1af0996ceSBarry Smith #include <petsc/private/sfimpl.h> /*I "petscsf.h" I*/ 2c4e6a40aSLawrence Mitchell #include <petsc/private/hashseti.h> 353dd6d7dSJunchao Zhang #include <petsc/private/viewerimpl.h> 4eec179cfSJacob Faibussowitsch #include <petsc/private/hashmapi.h> 595fce210SBarry Smith 67fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_CUDA) 77fd2d3dbSJunchao Zhang #include <cuda_runtime.h> 8715b587bSJunchao Zhang #include <petscdevice_cuda.h> 97fd2d3dbSJunchao Zhang #endif 107fd2d3dbSJunchao Zhang 117fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_HIP) 127fd2d3dbSJunchao Zhang #include <hip/hip_runtime.h> 137fd2d3dbSJunchao Zhang #endif 147fd2d3dbSJunchao Zhang 152abc8c78SJacob Faibussowitsch #if defined(PETSC_CLANG_STATIC_ANALYZER) 162abc8c78SJacob Faibussowitsch void PetscSFCheckGraphSet(PetscSF, int); 172abc8c78SJacob Faibussowitsch #else 1895fce210SBarry Smith #if defined(PETSC_USE_DEBUG) 19a8f51744SPierre Jolivet #define PetscSFCheckGraphSet(sf, arg) PetscCheck((sf)->graphset, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetGraph() or PetscSFSetGraphWithPattern() on argument %d \"%s\" before %s()", (arg), #sf, PETSC_FUNCTION_NAME) 2095fce210SBarry Smith #else 219371c9d4SSatish Balay #define PetscSFCheckGraphSet(sf, arg) \ 229371c9d4SSatish Balay do { \ 239371c9d4SSatish Balay } while (0) 2495fce210SBarry Smith #endif 252abc8c78SJacob Faibussowitsch #endif 2695fce210SBarry Smith 274c8fdceaSLisandro Dalcin const char *const PetscSFDuplicateOptions[] = {"CONFONLY", "RANKS", "GRAPH", "PetscSFDuplicateOption", "PETSCSF_DUPLICATE_", NULL}; 281f40158dSVaclav Hapla const char *const PetscSFConcatenateRootModes[] = {"local", "shared", "global", "PetscSFConcatenateRootMode", "PETSCSF_CONCATENATE_ROOTMODE_", NULL}; 2995fce210SBarry Smith 308af6ec1cSBarry Smith /*@ 3195fce210SBarry Smith PetscSFCreate - create a star forest communication context 3295fce210SBarry Smith 33d083f849SBarry Smith Collective 3495fce210SBarry Smith 354165533cSJose E. Roman Input Parameter: 3695fce210SBarry Smith . comm - communicator on which the star forest will operate 3795fce210SBarry Smith 384165533cSJose E. Roman Output Parameter: 3995fce210SBarry Smith . sf - new star forest context 4095fce210SBarry Smith 4120662ed9SBarry Smith Options Database Key: 42cab54364SBarry Smith . -sf_type type - value of type may be 43cab54364SBarry Smith .vb 44cab54364SBarry Smith basic -Use MPI persistent Isend/Irecv for communication (Default) 45cab54364SBarry Smith window -Use MPI-3 one-sided window for communication 46cab54364SBarry Smith neighbor -Use MPI-3 neighborhood collectives for communication 47cab54364SBarry Smith .ve 48dd5b3ca6SJunchao Zhang 4995fce210SBarry Smith Level: intermediate 5095fce210SBarry Smith 51cab54364SBarry Smith Note: 52cab54364SBarry Smith When one knows the communication graph is one of the predefined graph, such as `MPI_Alltoall()`, `MPI_Allgatherv()`, 53cab54364SBarry Smith `MPI_Gatherv()`, one can create a `PetscSF` and then set its graph with `PetscSFSetGraphWithPattern()`. These special 5420662ed9SBarry Smith `SF`s are optimized and they have better performance than the general `SF`s. 55dd5b3ca6SJunchao Zhang 5638b5cf2dSJacob Faibussowitsch .seealso: `PetscSF`, `PetscSFSetType`, `PetscSFSetGraph()`, `PetscSFSetGraphWithPattern()`, `PetscSFDestroy()` 5795fce210SBarry Smith @*/ 58d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreate(MPI_Comm comm, PetscSF *sf) 59d71ae5a4SJacob Faibussowitsch { 6095fce210SBarry Smith PetscSF b; 6195fce210SBarry Smith 6295fce210SBarry Smith PetscFunctionBegin; 6395fce210SBarry Smith PetscValidPointer(sf, 2); 649566063dSJacob Faibussowitsch PetscCall(PetscSFInitializePackage()); 6595fce210SBarry Smith 669566063dSJacob Faibussowitsch PetscCall(PetscHeaderCreate(b, PETSCSF_CLASSID, "PetscSF", "Star Forest", "PetscSF", comm, PetscSFDestroy, PetscSFView)); 6795fce210SBarry Smith 6895fce210SBarry Smith b->nroots = -1; 6995fce210SBarry Smith b->nleaves = -1; 7029046d53SLisandro Dalcin b->minleaf = PETSC_MAX_INT; 7129046d53SLisandro Dalcin b->maxleaf = PETSC_MIN_INT; 7295fce210SBarry Smith b->nranks = -1; 7395fce210SBarry Smith b->rankorder = PETSC_TRUE; 7495fce210SBarry Smith b->ingroup = MPI_GROUP_NULL; 7595fce210SBarry Smith b->outgroup = MPI_GROUP_NULL; 7695fce210SBarry Smith b->graphset = PETSC_FALSE; 7720c24465SJunchao Zhang #if defined(PETSC_HAVE_DEVICE) 7820c24465SJunchao Zhang b->use_gpu_aware_mpi = use_gpu_aware_mpi; 7920c24465SJunchao Zhang b->use_stream_aware_mpi = PETSC_FALSE; 8071438e86SJunchao Zhang b->unknown_input_stream = PETSC_FALSE; 8127f636e8SJunchao Zhang #if defined(PETSC_HAVE_KOKKOS) /* Prefer kokkos over cuda*/ 8220c24465SJunchao Zhang b->backend = PETSCSF_BACKEND_KOKKOS; 8327f636e8SJunchao Zhang #elif defined(PETSC_HAVE_CUDA) 8427f636e8SJunchao Zhang b->backend = PETSCSF_BACKEND_CUDA; 8559af0bd3SScott Kruger #elif defined(PETSC_HAVE_HIP) 8659af0bd3SScott Kruger b->backend = PETSCSF_BACKEND_HIP; 8720c24465SJunchao Zhang #endif 8871438e86SJunchao Zhang 8971438e86SJunchao Zhang #if defined(PETSC_HAVE_NVSHMEM) 9071438e86SJunchao Zhang b->use_nvshmem = PETSC_FALSE; /* Default is not to try NVSHMEM */ 9171438e86SJunchao Zhang b->use_nvshmem_get = PETSC_FALSE; /* Default is to use nvshmem_put based protocol */ 929566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL, NULL, "-use_nvshmem", &b->use_nvshmem, NULL)); 939566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL, NULL, "-use_nvshmem_get", &b->use_nvshmem_get, NULL)); 9471438e86SJunchao Zhang #endif 9520c24465SJunchao Zhang #endif 9660c22052SBarry Smith b->vscat.from_n = -1; 9760c22052SBarry Smith b->vscat.to_n = -1; 9860c22052SBarry Smith b->vscat.unit = MPIU_SCALAR; 9995fce210SBarry Smith *sf = b; 1003ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 10195fce210SBarry Smith } 10295fce210SBarry Smith 10329046d53SLisandro Dalcin /*@ 10495fce210SBarry Smith PetscSFReset - Reset a star forest so that different sizes or neighbors can be used 10595fce210SBarry Smith 10695fce210SBarry Smith Collective 10795fce210SBarry Smith 1084165533cSJose E. Roman Input Parameter: 10995fce210SBarry Smith . sf - star forest 11095fce210SBarry Smith 11195fce210SBarry Smith Level: advanced 11295fce210SBarry Smith 113cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`, `PetscSFSetGraph()`, `PetscSFDestroy()` 11495fce210SBarry Smith @*/ 115d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReset(PetscSF sf) 116d71ae5a4SJacob Faibussowitsch { 11795fce210SBarry Smith PetscFunctionBegin; 11895fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 119dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, Reset); 12029046d53SLisandro Dalcin sf->nroots = -1; 12129046d53SLisandro Dalcin sf->nleaves = -1; 12229046d53SLisandro Dalcin sf->minleaf = PETSC_MAX_INT; 12329046d53SLisandro Dalcin sf->maxleaf = PETSC_MIN_INT; 12495fce210SBarry Smith sf->mine = NULL; 12595fce210SBarry Smith sf->remote = NULL; 12629046d53SLisandro Dalcin sf->graphset = PETSC_FALSE; 1279566063dSJacob Faibussowitsch PetscCall(PetscFree(sf->mine_alloc)); 1289566063dSJacob Faibussowitsch PetscCall(PetscFree(sf->remote_alloc)); 12921c688dcSJed Brown sf->nranks = -1; 1309566063dSJacob Faibussowitsch PetscCall(PetscFree4(sf->ranks, sf->roffset, sf->rmine, sf->rremote)); 13129046d53SLisandro Dalcin sf->degreeknown = PETSC_FALSE; 1329566063dSJacob Faibussowitsch PetscCall(PetscFree(sf->degree)); 1339566063dSJacob Faibussowitsch if (sf->ingroup != MPI_GROUP_NULL) PetscCallMPI(MPI_Group_free(&sf->ingroup)); 1349566063dSJacob Faibussowitsch if (sf->outgroup != MPI_GROUP_NULL) PetscCallMPI(MPI_Group_free(&sf->outgroup)); 135013b3241SStefano Zampini if (sf->multi) sf->multi->multi = NULL; 1369566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&sf->multi)); 1379566063dSJacob Faibussowitsch PetscCall(PetscLayoutDestroy(&sf->map)); 13871438e86SJunchao Zhang 13971438e86SJunchao Zhang #if defined(PETSC_HAVE_DEVICE) 1409566063dSJacob Faibussowitsch for (PetscInt i = 0; i < 2; i++) PetscCall(PetscSFFree(sf, PETSC_MEMTYPE_DEVICE, sf->rmine_d[i])); 14171438e86SJunchao Zhang #endif 14271438e86SJunchao Zhang 14395fce210SBarry Smith sf->setupcalled = PETSC_FALSE; 1443ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 14595fce210SBarry Smith } 14695fce210SBarry Smith 14795fce210SBarry Smith /*@C 148cab54364SBarry Smith PetscSFSetType - Set the `PetscSF` communication implementation 14995fce210SBarry Smith 150c3339decSBarry Smith Collective 15195fce210SBarry Smith 15295fce210SBarry Smith Input Parameters: 153cab54364SBarry Smith + sf - the `PetscSF` context 15495fce210SBarry Smith - type - a known method 155cab54364SBarry Smith .vb 156cab54364SBarry Smith PETSCSFWINDOW - MPI-2/3 one-sided 157cab54364SBarry Smith PETSCSFBASIC - basic implementation using MPI-1 two-sided 158cab54364SBarry Smith .ve 15995fce210SBarry Smith 16095fce210SBarry Smith Options Database Key: 16120662ed9SBarry Smith . -sf_type <type> - Sets the method; for example `basic` or `window` use -help for a list of available methods 162cab54364SBarry Smith 163cab54364SBarry Smith Level: intermediate 16495fce210SBarry Smith 16595fce210SBarry Smith Notes: 16620662ed9SBarry Smith See `PetscSFType` for possible values 16795fce210SBarry Smith 16820662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFCreate()` 16995fce210SBarry Smith @*/ 170d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetType(PetscSF sf, PetscSFType type) 171d71ae5a4SJacob Faibussowitsch { 17295fce210SBarry Smith PetscBool match; 1735f80ce2aSJacob Faibussowitsch PetscErrorCode (*r)(PetscSF); 17495fce210SBarry Smith 17595fce210SBarry Smith PetscFunctionBegin; 17695fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 17795fce210SBarry Smith PetscValidCharPointer(type, 2); 17895fce210SBarry Smith 1799566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare((PetscObject)sf, type, &match)); 1803ba16761SJacob Faibussowitsch if (match) PetscFunctionReturn(PETSC_SUCCESS); 18195fce210SBarry Smith 1829566063dSJacob Faibussowitsch PetscCall(PetscFunctionListFind(PetscSFList, type, &r)); 1836adde796SStefano Zampini PetscCheck(r, PetscObjectComm((PetscObject)sf), PETSC_ERR_ARG_UNKNOWN_TYPE, "Unable to find requested PetscSF type %s", type); 18429046d53SLisandro Dalcin /* Destroy the previous PetscSF implementation context */ 185dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, Destroy); 1869566063dSJacob Faibussowitsch PetscCall(PetscMemzero(sf->ops, sizeof(*sf->ops))); 1879566063dSJacob Faibussowitsch PetscCall(PetscObjectChangeTypeName((PetscObject)sf, type)); 1889566063dSJacob Faibussowitsch PetscCall((*r)(sf)); 1893ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 19095fce210SBarry Smith } 19195fce210SBarry Smith 19229046d53SLisandro Dalcin /*@C 193cab54364SBarry Smith PetscSFGetType - Get the `PetscSF` communication implementation 19429046d53SLisandro Dalcin 19529046d53SLisandro Dalcin Not Collective 19629046d53SLisandro Dalcin 19729046d53SLisandro Dalcin Input Parameter: 198cab54364SBarry Smith . sf - the `PetscSF` context 19929046d53SLisandro Dalcin 20029046d53SLisandro Dalcin Output Parameter: 201cab54364SBarry Smith . type - the `PetscSF` type name 20229046d53SLisandro Dalcin 20329046d53SLisandro Dalcin Level: intermediate 20429046d53SLisandro Dalcin 20520662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFSetType()`, `PetscSFCreate()` 20629046d53SLisandro Dalcin @*/ 207d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetType(PetscSF sf, PetscSFType *type) 208d71ae5a4SJacob Faibussowitsch { 20929046d53SLisandro Dalcin PetscFunctionBegin; 21029046d53SLisandro Dalcin PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 21129046d53SLisandro Dalcin PetscValidPointer(type, 2); 21229046d53SLisandro Dalcin *type = ((PetscObject)sf)->type_name; 2133ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 21429046d53SLisandro Dalcin } 21529046d53SLisandro Dalcin 2161fb7b255SJunchao Zhang /*@C 21720662ed9SBarry Smith PetscSFDestroy - destroy a star forest 21895fce210SBarry Smith 21995fce210SBarry Smith Collective 22095fce210SBarry Smith 2214165533cSJose E. Roman Input Parameter: 22295fce210SBarry Smith . sf - address of star forest 22395fce210SBarry Smith 22495fce210SBarry Smith Level: intermediate 22595fce210SBarry Smith 22620662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFCreate()`, `PetscSFReset()` 22795fce210SBarry Smith @*/ 228d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFDestroy(PetscSF *sf) 229d71ae5a4SJacob Faibussowitsch { 23095fce210SBarry Smith PetscFunctionBegin; 2313ba16761SJacob Faibussowitsch if (!*sf) PetscFunctionReturn(PETSC_SUCCESS); 23295fce210SBarry Smith PetscValidHeaderSpecific((*sf), PETSCSF_CLASSID, 1); 2339371c9d4SSatish Balay if (--((PetscObject)(*sf))->refct > 0) { 2349371c9d4SSatish Balay *sf = NULL; 2353ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 2369371c9d4SSatish Balay } 2379566063dSJacob Faibussowitsch PetscCall(PetscSFReset(*sf)); 238dbbe0bcdSBarry Smith PetscTryTypeMethod((*sf), Destroy); 2399566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&(*sf)->vscat.lsf)); 2409566063dSJacob Faibussowitsch if ((*sf)->vscat.bs > 1) PetscCallMPI(MPI_Type_free(&(*sf)->vscat.unit)); 241715b587bSJunchao Zhang #if defined(PETSC_HAVE_MPIX_STREAM) 242715b587bSJunchao Zhang if ((*sf)->use_stream_aware_mpi) { 243715b587bSJunchao Zhang PetscCallMPI(MPIX_Stream_free(&(*sf)->mpi_stream)); 244715b587bSJunchao Zhang PetscCallMPI(MPI_Comm_free(&(*sf)->stream_comm)); 245715b587bSJunchao Zhang } 246715b587bSJunchao Zhang #endif 2479566063dSJacob Faibussowitsch PetscCall(PetscHeaderDestroy(sf)); 2483ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 24995fce210SBarry Smith } 25095fce210SBarry Smith 251d71ae5a4SJacob Faibussowitsch static PetscErrorCode PetscSFCheckGraphValid_Private(PetscSF sf) 252d71ae5a4SJacob Faibussowitsch { 253c4e6a40aSLawrence Mitchell PetscInt i, nleaves; 254c4e6a40aSLawrence Mitchell PetscMPIInt size; 255c4e6a40aSLawrence Mitchell const PetscInt *ilocal; 256c4e6a40aSLawrence Mitchell const PetscSFNode *iremote; 257c4e6a40aSLawrence Mitchell 258c4e6a40aSLawrence Mitchell PetscFunctionBegin; 2593ba16761SJacob Faibussowitsch if (!sf->graphset || !PetscDefined(USE_DEBUG)) PetscFunctionReturn(PETSC_SUCCESS); 2609566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, NULL, &nleaves, &ilocal, &iremote)); 2619566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)sf), &size)); 262c4e6a40aSLawrence Mitchell for (i = 0; i < nleaves; i++) { 263c4e6a40aSLawrence Mitchell const PetscInt rank = iremote[i].rank; 264c4e6a40aSLawrence Mitchell const PetscInt remote = iremote[i].index; 265c4e6a40aSLawrence Mitchell const PetscInt leaf = ilocal ? ilocal[i] : i; 266c9cc58a2SBarry Smith PetscCheck(rank >= 0 && rank < size, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided rank (%" PetscInt_FMT ") for remote %" PetscInt_FMT " is invalid, should be in [0, %d)", rank, i, size); 26708401ef6SPierre Jolivet PetscCheck(remote >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided index (%" PetscInt_FMT ") for remote %" PetscInt_FMT " is invalid, should be >= 0", remote, i); 26808401ef6SPierre Jolivet PetscCheck(leaf >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided location (%" PetscInt_FMT ") for leaf %" PetscInt_FMT " is invalid, should be >= 0", leaf, i); 269c4e6a40aSLawrence Mitchell } 2703ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 271c4e6a40aSLawrence Mitchell } 272c4e6a40aSLawrence Mitchell 27395fce210SBarry Smith /*@ 27420662ed9SBarry Smith PetscSFSetUp - set up communication structures for a `PetscSF`, after this is done it may be used to perform communication 27595fce210SBarry Smith 27695fce210SBarry Smith Collective 27795fce210SBarry Smith 2784165533cSJose E. Roman Input Parameter: 27995fce210SBarry Smith . sf - star forest communication object 28095fce210SBarry Smith 28195fce210SBarry Smith Level: beginner 28295fce210SBarry Smith 28320662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFSetFromOptions()`, `PetscSFSetType()` 28495fce210SBarry Smith @*/ 285d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetUp(PetscSF sf) 286d71ae5a4SJacob Faibussowitsch { 28795fce210SBarry Smith PetscFunctionBegin; 28829046d53SLisandro Dalcin PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 28929046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 2903ba16761SJacob Faibussowitsch if (sf->setupcalled) PetscFunctionReturn(PETSC_SUCCESS); 2919566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_SetUp, sf, 0, 0, 0)); 2929566063dSJacob Faibussowitsch PetscCall(PetscSFCheckGraphValid_Private(sf)); 2939566063dSJacob Faibussowitsch if (!((PetscObject)sf)->type_name) PetscCall(PetscSFSetType(sf, PETSCSFBASIC)); /* Zero all sf->ops */ 294dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, SetUp); 29520c24465SJunchao Zhang #if defined(PETSC_HAVE_CUDA) 29620c24465SJunchao Zhang if (sf->backend == PETSCSF_BACKEND_CUDA) { 29771438e86SJunchao Zhang sf->ops->Malloc = PetscSFMalloc_CUDA; 29871438e86SJunchao Zhang sf->ops->Free = PetscSFFree_CUDA; 29920c24465SJunchao Zhang } 30020c24465SJunchao Zhang #endif 30159af0bd3SScott Kruger #if defined(PETSC_HAVE_HIP) 30259af0bd3SScott Kruger if (sf->backend == PETSCSF_BACKEND_HIP) { 30359af0bd3SScott Kruger sf->ops->Malloc = PetscSFMalloc_HIP; 30459af0bd3SScott Kruger sf->ops->Free = PetscSFFree_HIP; 30559af0bd3SScott Kruger } 30659af0bd3SScott Kruger #endif 30720c24465SJunchao Zhang 30820c24465SJunchao Zhang #if defined(PETSC_HAVE_KOKKOS) 30920c24465SJunchao Zhang if (sf->backend == PETSCSF_BACKEND_KOKKOS) { 31020c24465SJunchao Zhang sf->ops->Malloc = PetscSFMalloc_Kokkos; 31120c24465SJunchao Zhang sf->ops->Free = PetscSFFree_Kokkos; 31220c24465SJunchao Zhang } 31320c24465SJunchao Zhang #endif 3149566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_SetUp, sf, 0, 0, 0)); 31595fce210SBarry Smith sf->setupcalled = PETSC_TRUE; 3163ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 31795fce210SBarry Smith } 31895fce210SBarry Smith 3198af6ec1cSBarry Smith /*@ 320cab54364SBarry Smith PetscSFSetFromOptions - set `PetscSF` options using the options database 32195fce210SBarry Smith 32295fce210SBarry Smith Logically Collective 32395fce210SBarry Smith 3244165533cSJose E. Roman Input Parameter: 32595fce210SBarry Smith . sf - star forest 32695fce210SBarry Smith 32795fce210SBarry Smith Options Database Keys: 32820662ed9SBarry Smith + -sf_type - implementation type, see `PetscSFSetType()` 32951ccb202SJunchao Zhang . -sf_rank_order - sort composite points for gathers and scatters in rank order, gathers are non-deterministic otherwise 33020662ed9SBarry Smith . -sf_use_default_stream - Assume callers of `PetscSF` computed the input root/leafdata with the default CUDA stream. `PetscSF` will also 33120662ed9SBarry Smith use the default stream to process data. Therefore, no stream synchronization is needed between `PetscSF` and its caller (default: true). 33220662ed9SBarry Smith If true, this option only works with `-use_gpu_aware_mpi 1`. 33320662ed9SBarry Smith . -sf_use_stream_aware_mpi - Assume the underlying MPI is CUDA-stream aware and `PetscSF` won't sync streams for send/recv buffers passed to MPI (default: false). 33420662ed9SBarry Smith If true, this option only works with `-use_gpu_aware_mpi 1`. 33595fce210SBarry Smith 33638b5cf2dSJacob Faibussowitsch - -sf_backend cuda | hip | kokkos -Select the device backend SF uses. Currently `PetscSF` has these backends: cuda - hip and Kokkos. 33759af0bd3SScott Kruger On CUDA (HIP) devices, one can choose cuda (hip) or kokkos with the default being kokkos. On other devices, 33820c24465SJunchao Zhang the only available is kokkos. 