1af0996ceSBarry Smith #include <petsc/private/sfimpl.h> /*I "petscsf.h" I*/ 2c4e6a40aSLawrence Mitchell #include <petsc/private/hashseti.h> 353dd6d7dSJunchao Zhang #include <petsc/private/viewerimpl.h> 495fce210SBarry Smith #include <petscctable.h> 595fce210SBarry Smith 67fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_CUDA) 77fd2d3dbSJunchao Zhang #include <cuda_runtime.h> 87fd2d3dbSJunchao Zhang #endif 97fd2d3dbSJunchao Zhang 107fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_HIP) 117fd2d3dbSJunchao Zhang #include <hip/hip_runtime.h> 127fd2d3dbSJunchao Zhang #endif 137fd2d3dbSJunchao Zhang 142abc8c78SJacob Faibussowitsch #if defined(PETSC_CLANG_STATIC_ANALYZER) 152abc8c78SJacob Faibussowitsch void PetscSFCheckGraphSet(PetscSF, int); 162abc8c78SJacob Faibussowitsch #else 1795fce210SBarry Smith #if defined(PETSC_USE_DEBUG) 187a46b595SBarry Smith #define PetscSFCheckGraphSet(sf, arg) PetscCheck((sf)->graphset, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetGraph() or PetscSFSetGraphWithPattern() on argument %d \"%s\" before %s()", (arg), #sf, PETSC_FUNCTION_NAME); 1995fce210SBarry Smith #else 209371c9d4SSatish Balay #define PetscSFCheckGraphSet(sf, arg) \ 219371c9d4SSatish Balay do { \ 229371c9d4SSatish Balay } while (0) 2395fce210SBarry Smith #endif 242abc8c78SJacob Faibussowitsch #endif 2595fce210SBarry Smith 264c8fdceaSLisandro Dalcin const char *const PetscSFDuplicateOptions[] = {"CONFONLY", "RANKS", "GRAPH", "PetscSFDuplicateOption", "PETSCSF_DUPLICATE_", NULL}; 27*1f40158dSVaclav Hapla const char *const PetscSFConcatenateRootModes[] = {"local", "shared", "global", "PetscSFConcatenateRootMode", "PETSCSF_CONCATENATE_ROOTMODE_", NULL}; 2895fce210SBarry Smith 298af6ec1cSBarry Smith /*@ 3095fce210SBarry Smith PetscSFCreate - create a star forest communication context 3195fce210SBarry Smith 32d083f849SBarry Smith Collective 3395fce210SBarry Smith 344165533cSJose E. Roman Input Parameter: 3595fce210SBarry Smith . comm - communicator on which the star forest will operate 3695fce210SBarry Smith 374165533cSJose E. Roman Output Parameter: 3895fce210SBarry Smith . sf - new star forest context 3995fce210SBarry Smith 40dd5b3ca6SJunchao Zhang Options Database Keys: 41cab54364SBarry Smith . -sf_type type - value of type may be 42cab54364SBarry Smith .vb 43cab54364SBarry Smith basic -Use MPI persistent Isend/Irecv for communication (Default) 44cab54364SBarry Smith window -Use MPI-3 one-sided window for communication 45cab54364SBarry Smith neighbor -Use MPI-3 neighborhood collectives for communication 46cab54364SBarry Smith .ve 47dd5b3ca6SJunchao Zhang 4895fce210SBarry Smith Level: intermediate 4995fce210SBarry Smith 50cab54364SBarry Smith Note: 51cab54364SBarry Smith When one knows the communication graph is one of the predefined graph, such as `MPI_Alltoall()`, `MPI_Allgatherv()`, 52cab54364SBarry Smith `MPI_Gatherv()`, one can create a `PetscSF` and then set its graph with `PetscSFSetGraphWithPattern()`. These special 53cab54364SBarry Smith `SF`s are optimized and they have better performance than general `SF`s. 54dd5b3ca6SJunchao Zhang 55cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFSetGraphWithPattern()`, `PetscSFDestroy()` 5695fce210SBarry Smith @*/ 57d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreate(MPI_Comm comm, PetscSF *sf) 58d71ae5a4SJacob Faibussowitsch { 5995fce210SBarry Smith PetscSF b; 6095fce210SBarry Smith 6195fce210SBarry Smith PetscFunctionBegin; 6295fce210SBarry Smith PetscValidPointer(sf, 2); 639566063dSJacob Faibussowitsch PetscCall(PetscSFInitializePackage()); 6495fce210SBarry Smith 659566063dSJacob Faibussowitsch PetscCall(PetscHeaderCreate(b, PETSCSF_CLASSID, "PetscSF", "Star Forest", "PetscSF", comm, PetscSFDestroy, PetscSFView)); 6695fce210SBarry Smith 6795fce210SBarry Smith b->nroots = -1; 6895fce210SBarry Smith b->nleaves = -1; 6929046d53SLisandro Dalcin b->minleaf = PETSC_MAX_INT; 7029046d53SLisandro Dalcin b->maxleaf = PETSC_MIN_INT; 7195fce210SBarry Smith b->nranks = -1; 7295fce210SBarry Smith b->rankorder = PETSC_TRUE; 7395fce210SBarry Smith b->ingroup = MPI_GROUP_NULL; 7495fce210SBarry Smith b->outgroup = MPI_GROUP_NULL; 7595fce210SBarry Smith b->graphset = PETSC_FALSE; 7620c24465SJunchao Zhang #if defined(PETSC_HAVE_DEVICE) 7720c24465SJunchao Zhang b->use_gpu_aware_mpi = use_gpu_aware_mpi; 7820c24465SJunchao Zhang b->use_stream_aware_mpi = PETSC_FALSE; 7971438e86SJunchao Zhang b->unknown_input_stream = PETSC_FALSE; 8027f636e8SJunchao Zhang #if defined(PETSC_HAVE_KOKKOS) /* Prefer kokkos over cuda*/ 8120c24465SJunchao Zhang b->backend = PETSCSF_BACKEND_KOKKOS; 8227f636e8SJunchao Zhang #elif defined(PETSC_HAVE_CUDA) 8327f636e8SJunchao Zhang b->backend = PETSCSF_BACKEND_CUDA; 8459af0bd3SScott Kruger #elif defined(PETSC_HAVE_HIP) 8559af0bd3SScott Kruger b->backend = PETSCSF_BACKEND_HIP; 8620c24465SJunchao Zhang #endif 8771438e86SJunchao Zhang 8871438e86SJunchao Zhang #if defined(PETSC_HAVE_NVSHMEM) 8971438e86SJunchao Zhang b->use_nvshmem = PETSC_FALSE; /* Default is not to try NVSHMEM */ 9071438e86SJunchao Zhang b->use_nvshmem_get = PETSC_FALSE; /* Default is to use nvshmem_put based protocol */ 919566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL, NULL, "-use_nvshmem", &b->use_nvshmem, NULL)); 929566063dSJacob Faibussowitsch PetscCall(PetscOptionsGetBool(NULL, NULL, "-use_nvshmem_get", &b->use_nvshmem_get, NULL)); 9371438e86SJunchao Zhang #endif 9420c24465SJunchao Zhang #endif 9560c22052SBarry Smith b->vscat.from_n = -1; 9660c22052SBarry Smith b->vscat.to_n = -1; 9760c22052SBarry Smith b->vscat.unit = MPIU_SCALAR; 9895fce210SBarry Smith *sf = b; 9995fce210SBarry Smith PetscFunctionReturn(0); 10095fce210SBarry Smith } 10195fce210SBarry Smith 10229046d53SLisandro Dalcin /*@ 10395fce210SBarry Smith PetscSFReset - Reset a star forest so that different sizes or neighbors can be used 10495fce210SBarry Smith 10595fce210SBarry Smith Collective 10695fce210SBarry Smith 1074165533cSJose E. Roman Input Parameter: 10895fce210SBarry Smith . sf - star forest 10995fce210SBarry Smith 11095fce210SBarry Smith Level: advanced 11195fce210SBarry Smith 112cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`, `PetscSFSetGraph()`, `PetscSFDestroy()` 11395fce210SBarry Smith @*/ 114d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReset(PetscSF sf) 115d71ae5a4SJacob Faibussowitsch { 11695fce210SBarry Smith PetscFunctionBegin; 11795fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 118dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, Reset); 11929046d53SLisandro Dalcin sf->nroots = -1; 12029046d53SLisandro Dalcin sf->nleaves = -1; 12129046d53SLisandro Dalcin sf->minleaf = PETSC_MAX_INT; 12229046d53SLisandro Dalcin sf->maxleaf = PETSC_MIN_INT; 12395fce210SBarry Smith sf->mine = NULL; 12495fce210SBarry Smith sf->remote = NULL; 12529046d53SLisandro Dalcin sf->graphset = PETSC_FALSE; 1269566063dSJacob Faibussowitsch PetscCall(PetscFree(sf->mine_alloc)); 1279566063dSJacob Faibussowitsch PetscCall(PetscFree(sf->remote_alloc)); 12821c688dcSJed Brown sf->nranks = -1; 1299566063dSJacob Faibussowitsch PetscCall(PetscFree4(sf->ranks, sf->roffset, sf->rmine, sf->rremote)); 13029046d53SLisandro Dalcin sf->degreeknown = PETSC_FALSE; 1319566063dSJacob Faibussowitsch PetscCall(PetscFree(sf->degree)); 1329566063dSJacob Faibussowitsch if (sf->ingroup != MPI_GROUP_NULL) PetscCallMPI(MPI_Group_free(&sf->ingroup)); 1339566063dSJacob Faibussowitsch if (sf->outgroup != MPI_GROUP_NULL) PetscCallMPI(MPI_Group_free(&sf->outgroup)); 134013b3241SStefano Zampini if (sf->multi) sf->multi->multi = NULL; 1359566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&sf->multi)); 1369566063dSJacob Faibussowitsch PetscCall(PetscLayoutDestroy(&sf->map)); 13771438e86SJunchao Zhang 13871438e86SJunchao Zhang #if defined(PETSC_HAVE_DEVICE) 1399566063dSJacob Faibussowitsch for (PetscInt i = 0; i < 2; i++) PetscCall(PetscSFFree(sf, PETSC_MEMTYPE_DEVICE, sf->rmine_d[i])); 14071438e86SJunchao Zhang #endif 14171438e86SJunchao Zhang 14295fce210SBarry Smith sf->setupcalled = PETSC_FALSE; 14395fce210SBarry Smith PetscFunctionReturn(0); 14495fce210SBarry Smith } 14595fce210SBarry Smith 14695fce210SBarry Smith /*@C 147cab54364SBarry Smith PetscSFSetType - Set the `PetscSF` communication implementation 14895fce210SBarry Smith 149cab54364SBarry Smith Collective on sf 15095fce210SBarry Smith 15195fce210SBarry Smith Input Parameters: 152cab54364SBarry Smith + sf - the `PetscSF` context 15395fce210SBarry Smith - type - a known method 154cab54364SBarry Smith .vb 155cab54364SBarry Smith PETSCSFWINDOW - MPI-2/3 one-sided 156cab54364SBarry Smith PETSCSFBASIC - basic implementation using MPI-1 two-sided 157cab54364SBarry Smith .ve 15895fce210SBarry Smith 15995fce210SBarry Smith Options Database Key: 160cab54364SBarry Smith . -sf_type <type> - Sets the method; for example basic or window use -help for a list of available methods (for instance, window, basic, neighbor) 161cab54364SBarry Smith 162cab54364SBarry Smith Level: intermediate 16395fce210SBarry Smith 16495fce210SBarry Smith Notes: 16595fce210SBarry Smith See "include/petscsf.h" for available methods (for instance) 16695fce210SBarry Smith 167db781477SPatrick Sanan .seealso: `PetscSFType`, `PetscSFCreate()` 16895fce210SBarry Smith @*/ 169d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetType(PetscSF sf, PetscSFType type) 170d71ae5a4SJacob Faibussowitsch { 17195fce210SBarry Smith PetscBool match; 1725f80ce2aSJacob Faibussowitsch PetscErrorCode (*r)(PetscSF); 17395fce210SBarry Smith 17495fce210SBarry Smith PetscFunctionBegin; 17595fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 17695fce210SBarry Smith PetscValidCharPointer(type, 2); 17795fce210SBarry Smith 1789566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare((PetscObject)sf, type, &match)); 17995fce210SBarry Smith if (match) PetscFunctionReturn(0); 18095fce210SBarry Smith 1819566063dSJacob Faibussowitsch PetscCall(PetscFunctionListFind(PetscSFList, type, &r)); 18228b400f6SJacob Faibussowitsch PetscCheck(r, PETSC_COMM_SELF, PETSC_ERR_ARG_UNKNOWN_TYPE, "Unable to find requested PetscSF type %s", type); 18329046d53SLisandro Dalcin /* Destroy the previous PetscSF implementation context */ 184dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, Destroy); 1859566063dSJacob Faibussowitsch PetscCall(PetscMemzero(sf->ops, sizeof(*sf->ops))); 1869566063dSJacob Faibussowitsch PetscCall(PetscObjectChangeTypeName((PetscObject)sf, type)); 1879566063dSJacob Faibussowitsch PetscCall((*r)(sf)); 18895fce210SBarry Smith PetscFunctionReturn(0); 18995fce210SBarry Smith } 19095fce210SBarry Smith 19129046d53SLisandro Dalcin /*@C 192cab54364SBarry Smith PetscSFGetType - Get the `PetscSF` communication implementation 19329046d53SLisandro Dalcin 19429046d53SLisandro Dalcin Not Collective 19529046d53SLisandro Dalcin 19629046d53SLisandro Dalcin Input Parameter: 197cab54364SBarry Smith . sf - the `PetscSF` context 19829046d53SLisandro Dalcin 19929046d53SLisandro Dalcin Output Parameter: 200cab54364SBarry Smith . type - the `PetscSF` type name 20129046d53SLisandro Dalcin 20229046d53SLisandro Dalcin Level: intermediate 20329046d53SLisandro Dalcin 204cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFSetType()`, `PetscSFCreate()` 20529046d53SLisandro Dalcin @*/ 206d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetType(PetscSF sf, PetscSFType *type) 207d71ae5a4SJacob Faibussowitsch { 20829046d53SLisandro Dalcin PetscFunctionBegin; 20929046d53SLisandro Dalcin PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 21029046d53SLisandro Dalcin PetscValidPointer(type, 2); 21129046d53SLisandro Dalcin *type = ((PetscObject)sf)->type_name; 21229046d53SLisandro Dalcin PetscFunctionReturn(0); 21329046d53SLisandro Dalcin } 21429046d53SLisandro Dalcin 2151fb7b255SJunchao Zhang /*@C 21695fce210SBarry Smith PetscSFDestroy - destroy star forest 21795fce210SBarry Smith 21895fce210SBarry Smith Collective 21995fce210SBarry Smith 2204165533cSJose E. Roman Input Parameter: 22195fce210SBarry Smith . sf - address of star forest 22295fce210SBarry Smith 22395fce210SBarry Smith Level: intermediate 22495fce210SBarry Smith 225cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFReset()` 22695fce210SBarry Smith @*/ 227d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFDestroy(PetscSF *sf) 228d71ae5a4SJacob Faibussowitsch { 22995fce210SBarry Smith PetscFunctionBegin; 23095fce210SBarry Smith if (!