1(lnleaves,&liremote);CHKERRQ(ierr);
20181c6ba672SJunchao Zhang 
20191c6ba672SJunchao Zhang     for (i=j=0; i<nleaves; i++) {
20201c6ba672SJunchao Zhang       if (iremote[i].rank == (PetscInt)myrank) {
20211c6ba672SJunchao Zhang         lilocal[j]        = ilocal? ilocal[i] : i; /* ilocal=NULL for contiguous storage */
20221c6ba672SJunchao Zhang         liremote[j].rank  = 0; /* rank in PETSC_COMM_SELF */
20231c6ba672SJunchao Zhang         liremote[j].index = iremote[i].index;
20241c6ba672SJunchao Zhang         j++;
20251c6ba672SJunchao Zhang       }
20261c6ba672SJunchao Zhang     }
20271c6ba672SJunchao Zhang     ierr = PetscSFCreate(PETSC_COMM_SELF,&lsf);CHKERRQ(ierr);
20281c6ba672SJunchao Zhang     ierr = PetscSFSetGraph(lsf,nroots,lnleaves,lilocal,PETSC_OWN_POINTER,liremote,PETSC_OWN_POINTER);CHKERRQ(ierr);
20291c6ba672SJunchao Zhang     ierr = PetscSFSetUp(lsf);CHKERRQ(ierr);
20301c6ba672SJunchao Zhang     *out = lsf;
20311c6ba672SJunchao Zhang   }
20321c6ba672SJunchao Zhang   PetscFunctionReturn(0);
20331c6ba672SJunchao Zhang }
2034dd5b3ca6SJunchao Zhang 
2035dd5b3ca6SJunchao Zhang /* Similar to PetscSFBcast, but only Bcast to leaves on rank 0 */
2036dd5b3ca6SJunchao Zhang PetscErrorCode PetscSFBcastToZero_Private(PetscSF sf,MPI_Datatype unit,const void *rootdata,void *leafdata)
2037dd5b3ca6SJunchao Zhang {
2038dd5b3ca6SJunchao Zhang   PetscErrorCode     ierr;
2039eb02082bSJunchao Zhang   PetscMemType       rootmtype,leafmtype;
2040dd5b3ca6SJunchao Zhang 
2041dd5b3ca6SJunchao Zhang   PetscFunctionBegin;
2042dd5b3ca6SJunchao Zhang   PetscValidHeaderSpecific(sf,PETSCSF_CLASSID,1);
2043dd5b3ca6SJunchao Zhang   ierr = PetscSFSetUp(sf);CHKERRQ(ierr);
2044dd5b3ca6SJunchao Zhang   ierr = PetscLogEventBegin(PETSCSF_BcastAndOpBegin,sf,0,0,0);CHKERRQ(ierr);
2045eb02082bSJunchao Zhang   ierr = PetscGetMemType(rootdata,&rootmtype);CHKERRQ(ierr);
2046eb02082bSJunchao Zhang   ierr = PetscGetMemType(leafdata,&leafmtype);CHKERRQ(ierr);
2047dd5b3ca6SJunchao Zhang   if (sf->ops->BcastToZero) {
2048eb02082bSJunchao Zhang     ierr = (*sf->ops->BcastToZero)(sf,unit,rootmtype,rootdata,leafmtype,leafdata);CHKERRQ(ierr);
2049dd5b3ca6SJunchao Zhang   } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"PetscSFBcastToZero_Private is not supported on this SF type");
2050dd5b3ca6SJunchao Zhang   ierr = PetscLogEventEnd(PETSCSF_BcastAndOpBegin,sf,0,0,0);CHKERRQ(ierr);
2051dd5b3ca6SJunchao Zhang   PetscFunctionReturn(0);
2052dd5b3ca6SJunchao Zhang }
2053dd5b3ca6SJunchao Zhang 
2054