33920c24465SJunchao Zhang 34095fce210SBarry Smith Level: intermediate 341cab54364SBarry Smith 342cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`, `PetscSFSetType()` 34395fce210SBarry Smith @*/ 344d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetFromOptions(PetscSF sf) 345d71ae5a4SJacob Faibussowitsch { 34695fce210SBarry Smith PetscSFType deft; 34795fce210SBarry Smith char type[256]; 34895fce210SBarry Smith PetscBool flg; 34995fce210SBarry Smith 35095fce210SBarry Smith PetscFunctionBegin; 35195fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 352d0609cedSBarry Smith PetscObjectOptionsBegin((PetscObject)sf); 35395fce210SBarry Smith deft = ((PetscObject)sf)->type_name ? ((PetscObject)sf)->type_name : PETSCSFBASIC; 3549566063dSJacob Faibussowitsch PetscCall(PetscOptionsFList("-sf_type", "PetscSF implementation type", "PetscSFSetType", PetscSFList, deft, type, sizeof(type), &flg)); 3559566063dSJacob Faibussowitsch PetscCall(PetscSFSetType(sf, flg ? type : deft)); 3569566063dSJacob Faibussowitsch PetscCall(PetscOptionsBool("-sf_rank_order", "sort composite points for gathers and scatters in rank order, gathers are non-deterministic otherwise", "PetscSFSetRankOrder", sf->rankorder, &sf->rankorder, NULL)); 3577fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE) 35820c24465SJunchao Zhang { 35920c24465SJunchao Zhang char backendstr[32] = {0}; 36059af0bd3SScott Kruger PetscBool isCuda = PETSC_FALSE, isHip = PETSC_FALSE, isKokkos = PETSC_FALSE, set; 36120c24465SJunchao Zhang /* Change the defaults set in PetscSFCreate() with command line options */ 362d5b43468SJose E. Roman PetscCall(PetscOptionsBool("-sf_unknown_input_stream", "SF root/leafdata is computed on arbitrary streams unknown to SF", "PetscSFSetFromOptions", sf->unknown_input_stream, &sf->unknown_input_stream, NULL)); 3639566063dSJacob Faibussowitsch PetscCall(PetscOptionsBool("-sf_use_stream_aware_mpi", "Assume the underlying MPI is cuda-stream aware", "PetscSFSetFromOptions", sf->use_stream_aware_mpi, &sf->use_stream_aware_mpi, NULL)); 3649566063dSJacob Faibussowitsch PetscCall(PetscOptionsString("-sf_backend", "Select the device backend SF uses", "PetscSFSetFromOptions", NULL, backendstr, sizeof(backendstr), &set)); 3659566063dSJacob Faibussowitsch PetscCall(PetscStrcasecmp("cuda", backendstr, &isCuda)); 3669566063dSJacob Faibussowitsch PetscCall(PetscStrcasecmp("kokkos", backendstr, &isKokkos)); 3679566063dSJacob Faibussowitsch PetscCall(PetscStrcasecmp("hip", backendstr, &isHip)); 36859af0bd3SScott Kruger #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) 36920c24465SJunchao Zhang if (isCuda) sf->backend = PETSCSF_BACKEND_CUDA; 37020c24465SJunchao Zhang else if (isKokkos) sf->backend = PETSCSF_BACKEND_KOKKOS; 37159af0bd3SScott Kruger else if (isHip) sf->backend = PETSCSF_BACKEND_HIP; 37228b400f6SJacob Faibussowitsch else PetscCheck(!set, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "-sf_backend %s is not supported. You may choose cuda, hip or kokkos (if installed)", backendstr); 37320c24465SJunchao Zhang #elif defined(PETSC_HAVE_KOKKOS) 37408401ef6SPierre Jolivet PetscCheck(!set || isKokkos, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "-sf_backend %s is not supported. You can only choose kokkos", backendstr); 37520c24465SJunchao Zhang #endif 376715b587bSJunchao Zhang 377715b587bSJunchao Zhang #if defined(PETSC_HAVE_CUDA) && defined(PETSC_HAVE_MPIX_STREAM) 378715b587bSJunchao Zhang if (sf->use_stream_aware_mpi) { 379715b587bSJunchao Zhang MPI_Info info; 380715b587bSJunchao Zhang 381715b587bSJunchao Zhang PetscCallMPI(MPI_Info_create(&info)); 382715b587bSJunchao Zhang PetscCallMPI(MPI_Info_set(info, "type", "cudaStream_t")); 383715b587bSJunchao Zhang PetscCallMPI(MPIX_Info_set_hex(info, "value", &PetscDefaultCudaStream, sizeof(PetscDefaultCudaStream))); 384715b587bSJunchao Zhang PetscCallMPI(MPIX_Stream_create(info, &sf->mpi_stream)); 385715b587bSJunchao Zhang PetscCallMPI(MPI_Info_free(&info)); 386715b587bSJunchao Zhang PetscCallMPI(MPIX_Stream_comm_create(PetscObjectComm((PetscObject)sf), sf->mpi_stream, &sf->stream_comm)); 387715b587bSJunchao Zhang } 388715b587bSJunchao Zhang #endif 38920c24465SJunchao Zhang } 390c2a741eeSJunchao Zhang #endif 391dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, SetFromOptions, PetscOptionsObject); 392d0609cedSBarry Smith PetscOptionsEnd(); 3933ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 39495fce210SBarry Smith } 39595fce210SBarry Smith 39629046d53SLisandro Dalcin /*@ 39795fce210SBarry Smith PetscSFSetRankOrder - sort multi-points for gathers and scatters by rank order 39895fce210SBarry Smith 39995fce210SBarry Smith Logically Collective 40095fce210SBarry Smith 4014165533cSJose E. Roman Input Parameters: 40295fce210SBarry Smith + sf - star forest 403cab54364SBarry Smith - flg - `PETSC_TRUE` to sort, `PETSC_FALSE` to skip sorting (lower setup cost, but non-deterministic) 40495fce210SBarry Smith 40595fce210SBarry Smith Level: advanced 40695fce210SBarry Smith 40720662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFGatherBegin()`, `PetscSFScatterBegin()` 40895fce210SBarry Smith @*/ 409d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetRankOrder(PetscSF sf, PetscBool flg) 410d71ae5a4SJacob Faibussowitsch { 41195fce210SBarry Smith PetscFunctionBegin; 41295fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 41395fce210SBarry Smith PetscValidLogicalCollectiveBool(sf, flg, 2); 41428b400f6SJacob Faibussowitsch PetscCheck(!sf->multi, PetscObjectComm((PetscObject)sf), PETSC_ERR_ARG_WRONGSTATE, "Rank ordering must be set before first call to PetscSFGatherBegin() or PetscSFScatterBegin()"); 41595fce210SBarry Smith sf->rankorder = flg; 4163ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 41795fce210SBarry Smith } 41895fce210SBarry Smith 4198dbb0df6SBarry Smith /*@C 42095fce210SBarry Smith PetscSFSetGraph - Set a parallel star forest 42195fce210SBarry Smith 42295fce210SBarry Smith Collective 42395fce210SBarry Smith 4244165533cSJose E. Roman Input Parameters: 42595fce210SBarry Smith + sf - star forest 42695fce210SBarry Smith . nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves) 42795fce210SBarry Smith . nleaves - number of leaf vertices on the current process, each of these references a root on any process 42820662ed9SBarry Smith . ilocal - locations of leaves in leafdata buffers, pass `NULL` for contiguous storage (locations must be >= 0, enforced 429c4e6a40aSLawrence Mitchell during setup in debug mode) 43020662ed9SBarry Smith . localmode - copy mode for `ilocal` 431c4e6a40aSLawrence Mitchell . iremote - remote locations of root vertices for each leaf on the current process (locations must be >= 0, enforced 432c4e6a40aSLawrence Mitchell during setup in debug mode) 43320662ed9SBarry Smith - remotemode - copy mode for `iremote` 43495fce210SBarry Smith 43595fce210SBarry Smith Level: intermediate 43695fce210SBarry Smith 43795452b02SPatrick Sanan Notes: 43820662ed9SBarry Smith Leaf indices in `ilocal` must be unique, otherwise an error occurs. 43938ab3f8aSBarry Smith 44020662ed9SBarry Smith Input arrays `ilocal` and `iremote` follow the `PetscCopyMode` semantics. 44120662ed9SBarry Smith In particular, if `localmode` or `remotemode` is `PETSC_OWN_POINTER` or `PETSC_USE_POINTER`, 442db2b9530SVaclav Hapla PETSc might modify the respective array; 44320662ed9SBarry Smith if `PETSC_USE_POINTER`, the user must delete the array after `PetscSFDestroy()`. 444cab54364SBarry Smith Only if `PETSC_COPY_VALUES` is used, the respective array is guaranteed to stay intact and a const array can be passed (but a cast to non-const is needed). 445db2b9530SVaclav Hapla 44638b5cf2dSJacob Faibussowitsch Fortran Notes: 44720662ed9SBarry Smith In Fortran you must use `PETSC_COPY_VALUES` for `localmode` and `remotemode`. 448c4e6a40aSLawrence Mitchell 44938b5cf2dSJacob Faibussowitsch Developer Notes: 450db2b9530SVaclav Hapla We sort leaves to check for duplicates and contiguousness and to find minleaf/maxleaf. 45120662ed9SBarry Smith This also allows to compare leaf sets of two `PetscSF`s easily. 45272bf8598SVaclav Hapla 45320662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFGetGraph()` 45495fce210SBarry Smith @*/ 455d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetGraph(PetscSF sf, PetscInt nroots, PetscInt nleaves, PetscInt *ilocal, PetscCopyMode localmode, PetscSFNode *iremote, PetscCopyMode remotemode) 456d71ae5a4SJacob Faibussowitsch { 457db2b9530SVaclav Hapla PetscBool unique, contiguous; 45895fce210SBarry Smith 45995fce210SBarry Smith PetscFunctionBegin; 46095fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 46129046d53SLisandro Dalcin if (nleaves > 0 && ilocal) PetscValidIntPointer(ilocal, 4); 46229046d53SLisandro Dalcin if (nleaves > 0) PetscValidPointer(iremote, 6); 46308401ef6SPierre Jolivet PetscCheck(nroots >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nroots %" PetscInt_FMT ", cannot be negative", nroots); 46408401ef6SPierre Jolivet PetscCheck(nleaves >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nleaves %" PetscInt_FMT ", cannot be negative", nleaves); 4658da24d32SBarry Smith /* enums may be handled as unsigned by some compilers, NVHPC for example, the int cast 4668da24d32SBarry Smith * below is to prevent NVHPC from warning about meaningless comparison of unsigned with zero */ 4678da24d32SBarry Smith PetscCheck((int)localmode >= PETSC_COPY_VALUES && localmode <= PETSC_USE_POINTER, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Wrong localmode %d", localmode); 4688da24d32SBarry Smith PetscCheck((int)remotemode >= PETSC_COPY_VALUES && remotemode <= PETSC_USE_POINTER, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Wrong remotemode %d", remotemode); 46929046d53SLisandro Dalcin 4702a67d2daSStefano Zampini if (sf->nroots >= 0) { /* Reset only if graph already set */ 4719566063dSJacob Faibussowitsch PetscCall(PetscSFReset(sf)); 4722a67d2daSStefano Zampini } 4732a67d2daSStefano Zampini 4749566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_SetGraph, sf, 0, 0, 0)); 47529046d53SLisandro Dalcin 47695fce210SBarry Smith sf->nroots = nroots; 47795fce210SBarry Smith sf->nleaves = nleaves; 47829046d53SLisandro Dalcin 479db2b9530SVaclav Hapla if (localmode == PETSC_COPY_VALUES && ilocal) { 480db2b9530SVaclav Hapla PetscInt *tlocal = NULL; 481db2b9530SVaclav Hapla 4829566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nleaves, &tlocal)); 4839566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(tlocal, ilocal, nleaves)); 484db2b9530SVaclav Hapla ilocal = tlocal; 485db2b9530SVaclav Hapla } 486db2b9530SVaclav Hapla if (remotemode == PETSC_COPY_VALUES) { 487db2b9530SVaclav Hapla PetscSFNode *tremote = NULL; 488db2b9530SVaclav Hapla 4899566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nleaves, &tremote)); 4909566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(tremote, iremote, nleaves)); 491db2b9530SVaclav Hapla iremote = tremote; 492db2b9530SVaclav Hapla } 493db2b9530SVaclav Hapla 49429046d53SLisandro Dalcin if (nleaves && ilocal) { 495db2b9530SVaclav Hapla PetscSFNode work; 496db2b9530SVaclav Hapla 4979566063dSJacob Faibussowitsch PetscCall(PetscSortIntWithDataArray(nleaves, ilocal, iremote, sizeof(PetscSFNode), &work)); 4989566063dSJacob Faibussowitsch PetscCall(PetscSortedCheckDupsInt(nleaves, ilocal, &unique)); 499db2b9530SVaclav Hapla unique = PetscNot(unique); 500db2b9530SVaclav Hapla PetscCheck(sf->allow_multi_leaves || unique, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Input ilocal has duplicate entries which is not allowed for this PetscSF"); 501db2b9530SVaclav Hapla sf->minleaf = ilocal[0]; 502db2b9530SVaclav Hapla sf->maxleaf = ilocal[nleaves - 1]; 503db2b9530SVaclav Hapla contiguous = (PetscBool)(unique && ilocal[0] == 0 && ilocal[nleaves - 1] == nleaves - 1); 50429046d53SLisandro Dalcin } else { 50529046d53SLisandro Dalcin sf->minleaf = 0; 50629046d53SLisandro Dalcin sf->maxleaf = nleaves - 1; 507db2b9530SVaclav Hapla unique = PETSC_TRUE; 508db2b9530SVaclav Hapla contiguous = PETSC_TRUE; 50929046d53SLisandro Dalcin } 51029046d53SLisandro Dalcin 511db2b9530SVaclav Hapla if (contiguous) { 512db2b9530SVaclav Hapla if (localmode == PETSC_USE_POINTER) { 513db2b9530SVaclav Hapla ilocal = NULL; 514db2b9530SVaclav Hapla } else { 5159566063dSJacob Faibussowitsch PetscCall(PetscFree(ilocal)); 516db2b9530SVaclav Hapla } 517db2b9530SVaclav Hapla } 518db2b9530SVaclav Hapla sf->mine = ilocal; 519db2b9530SVaclav Hapla if (localmode == PETSC_USE_POINTER) { 52029046d53SLisandro Dalcin sf->mine_alloc = NULL; 521db2b9530SVaclav Hapla } else { 522db2b9530SVaclav Hapla sf->mine_alloc = ilocal; 52395fce210SBarry Smith } 524db2b9530SVaclav Hapla sf->remote = iremote; 525db2b9530SVaclav Hapla if (remotemode == PETSC_USE_POINTER) { 52629046d53SLisandro Dalcin sf->remote_alloc = NULL; 527db2b9530SVaclav Hapla } else { 528db2b9530SVaclav Hapla sf->remote_alloc = iremote; 52995fce210SBarry Smith } 5309566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_SetGraph, sf, 0, 0, 0)); 53129046d53SLisandro Dalcin sf->graphset = PETSC_TRUE; 5323ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 53395fce210SBarry Smith } 53495fce210SBarry Smith 53529046d53SLisandro Dalcin /*@ 536cab54364SBarry Smith PetscSFSetGraphWithPattern - Sets the graph of a `PetscSF` with a specific pattern 537dd5b3ca6SJunchao Zhang 538dd5b3ca6SJunchao Zhang Collective 539dd5b3ca6SJunchao Zhang 540dd5b3ca6SJunchao Zhang Input Parameters: 541cab54364SBarry Smith + sf - The `PetscSF` 542cab54364SBarry Smith . map - Layout of roots over all processes (insignificant when pattern is `PETSCSF_PATTERN_ALLTOALL`) 543cab54364SBarry Smith - pattern - One of `PETSCSF_PATTERN_ALLGATHER`, `PETSCSF_PATTERN_GATHER`, `PETSCSF_PATTERN_ALLTOALL` 544cab54364SBarry Smith 545cab54364SBarry Smith Level: intermediate 546dd5b3ca6SJunchao Zhang 547dd5b3ca6SJunchao Zhang Notes: 54820662ed9SBarry Smith It is easier to explain `PetscSFPattern` using vectors. Suppose we have an MPI vector `x` and its `PetscLayout` is `map`. 54920662ed9SBarry Smith `n` and `N` are the local and global sizes of `x` respectively. 550dd5b3ca6SJunchao Zhang 55120662ed9SBarry Smith With `PETSCSF_PATTERN_ALLGATHER`, the routine creates a graph that if one does `PetscSFBcastBegin()`/`PetscSFBcastEnd()` on it, it will copy `x` to 55220662ed9SBarry Smith sequential vectors `y` on all MPI processes. 553dd5b3ca6SJunchao Zhang 55420662ed9SBarry Smith With `PETSCSF_PATTERN_GATHER`, the routine creates a graph that if one does `PetscSFBcastBegin()`/`PetscSFBcastEnd()` on it, it will copy `x` to a 55520662ed9SBarry Smith sequential vector `y` on rank 0. 556dd5b3ca6SJunchao Zhang 55720662ed9SBarry Smith In above cases, entries of `x` are roots and entries of `y` are leaves. 558dd5b3ca6SJunchao Zhang 55920662ed9SBarry Smith With `PETSCSF_PATTERN_ALLTOALL`, map is insignificant. Suppose NP is size of `sf`'s communicator. The routine 560dd5b3ca6SJunchao Zhang creates a graph that every rank has NP leaves and NP roots. On rank i, its leaf j is connected to root i 561cab54364SBarry Smith of rank j. Here 0 <=i,j<NP. It is a kind of `MPI_Alltoall()` with sendcount/recvcount being 1. Note that it does 562dd5b3ca6SJunchao Zhang not mean one can not send multiple items. One just needs to create a new MPI datatype for the mulptiple data 563cab54364SBarry Smith items with `MPI_Type_contiguous` and use that as the <unit> argument in SF routines. 564dd5b3ca6SJunchao Zhang 565dd5b3ca6SJunchao Zhang In this case, roots and leaves are symmetric. 566dd5b3ca6SJunchao Zhang 567cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFGetGraph()` 568dd5b3ca6SJunchao Zhang @*/ 569d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetGraphWithPattern(PetscSF sf, PetscLayout map, PetscSFPattern pattern) 570d71ae5a4SJacob Faibussowitsch { 571dd5b3ca6SJunchao Zhang MPI_Comm comm; 572dd5b3ca6SJunchao Zhang PetscInt n, N, res[2]; 573dd5b3ca6SJunchao Zhang PetscMPIInt rank, size; 574dd5b3ca6SJunchao Zhang PetscSFType type; 575dd5b3ca6SJunchao Zhang 576dd5b3ca6SJunchao Zhang PetscFunctionBegin; 5772abc8c78SJacob Faibussowitsch PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 5782abc8c78SJacob Faibussowitsch if (pattern != PETSCSF_PATTERN_ALLTOALL) PetscValidPointer(map, 2); 5799566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 5802c71b3e2SJacob Faibussowitsch PetscCheck(pattern >= PETSCSF_PATTERN_ALLGATHER && pattern <= PETSCSF_PATTERN_ALLTOALL, comm, PETSC_ERR_ARG_OUTOFRANGE, "Unsupported PetscSFPattern %d", pattern); 5819566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 5829566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 583dd5b3ca6SJunchao Zhang 584dd5b3ca6SJunchao Zhang if (pattern == PETSCSF_PATTERN_ALLTOALL) { 585dd5b3ca6SJunchao Zhang type = PETSCSFALLTOALL; 5869566063dSJacob Faibussowitsch PetscCall(PetscLayoutCreate(comm, &sf->map)); 5879566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetLocalSize(sf->map, size)); 5889566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetSize(sf->map, ((PetscInt)size) * size)); 5899566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetUp(sf->map)); 590dd5b3ca6SJunchao Zhang } else { 5919566063dSJacob Faibussowitsch PetscCall(PetscLayoutGetLocalSize(map, &n)); 5929566063dSJacob Faibussowitsch PetscCall(PetscLayoutGetSize(map, &N)); 593dd5b3ca6SJunchao Zhang res[0] = n; 594dd5b3ca6SJunchao Zhang res[1] = -n; 595dd5b3ca6SJunchao Zhang /* Check if n are same over all ranks so that we can optimize it */ 5961c2dc1cbSBarry Smith PetscCall(MPIU_Allreduce(MPI_IN_PLACE, res, 2, MPIU_INT, MPI_MAX, comm)); 597dd5b3ca6SJunchao Zhang if (res[0] == -res[1]) { /* same n */ 598dd5b3ca6SJunchao Zhang type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHER : PETSCSFGATHER; 599dd5b3ca6SJunchao Zhang } else { 600dd5b3ca6SJunchao Zhang type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHERV : PETSCSFGATHERV; 601dd5b3ca6SJunchao Zhang } 6029566063dSJacob Faibussowitsch PetscCall(PetscLayoutReference(map, &sf->map)); 603dd5b3ca6SJunchao Zhang } 6049566063dSJacob Faibussowitsch PetscCall(PetscSFSetType(sf, type)); 605dd5b3ca6SJunchao Zhang 606dd5b3ca6SJunchao Zhang sf->pattern = pattern; 607dd5b3ca6SJunchao Zhang sf->mine = NULL; /* Contiguous */ 608dd5b3ca6SJunchao Zhang 609dd5b3ca6SJunchao Zhang /* Set nleaves, nroots here in case user calls PetscSFGetGraph, which is legal to call even before PetscSFSetUp is called. 610dd5b3ca6SJunchao Zhang Also set other easy stuff. 611dd5b3ca6SJunchao Zhang */ 612dd5b3ca6SJunchao Zhang if (pattern == PETSCSF_PATTERN_ALLGATHER) { 613dd5b3ca6SJunchao Zhang sf->nleaves = N; 614dd5b3ca6SJunchao Zhang sf->nroots = n; 615dd5b3ca6SJunchao Zhang sf->nranks = size; 616dd5b3ca6SJunchao Zhang sf->minleaf = 0; 617dd5b3ca6SJunchao Zhang sf->maxleaf = N - 1; 618dd5b3ca6SJunchao Zhang } else if (pattern == PETSCSF_PATTERN_GATHER) { 619dd5b3ca6SJunchao Zhang sf->nleaves = rank ? 0 : N; 620dd5b3ca6SJunchao Zhang sf->nroots = n; 621dd5b3ca6SJunchao Zhang sf->nranks = rank ? 0 : size; 622dd5b3ca6SJunchao Zhang sf->minleaf = 0; 623dd5b3ca6SJunchao Zhang sf->maxleaf = rank ? -1 : N - 1; 624dd5b3ca6SJunchao Zhang } else if (pattern == PETSCSF_PATTERN_ALLTOALL) { 625dd5b3ca6SJunchao Zhang sf->nleaves = size; 626dd5b3ca6SJunchao Zhang sf->nroots = size; 627dd5b3ca6SJunchao Zhang sf->nranks = size; 628dd5b3ca6SJunchao Zhang sf->minleaf = 0; 629dd5b3ca6SJunchao Zhang sf->maxleaf = size - 1; 630dd5b3ca6SJunchao Zhang } 631dd5b3ca6SJunchao Zhang sf->ndranks = 0; /* We do not need to separate out distinguished ranks for patterned graphs to improve communication performance */ 632dd5b3ca6SJunchao Zhang sf->graphset = PETSC_TRUE; 6333ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 634dd5b3ca6SJunchao Zhang } 635dd5b3ca6SJunchao Zhang 636dd5b3ca6SJunchao Zhang /*@ 637cab54364SBarry Smith PetscSFCreateInverseSF - given a `PetscSF` in which all vertices have degree 1, creates the inverse map 63895fce210SBarry Smith 63995fce210SBarry Smith Collective 64095fce210SBarry Smith 6414165533cSJose E. Roman Input Parameter: 64295fce210SBarry Smith . sf - star forest to invert 64395fce210SBarry Smith 6444165533cSJose E. Roman Output Parameter: 64520662ed9SBarry Smith . isf - inverse of `sf` 6464165533cSJose E. Roman 64795fce210SBarry Smith Level: advanced 64895fce210SBarry Smith 64995fce210SBarry Smith Notes: 65095fce210SBarry Smith All roots must have degree 1. 65195fce210SBarry Smith 65295fce210SBarry Smith The local space may be a permutation, but cannot be sparse. 