*sf) PetscFunctionReturn(0); 23195fce210SBarry Smith PetscValidHeaderSpecific((*sf), PETSCSF_CLASSID, 1); 2329371c9d4SSatish Balay if (--((PetscObject)(*sf))->refct > 0) { 2339371c9d4SSatish Balay *sf = NULL; 2349371c9d4SSatish Balay PetscFunctionReturn(0); 2359371c9d4SSatish Balay } 2369566063dSJacob Faibussowitsch PetscCall(PetscSFReset(*sf)); 237dbbe0bcdSBarry Smith PetscTryTypeMethod((*sf), Destroy); 2389566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&(*sf)->vscat.lsf)); 2399566063dSJacob Faibussowitsch if ((*sf)->vscat.bs > 1) PetscCallMPI(MPI_Type_free(&(*sf)->vscat.unit)); 2409566063dSJacob Faibussowitsch PetscCall(PetscHeaderDestroy(sf)); 24195fce210SBarry Smith PetscFunctionReturn(0); 24295fce210SBarry Smith } 24395fce210SBarry Smith 244d71ae5a4SJacob Faibussowitsch static PetscErrorCode PetscSFCheckGraphValid_Private(PetscSF sf) 245d71ae5a4SJacob Faibussowitsch { 246c4e6a40aSLawrence Mitchell PetscInt i, nleaves; 247c4e6a40aSLawrence Mitchell PetscMPIInt size; 248c4e6a40aSLawrence Mitchell const PetscInt *ilocal; 249c4e6a40aSLawrence Mitchell const PetscSFNode *iremote; 250c4e6a40aSLawrence Mitchell 251c4e6a40aSLawrence Mitchell PetscFunctionBegin; 25276bd3646SJed Brown if (!sf->graphset || !PetscDefined(USE_DEBUG)) PetscFunctionReturn(0); 2539566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, NULL, &nleaves, &ilocal, &iremote)); 2549566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)sf), &size)); 255c4e6a40aSLawrence Mitchell for (i = 0; i < nleaves; i++) { 256c4e6a40aSLawrence Mitchell const PetscInt rank = iremote[i].rank; 257c4e6a40aSLawrence Mitchell const PetscInt remote = iremote[i].index; 258c4e6a40aSLawrence Mitchell const PetscInt leaf = ilocal ? ilocal[i] : i; 259c9cc58a2SBarry Smith PetscCheck(rank >= 0 && rank < size, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided rank (%" PetscInt_FMT ") for remote %" PetscInt_FMT " is invalid, should be in [0, %d)", rank, i, size); 26008401ef6SPierre Jolivet PetscCheck(remote >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided index (%" PetscInt_FMT ") for remote %" PetscInt_FMT " is invalid, should be >= 0", remote, i); 26108401ef6SPierre Jolivet PetscCheck(leaf >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided location (%" PetscInt_FMT ") for leaf %" PetscInt_FMT " is invalid, should be >= 0", leaf, i); 262c4e6a40aSLawrence Mitchell } 263c4e6a40aSLawrence Mitchell PetscFunctionReturn(0); 264c4e6a40aSLawrence Mitchell } 265c4e6a40aSLawrence Mitchell 26695fce210SBarry Smith /*@ 26795fce210SBarry Smith PetscSFSetUp - set up communication structures 26895fce210SBarry Smith 26995fce210SBarry Smith Collective 27095fce210SBarry Smith 2714165533cSJose E. Roman Input Parameter: 27295fce210SBarry Smith . sf - star forest communication object 27395fce210SBarry Smith 27495fce210SBarry Smith Level: beginner 27595fce210SBarry Smith 276cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFSetFromOptions()`, `PetscSFSetType()` 27795fce210SBarry Smith @*/ 278d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetUp(PetscSF sf) 279d71ae5a4SJacob Faibussowitsch { 28095fce210SBarry Smith PetscFunctionBegin; 28129046d53SLisandro Dalcin PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 28229046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 28395fce210SBarry Smith if (sf->setupcalled) PetscFunctionReturn(0); 2849566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_SetUp, sf, 0, 0, 0)); 2859566063dSJacob Faibussowitsch PetscCall(PetscSFCheckGraphValid_Private(sf)); 2869566063dSJacob Faibussowitsch if (!((PetscObject)sf)->type_name) PetscCall(PetscSFSetType(sf, PETSCSFBASIC)); /* Zero all sf->ops */ 287dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, SetUp); 28820c24465SJunchao Zhang #if defined(PETSC_HAVE_CUDA) 28920c24465SJunchao Zhang if (sf->backend == PETSCSF_BACKEND_CUDA) { 29071438e86SJunchao Zhang sf->ops->Malloc = PetscSFMalloc_CUDA; 29171438e86SJunchao Zhang sf->ops->Free = PetscSFFree_CUDA; 29220c24465SJunchao Zhang } 29320c24465SJunchao Zhang #endif 29459af0bd3SScott Kruger #if defined(PETSC_HAVE_HIP) 29559af0bd3SScott Kruger if (sf->backend == PETSCSF_BACKEND_HIP) { 29659af0bd3SScott Kruger sf->ops->Malloc = PetscSFMalloc_HIP; 29759af0bd3SScott Kruger sf->ops->Free = PetscSFFree_HIP; 29859af0bd3SScott Kruger } 29959af0bd3SScott Kruger #endif 30020c24465SJunchao Zhang 30159af0bd3SScott Kruger # 30220c24465SJunchao Zhang #if defined(PETSC_HAVE_KOKKOS) 30320c24465SJunchao Zhang if (sf->backend == PETSCSF_BACKEND_KOKKOS) { 30420c24465SJunchao Zhang sf->ops->Malloc = PetscSFMalloc_Kokkos; 30520c24465SJunchao Zhang sf->ops->Free = PetscSFFree_Kokkos; 30620c24465SJunchao Zhang } 30720c24465SJunchao Zhang #endif 3089566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_SetUp, sf, 0, 0, 0)); 30995fce210SBarry Smith sf->setupcalled = PETSC_TRUE; 31095fce210SBarry Smith PetscFunctionReturn(0); 31195fce210SBarry Smith } 31295fce210SBarry Smith 3138af6ec1cSBarry Smith /*@ 314cab54364SBarry Smith PetscSFSetFromOptions - set `PetscSF` options using the options database 31595fce210SBarry Smith 31695fce210SBarry Smith Logically Collective 31795fce210SBarry Smith 3184165533cSJose E. Roman Input Parameter: 31995fce210SBarry Smith . sf - star forest 32095fce210SBarry Smith 32195fce210SBarry Smith Options Database Keys: 32260263706SJed Brown + -sf_type - implementation type, see PetscSFSetType() 32351ccb202SJunchao Zhang . -sf_rank_order - sort composite points for gathers and scatters in rank order, gathers are non-deterministic otherwise 324b85e67b7SJunchao Zhang . -sf_use_default_stream - Assume callers of SF computed the input root/leafdata with the default cuda stream. SF will also 325c2a741eeSJunchao Zhang use the default stream to process data. Therefore, no stream synchronization is needed between SF and its caller (default: true). 326c06a8e02SRichard Tran Mills If true, this option only works with -use_gpu_aware_mpi 1. 32720c24465SJunchao Zhang . -sf_use_stream_aware_mpi - Assume the underlying MPI is cuda-stream aware and SF won't sync streams for send/recv buffers passed to MPI (default: false). 328c06a8e02SRichard Tran Mills If true, this option only works with -use_gpu_aware_mpi 1. 32995fce210SBarry Smith 33059af0bd3SScott Kruger - -sf_backend cuda | hip | kokkos -Select the device backend SF uses. Currently SF has these backends: cuda, hip and Kokkos. 33159af0bd3SScott Kruger On CUDA (HIP) devices, one can choose cuda (hip) or kokkos with the default being kokkos. On other devices, 33220c24465SJunchao Zhang the only available is kokkos. 33320c24465SJunchao Zhang 33495fce210SBarry Smith Level: intermediate 335cab54364SBarry Smith 336cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`, `PetscSFSetType()` 33795fce210SBarry Smith @*/ 338d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetFromOptions(PetscSF sf) 339d71ae5a4SJacob Faibussowitsch { 34095fce210SBarry Smith PetscSFType deft; 34195fce210SBarry Smith char type[256]; 34295fce210SBarry Smith PetscBool flg; 34395fce210SBarry Smith 34495fce210SBarry Smith PetscFunctionBegin; 34595fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 346d0609cedSBarry Smith PetscObjectOptionsBegin((PetscObject)sf); 34795fce210SBarry Smith deft = ((PetscObject)sf)->type_name ? ((PetscObject)sf)->type_name : PETSCSFBASIC; 3489566063dSJacob Faibussowitsch PetscCall(PetscOptionsFList("-sf_type", "PetscSF implementation type", "PetscSFSetType", PetscSFList, deft, type, sizeof(type), &flg)); 3499566063dSJacob Faibussowitsch PetscCall(PetscSFSetType(sf, flg ? type : deft)); 3509566063dSJacob Faibussowitsch PetscCall(PetscOptionsBool("-sf_rank_order", "sort composite points for gathers and scatters in rank order, gathers are non-deterministic otherwise", "PetscSFSetRankOrder", sf->rankorder, &sf->rankorder, NULL)); 3517fd2d3dbSJunchao Zhang #if defined(PETSC_HAVE_DEVICE) 35220c24465SJunchao Zhang { 35320c24465SJunchao Zhang char backendstr[32] = {0}; 35459af0bd3SScott Kruger PetscBool isCuda = PETSC_FALSE, isHip = PETSC_FALSE, isKokkos = PETSC_FALSE, set; 35520c24465SJunchao Zhang /* Change the defaults set in PetscSFCreate() with command line options */ 356d5b43468SJose E. Roman PetscCall(PetscOptionsBool("-sf_unknown_input_stream", "SF root/leafdata is computed on arbitrary streams unknown to SF", "PetscSFSetFromOptions", sf->unknown_input_stream, &sf->unknown_input_stream, NULL)); 3579566063dSJacob Faibussowitsch PetscCall(PetscOptionsBool("-sf_use_stream_aware_mpi", "Assume the underlying MPI is cuda-stream aware", "PetscSFSetFromOptions", sf->use_stream_aware_mpi, &sf->use_stream_aware_mpi, NULL)); 3589566063dSJacob Faibussowitsch PetscCall(PetscOptionsString("-sf_backend", "Select the device backend SF uses", "PetscSFSetFromOptions", NULL, backendstr, sizeof(backendstr), &set)); 3599566063dSJacob Faibussowitsch PetscCall(PetscStrcasecmp("cuda", backendstr, &isCuda)); 3609566063dSJacob Faibussowitsch PetscCall(PetscStrcasecmp("kokkos", backendstr, &isKokkos)); 3619566063dSJacob Faibussowitsch PetscCall(PetscStrcasecmp("hip", backendstr, &isHip)); 36259af0bd3SScott Kruger #if defined(PETSC_HAVE_CUDA) || defined(PETSC_HAVE_HIP) 36320c24465SJunchao Zhang if (isCuda) sf->backend = PETSCSF_BACKEND_CUDA; 36420c24465SJunchao Zhang else if (isKokkos) sf->backend = PETSCSF_BACKEND_KOKKOS; 36559af0bd3SScott Kruger else if (isHip) sf->backend = PETSCSF_BACKEND_HIP; 36628b400f6SJacob Faibussowitsch else PetscCheck(!set, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "-sf_backend %s is not supported. You may choose cuda, hip or kokkos (if installed)", backendstr); 36720c24465SJunchao Zhang #elif defined(PETSC_HAVE_KOKKOS) 36808401ef6SPierre Jolivet PetscCheck(!set || isKokkos, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "-sf_backend %s is not supported. You can only choose kokkos", backendstr); 36920c24465SJunchao Zhang #endif 37020c24465SJunchao Zhang } 371c2a741eeSJunchao Zhang #endif 372dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, SetFromOptions, PetscOptionsObject); 373d0609cedSBarry Smith PetscOptionsEnd(); 37495fce210SBarry Smith PetscFunctionReturn(0); 37595fce210SBarry Smith } 37695fce210SBarry Smith 37729046d53SLisandro Dalcin /*@ 37895fce210SBarry Smith PetscSFSetRankOrder - sort multi-points for gathers and scatters by rank order 37995fce210SBarry Smith 38095fce210SBarry Smith Logically Collective 38195fce210SBarry Smith 3824165533cSJose E. Roman Input Parameters: 38395fce210SBarry Smith + sf - star forest 384cab54364SBarry Smith - flg - `PETSC_TRUE` to sort, `PETSC_FALSE` to skip sorting (lower setup cost, but non-deterministic) 38595fce210SBarry Smith 38695fce210SBarry Smith Level: advanced 38795fce210SBarry Smith 388cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFGatherBegin()`, `PetscSFScatterBegin()` 38995fce210SBarry Smith @*/ 390d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetRankOrder(PetscSF sf, PetscBool flg) 391d71ae5a4SJacob Faibussowitsch { 39295fce210SBarry Smith PetscFunctionBegin; 39395fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 39495fce210SBarry Smith PetscValidLogicalCollectiveBool(sf, flg, 2); 39528b400f6SJacob Faibussowitsch PetscCheck(!sf->multi, PetscObjectComm((PetscObject)sf), PETSC_ERR_ARG_WRONGSTATE, "Rank ordering must be set before first call to PetscSFGatherBegin() or PetscSFScatterBegin()"); 39695fce210SBarry Smith sf->rankorder = flg; 39795fce210SBarry Smith PetscFunctionReturn(0); 39895fce210SBarry Smith } 39995fce210SBarry Smith 4008dbb0df6SBarry Smith /*@C 40195fce210SBarry Smith PetscSFSetGraph - Set a parallel star forest 40295fce210SBarry Smith 40395fce210SBarry Smith Collective 40495fce210SBarry Smith 4054165533cSJose E. Roman Input Parameters: 40695fce210SBarry Smith + sf - star forest 40795fce210SBarry Smith . nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves) 40895fce210SBarry Smith . nleaves - number of leaf vertices on the current process, each of these references a root on any process 409c4e6a40aSLawrence Mitchell . ilocal - locations of leaves in leafdata buffers, pass NULL for contiguous storage (locations must be >= 0, enforced 410c4e6a40aSLawrence Mitchell during setup in debug mode) 41195fce210SBarry Smith . localmode - copy mode for ilocal 412c4e6a40aSLawrence Mitchell . iremote - remote locations of root vertices for each leaf on the current process (locations must be >= 0, enforced 413c4e6a40aSLawrence Mitchell during setup in debug mode) 41495fce210SBarry Smith - remotemode - copy mode for iremote 41595fce210SBarry Smith 41695fce210SBarry Smith Level: intermediate 41795fce210SBarry Smith 41895452b02SPatrick Sanan Notes: 419db2b9530SVaclav Hapla Leaf indices in ilocal must be unique, otherwise an error occurs. 42038ab3f8aSBarry Smith 421db2b9530SVaclav Hapla Input arrays ilocal and iremote follow the PetscCopyMode semantics. 422cab54364SBarry Smith In particular, if localmode/remotemode is `PETSC_OWN_POINTER` or `PETSC_USE_POINTER`, 423db2b9530SVaclav Hapla PETSc might modify the respective array; 424cab54364SBarry Smith if `PETSC_USE_POINTER`, the user must delete the array after PetscSFDestroy(). 425cab54364SBarry Smith Only if `PETSC_COPY_VALUES` is used, the respective array is guaranteed to stay intact and a const array can be passed (but a cast to non-const is needed). 426db2b9530SVaclav Hapla 427cab54364SBarry Smith Fortran Note: 428db2b9530SVaclav Hapla In Fortran you must use PETSC_COPY_VALUES for localmode and remotemode. 