65395fce210SBarry Smith 65420662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFSetGraph()` 65595fce210SBarry Smith @*/ 656d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateInverseSF(PetscSF sf, PetscSF *isf) 657d71ae5a4SJacob Faibussowitsch { 65895fce210SBarry Smith PetscMPIInt rank; 65995fce210SBarry Smith PetscInt i, nroots, nleaves, maxlocal, count, *newilocal; 66095fce210SBarry Smith const PetscInt *ilocal; 66195fce210SBarry Smith PetscSFNode *roots, *leaves; 66295fce210SBarry Smith 66395fce210SBarry Smith PetscFunctionBegin; 66429046d53SLisandro Dalcin PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 66529046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 66629046d53SLisandro Dalcin PetscValidPointer(isf, 2); 66729046d53SLisandro Dalcin 6689566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, NULL)); 66929046d53SLisandro Dalcin maxlocal = sf->maxleaf + 1; /* TODO: We should use PetscSFGetLeafRange() */ 67029046d53SLisandro Dalcin 6719566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank)); 6729566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(nroots, &roots, maxlocal, &leaves)); 673ae9aee6dSMatthew G. Knepley for (i = 0; i < maxlocal; i++) { 67495fce210SBarry Smith leaves[i].rank = rank; 67595fce210SBarry Smith leaves[i].index = i; 67695fce210SBarry Smith } 67795fce210SBarry Smith for (i = 0; i < nroots; i++) { 67895fce210SBarry Smith roots[i].rank = -1; 67995fce210SBarry Smith roots[i].index = -1; 68095fce210SBarry Smith } 6819566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(sf, MPIU_2INT, leaves, roots, MPI_REPLACE)); 6829566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(sf, MPIU_2INT, leaves, roots, MPI_REPLACE)); 68395fce210SBarry Smith 68495fce210SBarry Smith /* Check whether our leaves are sparse */ 6859371c9d4SSatish Balay for (i = 0, count = 0; i < nroots; i++) 6869371c9d4SSatish Balay if (roots[i].rank >= 0) count++; 68795fce210SBarry Smith if (count == nroots) newilocal = NULL; 6889371c9d4SSatish Balay else { /* Index for sparse leaves and compact "roots" array (which is to become our leaves). */ PetscCall(PetscMalloc1(count, &newilocal)); 68995fce210SBarry Smith for (i = 0, count = 0; i < nroots; i++) { 69095fce210SBarry Smith if (roots[i].rank >= 0) { 69195fce210SBarry Smith newilocal[count] = i; 69295fce210SBarry Smith roots[count].rank = roots[i].rank; 69395fce210SBarry Smith roots[count].index = roots[i].index; 69495fce210SBarry Smith count++; 69595fce210SBarry Smith } 69695fce210SBarry Smith } 69795fce210SBarry Smith } 69895fce210SBarry Smith 6999566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_CONFONLY, isf)); 7009566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*isf, maxlocal, count, newilocal, PETSC_OWN_POINTER, roots, PETSC_COPY_VALUES)); 7019566063dSJacob Faibussowitsch PetscCall(PetscFree2(roots, leaves)); 7023ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 70395fce210SBarry Smith } 70495fce210SBarry Smith 70595fce210SBarry Smith /*@ 706cab54364SBarry Smith PetscSFDuplicate - duplicate a `PetscSF`, optionally preserving rank connectivity and graph 70795fce210SBarry Smith 70895fce210SBarry Smith Collective 70995fce210SBarry Smith 7104165533cSJose E. Roman Input Parameters: 71195fce210SBarry Smith + sf - communication object to duplicate 712cab54364SBarry Smith - opt - `PETSCSF_DUPLICATE_CONFONLY`, `PETSCSF_DUPLICATE_RANKS`, or `PETSCSF_DUPLICATE_GRAPH` (see `PetscSFDuplicateOption`) 71395fce210SBarry Smith 7144165533cSJose E. Roman Output Parameter: 71595fce210SBarry Smith . newsf - new communication object 71695fce210SBarry Smith 71795fce210SBarry Smith Level: beginner 71895fce210SBarry Smith 71920662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFCreate()`, `PetscSFSetType()`, `PetscSFSetGraph()` 72095fce210SBarry Smith @*/ 721d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFDuplicate(PetscSF sf, PetscSFDuplicateOption opt, PetscSF *newsf) 722d71ae5a4SJacob Faibussowitsch { 72329046d53SLisandro Dalcin PetscSFType type; 72497929ea7SJunchao Zhang MPI_Datatype dtype = MPIU_SCALAR; 72595fce210SBarry Smith 72695fce210SBarry Smith PetscFunctionBegin; 72729046d53SLisandro Dalcin PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 72829046d53SLisandro Dalcin PetscValidLogicalCollectiveEnum(sf, opt, 2); 72929046d53SLisandro Dalcin PetscValidPointer(newsf, 3); 7309566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sf), newsf)); 7319566063dSJacob Faibussowitsch PetscCall(PetscSFGetType(sf, &type)); 7329566063dSJacob Faibussowitsch if (type) PetscCall(PetscSFSetType(*newsf, type)); 73335cb6cd3SPierre Jolivet (*newsf)->allow_multi_leaves = sf->allow_multi_leaves; /* Dup this flag earlier since PetscSFSetGraph() below checks on this flag */ 73495fce210SBarry Smith if (opt == PETSCSF_DUPLICATE_GRAPH) { 735dd5b3ca6SJunchao Zhang PetscSFCheckGraphSet(sf, 1); 736dd5b3ca6SJunchao Zhang if (sf->pattern == PETSCSF_PATTERN_GENERAL) { 73795fce210SBarry Smith PetscInt nroots, nleaves; 73895fce210SBarry Smith const PetscInt *ilocal; 73995fce210SBarry Smith const PetscSFNode *iremote; 7409566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, &iremote)); 7419566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*newsf, nroots, nleaves, (PetscInt *)ilocal, PETSC_COPY_VALUES, (PetscSFNode *)iremote, PETSC_COPY_VALUES)); 742dd5b3ca6SJunchao Zhang } else { 7439566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraphWithPattern(*newsf, sf->map, sf->pattern)); 744dd5b3ca6SJunchao Zhang } 74595fce210SBarry Smith } 74697929ea7SJunchao Zhang /* Since oldtype is committed, so is newtype, according to MPI */ 7479566063dSJacob Faibussowitsch if (sf->vscat.bs > 1) PetscCallMPI(MPI_Type_dup(sf->vscat.unit, &dtype)); 74897929ea7SJunchao Zhang (*newsf)->vscat.bs = sf->vscat.bs; 74997929ea7SJunchao Zhang (*newsf)->vscat.unit = dtype; 75097929ea7SJunchao Zhang (*newsf)->vscat.to_n = sf->vscat.to_n; 75197929ea7SJunchao Zhang (*newsf)->vscat.from_n = sf->vscat.from_n; 75297929ea7SJunchao Zhang /* Do not copy lsf. Build it on demand since it is rarely used */ 75397929ea7SJunchao Zhang 75420c24465SJunchao Zhang #if defined(PETSC_HAVE_DEVICE) 75520c24465SJunchao Zhang (*newsf)->backend = sf->backend; 75671438e86SJunchao Zhang (*newsf)->unknown_input_stream = sf->unknown_input_stream; 75720c24465SJunchao Zhang (*newsf)->use_gpu_aware_mpi = sf->use_gpu_aware_mpi; 75820c24465SJunchao Zhang (*newsf)->use_stream_aware_mpi = sf->use_stream_aware_mpi; 75920c24465SJunchao Zhang #endif 760dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, Duplicate, opt, *newsf); 76120c24465SJunchao Zhang /* Don't do PetscSFSetUp() since the new sf's graph might have not been set. */ 7623ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 76395fce210SBarry Smith } 76495fce210SBarry Smith 76595fce210SBarry Smith /*@C 76695fce210SBarry Smith PetscSFGetGraph - Get the graph specifying a parallel star forest 76795fce210SBarry Smith 76895fce210SBarry Smith Not Collective 76995fce210SBarry Smith 7704165533cSJose E. Roman Input Parameter: 77195fce210SBarry Smith . sf - star forest 77295fce210SBarry Smith 7734165533cSJose E. Roman Output Parameters: 77495fce210SBarry Smith + nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves) 77595fce210SBarry Smith . nleaves - number of leaf vertices on the current process, each of these references a root on any process 77620662ed9SBarry Smith . ilocal - locations of leaves in leafdata buffers (if returned value is `NULL`, it means leaves are in contiguous storage) 77795fce210SBarry Smith - iremote - remote locations of root vertices for each leaf on the current process 77895fce210SBarry Smith 779cab54364SBarry Smith Level: intermediate 780cab54364SBarry Smith 781373e0d91SLisandro Dalcin Notes: 78220662ed9SBarry Smith We are not currently requiring that the graph is set, thus returning `nroots` = -1 if it has not been set yet 783373e0d91SLisandro Dalcin 78420662ed9SBarry Smith The returned `ilocal` and `iremote` might contain values in different order than the input ones in `PetscSFSetGraph()` 785db2b9530SVaclav Hapla 7868dbb0df6SBarry Smith Fortran Notes: 78720662ed9SBarry Smith The returned `iremote` array is a copy and must be deallocated after use. Consequently, if you 78820662ed9SBarry Smith want to update the graph, you must call `PetscSFSetGraph()` after modifying the `iremote` array. 7898dbb0df6SBarry Smith 79020662ed9SBarry Smith To check for a `NULL` `ilocal` use 7918dbb0df6SBarry Smith $ if (loc(ilocal) == loc(PETSC_NULL_INTEGER)) then 792ca797d7aSLawrence Mitchell 79320662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFSetGraph()` 79495fce210SBarry Smith @*/ 795d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetGraph(PetscSF sf, PetscInt *nroots, PetscInt *nleaves, const PetscInt **ilocal, const PetscSFNode **iremote) 796d71ae5a4SJacob Faibussowitsch { 79795fce210SBarry Smith PetscFunctionBegin; 79895fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 799b8dee149SJunchao Zhang if (sf->ops->GetGraph) { 8009566063dSJacob Faibussowitsch PetscCall((sf->ops->GetGraph)(sf, nroots, nleaves, ilocal, iremote)); 801b8dee149SJunchao Zhang } else { 80295fce210SBarry Smith if (nroots) *nroots = sf->nroots; 80395fce210SBarry Smith if (nleaves) *nleaves = sf->nleaves; 80495fce210SBarry Smith if (ilocal) *ilocal = sf->mine; 80595fce210SBarry Smith if (iremote) *iremote = sf->remote; 806b8dee149SJunchao Zhang } 8073ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 80895fce210SBarry Smith } 80995fce210SBarry Smith 81029046d53SLisandro Dalcin /*@ 81195fce210SBarry Smith PetscSFGetLeafRange - Get the active leaf ranges 81295fce210SBarry Smith 81395fce210SBarry Smith Not Collective 81495fce210SBarry Smith 8154165533cSJose E. Roman Input Parameter: 81695fce210SBarry Smith . sf - star forest 81795fce210SBarry Smith 8184165533cSJose E. Roman Output Parameters: 81920662ed9SBarry Smith + minleaf - minimum active leaf on this process. Returns 0 if there are no leaves. 82020662ed9SBarry Smith - maxleaf - maximum active leaf on this process. Returns -1 if there are no leaves. 82195fce210SBarry Smith 82295fce210SBarry Smith Level: developer 82395fce210SBarry Smith 82420662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFType`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFSetGraph()`, `PetscSFGetGraph()` 82595fce210SBarry Smith @*/ 826d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetLeafRange(PetscSF sf, PetscInt *minleaf, PetscInt *maxleaf) 827d71ae5a4SJacob Faibussowitsch { 82895fce210SBarry Smith PetscFunctionBegin; 82995fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 83029046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 83195fce210SBarry Smith if (minleaf) *minleaf = sf->minleaf; 83295fce210SBarry Smith if (maxleaf) *maxleaf = sf->maxleaf; 8333ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 83495fce210SBarry Smith } 83595fce210SBarry Smith 83695fce210SBarry Smith /*@C 837cab54364SBarry Smith PetscSFViewFromOptions - View a `PetscSF` based on arguments in the options database 838fe2efc57SMark 83920f4b53cSBarry Smith Collective 840fe2efc57SMark 841fe2efc57SMark Input Parameters: 842fe2efc57SMark + A - the star forest 843cab54364SBarry Smith . obj - Optional object that provides the prefix for the option names 844736c3998SJose E. Roman - name - command line option 845fe2efc57SMark 846fe2efc57SMark Level: intermediate 847cab54364SBarry Smith 84820662ed9SBarry Smith Note: 84920662ed9SBarry Smith See `PetscObjectViewFromOptions()` for possible `PetscViewer` and `PetscViewerFormat` 85020662ed9SBarry Smith 851db781477SPatrick Sanan .seealso: `PetscSF`, `PetscSFView`, `PetscObjectViewFromOptions()`, `PetscSFCreate()` 852fe2efc57SMark @*/ 853d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFViewFromOptions(PetscSF A, PetscObject obj, const char name[]) 854d71ae5a4SJacob Faibussowitsch { 855fe2efc57SMark PetscFunctionBegin; 856fe2efc57SMark PetscValidHeaderSpecific(A, PETSCSF_CLASSID, 1); 8579566063dSJacob Faibussowitsch PetscCall(PetscObjectViewFromOptions((PetscObject)A, obj, name)); 8583ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 859fe2efc57SMark } 860fe2efc57SMark 861fe2efc57SMark /*@C 86295fce210SBarry Smith PetscSFView - view a star forest 86395fce210SBarry Smith 86495fce210SBarry Smith Collective 86595fce210SBarry Smith 8664165533cSJose E. Roman Input Parameters: 86795fce210SBarry Smith + sf - star forest 868cab54364SBarry Smith - viewer - viewer to display graph, for example `PETSC_VIEWER_STDOUT_WORLD` 86995fce210SBarry Smith 87095fce210SBarry Smith Level: beginner 87195fce210SBarry Smith 872cab54364SBarry Smith .seealso: `PetscSF`, `PetscViewer`, `PetscSFCreate()`, `PetscSFSetGraph()` 87395fce210SBarry Smith @*/ 874d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFView(PetscSF sf, PetscViewer viewer) 875d71ae5a4SJacob Faibussowitsch { 87695fce210SBarry Smith PetscBool iascii; 87795fce210SBarry Smith PetscViewerFormat format; 87895fce210SBarry Smith 87995fce210SBarry Smith PetscFunctionBegin; 88095fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 8819566063dSJacob Faibussowitsch if (!viewer) PetscCall(PetscViewerASCIIGetStdout(PetscObjectComm((PetscObject)sf), &viewer)); 88295fce210SBarry Smith PetscValidHeaderSpecific(viewer, PETSC_VIEWER_CLASSID, 2); 88395fce210SBarry Smith PetscCheckSameComm(sf, 1, viewer, 2); 8849566063dSJacob Faibussowitsch if (sf->graphset) PetscCall(PetscSFSetUp(sf)); 8859566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii)); 88653dd6d7dSJunchao Zhang if (iascii && viewer->format != PETSC_VIEWER_ASCII_MATLAB) { 88795fce210SBarry Smith PetscMPIInt rank; 88881bfa7aaSJed Brown PetscInt ii, i, j; 88995fce210SBarry Smith 8909566063dSJacob Faibussowitsch PetscCall(PetscObjectPrintClassNamePrefixType((PetscObject)sf, viewer)); 8919566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPushTab(viewer)); 892dd5b3ca6SJunchao Zhang if (sf->pattern == PETSCSF_PATTERN_GENERAL) { 89380153354SVaclav Hapla if (!sf->graphset) { 8949566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(viewer, "PetscSFSetGraph() has not been called yet\n")); 8959566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPopTab(viewer)); 8963ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 89780153354SVaclav Hapla } 8989566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank)); 8999566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPushSynchronized(viewer)); 9009566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Number of roots=%" PetscInt_FMT ", leaves=%" PetscInt_FMT ", remote ranks=%" PetscInt_FMT "\n", rank, sf->nroots, sf->nleaves, sf->nranks)); 90148a46eb9SPierre Jolivet for (i = 0; i < sf->nleaves; i++) PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] %" PetscInt_FMT " <- (%" PetscInt_FMT ",%" PetscInt_FMT ")\n", rank, sf->mine ? sf->mine[i] : i, sf->remote[i].rank, sf->remote[i].index)); 9029566063dSJacob Faibussowitsch PetscCall(PetscViewerFlush(viewer)); 9039566063dSJacob Faibussowitsch PetscCall(PetscViewerGetFormat(viewer, &format)); 90495fce210SBarry Smith if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 90581bfa7aaSJed Brown PetscMPIInt *tmpranks, *perm; 9069566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(sf->nranks, &tmpranks, sf->nranks, &perm)); 9079566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(tmpranks, sf->ranks, sf->nranks)); 90881bfa7aaSJed Brown for (i = 0; i < sf->nranks; i++) perm[i] = i; 9099566063dSJacob Faibussowitsch PetscCall(PetscSortMPIIntWithArray(sf->nranks, tmpranks, perm)); 9109566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Roots referenced by my leaves, by rank\n", rank)); 91181bfa7aaSJed Brown for (ii = 0; ii < sf->nranks; ii++) { 91281bfa7aaSJed Brown i = perm[ii]; 9139566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] %d: %" PetscInt_FMT " edges\n", rank, sf->ranks[i], sf->roffset[i + 1] - sf->roffset[i])); 91448a46eb9SPierre Jolivet for (j = sf->roffset[i]; j < sf->roffset[i + 1]; j++) PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] %" PetscInt_FMT " <- %" PetscInt_FMT "\n", rank, sf->rmine[j], sf->rremote[j])); 91595fce210SBarry Smith } 9169566063dSJacob Faibussowitsch PetscCall(PetscFree2(tmpranks, perm)); 91795fce210SBarry Smith } 9189566063dSJacob Faibussowitsch PetscCall(PetscViewerFlush(viewer)); 9199566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPopSynchronized(viewer)); 920dd5b3ca6SJunchao Zhang } 9219566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPopTab(viewer)); 92295fce210SBarry Smith } 923dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, View, viewer); 9243ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 92595fce210SBarry Smith } 92695fce210SBarry Smith 92795fce210SBarry Smith /*@C 928dec1416fSJunchao Zhang PetscSFGetRootRanks - Get root ranks and number of vertices referenced by leaves on this process 92995fce210SBarry Smith 93095fce210SBarry Smith Not Collective 93195fce210SBarry Smith 9324165533cSJose E. Roman Input Parameter: 93395fce210SBarry Smith . sf - star forest 93495fce210SBarry Smith 9354165533cSJose E. Roman Output Parameters: 93695fce210SBarry Smith + nranks - number of ranks referenced by local part 93720662ed9SBarry Smith . ranks - [`nranks`] array of ranks 93820662ed9SBarry Smith . roffset - [`nranks`+1] offset in `rmine`/`rremote` for each rank 93920662ed9SBarry Smith . rmine - [`roffset`[`nranks`]] concatenated array holding local indices referencing each remote rank 94020662ed9SBarry Smith - rremote - [`roffset`[`nranks`]] concatenated array holding remote indices referenced for each remote rank 94195fce210SBarry Smith 94295fce210SBarry Smith Level: developer 94395fce210SBarry Smith 944cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetLeafRanks()` 94595fce210SBarry Smith @*/ 946d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetRootRanks(PetscSF sf, PetscInt *nranks, const PetscMPIInt **ranks, const PetscInt **roffset, const PetscInt **rmine, const PetscInt **rremote) 947d71ae5a4SJacob Faibussowitsch { 94895fce210SBarry Smith PetscFunctionBegin; 94995fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 95028b400f6SJacob Faibussowitsch PetscCheck(sf->setupcalled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetUp() before obtaining ranks"); 951dec1416fSJunchao Zhang if (sf->ops->GetRootRanks) { 9529566063dSJacob Faibussowitsch PetscCall((sf->ops->GetRootRanks)(sf, nranks, ranks, roffset, rmine, rremote)); 953dec1416fSJunchao Zhang } else { 954dec1416fSJunchao Zhang /* The generic implementation */ 95595fce210SBarry Smith if (nranks) *nranks = sf->nranks; 95695fce210SBarry Smith if (ranks) *ranks = sf->ranks; 95795fce210SBarry Smith if (roffset) *roffset = sf->roffset; 95895fce210SBarry Smith if (rmine) *rmine = sf->rmine; 95995fce210SBarry Smith if (rremote) *rremote = sf->rremote; 960dec1416fSJunchao Zhang } 9613ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 96295fce210SBarry Smith } 96395fce210SBarry Smith 9648750ddebSJunchao Zhang /*@C 9658750ddebSJunchao Zhang PetscSFGetLeafRanks - Get leaf ranks referencing roots on this process 9668750ddebSJunchao Zhang 9678750ddebSJunchao Zhang Not Collective 9688750ddebSJunchao Zhang 9694165533cSJose E. Roman Input Parameter: 9708750ddebSJunchao Zhang . sf - star forest 9718750ddebSJunchao Zhang 9724165533cSJose E. Roman Output Parameters: 9738750ddebSJunchao Zhang + niranks - number of leaf ranks referencing roots on this process 97420662ed9SBarry Smith . iranks - [`niranks`] array of ranks 97520662ed9SBarry Smith . ioffset - [`niranks`+1] offset in `irootloc` for each rank 97620662ed9SBarry Smith - irootloc - [`ioffset`[`niranks`]] concatenated array holding local indices of roots referenced by each leaf rank 9778750ddebSJunchao Zhang 9788750ddebSJunchao Zhang Level: developer 9798750ddebSJunchao Zhang 980cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetRootRanks()` 9818750ddebSJunchao Zhang @*/ 982d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetLeafRanks(PetscSF sf, PetscInt *niranks, const PetscMPIInt **iranks, const PetscInt **ioffset, const PetscInt **irootloc) 983d71ae5a4SJacob Faibussowitsch { 9848750ddebSJunchao Zhang PetscFunctionBegin; 9858750ddebSJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 98628b400f6SJacob Faibussowitsch PetscCheck(sf->setupcalled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetUp() before obtaining ranks"); 9878750ddebSJunchao Zhang if (sf->ops->GetLeafRanks) { 9889566063dSJacob Faibussowitsch PetscCall((sf->ops->GetLeafRanks)(sf, niranks, iranks, ioffset, irootloc)); 9898750ddebSJunchao Zhang } else { 9908750ddebSJunchao Zhang PetscSFType type; 9919566063dSJacob Faibussowitsch PetscCall(PetscSFGetType(sf, &type)); 99298921bdaSJacob Faibussowitsch SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "PetscSFGetLeafRanks() is not supported on this StarForest type: %s", type); 9938750ddebSJunchao Zhang } 9943ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 9958750ddebSJunchao Zhang } 9968750ddebSJunchao Zhang 997d71ae5a4SJacob Faibussowitsch static PetscBool InList(PetscMPIInt needle, PetscMPIInt n, const PetscMPIInt *list) 998d71ae5a4SJacob Faibussowitsch { 999b5a8e515SJed Brown PetscInt i; 1000b5a8e515SJed Brown for (i = 0; i < n; i++) { 1001b5a8e515SJed Brown if (needle == list[i]) return PETSC_TRUE; 1002b5a8e515SJed Brown } 1003b5a8e515SJed Brown return PETSC_FALSE; 1004b5a8e515SJed Brown } 1005b5a8e515SJed Brown 100695fce210SBarry Smith /*@C 1007cab54364SBarry Smith PetscSFSetUpRanks - Set up data structures associated with ranks; this is for internal use by `PetscSF` implementations. 100821c688dcSJed Brown 100921c688dcSJed Brown Collective 101021c688dcSJed Brown 10114165533cSJose E. Roman Input Parameters: 1012cab54364SBarry Smith + sf - `PetscSF` to set up; `PetscSFSetGraph()` must have been called 1013cab54364SBarry Smith - dgroup - `MPI_Group` of ranks to be distinguished (e.g., for self or shared memory exchange) 101421c688dcSJed Brown 101521c688dcSJed Brown Level: developer 101621c688dcSJed Brown 1017cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetRootRanks()` 101821c688dcSJed Brown @*/ 1019d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetUpRanks(PetscSF sf, MPI_Group dgroup) 1020d71ae5a4SJacob Faibussowitsch { 1021eec179cfSJacob Faibussowitsch PetscHMapI table; 1022eec179cfSJacob Faibussowitsch PetscHashIter pos; 1023b5a8e515SJed Brown PetscMPIInt size, groupsize, *groupranks; 1024247e8311SStefano Zampini PetscInt *rcount, *ranks; 1025247e8311SStefano Zampini PetscInt i, irank = -1, orank = -1; 102621c688dcSJed Brown 102721c688dcSJed Brown PetscFunctionBegin; 102821c688dcSJed Brown PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 102929046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 10309566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)sf), &size)); 1031eec179cfSJacob Faibussowitsch PetscCall(PetscHMapICreateWithSize(10, &table)); 103221c688dcSJed Brown for (i = 0; i < sf->nleaves; i++) { 103321c688dcSJed Brown /* Log 1-based rank */ 1034eec179cfSJacob Faibussowitsch PetscCall(PetscHMapISetWithMode(table, sf->remote[i].rank + 1, 1, ADD_VALUES)); 103521c688dcSJed Brown } 1036eec179cfSJacob Faibussowitsch PetscCall(PetscHMapIGetSize(table, &sf->nranks)); 10379566063dSJacob Faibussowitsch PetscCall(PetscMalloc4(sf->nranks, &sf->ranks, sf->nranks + 1, &sf->roffset, sf->nleaves, &sf->rmine, sf->nleaves, &sf->rremote)); 10389566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(sf->nranks, &rcount, sf->nranks, &ranks)); 1039eec179cfSJacob Faibussowitsch PetscHashIterBegin(table, pos); 104021c688dcSJed Brown for (i = 0; i < sf->nranks; i++) { 1041eec179cfSJacob Faibussowitsch PetscHashIterGetKey(table, pos, ranks[i]); 1042eec179cfSJacob Faibussowitsch PetscHashIterGetVal(table, pos, rcount[i]); 1043eec179cfSJacob Faibussowitsch PetscHashIterNext(table, pos); 104421c688dcSJed Brown ranks[i]--; /* Convert back to 0-based */ 104521c688dcSJed Brown } 1046eec179cfSJacob Faibussowitsch PetscCall(PetscHMapIDestroy(&table)); 1047b5a8e515SJed Brown 1048b5a8e515SJed Brown /* We expect that dgroup is reliably "small" while nranks could be large */ 1049b5a8e515SJed Brown { 10507fb8a5e4SKarl Rupp MPI_Group group = MPI_GROUP_NULL; 1051b5a8e515SJed Brown PetscMPIInt *dgroupranks; 10529566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf), &group)); 10539566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_size(dgroup, &groupsize)); 10549566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(groupsize, &dgroupranks)); 10559566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(groupsize, &groupranks)); 1056b5a8e515SJed Brown for (i = 0; i < groupsize; i++) dgroupranks[i] = i; 10579566063dSJacob Faibussowitsch if (groupsize) PetscCallMPI(MPI_Group_translate_ranks(dgroup, groupsize, dgroupranks, group, groupranks)); 10589566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_free(&group)); 10599566063dSJacob Faibussowitsch PetscCall(PetscFree(dgroupranks)); 1060b5a8e515SJed Brown } 1061b5a8e515SJed Brown 1062b5a8e515SJed Brown /* Partition ranks[] into distinguished (first sf->ndranks) followed by non-distinguished */ 1063b5a8e515SJed Brown for (sf->ndranks = 0, i = sf->nranks; sf->ndranks < i;) { 1064b5a8e515SJed Brown for (i--; sf->ndranks < i; i--) { /* Scan i backward looking for distinguished rank */ 1065b5a8e515SJed Brown if (InList(ranks[i], groupsize, groupranks)) break; 1066b5a8e515SJed Brown } 1067b5a8e515SJed Brown for (; sf->ndranks <= i; sf->ndranks++) { /* Scan sf->ndranks forward looking for non-distinguished rank */ 1068b5a8e515SJed Brown if (!InList(ranks[sf->ndranks], groupsize, groupranks)) break; 1069b5a8e515SJed Brown } 1070b5a8e515SJed Brown if (sf->ndranks < i) { /* Swap ranks[sf->ndranks] with ranks[i] */ 1071b5a8e515SJed Brown PetscInt tmprank, tmpcount; 1072247e8311SStefano Zampini 1073b5a8e515SJed Brown tmprank = ranks[i]; 1074b5a8e515SJed Brown tmpcount = rcount[i]; 1075b5a8e515SJed Brown ranks[i] = ranks[sf->ndranks]; 1076b5a8e515SJed Brown rcount[i] = rcount[sf->ndranks]; 1077b5a8e515SJed Brown ranks[sf->ndranks] = tmprank; 1078b5a8e515SJed Brown rcount[sf->ndranks] = tmpcount; 1079b5a8e515SJed Brown sf->ndranks++; 1080b5a8e515SJed Brown } 1081b5a8e515SJed Brown } 10829566063dSJacob Faibussowitsch PetscCall(PetscFree(groupranks)); 10839566063dSJacob Faibussowitsch PetscCall(PetscSortIntWithArray(sf->ndranks, ranks, rcount)); 10849566063dSJacob Faibussowitsch PetscCall(PetscSortIntWithArray(sf->nranks - sf->ndranks, ranks + sf->ndranks, rcount + sf->ndranks)); 108521c688dcSJed Brown sf->roffset[0] = 0; 108621c688dcSJed Brown for (i = 0; i < sf->nranks; i++) { 10879566063dSJacob Faibussowitsch PetscCall(PetscMPIIntCast(ranks[i], sf->ranks + i)); 108821c688dcSJed Brown sf->roffset[i + 1] = sf->roffset[i] + rcount[i]; 108921c688dcSJed Brown rcount[i] = 0; 109021c688dcSJed Brown } 1091247e8311SStefano Zampini for (i = 0, irank = -1, orank = -1; i < sf->nleaves; i++) { 1092247e8311SStefano Zampini /* short circuit */ 1093247e8311SStefano Zampini if (orank != sf->remote[i].rank) { 109421c688dcSJed Brown /* Search for index of iremote[i].rank in sf->ranks */ 10959566063dSJacob Faibussowitsch PetscCall(PetscFindMPIInt(sf->remote[i].rank, sf->ndranks, sf->ranks, &irank)); 1096b5a8e515SJed Brown if (irank < 0) { 10979566063dSJacob Faibussowitsch PetscCall(PetscFindMPIInt(sf->remote[i].rank, sf->nranks - sf->ndranks, sf->ranks + sf->ndranks, &irank)); 1098b5a8e515SJed Brown if (irank >= 0) irank += sf->ndranks; 109921c688dcSJed Brown } 1100247e8311SStefano Zampini orank = sf->remote[i].rank; 1101247e8311SStefano Zampini } 110208401ef6SPierre Jolivet PetscCheck(irank >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Could not find rank %" PetscInt_FMT " in array", sf->remote[i].rank); 110321c688dcSJed Brown sf->rmine[sf->roffset[irank] + rcount[irank]] = sf->mine ? sf->mine[i] : i; 110421c688dcSJed Brown sf->rremote[sf->roffset[irank] + rcount[irank]] = sf->remote[i].index; 110521c688dcSJed Brown rcount[irank]++; 110621c688dcSJed Brown } 11079566063dSJacob Faibussowitsch PetscCall(PetscFree2(rcount, ranks)); 11083ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 110921c688dcSJed Brown } 111021c688dcSJed Brown 111121c688dcSJed Brown /*@C 111295fce210SBarry Smith PetscSFGetGroups - gets incoming and outgoing process groups 111395fce210SBarry Smith 111495fce210SBarry Smith Collective 111595fce210SBarry Smith 11164165533cSJose E. Roman Input Parameter: 111795fce210SBarry Smith . sf - star forest 111895fce210SBarry Smith 11194165533cSJose E. Roman Output Parameters: 112095fce210SBarry Smith + incoming - group of origin processes for incoming edges (leaves that reference my roots) 112195fce210SBarry Smith - outgoing - group of destination processes for outgoing edges (roots that I reference) 112295fce210SBarry Smith 112395fce210SBarry Smith Level: developer 112495fce210SBarry Smith 1125cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetWindow()`, `PetscSFRestoreWindow()` 112695fce210SBarry Smith @*/ 1127d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetGroups(PetscSF sf, MPI_Group *incoming, MPI_Group *outgoing) 1128d71ae5a4SJacob Faibussowitsch { 11297fb8a5e4SKarl Rupp MPI_Group group = MPI_GROUP_NULL; 113095fce210SBarry Smith 113195fce210SBarry Smith PetscFunctionBegin; 113208401ef6SPierre Jolivet PetscCheck(sf->nranks >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetUpRanks() before obtaining groups"); 113395fce210SBarry Smith if (sf->ingroup == MPI_GROUP_NULL) { 113495fce210SBarry Smith PetscInt i; 113595fce210SBarry Smith const PetscInt *indegree; 113695fce210SBarry Smith PetscMPIInt rank, *outranks, *inranks; 113795fce210SBarry Smith PetscSFNode *remote; 113895fce210SBarry Smith PetscSF bgcount; 113995fce210SBarry Smith 114095fce210SBarry Smith /* Compute the number of incoming ranks */ 11419566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(sf->nranks, &remote)); 114295fce210SBarry Smith for (i = 0; i < sf->nranks; i++) { 114395fce210SBarry Smith remote[i].rank = sf->ranks[i]; 114495fce210SBarry Smith remote[i].index = 0; 114595fce210SBarry Smith } 11469566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_CONFONLY, &bgcount)); 11479566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(bgcount, 1, sf->nranks, NULL, PETSC_COPY_VALUES, remote, PETSC_OWN_POINTER)); 11489566063dSJacob Faibussowitsch PetscCall(PetscSFComputeDegreeBegin(bgcount, &indegree)); 11499566063dSJacob Faibussowitsch PetscCall(PetscSFComputeDegreeEnd(bgcount, &indegree)); 115095fce210SBarry Smith /* Enumerate the incoming ranks */ 11519566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(indegree[0], &inranks, sf->nranks, &outranks)); 11529566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank)); 115395fce210SBarry Smith for (i = 0; i < sf->nranks; i++) outranks[i] = rank; 11549566063dSJacob Faibussowitsch PetscCall(PetscSFGatherBegin(bgcount, MPI_INT, outranks, inranks)); 11559566063dSJacob Faibussowitsch PetscCall(PetscSFGatherEnd(bgcount, MPI_INT, outranks, inranks)); 11569566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf), &group)); 11579566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_incl(group, indegree[0], inranks, &sf->ingroup)); 11589566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_free(&group)); 11599566063dSJacob Faibussowitsch PetscCall(PetscFree2(inranks, outranks)); 11609566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&bgcount)); 116195fce210SBarry Smith } 116295fce210SBarry Smith *incoming = sf->ingroup; 116395fce210SBarry Smith 116495fce210SBarry Smith if (sf->outgroup == MPI_GROUP_NULL) { 11659566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf), &group)); 11669566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_incl(group, sf->nranks, sf->ranks, &sf->outgroup)); 11679566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_free(&group)); 116895fce210SBarry Smith } 116995fce210SBarry Smith *outgoing = sf->outgroup; 11703ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 117195fce210SBarry Smith } 117295fce210SBarry Smith 117329046d53SLisandro Dalcin /*@ 1174cab54364SBarry Smith PetscSFGetMultiSF - gets the inner `PetscSF` implementing gathers and scatters 117595fce210SBarry Smith 117695fce210SBarry Smith Collective 117795fce210SBarry Smith 11784165533cSJose E. Roman Input Parameter: 117995fce210SBarry Smith . sf - star forest that may contain roots with 0 or with more than 1 vertex 118095fce210SBarry Smith 11814165533cSJose E. Roman Output Parameter: 118295fce210SBarry Smith . multi - star forest with split roots, such that each root has degree exactly 1 118395fce210SBarry Smith 118495fce210SBarry Smith Level: developer 118595fce210SBarry Smith 1186cab54364SBarry Smith Note: 1187cab54364SBarry Smith In most cases, users should use `PetscSFGatherBegin()` and `PetscSFScatterBegin()` instead of manipulating multi 118895fce210SBarry Smith directly. Since multi satisfies the stronger condition that each entry in the global space has exactly one incoming 118995fce210SBarry Smith edge, it is a candidate for future optimization that might involve its removal. 119095fce210SBarry Smith 1191cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFGatherBegin()`, `PetscSFScatterBegin()`, `PetscSFComputeMultiRootOriginalNumbering()` 119295fce210SBarry Smith @*/ 1193d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetMultiSF(PetscSF sf, PetscSF *multi) 1194d71ae5a4SJacob Faibussowitsch { 119595fce210SBarry Smith PetscFunctionBegin; 119695fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 119795fce210SBarry Smith PetscValidPointer(multi, 2); 119895fce210SBarry Smith if (sf->nroots < 0) { /* Graph has not been set yet; why do we need this? */ 11999566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_RANKS, &sf->multi)); 120095fce210SBarry Smith *multi = sf->multi; 1201013b3241SStefano Zampini sf->multi->multi = sf->multi; 12023ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 120395fce210SBarry Smith } 120495fce210SBarry Smith if (!sf->multi) { 120595fce210SBarry Smith const PetscInt *indegree; 12069837ea96SMatthew G. Knepley PetscInt i, *inoffset, *outones, *outoffset, maxlocal; 120795fce210SBarry Smith PetscSFNode *remote; 120829046d53SLisandro Dalcin maxlocal = sf->maxleaf + 1; /* TODO: We should use PetscSFGetLeafRange() */ 12099566063dSJacob Faibussowitsch PetscCall(PetscSFComputeDegreeBegin(sf, &indegree)); 12109566063dSJacob Faibussowitsch PetscCall(PetscSFComputeDegreeEnd(sf, &indegree)); 12119566063dSJacob Faibussowitsch PetscCall(PetscMalloc3(sf->nroots + 1, &inoffset, maxlocal, &outones, maxlocal, &outoffset)); 121295fce210SBarry Smith inoffset[0] = 0; 121395fce210SBarry Smith for (i = 0; i < sf->nroots; i++) inoffset[i + 1] = inoffset[i] + indegree[i]; 12149837ea96SMatthew G. Knepley for (i = 0; i < maxlocal; i++) outones[i] = 1; 12159566063dSJacob Faibussowitsch PetscCall(PetscSFFetchAndOpBegin(sf, MPIU_INT, inoffset, outones, outoffset, MPI_SUM)); 12169566063dSJacob Faibussowitsch PetscCall(PetscSFFetchAndOpEnd(sf, MPIU_INT, inoffset, outones, outoffset, MPI_SUM)); 121795fce210SBarry Smith for (i = 0; i < sf->nroots; i++) inoffset[i] -= indegree[i]; /* Undo the increment */ 121876bd3646SJed Brown if (PetscDefined(USE_DEBUG)) { /* Check that the expected number of increments occurred */ 1219ad540459SPierre Jolivet for (i = 0; i < sf->nroots; i++) PetscCheck(inoffset[i] + indegree[i] == inoffset[i + 1], PETSC_COMM_SELF, PETSC_ERR_PLIB, "Incorrect result after PetscSFFetchAndOp"); 122076bd3646SJed Brown } 12219566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(sf->nleaves, &remote)); 122295fce210SBarry Smith for (i = 0; i < sf->nleaves; i++) { 122395fce210SBarry Smith remote[i].rank = sf->remote[i].rank; 122438e7336fSToby Isaac remote[i].index = outoffset[sf->mine ? sf->mine[i] : i]; 122595fce210SBarry Smith } 12269566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_RANKS, &sf->multi)); 1227013b3241SStefano Zampini sf->multi->multi = sf->multi; 12289566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(sf->multi, inoffset[sf->nroots], sf->nleaves, sf->mine, PETSC_COPY_VALUES, remote, PETSC_OWN_POINTER)); 122995fce210SBarry Smith if (sf->rankorder) { /* Sort the ranks */ 123095fce210SBarry Smith PetscMPIInt rank; 123195fce210SBarry Smith PetscInt *inranks, *newoffset, *outranks, *newoutoffset, *tmpoffset, maxdegree; 123295fce210SBarry Smith PetscSFNode *newremote; 12339566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank)); 123495fce210SBarry Smith for (i = 0, maxdegree = 0; i < sf->nroots; i++) maxdegree = PetscMax(maxdegree, indegree[i]); 12359566063dSJacob Faibussowitsch PetscCall(PetscMalloc5(sf->multi->nroots, &inranks, sf->multi->nroots, &newoffset, maxlocal, &outranks, maxlocal, &newoutoffset, maxdegree, &tmpoffset)); 12369837ea96SMatthew G. Knepley for (i = 0; i < maxlocal; i++) outranks[i] = rank; 12379566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(sf->multi, MPIU_INT, outranks, inranks, MPI_REPLACE)); 12389566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(sf->multi, MPIU_INT, outranks, inranks, MPI_REPLACE)); 123995fce210SBarry Smith /* Sort the incoming ranks at each vertex, build the inverse map */ 124095fce210SBarry Smith for (i = 0; i < sf->nroots; i++) { 124195fce210SBarry Smith PetscInt j; 124295fce210SBarry Smith for (j = 0; j < indegree[i]; j++) tmpoffset[j] = j; 12439566063dSJacob Faibussowitsch PetscCall(PetscSortIntWithArray(indegree[i], inranks + inoffset[i], tmpoffset)); 124495fce210SBarry Smith for (j = 0; j < indegree[i]; j++) newoffset[inoffset[i] + tmpoffset[j]] = inoffset[i] + j; 124595fce210SBarry Smith } 12469566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(sf->multi, MPIU_INT, newoffset, newoutoffset, MPI_REPLACE)); 12479566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(sf->multi, MPIU_INT, newoffset, newoutoffset, MPI_REPLACE)); 12489566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(sf->nleaves, &newremote)); 124995fce210SBarry Smith for (i = 0; i < sf->nleaves; i++) { 125095fce210SBarry Smith newremote[i].rank = sf->remote[i].rank; 125101365b40SToby Isaac newremote[i].index = newoutoffset[sf->mine ? sf->mine[i] : i]; 125295fce210SBarry Smith } 12539566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(sf->multi, inoffset[sf->nroots], sf->nleaves, sf->mine, PETSC_COPY_VALUES, newremote, PETSC_OWN_POINTER)); 12549566063dSJacob Faibussowitsch PetscCall(PetscFree5(inranks, newoffset, outranks, newoutoffset, tmpoffset)); 125595fce210SBarry Smith } 12569566063dSJacob Faibussowitsch PetscCall(PetscFree3(inoffset, outones, outoffset)); 125795fce210SBarry Smith } 125895fce210SBarry Smith *multi = sf->multi; 12593ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 126095fce210SBarry Smith } 126195fce210SBarry Smith 126295fce210SBarry Smith /*@C 126320662ed9SBarry Smith PetscSFCreateEmbeddedRootSF - removes edges from all but the selected roots of a `PetscSF`, does not remap indices 126495fce210SBarry Smith 126595fce210SBarry Smith Collective 126695fce210SBarry Smith 12674165533cSJose E. Roman Input Parameters: 126895fce210SBarry Smith + sf - original star forest 1269ba2a7774SJunchao Zhang . nselected - number of selected roots on this process 1270ba2a7774SJunchao Zhang - selected - indices of the selected roots on this process 127195fce210SBarry Smith 12724165533cSJose E. Roman Output Parameter: 1273cd620004SJunchao Zhang . esf - new star forest 127495fce210SBarry Smith 127595fce210SBarry Smith Level: advanced 127695fce210SBarry Smith 127795fce210SBarry Smith Note: 1278cab54364SBarry Smith To use the new `PetscSF`, it may be necessary to know the indices of the leaves that are still participating. This can 127995fce210SBarry Smith be done by calling PetscSFGetGraph(). 