429c4e6a40aSLawrence Mitchell 430cab54364SBarry Smith Developer Note: 431db2b9530SVaclav Hapla We sort leaves to check for duplicates and contiguousness and to find minleaf/maxleaf. 432db2b9530SVaclav Hapla This also allows to compare leaf sets of two SFs easily. 43372bf8598SVaclav Hapla 434cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFGetGraph()` 43595fce210SBarry Smith @*/ 436d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetGraph(PetscSF sf, PetscInt nroots, PetscInt nleaves, PetscInt *ilocal, PetscCopyMode localmode, PetscSFNode *iremote, PetscCopyMode remotemode) 437d71ae5a4SJacob Faibussowitsch { 438db2b9530SVaclav Hapla PetscBool unique, contiguous; 43995fce210SBarry Smith 44095fce210SBarry Smith PetscFunctionBegin; 44195fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 44229046d53SLisandro Dalcin if (nleaves > 0 && ilocal) PetscValidIntPointer(ilocal, 4); 44329046d53SLisandro Dalcin if (nleaves > 0) PetscValidPointer(iremote, 6); 44408401ef6SPierre Jolivet PetscCheck(nroots >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nroots %" PetscInt_FMT ", cannot be negative", nroots); 44508401ef6SPierre Jolivet PetscCheck(nleaves >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "nleaves %" PetscInt_FMT ", cannot be negative", nleaves); 4468da24d32SBarry Smith /* enums may be handled as unsigned by some compilers, NVHPC for example, the int cast 4478da24d32SBarry Smith * below is to prevent NVHPC from warning about meaningless comparison of unsigned with zero */ 4488da24d32SBarry Smith PetscCheck((int)localmode >= PETSC_COPY_VALUES && localmode <= PETSC_USE_POINTER, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Wrong localmode %d", localmode); 4498da24d32SBarry Smith PetscCheck((int)remotemode >= PETSC_COPY_VALUES && remotemode <= PETSC_USE_POINTER, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Wrong remotemode %d", remotemode); 45029046d53SLisandro Dalcin 4512a67d2daSStefano Zampini if (sf->nroots >= 0) { /* Reset only if graph already set */ 4529566063dSJacob Faibussowitsch PetscCall(PetscSFReset(sf)); 4532a67d2daSStefano Zampini } 4542a67d2daSStefano Zampini 4559566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_SetGraph, sf, 0, 0, 0)); 45629046d53SLisandro Dalcin 45795fce210SBarry Smith sf->nroots = nroots; 45895fce210SBarry Smith sf->nleaves = nleaves; 45929046d53SLisandro Dalcin 460db2b9530SVaclav Hapla if (localmode == PETSC_COPY_VALUES && ilocal) { 461db2b9530SVaclav Hapla PetscInt *tlocal = NULL; 462db2b9530SVaclav Hapla 4639566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nleaves, &tlocal)); 4649566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(tlocal, ilocal, nleaves)); 465db2b9530SVaclav Hapla ilocal = tlocal; 466db2b9530SVaclav Hapla } 467db2b9530SVaclav Hapla if (remotemode == PETSC_COPY_VALUES) { 468db2b9530SVaclav Hapla PetscSFNode *tremote = NULL; 469db2b9530SVaclav Hapla 4709566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nleaves, &tremote)); 4719566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(tremote, iremote, nleaves)); 472db2b9530SVaclav Hapla iremote = tremote; 473db2b9530SVaclav Hapla } 474db2b9530SVaclav Hapla 47529046d53SLisandro Dalcin if (nleaves && ilocal) { 476db2b9530SVaclav Hapla PetscSFNode work; 477db2b9530SVaclav Hapla 4789566063dSJacob Faibussowitsch PetscCall(PetscSortIntWithDataArray(nleaves, ilocal, iremote, sizeof(PetscSFNode), &work)); 4799566063dSJacob Faibussowitsch PetscCall(PetscSortedCheckDupsInt(nleaves, ilocal, &unique)); 480db2b9530SVaclav Hapla unique = PetscNot(unique); 481db2b9530SVaclav Hapla PetscCheck(sf->allow_multi_leaves || unique, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONG, "Input ilocal has duplicate entries which is not allowed for this PetscSF"); 482db2b9530SVaclav Hapla sf->minleaf = ilocal[0]; 483db2b9530SVaclav Hapla sf->maxleaf = ilocal[nleaves - 1]; 484db2b9530SVaclav Hapla contiguous = (PetscBool)(unique && ilocal[0] == 0 && ilocal[nleaves - 1] == nleaves - 1); 48529046d53SLisandro Dalcin } else { 48629046d53SLisandro Dalcin sf->minleaf = 0; 48729046d53SLisandro Dalcin sf->maxleaf = nleaves - 1; 488db2b9530SVaclav Hapla unique = PETSC_TRUE; 489db2b9530SVaclav Hapla contiguous = PETSC_TRUE; 49029046d53SLisandro Dalcin } 49129046d53SLisandro Dalcin 492db2b9530SVaclav Hapla if (contiguous) { 493db2b9530SVaclav Hapla if (localmode == PETSC_USE_POINTER) { 494db2b9530SVaclav Hapla ilocal = NULL; 495db2b9530SVaclav Hapla } else { 4969566063dSJacob Faibussowitsch PetscCall(PetscFree(ilocal)); 497db2b9530SVaclav Hapla } 498db2b9530SVaclav Hapla } 499db2b9530SVaclav Hapla sf->mine = ilocal; 500db2b9530SVaclav Hapla if (localmode == PETSC_USE_POINTER) { 50129046d53SLisandro Dalcin sf->mine_alloc = NULL; 502db2b9530SVaclav Hapla } else { 503db2b9530SVaclav Hapla sf->mine_alloc = ilocal; 50495fce210SBarry Smith } 505db2b9530SVaclav Hapla sf->remote = iremote; 506db2b9530SVaclav Hapla if (remotemode == PETSC_USE_POINTER) { 50729046d53SLisandro Dalcin sf->remote_alloc = NULL; 508db2b9530SVaclav Hapla } else { 509db2b9530SVaclav Hapla sf->remote_alloc = iremote; 51095fce210SBarry Smith } 5119566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_SetGraph, sf, 0, 0, 0)); 51229046d53SLisandro Dalcin sf->graphset = PETSC_TRUE; 51395fce210SBarry Smith PetscFunctionReturn(0); 51495fce210SBarry Smith } 51595fce210SBarry Smith 51629046d53SLisandro Dalcin /*@ 517cab54364SBarry Smith PetscSFSetGraphWithPattern - Sets the graph of a `PetscSF` with a specific pattern 518dd5b3ca6SJunchao Zhang 519dd5b3ca6SJunchao Zhang Collective 520dd5b3ca6SJunchao Zhang 521dd5b3ca6SJunchao Zhang Input Parameters: 522cab54364SBarry Smith + sf - The `PetscSF` 523cab54364SBarry Smith . map - Layout of roots over all processes (insignificant when pattern is `PETSCSF_PATTERN_ALLTOALL`) 524cab54364SBarry Smith - pattern - One of `PETSCSF_PATTERN_ALLGATHER`, `PETSCSF_PATTERN_GATHER`, `PETSCSF_PATTERN_ALLTOALL` 525cab54364SBarry Smith 526cab54364SBarry Smith Level: intermediate 527dd5b3ca6SJunchao Zhang 528dd5b3ca6SJunchao Zhang Notes: 529cab54364SBarry Smith It is easier to explain `PetscSFPattern` using vectors. Suppose we have an MPI vector x and its layout is map. 530dd5b3ca6SJunchao Zhang n and N are local and global sizes of x respectively. 531dd5b3ca6SJunchao Zhang 532cab54364SBarry Smith With `PETSCSF_PATTERN_ALLGATHER`, the routine creates a graph that if one does Bcast on it, it will copy x to 533dd5b3ca6SJunchao Zhang sequential vectors y on all ranks. 534dd5b3ca6SJunchao Zhang 535cab54364SBarry Smith With `PETSCSF_PATTERN_GATHER`, the routine creates a graph that if one does Bcast on it, it will copy x to a 536dd5b3ca6SJunchao Zhang sequential vector y on rank 0. 537dd5b3ca6SJunchao Zhang 538dd5b3ca6SJunchao Zhang In above cases, entries of x are roots and entries of y are leaves. 539dd5b3ca6SJunchao Zhang 540cab54364SBarry Smith With `PETSCSF_PATTERN_ALLTOALL`, map is insignificant. Suppose NP is size of sf's communicator. The routine 541dd5b3ca6SJunchao Zhang creates a graph that every rank has NP leaves and NP roots. On rank i, its leaf j is connected to root i 542cab54364SBarry Smith of rank j. Here 0 <=i,j<NP. It is a kind of `MPI_Alltoall()` with sendcount/recvcount being 1. Note that it does 543dd5b3ca6SJunchao Zhang not mean one can not send multiple items. One just needs to create a new MPI datatype for the mulptiple data 544cab54364SBarry Smith items with `MPI_Type_contiguous` and use that as the <unit> argument in SF routines. 545dd5b3ca6SJunchao Zhang 546dd5b3ca6SJunchao Zhang In this case, roots and leaves are symmetric. 547dd5b3ca6SJunchao Zhang 548cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFGetGraph()` 549dd5b3ca6SJunchao Zhang @*/ 550d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetGraphWithPattern(PetscSF sf, PetscLayout map, PetscSFPattern pattern) 551d71ae5a4SJacob Faibussowitsch { 552dd5b3ca6SJunchao Zhang MPI_Comm comm; 553dd5b3ca6SJunchao Zhang PetscInt n, N, res[2]; 554dd5b3ca6SJunchao Zhang PetscMPIInt rank, size; 555dd5b3ca6SJunchao Zhang PetscSFType type; 556dd5b3ca6SJunchao Zhang 557dd5b3ca6SJunchao Zhang PetscFunctionBegin; 5582abc8c78SJacob Faibussowitsch PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 5592abc8c78SJacob Faibussowitsch if (pattern != PETSCSF_PATTERN_ALLTOALL) PetscValidPointer(map, 2); 5609566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 5612c71b3e2SJacob Faibussowitsch PetscCheck(pattern >= PETSCSF_PATTERN_ALLGATHER && pattern <= PETSCSF_PATTERN_ALLTOALL, comm, PETSC_ERR_ARG_OUTOFRANGE, "Unsupported PetscSFPattern %d", pattern); 5629566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 5639566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(comm, &size)); 564dd5b3ca6SJunchao Zhang 565dd5b3ca6SJunchao Zhang if (pattern == PETSCSF_PATTERN_ALLTOALL) { 566dd5b3ca6SJunchao Zhang type = PETSCSFALLTOALL; 5679566063dSJacob Faibussowitsch PetscCall(PetscLayoutCreate(comm, &sf->map)); 5689566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetLocalSize(sf->map, size)); 5699566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetSize(sf->map, ((PetscInt)size) * size)); 5709566063dSJacob Faibussowitsch PetscCall(PetscLayoutSetUp(sf->map)); 571dd5b3ca6SJunchao Zhang } else { 5729566063dSJacob Faibussowitsch PetscCall(PetscLayoutGetLocalSize(map, &n)); 5739566063dSJacob Faibussowitsch PetscCall(PetscLayoutGetSize(map, &N)); 574dd5b3ca6SJunchao Zhang res[0] = n; 575dd5b3ca6SJunchao Zhang res[1] = -n; 576dd5b3ca6SJunchao Zhang /* Check if n are same over all ranks so that we can optimize it */ 5771c2dc1cbSBarry Smith PetscCall(MPIU_Allreduce(MPI_IN_PLACE, res, 2, MPIU_INT, MPI_MAX, comm)); 578dd5b3ca6SJunchao Zhang if (res[0] == -res[1]) { /* same n */ 579dd5b3ca6SJunchao Zhang type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHER : PETSCSFGATHER; 580dd5b3ca6SJunchao Zhang } else { 581dd5b3ca6SJunchao Zhang type = (pattern == PETSCSF_PATTERN_ALLGATHER) ? PETSCSFALLGATHERV : PETSCSFGATHERV; 582dd5b3ca6SJunchao Zhang } 5839566063dSJacob Faibussowitsch PetscCall(PetscLayoutReference(map, &sf->map)); 584dd5b3ca6SJunchao Zhang } 5859566063dSJacob Faibussowitsch PetscCall(PetscSFSetType(sf, type)); 586dd5b3ca6SJunchao Zhang 587dd5b3ca6SJunchao Zhang sf->pattern = pattern; 588dd5b3ca6SJunchao Zhang sf->mine = NULL; /* Contiguous */ 589dd5b3ca6SJunchao Zhang 590dd5b3ca6SJunchao Zhang /* Set nleaves, nroots here in case user calls PetscSFGetGraph, which is legal to call even before PetscSFSetUp is called. 591dd5b3ca6SJunchao Zhang Also set other easy stuff. 592dd5b3ca6SJunchao Zhang */ 593dd5b3ca6SJunchao Zhang if (pattern == PETSCSF_PATTERN_ALLGATHER) { 594dd5b3ca6SJunchao Zhang sf->nleaves = N; 595dd5b3ca6SJunchao Zhang sf->nroots = n; 596dd5b3ca6SJunchao Zhang sf->nranks = size; 597dd5b3ca6SJunchao Zhang sf->minleaf = 0; 598dd5b3ca6SJunchao Zhang sf->maxleaf = N - 1; 599dd5b3ca6SJunchao Zhang } else if (pattern == PETSCSF_PATTERN_GATHER) { 600dd5b3ca6SJunchao Zhang sf->nleaves = rank ? 0 : N; 601dd5b3ca6SJunchao Zhang sf->nroots = n; 602dd5b3ca6SJunchao Zhang sf->nranks = rank ? 0 : size; 603dd5b3ca6SJunchao Zhang sf->minleaf = 0; 604dd5b3ca6SJunchao Zhang sf->maxleaf = rank ? -1 : N - 1; 605dd5b3ca6SJunchao Zhang } else if (pattern == PETSCSF_PATTERN_ALLTOALL) { 606dd5b3ca6SJunchao Zhang sf->nleaves = size; 607dd5b3ca6SJunchao Zhang sf->nroots = size; 608dd5b3ca6SJunchao Zhang sf->nranks = size; 609dd5b3ca6SJunchao Zhang sf->minleaf = 0; 610dd5b3ca6SJunchao Zhang sf->maxleaf = size - 1; 611dd5b3ca6SJunchao Zhang } 612dd5b3ca6SJunchao Zhang sf->ndranks = 0; /* We do not need to separate out distinguished ranks for patterned graphs to improve communication performance */ 613dd5b3ca6SJunchao Zhang sf->graphset = PETSC_TRUE; 614dd5b3ca6SJunchao Zhang PetscFunctionReturn(0); 615dd5b3ca6SJunchao Zhang } 616dd5b3ca6SJunchao Zhang 617dd5b3ca6SJunchao Zhang /*@ 618cab54364SBarry Smith PetscSFCreateInverseSF - given a `PetscSF` in which all vertices have degree 1, creates the inverse map 61995fce210SBarry Smith 62095fce210SBarry Smith Collective 62195fce210SBarry Smith 6224165533cSJose E. Roman Input Parameter: 62395fce210SBarry Smith . sf - star forest to invert 62495fce210SBarry Smith 6254165533cSJose E. Roman Output Parameter: 62695fce210SBarry Smith . isf - inverse of sf 6274165533cSJose E. Roman 62895fce210SBarry Smith Level: advanced 62995fce210SBarry Smith 63095fce210SBarry Smith Notes: 63195fce210SBarry Smith All roots must have degree 1. 63295fce210SBarry Smith 63395fce210SBarry Smith The local space may be a permutation, but cannot be sparse. 