128095fce210SBarry Smith 1281cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFGetGraph()` 128295fce210SBarry Smith @*/ 1283d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateEmbeddedRootSF(PetscSF sf, PetscInt nselected, const PetscInt *selected, PetscSF *esf) 1284d71ae5a4SJacob Faibussowitsch { 1285cd620004SJunchao Zhang PetscInt i, j, n, nroots, nleaves, esf_nleaves, *new_ilocal, minleaf, maxleaf, maxlocal; 1286cd620004SJunchao Zhang const PetscInt *ilocal; 1287cd620004SJunchao Zhang signed char *rootdata, *leafdata, *leafmem; 1288ba2a7774SJunchao Zhang const PetscSFNode *iremote; 1289f659e5c7SJunchao Zhang PetscSFNode *new_iremote; 1290f659e5c7SJunchao Zhang MPI_Comm comm; 129195fce210SBarry Smith 129295fce210SBarry Smith PetscFunctionBegin; 129395fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 129429046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 1295dadcf809SJacob Faibussowitsch if (nselected) PetscValidIntPointer(selected, 3); 1296cd620004SJunchao Zhang PetscValidPointer(esf, 4); 12970511a646SMatthew G. Knepley 12989566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 12999566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_EmbedSF, sf, 0, 0, 0)); 13009566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 13019566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, &iremote)); 1302cd620004SJunchao Zhang 130376bd3646SJed Brown if (PetscDefined(USE_DEBUG)) { /* Error out if selected[] has dups or out of range indices */ 1304cd620004SJunchao Zhang PetscBool dups; 13059566063dSJacob Faibussowitsch PetscCall(PetscCheckDupsInt(nselected, selected, &dups)); 130628b400f6SJacob Faibussowitsch PetscCheck(!dups, comm, PETSC_ERR_ARG_WRONG, "selected[] has dups"); 13079371c9d4SSatish Balay for (i = 0; i < nselected; i++) PetscCheck(selected[i] >= 0 && selected[i] < nroots, comm, PETSC_ERR_ARG_OUTOFRANGE, "selected root indice %" PetscInt_FMT " is out of [0,%" PetscInt_FMT ")", selected[i], nroots); 1308cd620004SJunchao Zhang } 1309f659e5c7SJunchao Zhang 1310dbbe0bcdSBarry Smith if (sf->ops->CreateEmbeddedRootSF) PetscUseTypeMethod(sf, CreateEmbeddedRootSF, nselected, selected, esf); 1311dbbe0bcdSBarry Smith else { 1312cd620004SJunchao Zhang /* A generic version of creating embedded sf */ 13139566063dSJacob Faibussowitsch PetscCall(PetscSFGetLeafRange(sf, &minleaf, &maxleaf)); 1314cd620004SJunchao Zhang maxlocal = maxleaf - minleaf + 1; 13159566063dSJacob Faibussowitsch PetscCall(PetscCalloc2(nroots, &rootdata, maxlocal, &leafmem)); 1316cd620004SJunchao Zhang leafdata = leafmem - minleaf; 1317cd620004SJunchao Zhang /* Tag selected roots and bcast to leaves */ 1318cd620004SJunchao Zhang for (i = 0; i < nselected; i++) rootdata[selected[i]] = 1; 13199566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(sf, MPI_SIGNED_CHAR, rootdata, leafdata, MPI_REPLACE)); 13209566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(sf, MPI_SIGNED_CHAR, rootdata, leafdata, MPI_REPLACE)); 1321ba2a7774SJunchao Zhang 1322cd620004SJunchao Zhang /* Build esf with leaves that are still connected */ 1323cd620004SJunchao Zhang esf_nleaves = 0; 1324cd620004SJunchao Zhang for (i = 0; i < nleaves; i++) { 1325cd620004SJunchao Zhang j = ilocal ? ilocal[i] : i; 1326cd620004SJunchao Zhang /* esf_nleaves += leafdata[j] should work in theory, but failed with SFWindow bugs 1327cd620004SJunchao Zhang with PetscSFBcast. See https://gitlab.com/petsc/petsc/issues/555 1328cd620004SJunchao Zhang */ 1329cd620004SJunchao Zhang esf_nleaves += (leafdata[j] ? 1 : 0); 1330cd620004SJunchao Zhang } 13319566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(esf_nleaves, &new_ilocal)); 13329566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(esf_nleaves, &new_iremote)); 1333cd620004SJunchao Zhang for (i = n = 0; i < nleaves; i++) { 1334cd620004SJunchao Zhang j = ilocal ? ilocal[i] : i; 1335cd620004SJunchao Zhang if (leafdata[j]) { 1336cd620004SJunchao Zhang new_ilocal[n] = j; 1337cd620004SJunchao Zhang new_iremote[n].rank = iremote[i].rank; 1338cd620004SJunchao Zhang new_iremote[n].index = iremote[i].index; 1339fc1ede2bSMatthew G. Knepley ++n; 134095fce210SBarry Smith } 134195fce210SBarry Smith } 13429566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, esf)); 13439566063dSJacob Faibussowitsch PetscCall(PetscSFSetFromOptions(*esf)); 13449566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*esf, nroots, esf_nleaves, new_ilocal, PETSC_OWN_POINTER, new_iremote, PETSC_OWN_POINTER)); 13459566063dSJacob Faibussowitsch PetscCall(PetscFree2(rootdata, leafmem)); 1346f659e5c7SJunchao Zhang } 13479566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_EmbedSF, sf, 0, 0, 0)); 13483ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 134995fce210SBarry Smith } 135095fce210SBarry Smith 13512f5fb4c2SMatthew G. Knepley /*@C 135220662ed9SBarry Smith PetscSFCreateEmbeddedLeafSF - removes edges from all but the selected leaves of a `PetscSF`, does not remap indices 13532f5fb4c2SMatthew G. Knepley 13542f5fb4c2SMatthew G. Knepley Collective 13552f5fb4c2SMatthew G. Knepley 13564165533cSJose E. Roman Input Parameters: 13572f5fb4c2SMatthew G. Knepley + sf - original star forest 1358f659e5c7SJunchao Zhang . nselected - number of selected leaves on this process 1359f659e5c7SJunchao Zhang - selected - indices of the selected leaves on this process 13602f5fb4c2SMatthew G. Knepley 13614165533cSJose E. Roman Output Parameter: 13622f5fb4c2SMatthew G. Knepley . newsf - new star forest 13632f5fb4c2SMatthew G. Knepley 13642f5fb4c2SMatthew G. Knepley Level: advanced 13652f5fb4c2SMatthew G. Knepley 1366cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreateEmbeddedRootSF()`, `PetscSFSetGraph()`, `PetscSFGetGraph()` 13672f5fb4c2SMatthew G. Knepley @*/ 1368d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateEmbeddedLeafSF(PetscSF sf, PetscInt nselected, const PetscInt *selected, PetscSF *newsf) 1369d71ae5a4SJacob Faibussowitsch { 1370f659e5c7SJunchao Zhang const PetscSFNode *iremote; 1371f659e5c7SJunchao Zhang PetscSFNode *new_iremote; 1372f659e5c7SJunchao Zhang const PetscInt *ilocal; 1373f659e5c7SJunchao Zhang PetscInt i, nroots, *leaves, *new_ilocal; 1374f659e5c7SJunchao Zhang MPI_Comm comm; 13752f5fb4c2SMatthew G. Knepley 13762f5fb4c2SMatthew G. Knepley PetscFunctionBegin; 13772f5fb4c2SMatthew G. Knepley PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 137829046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 1379dadcf809SJacob Faibussowitsch if (nselected) PetscValidIntPointer(selected, 3); 13802f5fb4c2SMatthew G. Knepley PetscValidPointer(newsf, 4); 13812f5fb4c2SMatthew G. Knepley 1382f659e5c7SJunchao Zhang /* Uniq selected[] and put results in leaves[] */ 13839566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 13849566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nselected, &leaves)); 13859566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(leaves, selected, nselected)); 13869566063dSJacob Faibussowitsch PetscCall(PetscSortedRemoveDupsInt(&nselected, leaves)); 138708401ef6SPierre Jolivet PetscCheck(!nselected || !(leaves[0] < 0 || leaves[nselected - 1] >= sf->nleaves), comm, PETSC_ERR_ARG_OUTOFRANGE, "Min/Max leaf indices %" PetscInt_FMT "/%" PetscInt_FMT " are not in [0,%" PetscInt_FMT ")", leaves[0], leaves[nselected - 1], sf->nleaves); 1388f659e5c7SJunchao Zhang 1389f659e5c7SJunchao Zhang /* Optimize the routine only when sf is setup and hence we can reuse sf's communication pattern */ 1390dbbe0bcdSBarry Smith if (sf->setupcalled && sf->ops->CreateEmbeddedLeafSF) PetscUseTypeMethod(sf, CreateEmbeddedLeafSF, nselected, leaves, newsf); 1391dbbe0bcdSBarry Smith else { 13929566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, NULL, &ilocal, &iremote)); 13939566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nselected, &new_ilocal)); 13949566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nselected, &new_iremote)); 1395f659e5c7SJunchao Zhang for (i = 0; i < nselected; ++i) { 1396f659e5c7SJunchao Zhang const PetscInt l = leaves[i]; 1397f659e5c7SJunchao Zhang new_ilocal[i] = ilocal ? ilocal[l] : l; 1398f659e5c7SJunchao Zhang new_iremote[i].rank = iremote[l].rank; 1399f659e5c7SJunchao Zhang new_iremote[i].index = iremote[l].index; 14002f5fb4c2SMatthew G. Knepley } 14019566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_CONFONLY, newsf)); 14029566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*newsf, nroots, nselected, new_ilocal, PETSC_OWN_POINTER, new_iremote, PETSC_OWN_POINTER)); 1403f659e5c7SJunchao Zhang } 14049566063dSJacob Faibussowitsch PetscCall(PetscFree(leaves)); 14053ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 14062f5fb4c2SMatthew G. Knepley } 14072f5fb4c2SMatthew G. Knepley 140895fce210SBarry Smith /*@C 1409cab54364SBarry Smith PetscSFBcastBegin - begin pointwise broadcast with root value being reduced to leaf value, to be concluded with call to `PetscSFBcastEnd()` 14103482bfa8SJunchao Zhang 1411c3339decSBarry Smith Collective 14123482bfa8SJunchao Zhang 14134165533cSJose E. Roman Input Parameters: 14143482bfa8SJunchao Zhang + sf - star forest on which to communicate 14153482bfa8SJunchao Zhang . unit - data type associated with each node 14163482bfa8SJunchao Zhang . rootdata - buffer to broadcast 14173482bfa8SJunchao Zhang - op - operation to use for reduction 14183482bfa8SJunchao Zhang 14194165533cSJose E. Roman Output Parameter: 14203482bfa8SJunchao Zhang . leafdata - buffer to be reduced with values from each leaf's respective root 14213482bfa8SJunchao Zhang 14223482bfa8SJunchao Zhang Level: intermediate 14233482bfa8SJunchao Zhang 142420662ed9SBarry Smith Note: 142520662ed9SBarry Smith When PETSc is configured with device support, it will use its own mechanism to figure out whether the given data pointers 1426da81f932SPierre Jolivet are host pointers or device pointers, which may incur a noticeable cost. If you already knew the info, you should 1427cab54364SBarry Smith use `PetscSFBcastWithMemTypeBegin()` instead. 1428cab54364SBarry Smith 1429cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastEnd()`, `PetscSFBcastWithMemTypeBegin()` 14303482bfa8SJunchao Zhang @*/ 1431d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastBegin(PetscSF sf, MPI_Datatype unit, const void *rootdata, void *leafdata, MPI_Op op) 1432d71ae5a4SJacob Faibussowitsch { 1433eb02082bSJunchao Zhang PetscMemType rootmtype, leafmtype; 14343482bfa8SJunchao Zhang 14353482bfa8SJunchao Zhang PetscFunctionBegin; 14363482bfa8SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 14379566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 14389566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin, sf, 0, 0, 0)); 14399566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(rootdata, &rootmtype)); 14409566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafdata, &leafmtype)); 1441dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, BcastBegin, unit, rootmtype, rootdata, leafmtype, leafdata, op); 14429566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin, sf, 0, 0, 0)); 14433ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 14443482bfa8SJunchao Zhang } 14453482bfa8SJunchao Zhang 14463482bfa8SJunchao Zhang /*@C 144720662ed9SBarry Smith PetscSFBcastWithMemTypeBegin - begin pointwise broadcast with root value being reduced to leaf value with explicit memory types, to be concluded with call 144820662ed9SBarry Smith to `PetscSFBcastEnd()` 1449d0295fc0SJunchao Zhang 1450c3339decSBarry Smith Collective 1451d0295fc0SJunchao Zhang 14524165533cSJose E. Roman Input Parameters: 1453d0295fc0SJunchao Zhang + sf - star forest on which to communicate 1454d0295fc0SJunchao Zhang . unit - data type associated with each node 1455d0295fc0SJunchao Zhang . rootmtype - memory type of rootdata 1456d0295fc0SJunchao Zhang . rootdata - buffer to broadcast 1457d0295fc0SJunchao Zhang . leafmtype - memory type of leafdata 1458d0295fc0SJunchao Zhang - op - operation to use for reduction 1459d0295fc0SJunchao Zhang 14604165533cSJose E. Roman Output Parameter: 1461d0295fc0SJunchao Zhang . leafdata - buffer to be reduced with values from each leaf's respective root 1462d0295fc0SJunchao Zhang 1463d0295fc0SJunchao Zhang Level: intermediate 1464d0295fc0SJunchao Zhang 1465cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastEnd()`, `PetscSFBcastBegin()` 1466d0295fc0SJunchao Zhang @*/ 1467d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastWithMemTypeBegin(PetscSF sf, MPI_Datatype unit, PetscMemType rootmtype, const void *rootdata, PetscMemType leafmtype, void *leafdata, MPI_Op op) 1468d71ae5a4SJacob Faibussowitsch { 1469d0295fc0SJunchao Zhang PetscFunctionBegin; 1470d0295fc0SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 14719566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 14729566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin, sf, 0, 0, 0)); 1473dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, BcastBegin, unit, rootmtype, rootdata, leafmtype, leafdata, op); 14749566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin, sf, 0, 0, 0)); 14753ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 1476d0295fc0SJunchao Zhang } 1477d0295fc0SJunchao Zhang 1478d0295fc0SJunchao Zhang /*@C 147920662ed9SBarry Smith PetscSFBcastEnd - end a broadcast and reduce operation started with `PetscSFBcastBegin()` or `PetscSFBcastWithMemTypeBegin()` 14803482bfa8SJunchao Zhang 14813482bfa8SJunchao Zhang Collective 14823482bfa8SJunchao Zhang 14834165533cSJose E. Roman Input Parameters: 14843482bfa8SJunchao Zhang + sf - star forest 14853482bfa8SJunchao Zhang . unit - data type 14863482bfa8SJunchao Zhang . rootdata - buffer to broadcast 14873482bfa8SJunchao Zhang - op - operation to use for reduction 14883482bfa8SJunchao Zhang 14894165533cSJose E. Roman Output Parameter: 14903482bfa8SJunchao Zhang . leafdata - buffer to be reduced with values from each leaf's respective root 14913482bfa8SJunchao Zhang 14923482bfa8SJunchao Zhang Level: intermediate 14933482bfa8SJunchao Zhang 1494cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFReduceEnd()` 14953482bfa8SJunchao Zhang @*/ 1496d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastEnd(PetscSF sf, MPI_Datatype unit, const void *rootdata, void *leafdata, MPI_Op op) 1497d71ae5a4SJacob Faibussowitsch { 14983482bfa8SJunchao Zhang PetscFunctionBegin; 14993482bfa8SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 15009566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastEnd, sf, 0, 0, 0)); 1501dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, BcastEnd, unit, rootdata, leafdata, op); 15029566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastEnd, sf, 0, 0, 0)); 15033ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 15043482bfa8SJunchao Zhang } 15053482bfa8SJunchao Zhang 15063482bfa8SJunchao Zhang /*@C 1507cab54364SBarry Smith PetscSFReduceBegin - begin reduction of leafdata into rootdata, to be completed with call to `PetscSFReduceEnd()` 150895fce210SBarry Smith 150995fce210SBarry Smith Collective 151095fce210SBarry Smith 15114165533cSJose E. Roman Input Parameters: 151295fce210SBarry Smith + sf - star forest 151395fce210SBarry Smith . unit - data type 151495fce210SBarry Smith . leafdata - values to reduce 151595fce210SBarry Smith - op - reduction operation 151695fce210SBarry Smith 15174165533cSJose E. Roman Output Parameter: 151895fce210SBarry Smith . rootdata - result of reduction of values from all leaves of each root 151995fce210SBarry Smith 152095fce210SBarry Smith Level: intermediate 152195fce210SBarry Smith 152220662ed9SBarry Smith Note: 152320662ed9SBarry Smith When PETSc is configured with device support, it will use its own mechanism to figure out whether the given data pointers 1524da81f932SPierre Jolivet are host pointers or device pointers, which may incur a noticeable cost. If you already knew the info, you should 1525cab54364SBarry Smith use `PetscSFReduceWithMemTypeBegin()` instead. 1526d0295fc0SJunchao Zhang 152720662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFBcastBegin()`, `PetscSFReduceWithMemTypeBegin()`, `PetscSFReduceEnd()` 152895fce210SBarry Smith @*/ 1529d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReduceBegin(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *rootdata, MPI_Op op) 1530d71ae5a4SJacob Faibussowitsch { 1531eb02082bSJunchao Zhang PetscMemType rootmtype, leafmtype; 153295fce210SBarry Smith 153395fce210SBarry Smith PetscFunctionBegin; 153495fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 15359566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 15369566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceBegin, sf, 0, 0, 0)); 15379566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(rootdata, &rootmtype)); 15389566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafdata, &leafmtype)); 15399566063dSJacob Faibussowitsch PetscCall((sf->ops->ReduceBegin)(sf, unit, leafmtype, leafdata, rootmtype, rootdata, op)); 15409566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceBegin, sf, 0, 0, 0)); 15413ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 154295fce210SBarry Smith } 154395fce210SBarry Smith 154495fce210SBarry Smith /*@C 1545cab54364SBarry Smith PetscSFReduceWithMemTypeBegin - begin reduction of leafdata into rootdata with explicit memory types, to be completed with call to `PetscSFReduceEnd()` 1546d0295fc0SJunchao Zhang 1547d0295fc0SJunchao Zhang Collective 1548d0295fc0SJunchao Zhang 15494165533cSJose E. Roman Input Parameters: 1550d0295fc0SJunchao Zhang + sf - star forest 1551d0295fc0SJunchao Zhang . unit - data type 1552d0295fc0SJunchao Zhang . leafmtype - memory type of leafdata 1553d0295fc0SJunchao Zhang . leafdata - values to reduce 1554d0295fc0SJunchao Zhang . rootmtype - memory type of rootdata 1555d0295fc0SJunchao Zhang - op - reduction operation 1556d0295fc0SJunchao Zhang 15574165533cSJose E. Roman Output Parameter: 1558d0295fc0SJunchao Zhang . rootdata - result of reduction of values from all leaves of each root 1559d0295fc0SJunchao Zhang 1560d0295fc0SJunchao Zhang Level: intermediate 1561d0295fc0SJunchao Zhang 156220662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFBcastBegin()`, `PetscSFReduceBegin()`, `PetscSFReduceEnd()` 1563d0295fc0SJunchao Zhang @*/ 1564d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReduceWithMemTypeBegin(PetscSF sf, MPI_Datatype unit, PetscMemType leafmtype, const void *leafdata, PetscMemType rootmtype, void *rootdata, MPI_Op op) 1565d71ae5a4SJacob Faibussowitsch { 1566d0295fc0SJunchao Zhang PetscFunctionBegin; 1567d0295fc0SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 15689566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 15699566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceBegin, sf, 0, 0, 0)); 15709566063dSJacob Faibussowitsch PetscCall((sf->ops->ReduceBegin)(sf, unit, leafmtype, leafdata, rootmtype, rootdata, op)); 15719566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceBegin, sf, 0, 0, 0)); 15723ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 1573d0295fc0SJunchao Zhang } 1574d0295fc0SJunchao Zhang 1575d0295fc0SJunchao Zhang /*@C 157620662ed9SBarry Smith PetscSFReduceEnd - end a reduction operation started with `PetscSFReduceBegin()` or `PetscSFReduceWithMemTypeBegin()` 157795fce210SBarry Smith 157895fce210SBarry Smith Collective 157995fce210SBarry Smith 15804165533cSJose E. Roman Input Parameters: 158195fce210SBarry Smith + sf - star forest 158295fce210SBarry Smith . unit - data type 158395fce210SBarry Smith . leafdata - values to reduce 158495fce210SBarry Smith - op - reduction operation 158595fce210SBarry Smith 15864165533cSJose E. Roman Output Parameter: 158795fce210SBarry Smith . rootdata - result of reduction of values from all leaves of each root 158895fce210SBarry Smith 158995fce210SBarry Smith Level: intermediate 159095fce210SBarry Smith 159120662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFBcastEnd()`, `PetscSFReduceBegin()`, `PetscSFReduceWithMemTypeBegin()` 159295fce210SBarry Smith @*/ 1593d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReduceEnd(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *rootdata, MPI_Op op) 1594d71ae5a4SJacob Faibussowitsch { 159595fce210SBarry Smith PetscFunctionBegin; 159695fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 15979566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceEnd, sf, 0, 0, 0)); 1598dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, ReduceEnd, unit, leafdata, rootdata, op); 15999566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceEnd, sf, 0, 0, 0)); 16003ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 160195fce210SBarry Smith } 160295fce210SBarry Smith 160395fce210SBarry Smith /*@C 1604cab54364SBarry Smith PetscSFFetchAndOpBegin - begin operation that fetches values from root and updates atomically by applying operation using my leaf value, 1605cab54364SBarry Smith to be completed with `PetscSFFetchAndOpEnd()` 1606a1729e3fSJunchao Zhang 1607a1729e3fSJunchao Zhang Collective 1608a1729e3fSJunchao Zhang 16094165533cSJose E. Roman Input Parameters: 1610a1729e3fSJunchao Zhang + sf - star forest 1611a1729e3fSJunchao Zhang . unit - data type 1612a1729e3fSJunchao Zhang . leafdata - leaf values to use in reduction 1613a1729e3fSJunchao Zhang - op - operation to use for reduction 1614a1729e3fSJunchao Zhang 16154165533cSJose E. Roman Output Parameters: 1616a1729e3fSJunchao Zhang + rootdata - root values to be updated, input state is seen by first process to perform an update 1617a1729e3fSJunchao Zhang - leafupdate - state at each leaf's respective root immediately prior to my atomic update 1618a1729e3fSJunchao Zhang 1619a1729e3fSJunchao Zhang Level: advanced 1620a1729e3fSJunchao Zhang 1621a1729e3fSJunchao Zhang Note: 1622a1729e3fSJunchao Zhang The update is only atomic at the granularity provided by the hardware. Different roots referenced by the same process 1623a1729e3fSJunchao Zhang might be updated in a different order. Furthermore, if a composite type is used for the unit datatype, atomicity is 1624a1729e3fSJunchao Zhang not guaranteed across the whole vertex. Therefore, this function is mostly only used with primitive types such as 1625a1729e3fSJunchao Zhang integers. 