63495fce210SBarry Smith 635cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFSetGraph()` 63695fce210SBarry Smith @*/ 637d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateInverseSF(PetscSF sf, PetscSF *isf) 638d71ae5a4SJacob Faibussowitsch { 63995fce210SBarry Smith PetscMPIInt rank; 64095fce210SBarry Smith PetscInt i, nroots, nleaves, maxlocal, count, *newilocal; 64195fce210SBarry Smith const PetscInt *ilocal; 64295fce210SBarry Smith PetscSFNode *roots, *leaves; 64395fce210SBarry Smith 64495fce210SBarry Smith PetscFunctionBegin; 64529046d53SLisandro Dalcin PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 64629046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 64729046d53SLisandro Dalcin PetscValidPointer(isf, 2); 64829046d53SLisandro Dalcin 6499566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, NULL)); 65029046d53SLisandro Dalcin maxlocal = sf->maxleaf + 1; /* TODO: We should use PetscSFGetLeafRange() */ 65129046d53SLisandro Dalcin 6529566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank)); 6539566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(nroots, &roots, maxlocal, &leaves)); 654ae9aee6dSMatthew G. Knepley for (i = 0; i < maxlocal; i++) { 65595fce210SBarry Smith leaves[i].rank = rank; 65695fce210SBarry Smith leaves[i].index = i; 65795fce210SBarry Smith } 65895fce210SBarry Smith for (i = 0; i < nroots; i++) { 65995fce210SBarry Smith roots[i].rank = -1; 66095fce210SBarry Smith roots[i].index = -1; 66195fce210SBarry Smith } 6629566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(sf, MPIU_2INT, leaves, roots, MPI_REPLACE)); 6639566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(sf, MPIU_2INT, leaves, roots, MPI_REPLACE)); 66495fce210SBarry Smith 66595fce210SBarry Smith /* Check whether our leaves are sparse */ 6669371c9d4SSatish Balay for (i = 0, count = 0; i < nroots; i++) 6679371c9d4SSatish Balay if (roots[i].rank >= 0) count++; 66895fce210SBarry Smith if (count == nroots) newilocal = NULL; 6699371c9d4SSatish Balay else { /* Index for sparse leaves and compact "roots" array (which is to become our leaves). */ PetscCall(PetscMalloc1(count, &newilocal)); 67095fce210SBarry Smith for (i = 0, count = 0; i < nroots; i++) { 67195fce210SBarry Smith if (roots[i].rank >= 0) { 67295fce210SBarry Smith newilocal[count] = i; 67395fce210SBarry Smith roots[count].rank = roots[i].rank; 67495fce210SBarry Smith roots[count].index = roots[i].index; 67595fce210SBarry Smith count++; 67695fce210SBarry Smith } 67795fce210SBarry Smith } 67895fce210SBarry Smith } 67995fce210SBarry Smith 6809566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_CONFONLY, isf)); 6819566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*isf, maxlocal, count, newilocal, PETSC_OWN_POINTER, roots, PETSC_COPY_VALUES)); 6829566063dSJacob Faibussowitsch PetscCall(PetscFree2(roots, leaves)); 68395fce210SBarry Smith PetscFunctionReturn(0); 68495fce210SBarry Smith } 68595fce210SBarry Smith 68695fce210SBarry Smith /*@ 687cab54364SBarry Smith PetscSFDuplicate - duplicate a `PetscSF`, optionally preserving rank connectivity and graph 68895fce210SBarry Smith 68995fce210SBarry Smith Collective 69095fce210SBarry Smith 6914165533cSJose E. Roman Input Parameters: 69295fce210SBarry Smith + sf - communication object to duplicate 693cab54364SBarry Smith - opt - `PETSCSF_DUPLICATE_CONFONLY`, `PETSCSF_DUPLICATE_RANKS`, or `PETSCSF_DUPLICATE_GRAPH` (see `PetscSFDuplicateOption`) 69495fce210SBarry Smith 6954165533cSJose E. Roman Output Parameter: 69695fce210SBarry Smith . newsf - new communication object 69795fce210SBarry Smith 69895fce210SBarry Smith Level: beginner 69995fce210SBarry Smith 700cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFSetType()`, `PetscSFSetGraph()` 70195fce210SBarry Smith @*/ 702d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFDuplicate(PetscSF sf, PetscSFDuplicateOption opt, PetscSF *newsf) 703d71ae5a4SJacob Faibussowitsch { 70429046d53SLisandro Dalcin PetscSFType type; 70597929ea7SJunchao Zhang MPI_Datatype dtype = MPIU_SCALAR; 70695fce210SBarry Smith 70795fce210SBarry Smith PetscFunctionBegin; 70829046d53SLisandro Dalcin PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 70929046d53SLisandro Dalcin PetscValidLogicalCollectiveEnum(sf, opt, 2); 71029046d53SLisandro Dalcin PetscValidPointer(newsf, 3); 7119566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sf), newsf)); 7129566063dSJacob Faibussowitsch PetscCall(PetscSFGetType(sf, &type)); 7139566063dSJacob Faibussowitsch if (type) PetscCall(PetscSFSetType(*newsf, type)); 71454a24607SJunchao Zhang (*newsf)->allow_multi_leaves = sf->allow_multi_leaves; /* Dup this flag ealier since PetscSFSetGraph() below checks on this flag */ 71595fce210SBarry Smith if (opt == PETSCSF_DUPLICATE_GRAPH) { 716dd5b3ca6SJunchao Zhang PetscSFCheckGraphSet(sf, 1); 717dd5b3ca6SJunchao Zhang if (sf->pattern == PETSCSF_PATTERN_GENERAL) { 71895fce210SBarry Smith PetscInt nroots, nleaves; 71995fce210SBarry Smith const PetscInt *ilocal; 72095fce210SBarry Smith const PetscSFNode *iremote; 7219566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, &iremote)); 7229566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*newsf, nroots, nleaves, (PetscInt *)ilocal, PETSC_COPY_VALUES, (PetscSFNode *)iremote, PETSC_COPY_VALUES)); 723dd5b3ca6SJunchao Zhang } else { 7249566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraphWithPattern(*newsf, sf->map, sf->pattern)); 725dd5b3ca6SJunchao Zhang } 72695fce210SBarry Smith } 72797929ea7SJunchao Zhang /* Since oldtype is committed, so is newtype, according to MPI */ 7289566063dSJacob Faibussowitsch if (sf->vscat.bs > 1) PetscCallMPI(MPI_Type_dup(sf->vscat.unit, &dtype)); 72997929ea7SJunchao Zhang (*newsf)->vscat.bs = sf->vscat.bs; 73097929ea7SJunchao Zhang (*newsf)->vscat.unit = dtype; 73197929ea7SJunchao Zhang (*newsf)->vscat.to_n = sf->vscat.to_n; 73297929ea7SJunchao Zhang (*newsf)->vscat.from_n = sf->vscat.from_n; 73397929ea7SJunchao Zhang /* Do not copy lsf. Build it on demand since it is rarely used */ 73497929ea7SJunchao Zhang 73520c24465SJunchao Zhang #if defined(PETSC_HAVE_DEVICE) 73620c24465SJunchao Zhang (*newsf)->backend = sf->backend; 73771438e86SJunchao Zhang (*newsf)->unknown_input_stream = sf->unknown_input_stream; 73820c24465SJunchao Zhang (*newsf)->use_gpu_aware_mpi = sf->use_gpu_aware_mpi; 73920c24465SJunchao Zhang (*newsf)->use_stream_aware_mpi = sf->use_stream_aware_mpi; 74020c24465SJunchao Zhang #endif 741dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, Duplicate, opt, *newsf); 74220c24465SJunchao Zhang /* Don't do PetscSFSetUp() since the new sf's graph might have not been set. */ 74395fce210SBarry Smith PetscFunctionReturn(0); 74495fce210SBarry Smith } 74595fce210SBarry Smith 74695fce210SBarry Smith /*@C 74795fce210SBarry Smith PetscSFGetGraph - Get the graph specifying a parallel star forest 74895fce210SBarry Smith 74995fce210SBarry Smith Not Collective 75095fce210SBarry Smith 7514165533cSJose E. Roman Input Parameter: 75295fce210SBarry Smith . sf - star forest 75395fce210SBarry Smith 7544165533cSJose E. Roman Output Parameters: 75595fce210SBarry Smith + nroots - number of root vertices on the current process (these are possible targets for other process to attach leaves) 75695fce210SBarry Smith . nleaves - number of leaf vertices on the current process, each of these references a root on any process 757bc6585dcSJunchao Zhang . ilocal - locations of leaves in leafdata buffers (if returned value is NULL, it means leaves are in contiguous storage) 75895fce210SBarry Smith - iremote - remote locations of root vertices for each leaf on the current process 75995fce210SBarry Smith 760cab54364SBarry Smith Level: intermediate 761cab54364SBarry Smith 762373e0d91SLisandro Dalcin Notes: 763373e0d91SLisandro Dalcin We are not currently requiring that the graph is set, thus returning nroots=-1 if it has not been set yet 764373e0d91SLisandro Dalcin 765cab54364SBarry Smith The returned ilocal/iremote might contain values in different order than the input ones in `PetscSFSetGraph()`, 766db2b9530SVaclav Hapla see its manpage for details. 767db2b9530SVaclav Hapla 7688dbb0df6SBarry Smith Fortran Notes: 7698dbb0df6SBarry Smith The returned iremote array is a copy and must be deallocated after use. Consequently, if you 770cab54364SBarry Smith want to update the graph, you must call `PetscSFSetGraph()` after modifying the iremote array. 7718dbb0df6SBarry Smith 7728dbb0df6SBarry Smith To check for a NULL ilocal use 7738dbb0df6SBarry Smith $ if (loc(ilocal) == loc(PETSC_NULL_INTEGER)) then 774ca797d7aSLawrence Mitchell 775cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFSetGraph()` 77695fce210SBarry Smith @*/ 777d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetGraph(PetscSF sf, PetscInt *nroots, PetscInt *nleaves, const PetscInt **ilocal, const PetscSFNode **iremote) 778d71ae5a4SJacob Faibussowitsch { 77995fce210SBarry Smith PetscFunctionBegin; 78095fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 781b8dee149SJunchao Zhang if (sf->ops->GetGraph) { 7829566063dSJacob Faibussowitsch PetscCall((sf->ops->GetGraph)(sf, nroots, nleaves, ilocal, iremote)); 783b8dee149SJunchao Zhang } else { 78495fce210SBarry Smith if (nroots) *nroots = sf->nroots; 78595fce210SBarry Smith if (nleaves) *nleaves = sf->nleaves; 78695fce210SBarry Smith if (ilocal) *ilocal = sf->mine; 78795fce210SBarry Smith if (iremote) *iremote = sf->remote; 788b8dee149SJunchao Zhang } 78995fce210SBarry Smith PetscFunctionReturn(0); 79095fce210SBarry Smith } 79195fce210SBarry Smith 79229046d53SLisandro Dalcin /*@ 79395fce210SBarry Smith PetscSFGetLeafRange - Get the active leaf ranges 79495fce210SBarry Smith 79595fce210SBarry Smith Not Collective 79695fce210SBarry Smith 7974165533cSJose E. Roman Input Parameter: 79895fce210SBarry Smith . sf - star forest 79995fce210SBarry Smith 8004165533cSJose E. Roman Output Parameters: 801dd5b3ca6SJunchao Zhang + minleaf - minimum active leaf on this process. Return 0 if there are no leaves. 802dd5b3ca6SJunchao Zhang - maxleaf - maximum active leaf on this process. Return -1 if there are no leaves. 80395fce210SBarry Smith 80495fce210SBarry Smith Level: developer 80595fce210SBarry Smith 806cab54364SBarry Smith .seealso: `PetscSFType`, `PetscSFCreate()`, `PetscSFView()`, `PetscSFSetGraph()`, `PetscSFGetGraph()` 80795fce210SBarry Smith @*/ 808d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetLeafRange(PetscSF sf, PetscInt *minleaf, PetscInt *maxleaf) 809d71ae5a4SJacob Faibussowitsch { 81095fce210SBarry Smith PetscFunctionBegin; 81195fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 81229046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 81395fce210SBarry Smith if (minleaf) *minleaf = sf->minleaf; 81495fce210SBarry Smith if (maxleaf) *maxleaf = sf->maxleaf; 81595fce210SBarry Smith PetscFunctionReturn(0); 81695fce210SBarry Smith } 81795fce210SBarry Smith 81895fce210SBarry Smith /*@C 819cab54364SBarry Smith PetscSFViewFromOptions - View a `PetscSF` based on arguments in the options database 820fe2efc57SMark 821cab54364SBarry Smith Collective on A 822fe2efc57SMark 823fe2efc57SMark Input Parameters: 824fe2efc57SMark + A - the star forest 825cab54364SBarry Smith . obj - Optional object that provides the prefix for the option names 826736c3998SJose E. Roman - name - command line option 827fe2efc57SMark 828fe2efc57SMark Level: intermediate 829cab54364SBarry Smith 830db781477SPatrick Sanan .seealso: `PetscSF`, `PetscSFView`, `PetscObjectViewFromOptions()`, `PetscSFCreate()` 831fe2efc57SMark @*/ 832d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFViewFromOptions(PetscSF A, PetscObject obj, const char name[]) 833d71ae5a4SJacob Faibussowitsch { 834fe2efc57SMark PetscFunctionBegin; 835fe2efc57SMark PetscValidHeaderSpecific(A, PETSCSF_CLASSID, 1); 8369566063dSJacob Faibussowitsch PetscCall(PetscObjectViewFromOptions((PetscObject)A, obj, name)); 837fe2efc57SMark PetscFunctionReturn(0); 838fe2efc57SMark } 839fe2efc57SMark 840fe2efc57SMark /*@C 84195fce210SBarry Smith PetscSFView - view a star forest 84295fce210SBarry Smith 84395fce210SBarry Smith Collective 84495fce210SBarry Smith 8454165533cSJose E. Roman Input Parameters: 84695fce210SBarry Smith + sf - star forest 847cab54364SBarry Smith - viewer - viewer to display graph, for example `PETSC_VIEWER_STDOUT_WORLD` 84895fce210SBarry Smith 84995fce210SBarry Smith Level: beginner 85095fce210SBarry Smith 851cab54364SBarry Smith .seealso: `PetscSF`, `PetscViewer`, `PetscSFCreate()`, `PetscSFSetGraph()` 85295fce210SBarry Smith @*/ 853d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFView(PetscSF sf, PetscViewer viewer) 854d71ae5a4SJacob Faibussowitsch { 85595fce210SBarry Smith PetscBool iascii; 85695fce210SBarry Smith PetscViewerFormat format; 85795fce210SBarry Smith 85895fce210SBarry Smith PetscFunctionBegin; 85995fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 8609566063dSJacob Faibussowitsch if (!viewer) PetscCall(PetscViewerASCIIGetStdout(PetscObjectComm((PetscObject)sf), &viewer)); 86195fce210SBarry Smith PetscValidHeaderSpecific(viewer, PETSC_VIEWER_CLASSID, 2); 86295fce210SBarry Smith PetscCheckSameComm(sf, 1, viewer, 2); 8639566063dSJacob Faibussowitsch if (sf->graphset) PetscCall(PetscSFSetUp(sf)); 8649566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &iascii)); 86553dd6d7dSJunchao Zhang if (iascii && viewer->format != PETSC_VIEWER_ASCII_MATLAB) { 86695fce210SBarry Smith PetscMPIInt rank; 86781bfa7aaSJed Brown PetscInt ii, i, j; 86895fce210SBarry Smith 8699566063dSJacob Faibussowitsch PetscCall(PetscObjectPrintClassNamePrefixType((PetscObject)sf, viewer)); 8709566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPushTab(viewer)); 871dd5b3ca6SJunchao Zhang if (sf->pattern == PETSCSF_PATTERN_GENERAL) { 87280153354SVaclav Hapla if (!sf->graphset) { 8739566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPrintf(viewer, "PetscSFSetGraph() has not been called yet\n")); 8749566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPopTab(viewer)); 87580153354SVaclav Hapla PetscFunctionReturn(0); 87680153354SVaclav Hapla } 8779566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank)); 8789566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPushSynchronized(viewer)); 8799566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Number of roots=%" PetscInt_FMT ", leaves=%" PetscInt_FMT ", remote ranks=%" PetscInt_FMT "\n", rank, sf->nroots, sf->nleaves, sf->nranks)); 88048a46eb9SPierre Jolivet for (i = 0; i < sf->nleaves; i++) PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] %" PetscInt_FMT " <- (%" PetscInt_FMT ",%" PetscInt_FMT ")\n", rank, sf->mine ? sf->mine[i] : i, sf->remote[i].rank, sf->remote[i].index)); 8819566063dSJacob Faibussowitsch PetscCall(PetscViewerFlush(viewer)); 8829566063dSJacob Faibussowitsch PetscCall(PetscViewerGetFormat(viewer, &format)); 88395fce210SBarry Smith if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 88481bfa7aaSJed Brown PetscMPIInt *tmpranks, *perm; 8859566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(sf->nranks, &tmpranks, sf->nranks, &perm)); 8869566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(tmpranks, sf->ranks, sf->nranks)); 88781bfa7aaSJed Brown for (i = 0; i < sf->nranks; i++) perm[i] = i; 8889566063dSJacob Faibussowitsch PetscCall(PetscSortMPIIntWithArray(sf->nranks, tmpranks, perm)); 8899566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] Roots referenced by my leaves, by rank\n", rank)); 89081bfa7aaSJed Brown for (ii = 0; ii < sf->nranks; ii++) { 89181bfa7aaSJed Brown i = perm[ii]; 8929566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] %d: %" PetscInt_FMT " edges\n", rank, sf->ranks[i], sf->roffset[i + 1] - sf->roffset[i])); 89348a46eb9SPierre Jolivet for (j = sf->roffset[i]; j < sf->roffset[i + 1]; j++) PetscCall(PetscViewerASCIISynchronizedPrintf(viewer, "[%d] %" PetscInt_FMT " <- %" PetscInt_FMT "\n", rank, sf->rmine[j], sf->rremote[j])); 89495fce210SBarry Smith } 8959566063dSJacob Faibussowitsch PetscCall(PetscFree2(tmpranks, perm)); 89695fce210SBarry Smith } 8979566063dSJacob Faibussowitsch PetscCall(PetscViewerFlush(viewer)); 8989566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPopSynchronized(viewer)); 899dd5b3ca6SJunchao Zhang } 9009566063dSJacob Faibussowitsch PetscCall(PetscViewerASCIIPopTab(viewer)); 90195fce210SBarry Smith } 902dbbe0bcdSBarry Smith PetscTryTypeMethod(sf, View, viewer); 90395fce210SBarry Smith PetscFunctionReturn(0); 90495fce210SBarry Smith } 90595fce210SBarry Smith 90695fce210SBarry Smith /*@C 907dec1416fSJunchao Zhang PetscSFGetRootRanks - Get root ranks and number of vertices referenced by leaves on this process 90895fce210SBarry Smith 90995fce210SBarry Smith Not Collective 91095fce210SBarry Smith 9114165533cSJose E. Roman Input Parameter: 91295fce210SBarry Smith . sf - star forest 91395fce210SBarry Smith 9144165533cSJose E. Roman Output Parameters: 91595fce210SBarry Smith + nranks - number of ranks referenced by local part 91695fce210SBarry Smith . ranks - array of ranks 91795fce210SBarry Smith . roffset - offset in rmine/rremote for each rank (length nranks+1) 91895fce210SBarry Smith . rmine - concatenated array holding local indices referencing each remote rank 91995fce210SBarry Smith - rremote - concatenated array holding remote indices referenced for each remote rank 92095fce210SBarry Smith 92195fce210SBarry Smith Level: developer 92295fce210SBarry Smith 923cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetLeafRanks()` 92495fce210SBarry Smith @*/ 925d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetRootRanks(PetscSF sf, PetscInt *nranks, const PetscMPIInt **ranks, const PetscInt **roffset, const PetscInt **rmine, const PetscInt **rremote) 926d71ae5a4SJacob Faibussowitsch { 92795fce210SBarry Smith PetscFunctionBegin; 92895fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 92928b400f6SJacob Faibussowitsch PetscCheck(sf->setupcalled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetUp() before obtaining ranks"); 930dec1416fSJunchao Zhang if (sf->ops->GetRootRanks) { 9319566063dSJacob Faibussowitsch PetscCall((sf->ops->GetRootRanks)(sf, nranks, ranks, roffset, rmine, rremote)); 932dec1416fSJunchao Zhang } else { 933dec1416fSJunchao Zhang /* The generic implementation */ 93495fce210SBarry Smith if (nranks) *nranks = sf->nranks; 93595fce210SBarry Smith if (ranks) *ranks = sf->ranks; 93695fce210SBarry Smith if (roffset) *roffset = sf->roffset; 93795fce210SBarry Smith if (rmine) *rmine = sf->rmine; 93895fce210SBarry Smith if (rremote) *rremote = sf->rremote; 939dec1416fSJunchao Zhang } 94095fce210SBarry Smith PetscFunctionReturn(0); 94195fce210SBarry Smith } 94295fce210SBarry Smith 9438750ddebSJunchao Zhang /*@C 9448750ddebSJunchao Zhang PetscSFGetLeafRanks - Get leaf ranks referencing roots on this process 9458750ddebSJunchao Zhang 9468750ddebSJunchao Zhang Not Collective 9478750ddebSJunchao Zhang 9484165533cSJose E. Roman Input Parameter: 9498750ddebSJunchao Zhang . sf - star forest 9508750ddebSJunchao Zhang 9514165533cSJose E. Roman Output Parameters: 9528750ddebSJunchao Zhang + niranks - number of leaf ranks referencing roots on this process 9538750ddebSJunchao Zhang . iranks - array of ranks 9548750ddebSJunchao Zhang . ioffset - offset in irootloc for each rank (length niranks+1) 9558750ddebSJunchao Zhang - irootloc - concatenated array holding local indices of roots referenced by each leaf rank 9568750ddebSJunchao Zhang 9578750ddebSJunchao Zhang Level: developer 9588750ddebSJunchao Zhang 959cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetRootRanks()` 9608750ddebSJunchao Zhang @*/ 961d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetLeafRanks(PetscSF sf, PetscInt *niranks, const PetscMPIInt **iranks, const PetscInt **ioffset, const PetscInt **irootloc) 962d71ae5a4SJacob Faibussowitsch { 9638750ddebSJunchao Zhang PetscFunctionBegin; 9648750ddebSJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 96528b400f6SJacob Faibussowitsch PetscCheck(sf->setupcalled, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetUp() before obtaining ranks"); 9668750ddebSJunchao Zhang if (sf->ops->GetLeafRanks) { 9679566063dSJacob Faibussowitsch PetscCall((sf->ops->GetLeafRanks)(sf, niranks, iranks, ioffset, irootloc)); 9688750ddebSJunchao Zhang } else { 9698750ddebSJunchao Zhang PetscSFType type; 9709566063dSJacob Faibussowitsch PetscCall(PetscSFGetType(sf, &type)); 97198921bdaSJacob Faibussowitsch SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "PetscSFGetLeafRanks() is not supported on this StarForest type: %s", type); 9728750ddebSJunchao Zhang } 9738750ddebSJunchao Zhang PetscFunctionReturn(0); 9748750ddebSJunchao Zhang } 9758750ddebSJunchao Zhang 976d71ae5a4SJacob Faibussowitsch static PetscBool InList(PetscMPIInt needle, PetscMPIInt n, const PetscMPIInt *list) 977d71ae5a4SJacob Faibussowitsch { 978b5a8e515SJed Brown PetscInt i; 979b5a8e515SJed Brown for (i = 0; i < n; i++) { 980b5a8e515SJed Brown if (needle == list[i]) return PETSC_TRUE; 981b5a8e515SJed Brown } 982b5a8e515SJed Brown return PETSC_FALSE; 983b5a8e515SJed Brown } 984b5a8e515SJed Brown 98595fce210SBarry Smith /*@C 986cab54364SBarry Smith PetscSFSetUpRanks - Set up data structures associated with ranks; this is for internal use by `PetscSF` implementations. 98721c688dcSJed Brown 98821c688dcSJed Brown Collective 98921c688dcSJed Brown 9904165533cSJose E. Roman Input Parameters: 991cab54364SBarry Smith + sf - `PetscSF` to set up; `PetscSFSetGraph()` must have been called 992cab54364SBarry Smith - dgroup - `MPI_Group` of ranks to be distinguished (e.g., for self or shared memory exchange) 99321c688dcSJed Brown 99421c688dcSJed Brown Level: developer 99521c688dcSJed Brown 996cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetRootRanks()` 99721c688dcSJed Brown @*/ 998d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFSetUpRanks(PetscSF sf, MPI_Group dgroup) 999d71ae5a4SJacob Faibussowitsch { 100021c688dcSJed Brown PetscTable table; 100121c688dcSJed Brown PetscTablePosition pos; 1002b5a8e515SJed Brown PetscMPIInt size, groupsize, *groupranks; 1003247e8311SStefano Zampini PetscInt *rcount, *ranks; 1004247e8311SStefano Zampini PetscInt i, irank = -1, orank = -1; 100521c688dcSJed Brown 100621c688dcSJed Brown PetscFunctionBegin; 100721c688dcSJed Brown PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 100829046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 10099566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)sf), &size)); 10109566063dSJacob Faibussowitsch PetscCall(PetscTableCreate(10, size, &table)); 101121c688dcSJed Brown for (i = 0; i < sf->nleaves; i++) { 101221c688dcSJed Brown /* Log 1-based rank */ 10139566063dSJacob Faibussowitsch PetscCall(PetscTableAdd(table, sf->remote[i].rank + 1, 1, ADD_VALUES)); 101421c688dcSJed Brown } 10159566063dSJacob Faibussowitsch PetscCall(PetscTableGetCount(table, &sf->nranks)); 10169566063dSJacob Faibussowitsch PetscCall(PetscMalloc4(sf->nranks, &sf->ranks, sf->nranks + 1, &sf->roffset, sf->nleaves, &sf->rmine, sf->nleaves, &sf->rremote)); 10179566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(sf->nranks, &rcount, sf->nranks, &ranks)); 10189566063dSJacob Faibussowitsch PetscCall(PetscTableGetHeadPosition(table, &pos)); 101921c688dcSJed Brown for (i = 0; i < sf->nranks; i++) { 10209566063dSJacob Faibussowitsch PetscCall(PetscTableGetNext(table, &pos, &ranks[i], &rcount[i])); 102121c688dcSJed Brown ranks[i]--; /* Convert back to 0-based */ 102221c688dcSJed Brown } 10239566063dSJacob Faibussowitsch PetscCall(PetscTableDestroy(&table)); 1024b5a8e515SJed Brown 1025b5a8e515SJed Brown /* We expect that dgroup is reliably "small" while nranks could be large */ 1026b5a8e515SJed Brown { 10277fb8a5e4SKarl Rupp MPI_Group group = MPI_GROUP_NULL; 1028b5a8e515SJed Brown PetscMPIInt *dgroupranks; 10299566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf), &group)); 10309566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_size(dgroup, &groupsize)); 10319566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(groupsize, &dgroupranks)); 10329566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(groupsize, &groupranks)); 1033b5a8e515SJed Brown for (i = 0; i < groupsize; i++) dgroupranks[i] = i; 10349566063dSJacob Faibussowitsch if (groupsize) PetscCallMPI(MPI_Group_translate_ranks(dgroup, groupsize, dgroupranks, group, groupranks)); 10359566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_free(&group)); 10369566063dSJacob Faibussowitsch PetscCall(PetscFree(dgroupranks)); 1037b5a8e515SJed Brown } 1038b5a8e515SJed Brown 1039b5a8e515SJed Brown /* Partition ranks[] into distinguished (first sf->ndranks) followed by non-distinguished */ 1040b5a8e515SJed Brown for (sf->ndranks = 0, i = sf->nranks; sf->ndranks < i;) { 1041b5a8e515SJed Brown for (i--; sf->ndranks < i; i--) { /* Scan i backward looking for distinguished rank */ 1042b5a8e515SJed Brown if (InList(ranks[i], groupsize, groupranks)) break; 1043b5a8e515SJed Brown } 1044b5a8e515SJed Brown for (; sf->ndranks <= i; sf->ndranks++) { /* Scan sf->ndranks forward looking for non-distinguished rank */ 1045b5a8e515SJed Brown if (!InList(ranks[sf->ndranks], groupsize, groupranks)) break; 1046b5a8e515SJed Brown } 1047b5a8e515SJed Brown if (sf->ndranks < i) { /* Swap ranks[sf->ndranks] with ranks[i] */ 1048b5a8e515SJed Brown PetscInt tmprank, tmpcount; 1049247e8311SStefano Zampini 1050b5a8e515SJed Brown tmprank = ranks[i]; 1051b5a8e515SJed Brown tmpcount = rcount[i]; 1052b5a8e515SJed Brown ranks[i] = ranks[sf->ndranks]; 1053b5a8e515SJed Brown rcount[i] = rcount[sf->ndranks]; 1054b5a8e515SJed Brown ranks[sf->ndranks] = tmprank; 1055b5a8e515SJed Brown rcount[sf->ndranks] = tmpcount; 1056b5a8e515SJed Brown sf->ndranks++; 1057b5a8e515SJed Brown } 1058b5a8e515SJed Brown } 10599566063dSJacob Faibussowitsch PetscCall(PetscFree(groupranks)); 10609566063dSJacob Faibussowitsch PetscCall(PetscSortIntWithArray(sf->ndranks, ranks, rcount)); 10619566063dSJacob Faibussowitsch PetscCall(PetscSortIntWithArray(sf->nranks - sf->ndranks, ranks + sf->ndranks, rcount + sf->ndranks)); 106221c688dcSJed Brown sf->roffset[0] = 0; 106321c688dcSJed Brown for (i = 0; i < sf->nranks; i++) { 10649566063dSJacob Faibussowitsch