1626a1729e3fSJunchao Zhang 1627cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFReduceBegin()`, `PetscSFSetGraph()` 1628a1729e3fSJunchao Zhang @*/ 1629d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFFetchAndOpBegin(PetscSF sf, MPI_Datatype unit, void *rootdata, const void *leafdata, void *leafupdate, MPI_Op op) 1630d71ae5a4SJacob Faibussowitsch { 1631eb02082bSJunchao Zhang PetscMemType rootmtype, leafmtype, leafupdatemtype; 1632a1729e3fSJunchao Zhang 1633a1729e3fSJunchao Zhang PetscFunctionBegin; 1634a1729e3fSJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 16359566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 16369566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0)); 16379566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(rootdata, &rootmtype)); 16389566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafdata, &leafmtype)); 16399566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafupdate, &leafupdatemtype)); 164008401ef6SPierre Jolivet PetscCheck(leafmtype == leafupdatemtype, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for leafdata and leafupdate in different memory types"); 1641dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, FetchAndOpBegin, unit, rootmtype, rootdata, leafmtype, leafdata, leafupdate, op); 16429566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0)); 16433ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 1644a1729e3fSJunchao Zhang } 1645a1729e3fSJunchao Zhang 1646a1729e3fSJunchao Zhang /*@C 1647cab54364SBarry Smith PetscSFFetchAndOpWithMemTypeBegin - begin operation with explicit memory types that fetches values from root and updates atomically by 1648cab54364SBarry Smith applying operation using my leaf value, to be completed with `PetscSFFetchAndOpEnd()` 1649d3b3e55cSJunchao Zhang 1650d3b3e55cSJunchao Zhang Collective 1651d3b3e55cSJunchao Zhang 1652d3b3e55cSJunchao Zhang Input Parameters: 1653d3b3e55cSJunchao Zhang + sf - star forest 1654d3b3e55cSJunchao Zhang . unit - data type 1655d3b3e55cSJunchao Zhang . rootmtype - memory type of rootdata 1656d3b3e55cSJunchao Zhang . leafmtype - memory type of leafdata 1657d3b3e55cSJunchao Zhang . leafdata - leaf values to use in reduction 1658d3b3e55cSJunchao Zhang . leafupdatemtype - memory type of leafupdate 1659d3b3e55cSJunchao Zhang - op - operation to use for reduction 1660d3b3e55cSJunchao Zhang 1661d3b3e55cSJunchao Zhang Output Parameters: 1662d3b3e55cSJunchao Zhang + rootdata - root values to be updated, input state is seen by first process to perform an update 1663d3b3e55cSJunchao Zhang - leafupdate - state at each leaf's respective root immediately prior to my atomic update 1664d3b3e55cSJunchao Zhang 1665d3b3e55cSJunchao Zhang Level: advanced 1666d3b3e55cSJunchao Zhang 1667cab54364SBarry Smith Note: 1668cab54364SBarry Smith See `PetscSFFetchAndOpBegin()` for more details. 1669d3b3e55cSJunchao Zhang 167020662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFFetchAndOpBegin()`, `PetscSFComputeDegreeBegin()`, `PetscSFReduceBegin()`, `PetscSFSetGraph()`, `PetscSFFetchAndOpEnd()` 1671d3b3e55cSJunchao Zhang @*/ 1672d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFFetchAndOpWithMemTypeBegin(PetscSF sf, MPI_Datatype unit, PetscMemType rootmtype, void *rootdata, PetscMemType leafmtype, const void *leafdata, PetscMemType leafupdatemtype, void *leafupdate, MPI_Op op) 1673d71ae5a4SJacob Faibussowitsch { 1674d3b3e55cSJunchao Zhang PetscFunctionBegin; 1675d3b3e55cSJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 16769566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 16779566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0)); 167808401ef6SPierre Jolivet PetscCheck(leafmtype == leafupdatemtype, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for leafdata and leafupdate in different memory types"); 1679dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, FetchAndOpBegin, unit, rootmtype, rootdata, leafmtype, leafdata, leafupdate, op); 16809566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0)); 16813ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 1682d3b3e55cSJunchao Zhang } 1683d3b3e55cSJunchao Zhang 1684d3b3e55cSJunchao Zhang /*@C 168520662ed9SBarry Smith PetscSFFetchAndOpEnd - end operation started in matching call to `PetscSFFetchAndOpBegin()` or `PetscSFFetchAndOpWithMemTypeBegin()` 168620662ed9SBarry Smith to fetch values from roots and update atomically by applying operation using my leaf value 1687a1729e3fSJunchao Zhang 1688a1729e3fSJunchao Zhang Collective 1689a1729e3fSJunchao Zhang 16904165533cSJose E. Roman Input Parameters: 1691a1729e3fSJunchao Zhang + sf - star forest 1692a1729e3fSJunchao Zhang . unit - data type 1693a1729e3fSJunchao Zhang . leafdata - leaf values to use in reduction 1694a1729e3fSJunchao Zhang - op - operation to use for reduction 1695a1729e3fSJunchao Zhang 16964165533cSJose E. Roman Output Parameters: 1697a1729e3fSJunchao Zhang + rootdata - root values to be updated, input state is seen by first process to perform an update 1698a1729e3fSJunchao Zhang - leafupdate - state at each leaf's respective root immediately prior to my atomic update 1699a1729e3fSJunchao Zhang 1700a1729e3fSJunchao Zhang Level: advanced 1701a1729e3fSJunchao Zhang 170220662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeEnd()`, `PetscSFReduceEnd()`, `PetscSFSetGraph()`, `PetscSFFetchAndOpBegin()`, `PetscSFFetchAndOpWithMemTypeBegin()` 1703a1729e3fSJunchao Zhang @*/ 1704d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFFetchAndOpEnd(PetscSF sf, MPI_Datatype unit, void *rootdata, const void *leafdata, void *leafupdate, MPI_Op op) 1705d71ae5a4SJacob Faibussowitsch { 1706a1729e3fSJunchao Zhang PetscFunctionBegin; 1707a1729e3fSJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 17089566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpEnd, sf, 0, 0, 0)); 1709dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, FetchAndOpEnd, unit, rootdata, leafdata, leafupdate, op); 17109566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpEnd, sf, 0, 0, 0)); 17113ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 1712a1729e3fSJunchao Zhang } 1713a1729e3fSJunchao Zhang 1714a1729e3fSJunchao Zhang /*@C 1715cab54364SBarry Smith PetscSFComputeDegreeBegin - begin computation of degree for each root vertex, to be completed with `PetscSFComputeDegreeEnd()` 171695fce210SBarry Smith 171795fce210SBarry Smith Collective 171895fce210SBarry Smith 17194165533cSJose E. Roman Input Parameter: 172095fce210SBarry Smith . sf - star forest 172195fce210SBarry Smith 17224165533cSJose E. Roman Output Parameter: 172395fce210SBarry Smith . degree - degree of each root vertex 172495fce210SBarry Smith 172595fce210SBarry Smith Level: advanced 172695fce210SBarry Smith 1727cab54364SBarry Smith Note: 172820662ed9SBarry Smith The returned array is owned by `PetscSF` and automatically freed by `PetscSFDestroy()`. Hence there is no need to call `PetscFree()` on it. 1729ffe67aa5SVáclav Hapla 1730cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGatherBegin()`, `PetscSFComputeDegreeEnd()` 173195fce210SBarry Smith @*/ 1732d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComputeDegreeBegin(PetscSF sf, const PetscInt **degree) 1733d71ae5a4SJacob Faibussowitsch { 173495fce210SBarry Smith PetscFunctionBegin; 173595fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 173695fce210SBarry Smith PetscSFCheckGraphSet(sf, 1); 173795fce210SBarry Smith PetscValidPointer(degree, 2); 1738803bd9e8SMatthew G. Knepley if (!sf->degreeknown) { 17395b0d146aSStefano Zampini PetscInt i, nroots = sf->nroots, maxlocal; 174028b400f6SJacob Faibussowitsch PetscCheck(!sf->degree, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Calls to PetscSFComputeDegreeBegin() cannot be nested."); 17415b0d146aSStefano Zampini maxlocal = sf->maxleaf - sf->minleaf + 1; 17429566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nroots, &sf->degree)); 17439566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(PetscMax(maxlocal, 1), &sf->degreetmp)); /* allocate at least one entry, see check in PetscSFComputeDegreeEnd() */ 174429046d53SLisandro Dalcin for (i = 0; i < nroots; i++) sf->degree[i] = 0; 17459837ea96SMatthew G. Knepley for (i = 0; i < maxlocal; i++) sf->degreetmp[i] = 1; 17469566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(sf, MPIU_INT, sf->degreetmp - sf->minleaf, sf->degree, MPI_SUM)); 174795fce210SBarry Smith } 174895fce210SBarry Smith *degree = NULL; 17493ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 175095fce210SBarry Smith } 175195fce210SBarry Smith 175295fce210SBarry Smith /*@C 1753cab54364SBarry Smith PetscSFComputeDegreeEnd - complete computation of degree for each root vertex, started with `PetscSFComputeDegreeBegin()` 175495fce210SBarry Smith 175595fce210SBarry Smith Collective 175695fce210SBarry Smith 17574165533cSJose E. Roman Input Parameter: 175895fce210SBarry Smith . sf - star forest 175995fce210SBarry Smith 17604165533cSJose E. Roman Output Parameter: 176195fce210SBarry Smith . degree - degree of each root vertex 176295fce210SBarry Smith 176395fce210SBarry Smith Level: developer 176495fce210SBarry Smith 1765cab54364SBarry Smith Note: 176620662ed9SBarry Smith The returned array is owned by `PetscSF` and automatically freed by `PetscSFDestroy()`. Hence there is no need to call `PetscFree()` on it. 1767ffe67aa5SVáclav Hapla 1768cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGatherBegin()`, `PetscSFComputeDegreeBegin()` 176995fce210SBarry Smith @*/ 1770d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComputeDegreeEnd(PetscSF sf, const PetscInt **degree) 1771d71ae5a4SJacob Faibussowitsch { 177295fce210SBarry Smith PetscFunctionBegin; 177395fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 177495fce210SBarry Smith PetscSFCheckGraphSet(sf, 1); 177529046d53SLisandro Dalcin PetscValidPointer(degree, 2); 177695fce210SBarry Smith if (!sf->degreeknown) { 177728b400f6SJacob Faibussowitsch PetscCheck(sf->degreetmp, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFComputeDegreeBegin() before PetscSFComputeDegreeEnd()"); 17789566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(sf, MPIU_INT, sf->degreetmp - sf->minleaf, sf->degree, MPI_SUM)); 17799566063dSJacob Faibussowitsch PetscCall(PetscFree(sf->degreetmp)); 178095fce210SBarry Smith sf->degreeknown = PETSC_TRUE; 178195fce210SBarry Smith } 178295fce210SBarry Smith *degree = sf->degree; 17833ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 178495fce210SBarry Smith } 178595fce210SBarry Smith 1786673100f5SVaclav Hapla /*@C 178720662ed9SBarry Smith PetscSFComputeMultiRootOriginalNumbering - Returns original numbering of multi-roots (roots of multi-`PetscSF` returned by `PetscSFGetMultiSF()`). 178866dfcd1aSVaclav Hapla Each multi-root is assigned index of the corresponding original root. 1789673100f5SVaclav Hapla 1790673100f5SVaclav Hapla Collective 1791673100f5SVaclav Hapla 17924165533cSJose E. Roman Input Parameters: 1793673100f5SVaclav Hapla + sf - star forest 1794cab54364SBarry Smith - degree - degree of each root vertex, computed with `PetscSFComputeDegreeBegin()`/`PetscSFComputeDegreeEnd()` 1795673100f5SVaclav Hapla 17964165533cSJose E. Roman Output Parameters: 179720662ed9SBarry Smith + nMultiRoots - (optional) number of multi-roots (roots of multi-`PetscSF`) 179820662ed9SBarry Smith - multiRootsOrigNumbering - original indices of multi-roots; length of this array is `nMultiRoots` 1799673100f5SVaclav Hapla 1800673100f5SVaclav Hapla Level: developer 1801673100f5SVaclav Hapla 1802cab54364SBarry Smith Note: 180320662ed9SBarry Smith The returned array `multiRootsOrigNumbering` is newly allocated and should be destroyed with `PetscFree()` when no longer needed. 1804ffe67aa5SVáclav Hapla 1805cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFComputeDegreeEnd()`, `PetscSFGetMultiSF()` 1806673100f5SVaclav Hapla @*/ 1807d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComputeMultiRootOriginalNumbering(PetscSF sf, const PetscInt degree[], PetscInt *nMultiRoots, PetscInt *multiRootsOrigNumbering[]) 1808d71ae5a4SJacob Faibussowitsch { 1809673100f5SVaclav Hapla PetscSF msf; 1810673100f5SVaclav Hapla PetscInt i, j, k, nroots, nmroots; 1811673100f5SVaclav Hapla 1812673100f5SVaclav Hapla PetscFunctionBegin; 1813673100f5SVaclav Hapla PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 18149566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, NULL, NULL, NULL)); 181529328920SVaclav Hapla if (nroots) PetscValidIntPointer(degree, 2); 181666dfcd1aSVaclav Hapla if (nMultiRoots) PetscValidIntPointer(nMultiRoots, 3); 181766dfcd1aSVaclav Hapla PetscValidPointer(multiRootsOrigNumbering, 4); 18189566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &msf)); 18199566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(msf, &nmroots, NULL, NULL, NULL)); 18209566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nmroots, multiRootsOrigNumbering)); 1821673100f5SVaclav Hapla for (i = 0, j = 0, k = 0; i < nroots; i++) { 1822673100f5SVaclav Hapla if (!degree[i]) continue; 1823ad540459SPierre Jolivet for (j = 0; j < degree[i]; j++, k++) (*multiRootsOrigNumbering)[k] = i; 1824673100f5SVaclav Hapla } 182508401ef6SPierre Jolivet PetscCheck(k == nmroots, PETSC_COMM_SELF, PETSC_ERR_PLIB, "sanity check fail"); 182666dfcd1aSVaclav Hapla if (nMultiRoots) *nMultiRoots = nmroots; 18273ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 1828673100f5SVaclav Hapla } 1829673100f5SVaclav Hapla 183095fce210SBarry Smith /*@C 1831cab54364SBarry Smith PetscSFGatherBegin - begin pointwise gather of all leaves into multi-roots, to be completed with `PetscSFGatherEnd()` 183295fce210SBarry Smith 183395fce210SBarry Smith Collective 183495fce210SBarry Smith 18354165533cSJose E. Roman Input Parameters: 183695fce210SBarry Smith + sf - star forest 183795fce210SBarry Smith . unit - data type 183895fce210SBarry Smith - leafdata - leaf data to gather to roots 183995fce210SBarry Smith 18404165533cSJose E. Roman Output Parameter: 184195fce210SBarry Smith . multirootdata - root buffer to gather into, amount of space per root is equal to its degree 184295fce210SBarry Smith 184395fce210SBarry Smith Level: intermediate 184495fce210SBarry Smith 1845cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFScatterBegin()` 184695fce210SBarry Smith @*/ 1847d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGatherBegin(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *multirootdata) 1848d71ae5a4SJacob Faibussowitsch { 1849a5526d50SJunchao Zhang PetscSF multi = NULL; 185095fce210SBarry Smith 185195fce210SBarry Smith PetscFunctionBegin; 185295fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 18539566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 18549566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &multi)); 18559566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(multi, unit, leafdata, multirootdata, MPI_REPLACE)); 18563ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 185795fce210SBarry Smith } 185895fce210SBarry Smith 185995fce210SBarry Smith /*@C 1860cab54364SBarry Smith PetscSFGatherEnd - ends pointwise gather operation that was started with `PetscSFGatherBegin()` 186195fce210SBarry Smith 186295fce210SBarry Smith Collective 186395fce210SBarry Smith 18644165533cSJose E. Roman Input Parameters: 186595fce210SBarry Smith + sf - star forest 186695fce210SBarry Smith . unit - data type 186795fce210SBarry Smith - leafdata - leaf data to gather to roots 186895fce210SBarry Smith 18694165533cSJose E. Roman Output Parameter: 187095fce210SBarry Smith . multirootdata - root buffer to gather into, amount of space per root is equal to its degree 187195fce210SBarry Smith 187295fce210SBarry Smith Level: intermediate 187395fce210SBarry Smith 1874cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeEnd()`, `PetscSFScatterEnd()` 187595fce210SBarry Smith @*/ 1876d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGatherEnd(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *multirootdata) 1877d71ae5a4SJacob Faibussowitsch { 1878a5526d50SJunchao Zhang PetscSF multi = NULL; 187995fce210SBarry Smith 188095fce210SBarry Smith PetscFunctionBegin; 188195fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 18829566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &multi)); 18839566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(multi, unit, leafdata, multirootdata, MPI_REPLACE)); 18843ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 188595fce210SBarry Smith } 188695fce210SBarry Smith 188795fce210SBarry Smith /*@C 1888cab54364SBarry Smith PetscSFScatterBegin - begin pointwise scatter operation from multi-roots to leaves, to be completed with `PetscSFScatterEnd()` 188995fce210SBarry Smith 189095fce210SBarry Smith Collective 189195fce210SBarry Smith 18924165533cSJose E. Roman Input Parameters: 189395fce210SBarry Smith + sf - star forest 189495fce210SBarry Smith . unit - data type 189595fce210SBarry Smith - multirootdata - root buffer to send to each leaf, one unit of data per leaf 189695fce210SBarry Smith 18974165533cSJose E. Roman Output Parameter: 189895fce210SBarry Smith . leafdata - leaf data to be update with personal data from each respective root 189995fce210SBarry Smith 190095fce210SBarry Smith Level: intermediate 190195fce210SBarry Smith 190220662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFScatterEnd()` 190395fce210SBarry Smith @*/ 1904d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFScatterBegin(PetscSF sf, MPI_Datatype unit, const void *multirootdata, void *leafdata) 1905d71ae5a4SJacob Faibussowitsch { 1906a5526d50SJunchao Zhang PetscSF multi = NULL; 190795fce210SBarry Smith 190895fce210SBarry Smith PetscFunctionBegin; 190995fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 19109566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 19119566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &multi)); 19129566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(multi, unit, multirootdata, leafdata, MPI_REPLACE)); 19133ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 191495fce210SBarry Smith } 191595fce210SBarry Smith 191695fce210SBarry Smith /*@C 1917cab54364SBarry Smith PetscSFScatterEnd - ends pointwise scatter operation that was started with `PetscSFScatterBegin()` 191895fce210SBarry Smith 191995fce210SBarry Smith Collective 192095fce210SBarry Smith 19214165533cSJose E. Roman Input Parameters: 192295fce210SBarry Smith + sf - star forest 192395fce210SBarry Smith . unit - data type 192495fce210SBarry Smith - multirootdata - root buffer to send to each leaf, one unit of data per leaf 192595fce210SBarry Smith 19264165533cSJose E. Roman Output Parameter: 192795fce210SBarry Smith . leafdata - leaf data to be update with personal data from each respective root 192895fce210SBarry Smith 192995fce210SBarry Smith Level: intermediate 193095fce210SBarry Smith 193120662ed9SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeEnd()`, `PetscSFScatterBegin()` 193295fce210SBarry Smith @*/ 1933d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFScatterEnd(PetscSF sf, MPI_Datatype unit, const void *multirootdata, void *leafdata) 1934d71ae5a4SJacob Faibussowitsch { 1935a5526d50SJunchao Zhang PetscSF multi = NULL; 193695fce210SBarry Smith 193795fce210SBarry Smith PetscFunctionBegin; 193895fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 19399566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &multi)); 19409566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(multi, unit, multirootdata, leafdata, MPI_REPLACE)); 19413ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 194295fce210SBarry Smith } 1943a7b3aa13SAta Mesgarnejad 1944d71ae5a4SJacob Faibussowitsch static PetscErrorCode PetscSFCheckLeavesUnique_Private(PetscSF sf) 1945d71ae5a4SJacob Faibussowitsch { 1946a072220fSLawrence Mitchell PetscInt i, n, nleaves; 1947a072220fSLawrence Mitchell const PetscInt *ilocal = NULL; 1948a072220fSLawrence Mitchell PetscHSetI seen; 1949a072220fSLawrence Mitchell 1950a072220fSLawrence Mitchell PetscFunctionBegin; 1951b458e8f1SJose E. Roman if (PetscDefined(USE_DEBUG)) { 19529566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, NULL, &nleaves, &ilocal, NULL)); 19539566063dSJacob Faibussowitsch PetscCall(PetscHSetICreate(&seen)); 1954a072220fSLawrence Mitchell for (i = 0; i < nleaves; i++) { 1955a072220fSLawrence Mitchell const PetscInt leaf = ilocal ? ilocal[i] : i; 19569566063dSJacob Faibussowitsch PetscCall(PetscHSetIAdd(seen, leaf)); 1957a072220fSLawrence Mitchell } 19589566063dSJacob Faibussowitsch PetscCall(PetscHSetIGetSize(seen, &n)); 195908401ef6SPierre Jolivet PetscCheck(n == nleaves, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided leaves have repeated values: all leaves must be unique"); 19609566063dSJacob Faibussowitsch PetscCall(PetscHSetIDestroy(&seen)); 1961b458e8f1SJose E. Roman } 19623ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 1963a072220fSLawrence Mitchell } 196454729392SStefano Zampini 1965a7b3aa13SAta Mesgarnejad /*@ 1966cab54364SBarry Smith PetscSFCompose - Compose a new `PetscSF` by putting the second `PetscSF` under the first one in a top (roots) down (leaves) view 1967a7b3aa13SAta Mesgarnejad 1968a7b3aa13SAta Mesgarnejad Input Parameters: 1969cab54364SBarry Smith + sfA - The first `PetscSF` 1970cab54364SBarry Smith - sfB - The second `PetscSF` 1971a7b3aa13SAta Mesgarnejad 19722fe279fdSBarry Smith Output Parameter: 1973cab54364SBarry Smith . sfBA - The composite `PetscSF` 1974a7b3aa13SAta Mesgarnejad 1975a7b3aa13SAta Mesgarnejad Level: developer 1976a7b3aa13SAta Mesgarnejad 1977a072220fSLawrence Mitchell Notes: 1978cab54364SBarry Smith Currently, the two `PetscSF`s must be defined on congruent communicators and they must be true star 197954729392SStefano Zampini forests, i.e. the same leaf is not connected with different roots. 