PetscCall(PetscMPIIntCast(ranks[i], sf->ranks + i)); 106521c688dcSJed Brown sf->roffset[i + 1] = sf->roffset[i] + rcount[i]; 106621c688dcSJed Brown rcount[i] = 0; 106721c688dcSJed Brown } 1068247e8311SStefano Zampini for (i = 0, irank = -1, orank = -1; i < sf->nleaves; i++) { 1069247e8311SStefano Zampini /* short circuit */ 1070247e8311SStefano Zampini if (orank != sf->remote[i].rank) { 107121c688dcSJed Brown /* Search for index of iremote[i].rank in sf->ranks */ 10729566063dSJacob Faibussowitsch PetscCall(PetscFindMPIInt(sf->remote[i].rank, sf->ndranks, sf->ranks, &irank)); 1073b5a8e515SJed Brown if (irank < 0) { 10749566063dSJacob Faibussowitsch PetscCall(PetscFindMPIInt(sf->remote[i].rank, sf->nranks - sf->ndranks, sf->ranks + sf->ndranks, &irank)); 1075b5a8e515SJed Brown if (irank >= 0) irank += sf->ndranks; 107621c688dcSJed Brown } 1077247e8311SStefano Zampini orank = sf->remote[i].rank; 1078247e8311SStefano Zampini } 107908401ef6SPierre Jolivet PetscCheck(irank >= 0, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Could not find rank %" PetscInt_FMT " in array", sf->remote[i].rank); 108021c688dcSJed Brown sf->rmine[sf->roffset[irank] + rcount[irank]] = sf->mine ? sf->mine[i] : i; 108121c688dcSJed Brown sf->rremote[sf->roffset[irank] + rcount[irank]] = sf->remote[i].index; 108221c688dcSJed Brown rcount[irank]++; 108321c688dcSJed Brown } 10849566063dSJacob Faibussowitsch PetscCall(PetscFree2(rcount, ranks)); 108521c688dcSJed Brown PetscFunctionReturn(0); 108621c688dcSJed Brown } 108721c688dcSJed Brown 108821c688dcSJed Brown /*@C 108995fce210SBarry Smith PetscSFGetGroups - gets incoming and outgoing process groups 109095fce210SBarry Smith 109195fce210SBarry Smith Collective 109295fce210SBarry Smith 10934165533cSJose E. Roman Input Parameter: 109495fce210SBarry Smith . sf - star forest 109595fce210SBarry Smith 10964165533cSJose E. Roman Output Parameters: 109795fce210SBarry Smith + incoming - group of origin processes for incoming edges (leaves that reference my roots) 109895fce210SBarry Smith - outgoing - group of destination processes for outgoing edges (roots that I reference) 109995fce210SBarry Smith 110095fce210SBarry Smith Level: developer 110195fce210SBarry Smith 1102cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGetWindow()`, `PetscSFRestoreWindow()` 110395fce210SBarry Smith @*/ 1104d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetGroups(PetscSF sf, MPI_Group *incoming, MPI_Group *outgoing) 1105d71ae5a4SJacob Faibussowitsch { 11067fb8a5e4SKarl Rupp MPI_Group group = MPI_GROUP_NULL; 110795fce210SBarry Smith 110895fce210SBarry Smith PetscFunctionBegin; 110908401ef6SPierre Jolivet PetscCheck(sf->nranks >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFSetUpRanks() before obtaining groups"); 111095fce210SBarry Smith if (sf->ingroup == MPI_GROUP_NULL) { 111195fce210SBarry Smith PetscInt i; 111295fce210SBarry Smith const PetscInt *indegree; 111395fce210SBarry Smith PetscMPIInt rank, *outranks, *inranks; 111495fce210SBarry Smith PetscSFNode *remote; 111595fce210SBarry Smith PetscSF bgcount; 111695fce210SBarry Smith 111795fce210SBarry Smith /* Compute the number of incoming ranks */ 11189566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(sf->nranks, &remote)); 111995fce210SBarry Smith for (i = 0; i < sf->nranks; i++) { 112095fce210SBarry Smith remote[i].rank = sf->ranks[i]; 112195fce210SBarry Smith remote[i].index = 0; 112295fce210SBarry Smith } 11239566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_CONFONLY, &bgcount)); 11249566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(bgcount, 1, sf->nranks, NULL, PETSC_COPY_VALUES, remote, PETSC_OWN_POINTER)); 11259566063dSJacob Faibussowitsch PetscCall(PetscSFComputeDegreeBegin(bgcount, &indegree)); 11269566063dSJacob Faibussowitsch PetscCall(PetscSFComputeDegreeEnd(bgcount, &indegree)); 112795fce210SBarry Smith /* Enumerate the incoming ranks */ 11289566063dSJacob Faibussowitsch PetscCall(PetscMalloc2(indegree[0], &inranks, sf->nranks, &outranks)); 11299566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank)); 113095fce210SBarry Smith for (i = 0; i < sf->nranks; i++) outranks[i] = rank; 11319566063dSJacob Faibussowitsch PetscCall(PetscSFGatherBegin(bgcount, MPI_INT, outranks, inranks)); 11329566063dSJacob Faibussowitsch PetscCall(PetscSFGatherEnd(bgcount, MPI_INT, outranks, inranks)); 11339566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf), &group)); 11349566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_incl(group, indegree[0], inranks, &sf->ingroup)); 11359566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_free(&group)); 11369566063dSJacob Faibussowitsch PetscCall(PetscFree2(inranks, outranks)); 11379566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&bgcount)); 113895fce210SBarry Smith } 113995fce210SBarry Smith *incoming = sf->ingroup; 114095fce210SBarry Smith 114195fce210SBarry Smith if (sf->outgroup == MPI_GROUP_NULL) { 11429566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_group(PetscObjectComm((PetscObject)sf), &group)); 11439566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_incl(group, sf->nranks, sf->ranks, &sf->outgroup)); 11449566063dSJacob Faibussowitsch PetscCallMPI(MPI_Group_free(&group)); 114595fce210SBarry Smith } 114695fce210SBarry Smith *outgoing = sf->outgroup; 114795fce210SBarry Smith PetscFunctionReturn(0); 114895fce210SBarry Smith } 114995fce210SBarry Smith 115029046d53SLisandro Dalcin /*@ 1151cab54364SBarry Smith PetscSFGetMultiSF - gets the inner `PetscSF` implementing gathers and scatters 115295fce210SBarry Smith 115395fce210SBarry Smith Collective 115495fce210SBarry Smith 11554165533cSJose E. Roman Input Parameter: 115695fce210SBarry Smith . sf - star forest that may contain roots with 0 or with more than 1 vertex 115795fce210SBarry Smith 11584165533cSJose E. Roman Output Parameter: 115995fce210SBarry Smith . multi - star forest with split roots, such that each root has degree exactly 1 116095fce210SBarry Smith 116195fce210SBarry Smith Level: developer 116295fce210SBarry Smith 1163cab54364SBarry Smith Note: 1164cab54364SBarry Smith In most cases, users should use `PetscSFGatherBegin()` and `PetscSFScatterBegin()` instead of manipulating multi 116595fce210SBarry Smith directly. Since multi satisfies the stronger condition that each entry in the global space has exactly one incoming 116695fce210SBarry Smith edge, it is a candidate for future optimization that might involve its removal. 116795fce210SBarry Smith 1168cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFGatherBegin()`, `PetscSFScatterBegin()`, `PetscSFComputeMultiRootOriginalNumbering()` 116995fce210SBarry Smith @*/ 1170d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGetMultiSF(PetscSF sf, PetscSF *multi) 1171d71ae5a4SJacob Faibussowitsch { 117295fce210SBarry Smith PetscFunctionBegin; 117395fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 117495fce210SBarry Smith PetscValidPointer(multi, 2); 117595fce210SBarry Smith if (sf->nroots < 0) { /* Graph has not been set yet; why do we need this? */ 11769566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_RANKS, &sf->multi)); 117795fce210SBarry Smith *multi = sf->multi; 1178013b3241SStefano Zampini sf->multi->multi = sf->multi; 117995fce210SBarry Smith PetscFunctionReturn(0); 118095fce210SBarry Smith } 118195fce210SBarry Smith if (!sf->multi) { 118295fce210SBarry Smith const PetscInt *indegree; 11839837ea96SMatthew G. Knepley PetscInt i, *inoffset, *outones, *outoffset, maxlocal; 118495fce210SBarry Smith PetscSFNode *remote; 118529046d53SLisandro Dalcin maxlocal = sf->maxleaf + 1; /* TODO: We should use PetscSFGetLeafRange() */ 11869566063dSJacob Faibussowitsch PetscCall(PetscSFComputeDegreeBegin(sf, &indegree)); 11879566063dSJacob Faibussowitsch PetscCall(PetscSFComputeDegreeEnd(sf, &indegree)); 11889566063dSJacob Faibussowitsch PetscCall(PetscMalloc3(sf->nroots + 1, &inoffset, maxlocal, &outones, maxlocal, &outoffset)); 118995fce210SBarry Smith inoffset[0] = 0; 119095fce210SBarry Smith for (i = 0; i < sf->nroots; i++) inoffset[i + 1] = inoffset[i] + indegree[i]; 11919837ea96SMatthew G. Knepley for (i = 0; i < maxlocal; i++) outones[i] = 1; 11929566063dSJacob Faibussowitsch PetscCall(PetscSFFetchAndOpBegin(sf, MPIU_INT, inoffset, outones, outoffset, MPI_SUM)); 11939566063dSJacob Faibussowitsch PetscCall(PetscSFFetchAndOpEnd(sf, MPIU_INT, inoffset, outones, outoffset, MPI_SUM)); 119495fce210SBarry Smith for (i = 0; i < sf->nroots; i++) inoffset[i] -= indegree[i]; /* Undo the increment */ 119576bd3646SJed Brown if (PetscDefined(USE_DEBUG)) { /* Check that the expected number of increments occurred */ 1196ad540459SPierre 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"); 119776bd3646SJed Brown } 11989566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(sf->nleaves, &remote)); 119995fce210SBarry Smith for (i = 0; i < sf->nleaves; i++) { 120095fce210SBarry Smith remote[i].rank = sf->remote[i].rank; 120138e7336fSToby Isaac remote[i].index = outoffset[sf->mine ? sf->mine[i] : i]; 120295fce210SBarry Smith } 12039566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_RANKS, &sf->multi)); 1204013b3241SStefano Zampini sf->multi->multi = sf->multi; 12059566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(sf->multi, inoffset[sf->nroots], sf->nleaves, sf->mine, PETSC_COPY_VALUES, remote, PETSC_OWN_POINTER)); 120695fce210SBarry Smith if (sf->rankorder) { /* Sort the ranks */ 120795fce210SBarry Smith PetscMPIInt rank; 120895fce210SBarry Smith PetscInt *inranks, *newoffset, *outranks, *newoutoffset, *tmpoffset, maxdegree; 120995fce210SBarry Smith PetscSFNode *newremote; 12109566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)sf), &rank)); 121195fce210SBarry Smith for (i = 0, maxdegree = 0; i < sf->nroots; i++) maxdegree = PetscMax(maxdegree, indegree[i]); 12129566063dSJacob Faibussowitsch PetscCall(PetscMalloc5(sf->multi->nroots, &inranks, sf->multi->nroots, &newoffset, maxlocal, &outranks, maxlocal, &newoutoffset, maxdegree, &tmpoffset)); 12139837ea96SMatthew G. Knepley for (i = 0; i < maxlocal; i++) outranks[i] = rank; 12149566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(sf->multi, MPIU_INT, outranks, inranks, MPI_REPLACE)); 12159566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(sf->multi, MPIU_INT, outranks, inranks, MPI_REPLACE)); 121695fce210SBarry Smith /* Sort the incoming ranks at each vertex, build the inverse map */ 121795fce210SBarry Smith for (i = 0; i < sf->nroots; i++) { 121895fce210SBarry Smith PetscInt j; 121995fce210SBarry Smith for (j = 0; j < indegree[i]; j++) tmpoffset[j] = j; 12209566063dSJacob Faibussowitsch PetscCall(PetscSortIntWithArray(indegree[i], inranks + inoffset[i], tmpoffset)); 122195fce210SBarry Smith for (j = 0; j < indegree[i]; j++) newoffset[inoffset[i] + tmpoffset[j]] = inoffset[i] + j; 122295fce210SBarry Smith } 12239566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(sf->multi, MPIU_INT, newoffset, newoutoffset, MPI_REPLACE)); 12249566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(sf->multi, MPIU_INT, newoffset, newoutoffset, MPI_REPLACE)); 12259566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(sf->nleaves, &newremote)); 122695fce210SBarry Smith for (i = 0; i < sf->nleaves; i++) { 122795fce210SBarry Smith newremote[i].rank = sf->remote[i].rank; 122801365b40SToby Isaac newremote[i].index = newoutoffset[sf->mine ? sf->mine[i] : i]; 122995fce210SBarry Smith } 12309566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(sf->multi, inoffset[sf->nroots], sf->nleaves, sf->mine, PETSC_COPY_VALUES, newremote, PETSC_OWN_POINTER)); 12319566063dSJacob Faibussowitsch PetscCall(PetscFree5(inranks, newoffset, outranks, newoutoffset, tmpoffset)); 123295fce210SBarry Smith } 12339566063dSJacob Faibussowitsch PetscCall(PetscFree3(inoffset, outones, outoffset)); 123495fce210SBarry Smith } 123595fce210SBarry Smith *multi = sf->multi; 123695fce210SBarry Smith PetscFunctionReturn(0); 123795fce210SBarry Smith } 123895fce210SBarry Smith 123995fce210SBarry Smith /*@C 124072502a1fSJunchao Zhang PetscSFCreateEmbeddedRootSF - removes edges from all but the selected roots, does not remap indices 124195fce210SBarry Smith 124295fce210SBarry Smith Collective 124395fce210SBarry Smith 12444165533cSJose E. Roman Input Parameters: 124595fce210SBarry Smith + sf - original star forest 1246ba2a7774SJunchao Zhang . nselected - number of selected roots on this process 1247ba2a7774SJunchao Zhang - selected - indices of the selected roots on this process 124895fce210SBarry Smith 12494165533cSJose E. Roman Output Parameter: 1250cd620004SJunchao Zhang . esf - new star forest 125195fce210SBarry Smith 125295fce210SBarry Smith Level: advanced 125395fce210SBarry Smith 125495fce210SBarry Smith Note: 1255cab54364SBarry Smith To use the new `PetscSF`, it may be necessary to know the indices of the leaves that are still participating. This can 125695fce210SBarry Smith be done by calling PetscSFGetGraph(). 