198054729392SStefano Zampini 198120662ed9SBarry Smith `sfA`'s leaf space and `sfB`'s root space might be partially overlapped. The composition builds 198220662ed9SBarry Smith a graph with `sfA`'s roots and `sfB`'s leaves only when there is a path between them. Unconnected 198320662ed9SBarry Smith nodes (roots or leaves) are not in `sfBA`. Doing a `PetscSFBcastBegin()`/`PetscSFBcastEnd()` on the new `PetscSF` is equivalent to doing a 198420662ed9SBarry Smith `PetscSFBcastBegin()`/`PetscSFBcastEnd()` on `sfA`, then a `PetscSFBcastBegin()`/`PetscSFBcastEnd()` on `sfB`, on connected nodes. 1985a072220fSLawrence Mitchell 1986db781477SPatrick Sanan .seealso: `PetscSF`, `PetscSFComposeInverse()`, `PetscSFGetGraph()`, `PetscSFSetGraph()` 1987a7b3aa13SAta Mesgarnejad @*/ 1988d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCompose(PetscSF sfA, PetscSF sfB, PetscSF *sfBA) 1989d71ae5a4SJacob Faibussowitsch { 1990a7b3aa13SAta Mesgarnejad const PetscSFNode *remotePointsA, *remotePointsB; 1991d41018fbSJunchao Zhang PetscSFNode *remotePointsBA = NULL, *reorderedRemotePointsA = NULL, *leafdataB; 199254729392SStefano Zampini const PetscInt *localPointsA, *localPointsB; 199354729392SStefano Zampini PetscInt *localPointsBA; 199454729392SStefano Zampini PetscInt i, numRootsA, numLeavesA, numRootsB, numLeavesB, minleaf, maxleaf, numLeavesBA; 199554729392SStefano Zampini PetscBool denseB; 1996a7b3aa13SAta Mesgarnejad 1997a7b3aa13SAta Mesgarnejad PetscFunctionBegin; 1998a7b3aa13SAta Mesgarnejad PetscValidHeaderSpecific(sfA, PETSCSF_CLASSID, 1); 199929046d53SLisandro Dalcin PetscSFCheckGraphSet(sfA, 1); 200029046d53SLisandro Dalcin PetscValidHeaderSpecific(sfB, PETSCSF_CLASSID, 2); 200129046d53SLisandro Dalcin PetscSFCheckGraphSet(sfB, 2); 200254729392SStefano Zampini PetscCheckSameComm(sfA, 1, sfB, 2); 200329046d53SLisandro Dalcin PetscValidPointer(sfBA, 3); 20049566063dSJacob Faibussowitsch PetscCall(PetscSFCheckLeavesUnique_Private(sfA)); 20059566063dSJacob Faibussowitsch PetscCall(PetscSFCheckLeavesUnique_Private(sfB)); 200654729392SStefano Zampini 20079566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfA, &numRootsA, &numLeavesA, &localPointsA, &remotePointsA)); 20089566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfB, &numRootsB, &numLeavesB, &localPointsB, &remotePointsB)); 200920662ed9SBarry Smith /* Make sure that PetscSFBcast{Begin, End}(sfB, ...) works with root data of size 201020662ed9SBarry Smith numRootsB; otherwise, garbage will be broadcasted. 201120662ed9SBarry Smith Example (comm size = 1): 201220662ed9SBarry Smith sfA: 0 <- (0, 0) 201320662ed9SBarry Smith sfB: 100 <- (0, 0) 201420662ed9SBarry Smith 101 <- (0, 1) 201520662ed9SBarry Smith Here, we have remotePointsA = [(0, 0)], but for remotePointsA to be a valid tartget 201620662ed9SBarry Smith of sfB, it has to be recasted as [(0, 0), (-1, -1)] so that points 100 and 101 would 201720662ed9SBarry Smith receive (0, 0) and (-1, -1), respectively, when PetscSFBcast(sfB, ...) is called on 201820662ed9SBarry Smith remotePointsA; if not recasted, point 101 would receive a garbage value. */ 20199566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numRootsB, &reorderedRemotePointsA)); 202054729392SStefano Zampini for (i = 0; i < numRootsB; i++) { 202154729392SStefano Zampini reorderedRemotePointsA[i].rank = -1; 202254729392SStefano Zampini reorderedRemotePointsA[i].index = -1; 202354729392SStefano Zampini } 202454729392SStefano Zampini for (i = 0; i < numLeavesA; i++) { 20250ea77edaSksagiyam PetscInt localp = localPointsA ? localPointsA[i] : i; 20260ea77edaSksagiyam 20270ea77edaSksagiyam if (localp >= numRootsB) continue; 20280ea77edaSksagiyam reorderedRemotePointsA[localp] = remotePointsA[i]; 202954729392SStefano Zampini } 2030d41018fbSJunchao Zhang remotePointsA = reorderedRemotePointsA; 20319566063dSJacob Faibussowitsch PetscCall(PetscSFGetLeafRange(sfB, &minleaf, &maxleaf)); 20329566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(maxleaf - minleaf + 1, &leafdataB)); 20330ea77edaSksagiyam for (i = 0; i < maxleaf - minleaf + 1; i++) { 20340ea77edaSksagiyam leafdataB[i].rank = -1; 20350ea77edaSksagiyam leafdataB[i].index = -1; 20360ea77edaSksagiyam } 20379566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(sfB, MPIU_2INT, remotePointsA, leafdataB - minleaf, MPI_REPLACE)); 20389566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(sfB, MPIU_2INT, remotePointsA, leafdataB - minleaf, MPI_REPLACE)); 20399566063dSJacob Faibussowitsch PetscCall(PetscFree(reorderedRemotePointsA)); 2040d41018fbSJunchao Zhang 204154729392SStefano Zampini denseB = (PetscBool)!localPointsB; 204254729392SStefano Zampini for (i = 0, numLeavesBA = 0; i < numLeavesB; i++) { 204354729392SStefano Zampini if (leafdataB[localPointsB ? localPointsB[i] - minleaf : i].rank == -1) denseB = PETSC_FALSE; 204454729392SStefano Zampini else numLeavesBA++; 204554729392SStefano Zampini } 204654729392SStefano Zampini if (denseB) { 2047d41018fbSJunchao Zhang localPointsBA = NULL; 2048d41018fbSJunchao Zhang remotePointsBA = leafdataB; 2049d41018fbSJunchao Zhang } else { 20509566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numLeavesBA, &localPointsBA)); 20519566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numLeavesBA, &remotePointsBA)); 205254729392SStefano Zampini for (i = 0, numLeavesBA = 0; i < numLeavesB; i++) { 205354729392SStefano Zampini const PetscInt l = localPointsB ? localPointsB[i] : i; 205454729392SStefano Zampini 205554729392SStefano Zampini if (leafdataB[l - minleaf].rank == -1) continue; 205654729392SStefano Zampini remotePointsBA[numLeavesBA] = leafdataB[l - minleaf]; 205754729392SStefano Zampini localPointsBA[numLeavesBA] = l; 205854729392SStefano Zampini numLeavesBA++; 205954729392SStefano Zampini } 20609566063dSJacob Faibussowitsch PetscCall(PetscFree(leafdataB)); 2061d41018fbSJunchao Zhang } 20629566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sfA), sfBA)); 20639566063dSJacob Faibussowitsch PetscCall(PetscSFSetFromOptions(*sfBA)); 20649566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*sfBA, numRootsA, numLeavesBA, localPointsBA, PETSC_OWN_POINTER, remotePointsBA, PETSC_OWN_POINTER)); 20653ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 2066a7b3aa13SAta Mesgarnejad } 20671c6ba672SJunchao Zhang 206804c0ada0SJunchao Zhang /*@ 2069cab54364SBarry Smith PetscSFComposeInverse - Compose a new `PetscSF` by putting the inverse of the second `PetscSF` under the first one 207004c0ada0SJunchao Zhang 207104c0ada0SJunchao Zhang Input Parameters: 2072cab54364SBarry Smith + sfA - The first `PetscSF` 2073cab54364SBarry Smith - sfB - The second `PetscSF` 207404c0ada0SJunchao Zhang 20752fe279fdSBarry Smith Output Parameter: 2076cab54364SBarry Smith . sfBA - The composite `PetscSF`. 207704c0ada0SJunchao Zhang 207804c0ada0SJunchao Zhang Level: developer 207904c0ada0SJunchao Zhang 208054729392SStefano Zampini Notes: 208120662ed9SBarry Smith Currently, the two `PetscSF`s must be defined on congruent communicators and they must be true star 208254729392SStefano Zampini forests, i.e. the same leaf is not connected with different roots. Even more, all roots of the 208320662ed9SBarry Smith second `PetscSF` must have a degree of 1, i.e., no roots have more than one leaf connected. 208454729392SStefano Zampini 208520662ed9SBarry Smith `sfA`'s leaf space and `sfB`'s leaf space might be partially overlapped. The composition builds 208620662ed9SBarry Smith a graph with `sfA`'s roots and `sfB`'s roots only when there is a path between them. Unconnected 208720662ed9SBarry Smith roots are not in `sfBA`. Doing a `PetscSFBcastBegin()`/`PetscSFBcastEnd()` on the new `PetscSF` is equivalent to doing a `PetscSFBcastBegin()`/`PetscSFBcastEnd()` 208820662ed9SBarry Smith on `sfA`, then 208920662ed9SBarry Smith a `PetscSFReduceBegin()`/`PetscSFReduceEnd()` on `sfB`, on connected roots. 209054729392SStefano Zampini 2091db781477SPatrick Sanan .seealso: `PetscSF`, `PetscSFCompose()`, `PetscSFGetGraph()`, `PetscSFSetGraph()`, `PetscSFCreateInverseSF()` 209204c0ada0SJunchao Zhang @*/ 2093d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComposeInverse(PetscSF sfA, PetscSF sfB, PetscSF *sfBA) 2094d71ae5a4SJacob Faibussowitsch { 209504c0ada0SJunchao Zhang const PetscSFNode *remotePointsA, *remotePointsB; 209604c0ada0SJunchao Zhang PetscSFNode *remotePointsBA; 209704c0ada0SJunchao Zhang const PetscInt *localPointsA, *localPointsB; 209854729392SStefano Zampini PetscSFNode *reorderedRemotePointsA = NULL; 209954729392SStefano Zampini PetscInt i, numRootsA, numLeavesA, numLeavesBA, numRootsB, numLeavesB, minleaf, maxleaf, *localPointsBA; 21005b0d146aSStefano Zampini MPI_Op op; 21015b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES) 21025b0d146aSStefano Zampini PetscBool iswin; 21035b0d146aSStefano Zampini #endif 210404c0ada0SJunchao Zhang 210504c0ada0SJunchao Zhang PetscFunctionBegin; 210604c0ada0SJunchao Zhang PetscValidHeaderSpecific(sfA, PETSCSF_CLASSID, 1); 210704c0ada0SJunchao Zhang PetscSFCheckGraphSet(sfA, 1); 210804c0ada0SJunchao Zhang PetscValidHeaderSpecific(sfB, PETSCSF_CLASSID, 2); 210904c0ada0SJunchao Zhang PetscSFCheckGraphSet(sfB, 2); 211054729392SStefano Zampini PetscCheckSameComm(sfA, 1, sfB, 2); 211104c0ada0SJunchao Zhang PetscValidPointer(sfBA, 3); 21129566063dSJacob Faibussowitsch PetscCall(PetscSFCheckLeavesUnique_Private(sfA)); 21139566063dSJacob Faibussowitsch PetscCall(PetscSFCheckLeavesUnique_Private(sfB)); 211454729392SStefano Zampini 21159566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfA, &numRootsA, &numLeavesA, &localPointsA, &remotePointsA)); 21169566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfB, &numRootsB, &numLeavesB, &localPointsB, &remotePointsB)); 21175b0d146aSStefano Zampini 21185b0d146aSStefano Zampini /* TODO: Check roots of sfB have degree of 1 */ 21195b0d146aSStefano Zampini /* Once we implement it, we can replace the MPI_MAXLOC 212083df288dSJunchao Zhang with MPI_REPLACE. In that case, MPI_MAXLOC and MPI_REPLACE have the same effect. 21215b0d146aSStefano Zampini We use MPI_MAXLOC only to have a deterministic output from this routine if 21225b0d146aSStefano Zampini the root condition is not meet. 21235b0d146aSStefano Zampini */ 21245b0d146aSStefano Zampini op = MPI_MAXLOC; 21255b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES) 21265b0d146aSStefano Zampini /* we accept a non-deterministic output (if any) with PETSCSFWINDOW, since MPI_MAXLOC cannot operate on MPIU_2INT with MPI_Accumulate */ 21279566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare((PetscObject)sfB, PETSCSFWINDOW, &iswin)); 212883df288dSJunchao Zhang if (iswin) op = MPI_REPLACE; 21295b0d146aSStefano Zampini #endif 21305b0d146aSStefano Zampini 21319566063dSJacob Faibussowitsch PetscCall(PetscSFGetLeafRange(sfB, &minleaf, &maxleaf)); 21329566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(maxleaf - minleaf + 1, &reorderedRemotePointsA)); 213354729392SStefano Zampini for (i = 0; i < maxleaf - minleaf + 1; i++) { 213454729392SStefano Zampini reorderedRemotePointsA[i].rank = -1; 213554729392SStefano Zampini reorderedRemotePointsA[i].index = -1; 213654729392SStefano Zampini } 213754729392SStefano Zampini if (localPointsA) { 213854729392SStefano Zampini for (i = 0; i < numLeavesA; i++) { 213954729392SStefano Zampini if (localPointsA[i] > maxleaf || localPointsA[i] < minleaf) continue; 214054729392SStefano Zampini reorderedRemotePointsA[localPointsA[i] - minleaf] = remotePointsA[i]; 214154729392SStefano Zampini } 214254729392SStefano Zampini } else { 214354729392SStefano Zampini for (i = 0; i < numLeavesA; i++) { 214454729392SStefano Zampini if (i > maxleaf || i < minleaf) continue; 214554729392SStefano Zampini reorderedRemotePointsA[i - minleaf] = remotePointsA[i]; 214654729392SStefano Zampini } 214754729392SStefano Zampini } 214854729392SStefano Zampini 21499566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numRootsB, &localPointsBA)); 21509566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numRootsB, &remotePointsBA)); 215154729392SStefano Zampini for (i = 0; i < numRootsB; i++) { 215254729392SStefano Zampini remotePointsBA[i].rank = -1; 215354729392SStefano Zampini remotePointsBA[i].index = -1; 215454729392SStefano Zampini } 215554729392SStefano Zampini 21569566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(sfB, MPIU_2INT, reorderedRemotePointsA - minleaf, remotePointsBA, op)); 21579566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(sfB, MPIU_2INT, reorderedRemotePointsA - minleaf, remotePointsBA, op)); 21589566063dSJacob Faibussowitsch PetscCall(PetscFree(reorderedRemotePointsA)); 215954729392SStefano Zampini for (i = 0, numLeavesBA = 0; i < numRootsB; i++) { 216054729392SStefano Zampini if (remotePointsBA[i].rank == -1) continue; 216154729392SStefano Zampini remotePointsBA[numLeavesBA].rank = remotePointsBA[i].rank; 216254729392SStefano Zampini remotePointsBA[numLeavesBA].index = remotePointsBA[i].index; 216354729392SStefano Zampini localPointsBA[numLeavesBA] = i; 216454729392SStefano Zampini numLeavesBA++; 216554729392SStefano Zampini } 21669566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sfA), sfBA)); 21679566063dSJacob Faibussowitsch PetscCall(PetscSFSetFromOptions(*sfBA)); 21689566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*sfBA, numRootsA, numLeavesBA, localPointsBA, PETSC_OWN_POINTER, remotePointsBA, PETSC_OWN_POINTER)); 21693ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 217004c0ada0SJunchao Zhang } 217104c0ada0SJunchao Zhang 21721c6ba672SJunchao Zhang /* 2173cab54364SBarry Smith PetscSFCreateLocalSF_Private - Creates a local `PetscSF` that only has intra-process edges of the global `PetscSF` 21741c6ba672SJunchao Zhang 21752fe279fdSBarry Smith Input Parameter: 2176cab54364SBarry Smith . sf - The global `PetscSF` 21771c6ba672SJunchao Zhang 21782fe279fdSBarry Smith Output Parameter: 2179cab54364SBarry Smith . out - The local `PetscSF` 2180cab54364SBarry Smith 2181cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()` 21821c6ba672SJunchao Zhang */ 2183d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateLocalSF_Private(PetscSF sf, PetscSF *out) 2184d71ae5a4SJacob Faibussowitsch { 21851c6ba672SJunchao Zhang MPI_Comm comm; 21861c6ba672SJunchao Zhang PetscMPIInt myrank; 21871c6ba672SJunchao Zhang const PetscInt *ilocal; 21881c6ba672SJunchao Zhang const PetscSFNode *iremote; 21891c6ba672SJunchao Zhang PetscInt i, j, nroots, nleaves, lnleaves, *lilocal; 21901c6ba672SJunchao Zhang PetscSFNode *liremote; 21911c6ba672SJunchao Zhang PetscSF lsf; 21921c6ba672SJunchao Zhang 21931c6ba672SJunchao Zhang PetscFunctionBegin; 21941c6ba672SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 2195dbbe0bcdSBarry Smith if (sf->ops->CreateLocalSF) PetscUseTypeMethod(sf, CreateLocalSF, out); 2196dbbe0bcdSBarry Smith else { 21971c6ba672SJunchao Zhang /* Could use PetscSFCreateEmbeddedLeafSF, but since we know the comm is PETSC_COMM_SELF, we can make it fast */ 21989566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 21999566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &myrank)); 22001c6ba672SJunchao Zhang 22011c6ba672SJunchao Zhang /* Find out local edges and build a local SF */ 22029566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, &iremote)); 22039371c9d4SSatish Balay for (i = lnleaves = 0; i < nleaves; i++) { 22049371c9d4SSatish Balay if (iremote[i].rank == (PetscInt)myrank) lnleaves++; 22059371c9d4SSatish Balay } 22069566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(lnleaves, &lilocal)); 22079566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(lnleaves, &liremote)); 22081c6ba672SJunchao Zhang 22091c6ba672SJunchao Zhang for (i = j = 0; i < nleaves; i++) { 22101c6ba672SJunchao Zhang if (iremote[i].rank == (PetscInt)myrank) { 22111c6ba672SJunchao Zhang lilocal[j] = ilocal ? ilocal[i] : i; /* ilocal=NULL for contiguous storage */ 22121c6ba672SJunchao Zhang liremote[j].rank = 0; /* rank in PETSC_COMM_SELF */ 22131c6ba672SJunchao Zhang liremote[j].index = iremote[i].index; 22141c6ba672SJunchao Zhang j++; 22151c6ba672SJunchao Zhang } 22161c6ba672SJunchao Zhang } 22179566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(PETSC_COMM_SELF, &lsf)); 22189566063dSJacob Faibussowitsch PetscCall(PetscSFSetFromOptions(lsf)); 22199566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(lsf, nroots, lnleaves, lilocal, PETSC_OWN_POINTER, liremote, PETSC_OWN_POINTER)); 22209566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(lsf)); 22211c6ba672SJunchao Zhang *out = lsf; 22221c6ba672SJunchao Zhang } 22233ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 22241c6ba672SJunchao Zhang } 2225dd5b3ca6SJunchao Zhang 2226dd5b3ca6SJunchao Zhang /* Similar to PetscSFBcast, but only Bcast to leaves on rank 0 */ 2227d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastToZero_Private(PetscSF sf, MPI_Datatype unit, const void *rootdata, void *leafdata) 2228d71ae5a4SJacob Faibussowitsch { 2229eb02082bSJunchao Zhang PetscMemType rootmtype, leafmtype; 2230dd5b3ca6SJunchao Zhang 2231dd5b3ca6SJunchao Zhang PetscFunctionBegin; 2232dd5b3ca6SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 22339566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 22349566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin, sf, 0, 0, 0)); 22359566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(rootdata, &rootmtype)); 22369566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafdata, &leafmtype)); 2237dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, BcastToZero, unit, rootmtype, rootdata, leafmtype, leafdata); 22389566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin, sf, 0, 0, 0)); 22393ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 2240dd5b3ca6SJunchao Zhang } 2241dd5b3ca6SJunchao Zhang 2242157edd7aSVaclav Hapla /*@ 2243cab54364SBarry Smith PetscSFConcatenate - concatenate multiple `PetscSF` into one 2244157edd7aSVaclav Hapla 2245157edd7aSVaclav Hapla Input Parameters: 2246157edd7aSVaclav Hapla + comm - the communicator 2247cab54364SBarry Smith . nsfs - the number of input `PetscSF` 2248cab54364SBarry Smith . sfs - the array of input `PetscSF` 22491f40158dSVaclav Hapla . rootMode - the root mode specifying how roots are handled 225020662ed9SBarry Smith - leafOffsets - the array of local leaf offsets, one for each input `PetscSF`, or `NULL` for contiguous storage 2251157edd7aSVaclav Hapla 22522fe279fdSBarry Smith Output Parameter: 2253cab54364SBarry Smith . newsf - The resulting `PetscSF` 2254157edd7aSVaclav Hapla 22551f40158dSVaclav Hapla Level: advanced 2256157edd7aSVaclav Hapla 2257157edd7aSVaclav Hapla Notes: 225820662ed9SBarry Smith The communicator of all `PetscSF`s in `sfs` must be comm. 2259157edd7aSVaclav Hapla 226020662ed9SBarry Smith Leaves are always concatenated locally, keeping them ordered by the input `PetscSF` index and original local order. 