125795fce210SBarry Smith 1258cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFGetGraph()` 125995fce210SBarry Smith @*/ 1260d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateEmbeddedRootSF(PetscSF sf, PetscInt nselected, const PetscInt *selected, PetscSF *esf) 1261d71ae5a4SJacob Faibussowitsch { 1262cd620004SJunchao Zhang PetscInt i, j, n, nroots, nleaves, esf_nleaves, *new_ilocal, minleaf, maxleaf, maxlocal; 1263cd620004SJunchao Zhang const PetscInt *ilocal; 1264cd620004SJunchao Zhang signed char *rootdata, *leafdata, *leafmem; 1265ba2a7774SJunchao Zhang const PetscSFNode *iremote; 1266f659e5c7SJunchao Zhang PetscSFNode *new_iremote; 1267f659e5c7SJunchao Zhang MPI_Comm comm; 126895fce210SBarry Smith 126995fce210SBarry Smith PetscFunctionBegin; 127095fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 127129046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 1272dadcf809SJacob Faibussowitsch if (nselected) PetscValidIntPointer(selected, 3); 1273cd620004SJunchao Zhang PetscValidPointer(esf, 4); 12740511a646SMatthew G. Knepley 12759566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 12769566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_EmbedSF, sf, 0, 0, 0)); 12779566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 12789566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, &iremote)); 1279cd620004SJunchao Zhang 128076bd3646SJed Brown if (PetscDefined(USE_DEBUG)) { /* Error out if selected[] has dups or out of range indices */ 1281cd620004SJunchao Zhang PetscBool dups; 12829566063dSJacob Faibussowitsch PetscCall(PetscCheckDupsInt(nselected, selected, &dups)); 128328b400f6SJacob Faibussowitsch PetscCheck(!dups, comm, PETSC_ERR_ARG_WRONG, "selected[] has dups"); 12849371c9d4SSatish 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); 1285cd620004SJunchao Zhang } 1286f659e5c7SJunchao Zhang 1287dbbe0bcdSBarry Smith if (sf->ops->CreateEmbeddedRootSF) PetscUseTypeMethod(sf, CreateEmbeddedRootSF, nselected, selected, esf); 1288dbbe0bcdSBarry Smith else { 1289cd620004SJunchao Zhang /* A generic version of creating embedded sf */ 12909566063dSJacob Faibussowitsch PetscCall(PetscSFGetLeafRange(sf, &minleaf, &maxleaf)); 1291cd620004SJunchao Zhang maxlocal = maxleaf - minleaf + 1; 12929566063dSJacob Faibussowitsch PetscCall(PetscCalloc2(nroots, &rootdata, maxlocal, &leafmem)); 1293cd620004SJunchao Zhang leafdata = leafmem - minleaf; 1294cd620004SJunchao Zhang /* Tag selected roots and bcast to leaves */ 1295cd620004SJunchao Zhang for (i = 0; i < nselected; i++) rootdata[selected[i]] = 1; 12969566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(sf, MPI_SIGNED_CHAR, rootdata, leafdata, MPI_REPLACE)); 12979566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(sf, MPI_SIGNED_CHAR, rootdata, leafdata, MPI_REPLACE)); 1298ba2a7774SJunchao Zhang 1299cd620004SJunchao Zhang /* Build esf with leaves that are still connected */ 1300cd620004SJunchao Zhang esf_nleaves = 0; 1301cd620004SJunchao Zhang for (i = 0; i < nleaves; i++) { 1302cd620004SJunchao Zhang j = ilocal ? ilocal[i] : i; 1303cd620004SJunchao Zhang /* esf_nleaves += leafdata[j] should work in theory, but failed with SFWindow bugs 1304cd620004SJunchao Zhang with PetscSFBcast. See https://gitlab.com/petsc/petsc/issues/555 1305cd620004SJunchao Zhang */ 1306cd620004SJunchao Zhang esf_nleaves += (leafdata[j] ? 1 : 0); 1307cd620004SJunchao Zhang } 13089566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(esf_nleaves, &new_ilocal)); 13099566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(esf_nleaves, &new_iremote)); 1310cd620004SJunchao Zhang for (i = n = 0; i < nleaves; i++) { 1311cd620004SJunchao Zhang j = ilocal ? ilocal[i] : i; 1312cd620004SJunchao Zhang if (leafdata[j]) { 1313cd620004SJunchao Zhang new_ilocal[n] = j; 1314cd620004SJunchao Zhang new_iremote[n].rank = iremote[i].rank; 1315cd620004SJunchao Zhang new_iremote[n].index = iremote[i].index; 1316fc1ede2bSMatthew G. Knepley ++n; 131795fce210SBarry Smith } 131895fce210SBarry Smith } 13199566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, esf)); 13209566063dSJacob Faibussowitsch PetscCall(PetscSFSetFromOptions(*esf)); 13219566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*esf, nroots, esf_nleaves, new_ilocal, PETSC_OWN_POINTER, new_iremote, PETSC_OWN_POINTER)); 13229566063dSJacob Faibussowitsch PetscCall(PetscFree2(rootdata, leafmem)); 1323f659e5c7SJunchao Zhang } 13249566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_EmbedSF, sf, 0, 0, 0)); 132595fce210SBarry Smith PetscFunctionReturn(0); 132695fce210SBarry Smith } 132795fce210SBarry Smith 13282f5fb4c2SMatthew G. Knepley /*@C 13292f5fb4c2SMatthew G. Knepley PetscSFCreateEmbeddedLeafSF - removes edges from all but the selected leaves, does not remap indices 13302f5fb4c2SMatthew G. Knepley 13312f5fb4c2SMatthew G. Knepley Collective 13322f5fb4c2SMatthew G. Knepley 13334165533cSJose E. Roman Input Parameters: 13342f5fb4c2SMatthew G. Knepley + sf - original star forest 1335f659e5c7SJunchao Zhang . nselected - number of selected leaves on this process 1336f659e5c7SJunchao Zhang - selected - indices of the selected leaves on this process 13372f5fb4c2SMatthew G. Knepley 13384165533cSJose E. Roman Output Parameter: 13392f5fb4c2SMatthew G. Knepley . newsf - new star forest 13402f5fb4c2SMatthew G. Knepley 13412f5fb4c2SMatthew G. Knepley Level: advanced 13422f5fb4c2SMatthew G. Knepley 1343cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreateEmbeddedRootSF()`, `PetscSFSetGraph()`, `PetscSFGetGraph()` 13442f5fb4c2SMatthew G. Knepley @*/ 1345d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateEmbeddedLeafSF(PetscSF sf, PetscInt nselected, const PetscInt *selected, PetscSF *newsf) 1346d71ae5a4SJacob Faibussowitsch { 1347f659e5c7SJunchao Zhang const PetscSFNode *iremote; 1348f659e5c7SJunchao Zhang PetscSFNode *new_iremote; 1349f659e5c7SJunchao Zhang const PetscInt *ilocal; 1350f659e5c7SJunchao Zhang PetscInt i, nroots, *leaves, *new_ilocal; 1351f659e5c7SJunchao Zhang MPI_Comm comm; 13522f5fb4c2SMatthew G. Knepley 13532f5fb4c2SMatthew G. Knepley PetscFunctionBegin; 13542f5fb4c2SMatthew G. Knepley PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 135529046d53SLisandro Dalcin PetscSFCheckGraphSet(sf, 1); 1356dadcf809SJacob Faibussowitsch if (nselected) PetscValidIntPointer(selected, 3); 13572f5fb4c2SMatthew G. Knepley PetscValidPointer(newsf, 4); 13582f5fb4c2SMatthew G. Knepley 1359f659e5c7SJunchao Zhang /* Uniq selected[] and put results in leaves[] */ 13609566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 13619566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nselected, &leaves)); 13629566063dSJacob Faibussowitsch PetscCall(PetscArraycpy(leaves, selected, nselected)); 13639566063dSJacob Faibussowitsch PetscCall(PetscSortedRemoveDupsInt(&nselected, leaves)); 136408401ef6SPierre 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); 1365f659e5c7SJunchao Zhang 1366f659e5c7SJunchao Zhang /* Optimize the routine only when sf is setup and hence we can reuse sf's communication pattern */ 1367dbbe0bcdSBarry Smith if (sf->setupcalled && sf->ops->CreateEmbeddedLeafSF) PetscUseTypeMethod(sf, CreateEmbeddedLeafSF, nselected, leaves, newsf); 1368dbbe0bcdSBarry Smith else { 13699566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, NULL, &ilocal, &iremote)); 13709566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nselected, &new_ilocal)); 13719566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nselected, &new_iremote)); 1372f659e5c7SJunchao Zhang for (i = 0; i < nselected; ++i) { 1373f659e5c7SJunchao Zhang const PetscInt l = leaves[i]; 1374f659e5c7SJunchao Zhang new_ilocal[i] = ilocal ? ilocal[l] : l; 1375f659e5c7SJunchao Zhang new_iremote[i].rank = iremote[l].rank; 1376f659e5c7SJunchao Zhang new_iremote[i].index = iremote[l].index; 13772f5fb4c2SMatthew G. Knepley } 13789566063dSJacob Faibussowitsch PetscCall(PetscSFDuplicate(sf, PETSCSF_DUPLICATE_CONFONLY, newsf)); 13799566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*newsf, nroots, nselected, new_ilocal, PETSC_OWN_POINTER, new_iremote, PETSC_OWN_POINTER)); 1380f659e5c7SJunchao Zhang } 13819566063dSJacob Faibussowitsch PetscCall(PetscFree(leaves)); 13822f5fb4c2SMatthew G. Knepley PetscFunctionReturn(0); 13832f5fb4c2SMatthew G. Knepley } 13842f5fb4c2SMatthew G. Knepley 138595fce210SBarry Smith /*@C 1386cab54364SBarry Smith PetscSFBcastBegin - begin pointwise broadcast with root value being reduced to leaf value, to be concluded with call to `PetscSFBcastEnd()` 13873482bfa8SJunchao Zhang 1388cab54364SBarry Smith Collective on sf 13893482bfa8SJunchao Zhang 13904165533cSJose E. Roman Input Parameters: 13913482bfa8SJunchao Zhang + sf - star forest on which to communicate 13923482bfa8SJunchao Zhang . unit - data type associated with each node 13933482bfa8SJunchao Zhang . rootdata - buffer to broadcast 13943482bfa8SJunchao Zhang - op - operation to use for reduction 13953482bfa8SJunchao Zhang 13964165533cSJose E. Roman Output Parameter: 13973482bfa8SJunchao Zhang . leafdata - buffer to be reduced with values from each leaf's respective root 13983482bfa8SJunchao Zhang 13993482bfa8SJunchao Zhang Level: intermediate 14003482bfa8SJunchao Zhang 1401d0295fc0SJunchao Zhang Notes: 1402d0295fc0SJunchao Zhang When petsc is configured with device support, it will use its own mechanism to figure out whether the given data pointers 1403d0295fc0SJunchao Zhang are host pointers or device pointers, which may incur a noticable cost. If you already knew the info, you should 1404cab54364SBarry Smith use `PetscSFBcastWithMemTypeBegin()` instead. 1405cab54364SBarry Smith 1406cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastEnd()`, `PetscSFBcastWithMemTypeBegin()` 14073482bfa8SJunchao Zhang @*/ 1408d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastBegin(PetscSF sf, MPI_Datatype unit, const void *rootdata, void *leafdata, MPI_Op op) 1409d71ae5a4SJacob Faibussowitsch { 1410eb02082bSJunchao Zhang PetscMemType rootmtype, leafmtype; 14113482bfa8SJunchao Zhang 14123482bfa8SJunchao Zhang PetscFunctionBegin; 14133482bfa8SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 14149566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 14159566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin, sf, 0, 0, 0)); 14169566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(rootdata, &rootmtype)); 14179566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafdata, &leafmtype)); 1418dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, BcastBegin, unit, rootmtype, rootdata, leafmtype, leafdata, op); 14199566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin, sf, 0, 0, 0)); 14203482bfa8SJunchao Zhang PetscFunctionReturn(0); 14213482bfa8SJunchao Zhang } 14223482bfa8SJunchao Zhang 14233482bfa8SJunchao Zhang /*@C 1424cab54364SBarry Smith PetscSFBcastWithMemTypeBegin - begin pointwise broadcast with root value being reduced to leaf value with explicit memory types, to be concluded with call to `PetscSFBcastEnd()` 1425d0295fc0SJunchao Zhang 1426cab54364SBarry Smith Collective on sf 1427d0295fc0SJunchao Zhang 14284165533cSJose E. Roman Input Parameters: 1429d0295fc0SJunchao Zhang + sf - star forest on which to communicate 1430d0295fc0SJunchao Zhang . unit - data type associated with each node 1431d0295fc0SJunchao Zhang . rootmtype - memory type of rootdata 1432d0295fc0SJunchao Zhang . rootdata - buffer to broadcast 1433d0295fc0SJunchao Zhang . leafmtype - memory type of leafdata 1434d0295fc0SJunchao Zhang - op - operation to use for reduction 1435d0295fc0SJunchao Zhang 14364165533cSJose E. Roman Output Parameter: 1437d0295fc0SJunchao Zhang . leafdata - buffer to be reduced with values from each leaf's respective root 1438d0295fc0SJunchao Zhang 1439d0295fc0SJunchao Zhang Level: intermediate 1440d0295fc0SJunchao Zhang 1441cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastEnd()`, `PetscSFBcastBegin()` 1442d0295fc0SJunchao Zhang @*/ 1443d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastWithMemTypeBegin(PetscSF sf, MPI_Datatype unit, PetscMemType rootmtype, const void *rootdata, PetscMemType leafmtype, void *leafdata, MPI_Op op) 1444d71ae5a4SJacob Faibussowitsch { 1445d0295fc0SJunchao Zhang PetscFunctionBegin; 1446d0295fc0SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 14479566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 14489566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin, sf, 0, 0, 0)); 1449dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, BcastBegin, unit, rootmtype, rootdata, leafmtype, leafdata, op); 14509566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin, sf, 0, 0, 0)); 1451d0295fc0SJunchao Zhang PetscFunctionReturn(0); 1452d0295fc0SJunchao Zhang } 1453d0295fc0SJunchao Zhang 1454d0295fc0SJunchao Zhang /*@C 1455cab54364SBarry Smith PetscSFBcastEnd - end a broadcast & reduce operation started with `PetscSFBcastBegin()` 14563482bfa8SJunchao Zhang 14573482bfa8SJunchao