226120662ed9SBarry Smith 226220662ed9SBarry Smith The offsets in `leafOffsets` are added to the original leaf indices. 226320662ed9SBarry Smith 226420662ed9SBarry Smith If all input SFs use contiguous leaf storage (`ilocal` = `NULL`), `leafOffsets` can be passed as `NULL` as well. 226520662ed9SBarry Smith In this case, `NULL` is also passed as `ilocal` to the resulting `PetscSF`. 226620662ed9SBarry Smith 226720662ed9SBarry Smith If any input `PetscSF` has non-null `ilocal`, `leafOffsets` is needed to distinguish leaves from different input `PetscSF`s. 2268157edd7aSVaclav Hapla In this case, user is responsible to provide correct offsets so that the resulting leaves are unique (otherwise an error occurs). 2269157edd7aSVaclav Hapla 227020662ed9SBarry Smith All root modes retain the essential connectivity condition. 227120662ed9SBarry Smith If two leaves of the same input `PetscSF` are connected (sharing the same root), they are also connected in the output `PetscSF`. 227220662ed9SBarry Smith Parameter `rootMode` controls how the input root spaces are combined. 227320662ed9SBarry Smith For `PETSCSF_CONCATENATE_ROOTMODE_SHARED`, the root space is considered the same for each input `PetscSF` (checked in debug mode) 227420662ed9SBarry Smith and is also the same in the output `PetscSF`. 22751f40158dSVaclav Hapla For `PETSCSF_CONCATENATE_ROOTMODE_LOCAL` and `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL`, the input root spaces are taken as separate and joined. 22761f40158dSVaclav Hapla `PETSCSF_CONCATENATE_ROOTMODE_LOCAL` joins the root spaces locally; 227720662ed9SBarry Smith roots of sfs[0], sfs[1], sfs[2], ... are joined on each rank separately, ordered by input `PetscSF` and original local index, and renumbered contiguously. 22781f40158dSVaclav Hapla `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL` joins the root spaces globally; 22791593df67SStefano Zampini roots of sfs[0], sfs[1], sfs[2], ... are joined globally, ordered by input `PetscSF` index and original global index, and renumbered contiguously; 22801f40158dSVaclav Hapla the original root ranks are ignored. 22811f40158dSVaclav Hapla For both `PETSCSF_CONCATENATE_ROOTMODE_LOCAL` and `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL`, 228220662ed9SBarry Smith the output `PetscSF`'s root layout is such that the local number of roots is a sum of the input `PetscSF`'s local numbers of roots on each rank 228320662ed9SBarry Smith to keep the load balancing. 228420662ed9SBarry Smith However, for `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL`, roots can move to different ranks. 22851f40158dSVaclav Hapla 22861f40158dSVaclav Hapla Example: 22871f40158dSVaclav Hapla We can use src/vec/is/sf/tests/ex18.c to compare the root modes. By running 228820662ed9SBarry Smith .vb 228920662ed9SBarry Smith make -C $PETSC_DIR/src/vec/is/sf/tests ex18 229020662ed9SBarry Smith for m in {local,global,shared}; do 229120662ed9SBarry Smith mpirun -n 2 $PETSC_DIR/src/vec/is/sf/tests/ex18 -nsfs 2 -n 2 -root_mode $m -sf_view 229220662ed9SBarry Smith done 229320662ed9SBarry Smith .ve 229420662ed9SBarry Smith we generate two identical `PetscSF`s sf_0 and sf_1, 229520662ed9SBarry Smith .vb 229620662ed9SBarry Smith PetscSF Object: sf_0 2 MPI processes 229720662ed9SBarry Smith type: basic 229820662ed9SBarry Smith rank #leaves #roots 229920662ed9SBarry Smith [ 0] 4 2 230020662ed9SBarry Smith [ 1] 4 2 230120662ed9SBarry Smith leaves roots roots in global numbering 230220662ed9SBarry Smith ( 0, 0) <- ( 0, 0) = 0 230320662ed9SBarry Smith ( 0, 1) <- ( 0, 1) = 1 230420662ed9SBarry Smith ( 0, 2) <- ( 1, 0) = 2 230520662ed9SBarry Smith ( 0, 3) <- ( 1, 1) = 3 230620662ed9SBarry Smith ( 1, 0) <- ( 0, 0) = 0 230720662ed9SBarry Smith ( 1, 1) <- ( 0, 1) = 1 230820662ed9SBarry Smith ( 1, 2) <- ( 1, 0) = 2 230920662ed9SBarry Smith ( 1, 3) <- ( 1, 1) = 3 231020662ed9SBarry Smith .ve 2311*e33f79d8SJacob Faibussowitsch and pass them to `PetscSFConcatenate()` along with different choices of `rootMode`, yielding different result_sf\: 231220662ed9SBarry Smith .vb 231320662ed9SBarry Smith rootMode = local: 231420662ed9SBarry Smith PetscSF Object: result_sf 2 MPI processes 231520662ed9SBarry Smith type: basic 231620662ed9SBarry Smith rank #leaves #roots 231720662ed9SBarry Smith [ 0] 8 4 231820662ed9SBarry Smith [ 1] 8 4 231920662ed9SBarry Smith leaves roots roots in global numbering 232020662ed9SBarry Smith ( 0, 0) <- ( 0, 0) = 0 232120662ed9SBarry Smith ( 0, 1) <- ( 0, 1) = 1 232220662ed9SBarry Smith ( 0, 2) <- ( 1, 0) = 4 232320662ed9SBarry Smith ( 0, 3) <- ( 1, 1) = 5 232420662ed9SBarry Smith ( 0, 4) <- ( 0, 2) = 2 232520662ed9SBarry Smith ( 0, 5) <- ( 0, 3) = 3 232620662ed9SBarry Smith ( 0, 6) <- ( 1, 2) = 6 232720662ed9SBarry Smith ( 0, 7) <- ( 1, 3) = 7 232820662ed9SBarry Smith ( 1, 0) <- ( 0, 0) = 0 232920662ed9SBarry Smith ( 1, 1) <- ( 0, 1) = 1 233020662ed9SBarry Smith ( 1, 2) <- ( 1, 0) = 4 233120662ed9SBarry Smith ( 1, 3) <- ( 1, 1) = 5 233220662ed9SBarry Smith ( 1, 4) <- ( 0, 2) = 2 233320662ed9SBarry Smith ( 1, 5) <- ( 0, 3) = 3 233420662ed9SBarry Smith ( 1, 6) <- ( 1, 2) = 6 233520662ed9SBarry Smith ( 1, 7) <- ( 1, 3) = 7 233620662ed9SBarry Smith 233720662ed9SBarry Smith rootMode = global: 233820662ed9SBarry Smith PetscSF Object: result_sf 2 MPI processes 233920662ed9SBarry Smith type: basic 234020662ed9SBarry Smith rank #leaves #roots 234120662ed9SBarry Smith [ 0] 8 4 234220662ed9SBarry Smith [ 1] 8 4 234320662ed9SBarry Smith leaves roots roots in global numbering 234420662ed9SBarry Smith ( 0, 0) <- ( 0, 0) = 0 234520662ed9SBarry Smith ( 0, 1) <- ( 0, 1) = 1 234620662ed9SBarry Smith ( 0, 2) <- ( 0, 2) = 2 234720662ed9SBarry Smith ( 0, 3) <- ( 0, 3) = 3 234820662ed9SBarry Smith ( 0, 4) <- ( 1, 0) = 4 234920662ed9SBarry Smith ( 0, 5) <- ( 1, 1) = 5 235020662ed9SBarry Smith ( 0, 6) <- ( 1, 2) = 6 235120662ed9SBarry Smith ( 0, 7) <- ( 1, 3) = 7 235220662ed9SBarry Smith ( 1, 0) <- ( 0, 0) = 0 235320662ed9SBarry Smith ( 1, 1) <- ( 0, 1) = 1 235420662ed9SBarry Smith ( 1, 2) <- ( 0, 2) = 2 235520662ed9SBarry Smith ( 1, 3) <- ( 0, 3) = 3 235620662ed9SBarry Smith ( 1, 4) <- ( 1, 0) = 4 235720662ed9SBarry Smith ( 1, 5) <- ( 1, 1) = 5 235820662ed9SBarry Smith ( 1, 6) <- ( 1, 2) = 6 235920662ed9SBarry Smith ( 1, 7) <- ( 1, 3) = 7 236020662ed9SBarry Smith 236120662ed9SBarry Smith rootMode = shared: 236220662ed9SBarry Smith PetscSF Object: result_sf 2 MPI processes 236320662ed9SBarry Smith type: basic 236420662ed9SBarry Smith rank #leaves #roots 236520662ed9SBarry Smith [ 0] 8 2 236620662ed9SBarry Smith [ 1] 8 2 236720662ed9SBarry Smith leaves roots roots in global numbering 236820662ed9SBarry Smith ( 0, 0) <- ( 0, 0) = 0 236920662ed9SBarry Smith ( 0, 1) <- ( 0, 1) = 1 237020662ed9SBarry Smith ( 0, 2) <- ( 1, 0) = 2 237120662ed9SBarry Smith ( 0, 3) <- ( 1, 1) = 3 237220662ed9SBarry Smith ( 0, 4) <- ( 0, 0) = 0 237320662ed9SBarry Smith ( 0, 5) <- ( 0, 1) = 1 237420662ed9SBarry Smith ( 0, 6) <- ( 1, 0) = 2 237520662ed9SBarry Smith ( 0, 7) <- ( 1, 1) = 3 237620662ed9SBarry Smith ( 1, 0) <- ( 0, 0) = 0 237720662ed9SBarry Smith ( 1, 1) <- ( 0, 1) = 1 237820662ed9SBarry Smith ( 1, 2) <- ( 1, 0) = 2 237920662ed9SBarry Smith ( 1, 3) <- ( 1, 1) = 3 238020662ed9SBarry Smith ( 1, 4) <- ( 0, 0) = 0 238120662ed9SBarry Smith ( 1, 5) <- ( 0, 1) = 1 238220662ed9SBarry Smith ( 1, 6) <- ( 1, 0) = 2 238320662ed9SBarry Smith ( 1, 7) <- ( 1, 1) = 3 238420662ed9SBarry Smith .ve 23851f40158dSVaclav Hapla 23861f40158dSVaclav Hapla .seealso: `PetscSF`, `PetscSFCompose()`, `PetscSFGetGraph()`, `PetscSFSetGraph()`, `PetscSFConcatenateRootMode` 2387157edd7aSVaclav Hapla @*/ 23881f40158dSVaclav Hapla PetscErrorCode PetscSFConcatenate(MPI_Comm comm, PetscInt nsfs, PetscSF sfs[], PetscSFConcatenateRootMode rootMode, PetscInt leafOffsets[], PetscSF *newsf) 2389d71ae5a4SJacob Faibussowitsch { 2390157edd7aSVaclav Hapla PetscInt i, s, nLeaves, nRoots; 2391157edd7aSVaclav Hapla PetscInt *leafArrayOffsets; 2392157edd7aSVaclav Hapla PetscInt *ilocal_new; 2393157edd7aSVaclav Hapla PetscSFNode *iremote_new; 2394157edd7aSVaclav Hapla PetscBool all_ilocal_null = PETSC_FALSE; 23951f40158dSVaclav Hapla PetscLayout glayout = NULL; 23961f40158dSVaclav Hapla PetscInt *gremote = NULL; 23971f40158dSVaclav Hapla PetscMPIInt rank, size; 2398157edd7aSVaclav Hapla 2399157edd7aSVaclav Hapla PetscFunctionBegin; 240012f479c1SVaclav Hapla if (PetscDefined(USE_DEBUG)) { 2401157edd7aSVaclav Hapla PetscSF dummy; /* just to have a PetscObject on comm for input validation */ 2402157edd7aSVaclav Hapla 24039566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, &dummy)); 2404157edd7aSVaclav Hapla PetscValidLogicalCollectiveInt(dummy, nsfs, 2); 2405157edd7aSVaclav Hapla PetscValidPointer(sfs, 3); 2406157edd7aSVaclav Hapla for (i = 0; i < nsfs; i++) { 2407157edd7aSVaclav Hapla PetscValidHeaderSpecific(sfs[i], PETSCSF_CLASSID, 3); 2408157edd7aSVaclav Hapla PetscCheckSameComm(dummy, 1, sfs[i], 3); 2409157edd7aSVaclav Hapla } 24101f40158dSVaclav Hapla PetscValidLogicalCollectiveEnum(dummy, rootMode, 4); 2411157edd7aSVaclav Hapla if (leafOffsets) PetscValidIntPointer(leafOffsets, 5); 2412157edd7aSVaclav Hapla PetscValidPointer(newsf, 6); 24139566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&dummy)); 2414157edd7aSVaclav Hapla } 2415157edd7aSVaclav Hapla if (!nsfs) { 24169566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, newsf)); 24179566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*newsf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER)); 24183ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 2419157edd7aSVaclav Hapla } 24209566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 24211f40158dSVaclav Hapla PetscCallMPI(MPI_Comm_size(comm, &size)); 2422157edd7aSVaclav Hapla 24231f40158dSVaclav Hapla /* Calculate leaf array offsets */ 24249566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nsfs + 1, &leafArrayOffsets)); 2425157edd7aSVaclav Hapla leafArrayOffsets[0] = 0; 2426157edd7aSVaclav Hapla for (s = 0; s < nsfs; s++) { 2427157edd7aSVaclav Hapla PetscInt nl; 2428157edd7aSVaclav Hapla 24299566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfs[s], NULL, &nl, NULL, NULL)); 2430157edd7aSVaclav Hapla leafArrayOffsets[s + 1] = leafArrayOffsets[s] + nl; 2431157edd7aSVaclav Hapla } 2432157edd7aSVaclav Hapla nLeaves = leafArrayOffsets[nsfs]; 2433157edd7aSVaclav Hapla 24341f40158dSVaclav Hapla /* Calculate number of roots */ 24351f40158dSVaclav Hapla switch (rootMode) { 24361f40158dSVaclav Hapla case PETSCSF_CONCATENATE_ROOTMODE_SHARED: { 24371f40158dSVaclav Hapla PetscCall(PetscSFGetGraph(sfs[0], &nRoots, NULL, NULL, NULL)); 24381f40158dSVaclav Hapla if (PetscDefined(USE_DEBUG)) { 24391f40158dSVaclav Hapla for (s = 1; s < nsfs; s++) { 24401f40158dSVaclav Hapla PetscInt nr; 24411f40158dSVaclav Hapla 24421f40158dSVaclav Hapla PetscCall(PetscSFGetGraph(sfs[s], &nr, NULL, NULL, NULL)); 24431f40158dSVaclav Hapla PetscCheck(nr == nRoots, comm, PETSC_ERR_ARG_SIZ, "rootMode = %s but sfs[%" PetscInt_FMT "] has a different number of roots (%" PetscInt_FMT ") than sfs[0] (%" PetscInt_FMT ")", PetscSFConcatenateRootModes[rootMode], s, nr, nRoots); 24441f40158dSVaclav Hapla } 24451f40158dSVaclav Hapla } 24461f40158dSVaclav Hapla } break; 24471f40158dSVaclav Hapla case PETSCSF_CONCATENATE_ROOTMODE_GLOBAL: { 24481f40158dSVaclav Hapla /* Calculate also global layout in this case */ 24491f40158dSVaclav Hapla PetscInt *nls; 24501f40158dSVaclav Hapla PetscLayout *lts; 24511f40158dSVaclav Hapla PetscInt **inds; 24521f40158dSVaclav Hapla PetscInt j; 24531f40158dSVaclav Hapla PetscInt rootOffset = 0; 24541f40158dSVaclav Hapla 24551f40158dSVaclav Hapla PetscCall(PetscCalloc3(nsfs, <s, nsfs, &nls, nsfs, &inds)); 24561f40158dSVaclav Hapla PetscCall(PetscLayoutCreate(comm, &glayout)); 24571f40158dSVaclav Hapla glayout->bs = 1; 24581f40158dSVaclav Hapla glayout->n = 0; 24591f40158dSVaclav Hapla glayout->N = 0; 24601f40158dSVaclav Hapla for (s = 0; s < nsfs; s++) { 24611f40158dSVaclav Hapla PetscCall(PetscSFGetGraphLayout(sfs[s], <s[s], &nls[s], NULL, &inds[s])); 24621f40158dSVaclav Hapla glayout->n += lts[s]->n; 24631f40158dSVaclav Hapla glayout->N += lts[s]->N; 24641f40158dSVaclav Hapla } 24651f40158dSVaclav Hapla PetscCall(PetscLayoutSetUp(glayout)); 24661f40158dSVaclav Hapla PetscCall(PetscMalloc1(nLeaves, &gremote)); 24671f40158dSVaclav Hapla for (s = 0, j = 0; s < nsfs; s++) { 24681f40158dSVaclav Hapla for (i = 0; i < nls[s]; i++, j++) gremote[j] = inds[s][i] + rootOffset; 24691f40158dSVaclav Hapla rootOffset += lts[s]->N; 24701f40158dSVaclav Hapla PetscCall(PetscLayoutDestroy(<s[s])); 24711f40158dSVaclav Hapla PetscCall(PetscFree(inds[s])); 24721f40158dSVaclav Hapla } 24731f40158dSVaclav Hapla PetscCall(PetscFree3(lts, nls, inds)); 24741f40158dSVaclav Hapla nRoots = glayout->N; 24751f40158dSVaclav Hapla } break; 24761f40158dSVaclav Hapla case PETSCSF_CONCATENATE_ROOTMODE_LOCAL: 24771f40158dSVaclav Hapla /* nRoots calculated later in this case */ 24781f40158dSVaclav Hapla break; 24791f40158dSVaclav Hapla default: 24801f40158dSVaclav Hapla SETERRQ(comm, PETSC_ERR_ARG_WRONG, "Invalid PetscSFConcatenateRootMode %d", rootMode); 24811f40158dSVaclav Hapla } 24821f40158dSVaclav Hapla 2483157edd7aSVaclav Hapla if (!leafOffsets) { 2484157edd7aSVaclav Hapla all_ilocal_null = PETSC_TRUE; 2485157edd7aSVaclav Hapla for (s = 0; s < nsfs; s++) { 2486157edd7aSVaclav Hapla const PetscInt *ilocal; 2487157edd7aSVaclav Hapla 24889566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfs[s], NULL, NULL, &ilocal, NULL)); 2489157edd7aSVaclav Hapla if (ilocal) { 2490157edd7aSVaclav Hapla all_ilocal_null = PETSC_FALSE; 2491157edd7aSVaclav Hapla break; 2492157edd7aSVaclav Hapla } 2493157edd7aSVaclav Hapla } 2494157edd7aSVaclav Hapla PetscCheck(all_ilocal_null, PETSC_COMM_SELF, PETSC_ERR_ARG_NULL, "leafOffsets can be passed as NULL only if all SFs have ilocal = NULL"); 2495157edd7aSVaclav Hapla } 2496157edd7aSVaclav Hapla 2497157edd7aSVaclav Hapla /* Renumber and concatenate local leaves */ 2498157edd7aSVaclav Hapla ilocal_new = NULL; 2499157edd7aSVaclav Hapla if (!all_ilocal_null) { 25009566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nLeaves, &ilocal_new)); 2501157edd7aSVaclav Hapla for (i = 0; i < nLeaves; i++) ilocal_new[i] = -1; 2502157edd7aSVaclav Hapla for (s = 0; s < nsfs; s++) { 2503157edd7aSVaclav Hapla const PetscInt *ilocal; 2504157edd7aSVaclav Hapla PetscInt *ilocal_l = &ilocal_new[leafArrayOffsets[s]]; 2505157edd7aSVaclav Hapla PetscInt i, nleaves_l; 2506157edd7aSVaclav Hapla 25079566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfs[s], NULL, &nleaves_l, &ilocal, NULL)); 2508157edd7aSVaclav Hapla for (i = 0; i < nleaves_l; i++) ilocal_l[i] = (ilocal ? ilocal[i] : i) + leafOffsets[s]; 2509157edd7aSVaclav Hapla } 2510157edd7aSVaclav Hapla } 2511157edd7aSVaclav Hapla 2512157edd7aSVaclav Hapla /* Renumber and concatenate remote roots */ 25131f40158dSVaclav Hapla if (rootMode == PETSCSF_CONCATENATE_ROOTMODE_LOCAL || rootMode == PETSCSF_CONCATENATE_ROOTMODE_SHARED) { 25141f40158dSVaclav Hapla PetscInt rootOffset = 0; 25151f40158dSVaclav Hapla 25169566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nLeaves, &iremote_new)); 2517157edd7aSVaclav Hapla for (i = 0; i < nLeaves; i++) { 2518157edd7aSVaclav Hapla iremote_new[i].rank = -1; 2519157edd7aSVaclav Hapla iremote_new[i].index = -1; 2520157edd7aSVaclav Hapla } 2521157edd7aSVaclav Hapla for (s = 0; s < nsfs; s++) { 2522157edd7aSVaclav Hapla PetscInt i, nl, nr; 2523157edd7aSVaclav Hapla PetscSF tmp_sf; 2524157edd7aSVaclav Hapla const PetscSFNode *iremote; 2525157edd7aSVaclav Hapla PetscSFNode *tmp_rootdata; 2526157edd7aSVaclav Hapla PetscSFNode *tmp_leafdata = &iremote_new[leafArrayOffsets[s]]; 2527157edd7aSVaclav Hapla 25289566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfs[s], &nr, &nl, NULL, &iremote)); 25299566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, &tmp_sf)); 2530157edd7aSVaclav Hapla /* create helper SF with contiguous leaves */ 25319566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(tmp_sf, nr, nl, NULL, PETSC_USE_POINTER, (PetscSFNode *)iremote, PETSC_COPY_VALUES)); 25329566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(tmp_sf)); 25339566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nr, &tmp_rootdata)); 25341f40158dSVaclav Hapla if (rootMode == PETSCSF_CONCATENATE_ROOTMODE_LOCAL) { 2535157edd7aSVaclav Hapla for (i = 0; i < nr; i++) { 25361f40158dSVaclav Hapla tmp_rootdata[i].index = i + rootOffset; 2537157edd7aSVaclav Hapla tmp_rootdata[i].rank = (PetscInt)rank; 2538157edd7aSVaclav Hapla } 25391f40158dSVaclav Hapla rootOffset += nr; 25401f40158dSVaclav Hapla } else { 25411f40158dSVaclav Hapla for (i = 0; i < nr; i++) { 25421f40158dSVaclav Hapla tmp_rootdata[i].index = i; 25431f40158dSVaclav Hapla tmp_rootdata[i].rank = (PetscInt)rank; 25441f40158dSVaclav Hapla } 25451f40158dSVaclav Hapla } 25469566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(tmp_sf, MPIU_2INT, tmp_rootdata, tmp_leafdata, MPI_REPLACE)); 25479566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(tmp_sf, MPIU_2INT, tmp_rootdata, tmp_leafdata, MPI_REPLACE)); 25489566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&tmp_sf)); 25499566063dSJacob Faibussowitsch PetscCall(PetscFree(tmp_rootdata)); 2550157edd7aSVaclav Hapla } 2551aa624791SPierre Jolivet if (rootMode == PETSCSF_CONCATENATE_ROOTMODE_LOCAL) nRoots = rootOffset; // else nRoots already calculated above 2552157edd7aSVaclav Hapla 2553157edd7aSVaclav Hapla /* Build the new SF */ 25549566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, newsf)); 25559566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*newsf, nRoots, nLeaves, ilocal_new, PETSC_OWN_POINTER, iremote_new, PETSC_OWN_POINTER)); 25561f40158dSVaclav Hapla } else { 25571f40158dSVaclav Hapla /* Build the new SF */ 25581f40158dSVaclav Hapla PetscCall(PetscSFCreate(comm, newsf)); 25591f40158dSVaclav Hapla PetscCall(PetscSFSetGraphLayout(*newsf, glayout, nLeaves, ilocal_new, PETSC_OWN_POINTER, gremote)); 25601f40158dSVaclav Hapla } 25619566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(*newsf)); 25621f40158dSVaclav Hapla PetscCall(PetscSFViewFromOptions(*newsf, NULL, "-sf_concat_view")); 25631f40158dSVaclav Hapla PetscCall(PetscLayoutDestroy(&glayout)); 25641f40158dSVaclav Hapla PetscCall(PetscFree(gremote)); 25659566063dSJacob Faibussowitsch PetscCall(PetscFree(leafArrayOffsets)); 25663ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 2567157edd7aSVaclav Hapla } 2568