Zhang Collective 14583482bfa8SJunchao Zhang 14594165533cSJose E. Roman Input Parameters: 14603482bfa8SJunchao Zhang + sf - star forest 14613482bfa8SJunchao Zhang . unit - data type 14623482bfa8SJunchao Zhang . rootdata - buffer to broadcast 14633482bfa8SJunchao Zhang - op - operation to use for reduction 14643482bfa8SJunchao Zhang 14654165533cSJose E. Roman Output Parameter: 14663482bfa8SJunchao Zhang . leafdata - buffer to be reduced with values from each leaf's respective root 14673482bfa8SJunchao Zhang 14683482bfa8SJunchao Zhang Level: intermediate 14693482bfa8SJunchao Zhang 1470cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFReduceEnd()` 14713482bfa8SJunchao Zhang @*/ 1472d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastEnd(PetscSF sf, MPI_Datatype unit, const void *rootdata, void *leafdata, MPI_Op op) 1473d71ae5a4SJacob Faibussowitsch { 14743482bfa8SJunchao Zhang PetscFunctionBegin; 14753482bfa8SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 14769566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_BcastEnd, sf, 0, 0, 0)); 1477dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, BcastEnd, unit, rootdata, leafdata, op); 14789566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_BcastEnd, sf, 0, 0, 0)); 14793482bfa8SJunchao Zhang PetscFunctionReturn(0); 14803482bfa8SJunchao Zhang } 14813482bfa8SJunchao Zhang 14823482bfa8SJunchao Zhang /*@C 1483cab54364SBarry Smith PetscSFReduceBegin - begin reduction of leafdata into rootdata, to be completed with call to `PetscSFReduceEnd()` 148495fce210SBarry Smith 148595fce210SBarry Smith Collective 148695fce210SBarry Smith 14874165533cSJose E. Roman Input Parameters: 148895fce210SBarry Smith + sf - star forest 148995fce210SBarry Smith . unit - data type 149095fce210SBarry Smith . leafdata - values to reduce 149195fce210SBarry Smith - op - reduction operation 149295fce210SBarry Smith 14934165533cSJose E. Roman Output Parameter: 149495fce210SBarry Smith . rootdata - result of reduction of values from all leaves of each root 149595fce210SBarry Smith 149695fce210SBarry Smith Level: intermediate 149795fce210SBarry Smith 1498d0295fc0SJunchao Zhang Notes: 1499d0295fc0SJunchao Zhang When petsc is configured with device support, it will use its own mechanism to figure out whether the given data pointers 1500d0295fc0SJunchao Zhang are host pointers or device pointers, which may incur a noticable cost. If you already knew the info, you should 1501cab54364SBarry Smith use `PetscSFReduceWithMemTypeBegin()` instead. 1502d0295fc0SJunchao Zhang 1503cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastBegin()`, `PetscSFReduceWithMemTypeBegin()` 150495fce210SBarry Smith @*/ 1505d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReduceBegin(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *rootdata, MPI_Op op) 1506d71ae5a4SJacob Faibussowitsch { 1507eb02082bSJunchao Zhang PetscMemType rootmtype, leafmtype; 150895fce210SBarry Smith 150995fce210SBarry Smith PetscFunctionBegin; 151095fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 15119566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 15129566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceBegin, sf, 0, 0, 0)); 15139566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(rootdata, &rootmtype)); 15149566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafdata, &leafmtype)); 15159566063dSJacob Faibussowitsch PetscCall((sf->ops->ReduceBegin)(sf, unit, leafmtype, leafdata, rootmtype, rootdata, op)); 15169566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceBegin, sf, 0, 0, 0)); 151795fce210SBarry Smith PetscFunctionReturn(0); 151895fce210SBarry Smith } 151995fce210SBarry Smith 152095fce210SBarry Smith /*@C 1521cab54364SBarry Smith PetscSFReduceWithMemTypeBegin - begin reduction of leafdata into rootdata with explicit memory types, to be completed with call to `PetscSFReduceEnd()` 1522d0295fc0SJunchao Zhang 1523d0295fc0SJunchao Zhang Collective 1524d0295fc0SJunchao Zhang 15254165533cSJose E. Roman Input Parameters: 1526d0295fc0SJunchao Zhang + sf - star forest 1527d0295fc0SJunchao Zhang . unit - data type 1528d0295fc0SJunchao Zhang . leafmtype - memory type of leafdata 1529d0295fc0SJunchao Zhang . leafdata - values to reduce 1530d0295fc0SJunchao Zhang . rootmtype - memory type of rootdata 1531d0295fc0SJunchao Zhang - op - reduction operation 1532d0295fc0SJunchao Zhang 15334165533cSJose E. Roman Output Parameter: 1534d0295fc0SJunchao Zhang . rootdata - result of reduction of values from all leaves of each root 1535d0295fc0SJunchao Zhang 1536d0295fc0SJunchao Zhang Level: intermediate 1537d0295fc0SJunchao Zhang 1538cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFBcastBegin()`, `PetscSFReduceBegin()` 1539d0295fc0SJunchao Zhang @*/ 1540d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReduceWithMemTypeBegin(PetscSF sf, MPI_Datatype unit, PetscMemType leafmtype, const void *leafdata, PetscMemType rootmtype, void *rootdata, MPI_Op op) 1541d71ae5a4SJacob Faibussowitsch { 1542d0295fc0SJunchao Zhang PetscFunctionBegin; 1543d0295fc0SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 15449566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 15459566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceBegin, sf, 0, 0, 0)); 15469566063dSJacob Faibussowitsch PetscCall((sf->ops->ReduceBegin)(sf, unit, leafmtype, leafdata, rootmtype, rootdata, op)); 15479566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceBegin, sf, 0, 0, 0)); 1548d0295fc0SJunchao Zhang PetscFunctionReturn(0); 1549d0295fc0SJunchao Zhang } 1550d0295fc0SJunchao Zhang 1551d0295fc0SJunchao Zhang /*@C 1552cab54364SBarry Smith PetscSFReduceEnd - end a reduction operation started with `PetscSFReduceBegin()` 155395fce210SBarry Smith 155495fce210SBarry Smith Collective 155595fce210SBarry Smith 15564165533cSJose E. Roman Input Parameters: 155795fce210SBarry Smith + sf - star forest 155895fce210SBarry Smith . unit - data type 155995fce210SBarry Smith . leafdata - values to reduce 156095fce210SBarry Smith - op - reduction operation 156195fce210SBarry Smith 15624165533cSJose E. Roman Output Parameter: 156395fce210SBarry Smith . rootdata - result of reduction of values from all leaves of each root 156495fce210SBarry Smith 156595fce210SBarry Smith Level: intermediate 156695fce210SBarry Smith 1567cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFSetGraph()`, `PetscSFBcastEnd()` 156895fce210SBarry Smith @*/ 1569d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFReduceEnd(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *rootdata, MPI_Op op) 1570d71ae5a4SJacob Faibussowitsch { 157195fce210SBarry Smith PetscFunctionBegin; 157295fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 15739566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventBegin(PETSCSF_ReduceEnd, sf, 0, 0, 0)); 1574dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, ReduceEnd, unit, leafdata, rootdata, op); 15759566063dSJacob Faibussowitsch if (!sf->vscat.logging) PetscCall(PetscLogEventEnd(PETSCSF_ReduceEnd, sf, 0, 0, 0)); 157695fce210SBarry Smith PetscFunctionReturn(0); 157795fce210SBarry Smith } 157895fce210SBarry Smith 157995fce210SBarry Smith /*@C 1580cab54364SBarry Smith PetscSFFetchAndOpBegin - begin operation that fetches values from root and updates atomically by applying operation using my leaf value, 1581cab54364SBarry Smith to be completed with `PetscSFFetchAndOpEnd()` 1582a1729e3fSJunchao Zhang 1583a1729e3fSJunchao Zhang Collective 1584a1729e3fSJunchao Zhang 15854165533cSJose E. Roman Input Parameters: 1586a1729e3fSJunchao Zhang + sf - star forest 1587a1729e3fSJunchao Zhang . unit - data type 1588a1729e3fSJunchao Zhang . leafdata - leaf values to use in reduction 1589a1729e3fSJunchao Zhang - op - operation to use for reduction 1590a1729e3fSJunchao Zhang 15914165533cSJose E. Roman Output Parameters: 1592a1729e3fSJunchao Zhang + rootdata - root values to be updated, input state is seen by first process to perform an update 1593a1729e3fSJunchao Zhang - leafupdate - state at each leaf's respective root immediately prior to my atomic update 1594a1729e3fSJunchao Zhang 1595a1729e3fSJunchao Zhang Level: advanced 1596a1729e3fSJunchao Zhang 1597a1729e3fSJunchao Zhang Note: 1598a1729e3fSJunchao Zhang The update is only atomic at the granularity provided by the hardware. Different roots referenced by the same process 1599a1729e3fSJunchao Zhang might be updated in a different order. Furthermore, if a composite type is used for the unit datatype, atomicity is 1600a1729e3fSJunchao Zhang not guaranteed across the whole vertex. Therefore, this function is mostly only used with primitive types such as 1601a1729e3fSJunchao Zhang integers. 1602a1729e3fSJunchao Zhang 1603cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFReduceBegin()`, `PetscSFSetGraph()` 1604a1729e3fSJunchao Zhang @*/ 1605d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFFetchAndOpBegin(PetscSF sf, MPI_Datatype unit, void *rootdata, const void *leafdata, void *leafupdate, MPI_Op op) 1606d71ae5a4SJacob Faibussowitsch { 1607eb02082bSJunchao Zhang PetscMemType rootmtype, leafmtype, leafupdatemtype; 1608a1729e3fSJunchao Zhang 1609a1729e3fSJunchao Zhang PetscFunctionBegin; 1610a1729e3fSJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 16119566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 16129566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0)); 16139566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(rootdata, &rootmtype)); 16149566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafdata, &leafmtype)); 16159566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafupdate, &leafupdatemtype)); 161608401ef6SPierre Jolivet PetscCheck(leafmtype == leafupdatemtype, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for leafdata and leafupdate in different memory types"); 1617dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, FetchAndOpBegin, unit, rootmtype, rootdata, leafmtype, leafdata, leafupdate, op); 16189566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0)); 1619a1729e3fSJunchao Zhang PetscFunctionReturn(0); 1620a1729e3fSJunchao Zhang } 1621a1729e3fSJunchao Zhang 1622a1729e3fSJunchao Zhang /*@C 1623cab54364SBarry Smith PetscSFFetchAndOpWithMemTypeBegin - begin operation with explicit memory types that fetches values from root and updates atomically by 1624cab54364SBarry Smith applying operation using my leaf value, to be completed with `PetscSFFetchAndOpEnd()` 1625d3b3e55cSJunchao Zhang 1626d3b3e55cSJunchao Zhang Collective 1627d3b3e55cSJunchao Zhang 1628d3b3e55cSJunchao Zhang Input Parameters: 1629d3b3e55cSJunchao Zhang + sf - star forest 1630d3b3e55cSJunchao Zhang . unit - data type 1631d3b3e55cSJunchao Zhang . rootmtype - memory type of rootdata 1632d3b3e55cSJunchao Zhang . leafmtype - memory type of leafdata 1633d3b3e55cSJunchao Zhang . leafdata - leaf values to use in reduction 1634d3b3e55cSJunchao Zhang . leafupdatemtype - memory type of leafupdate 1635d3b3e55cSJunchao Zhang - op - operation to use for reduction 1636d3b3e55cSJunchao Zhang 1637d3b3e55cSJunchao Zhang Output Parameters: 1638d3b3e55cSJunchao Zhang + rootdata - root values to be updated, input state is seen by first process to perform an update 1639d3b3e55cSJunchao Zhang - leafupdate - state at each leaf's respective root immediately prior to my atomic update 1640d3b3e55cSJunchao Zhang 1641d3b3e55cSJunchao Zhang Level: advanced 1642d3b3e55cSJunchao Zhang 1643cab54364SBarry Smith Note: 1644cab54364SBarry Smith See `PetscSFFetchAndOpBegin()` for more details. 1645d3b3e55cSJunchao Zhang 1646cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFFetchAndOpBegin()`, `PetscSFComputeDegreeBegin()`, `PetscSFReduceBegin()`, `PetscSFSetGraph()` 1647d3b3e55cSJunchao Zhang @*/ 1648d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFFetchAndOpWithMemTypeBegin(PetscSF sf, MPI_Datatype unit, PetscMemType rootmtype, void *rootdata, PetscMemType leafmtype, const void *leafdata, PetscMemType leafupdatemtype, void *leafupdate, MPI_Op op) 1649d71ae5a4SJacob Faibussowitsch { 1650d3b3e55cSJunchao Zhang PetscFunctionBegin; 1651d3b3e55cSJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 16529566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 16539566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0)); 165408401ef6SPierre Jolivet PetscCheck(leafmtype == leafupdatemtype, PETSC_COMM_SELF, PETSC_ERR_SUP, "No support for leafdata and leafupdate in different memory types"); 1655dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, FetchAndOpBegin, unit, rootmtype, rootdata, leafmtype, leafdata, leafupdate, op); 16569566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpBegin, sf, 0, 0, 0)); 1657d3b3e55cSJunchao Zhang PetscFunctionReturn(0); 1658d3b3e55cSJunchao Zhang } 1659d3b3e55cSJunchao Zhang 1660d3b3e55cSJunchao Zhang /*@C 1661a1729e3fSJunchao Zhang PetscSFFetchAndOpEnd - end operation started in matching call to PetscSFFetchAndOpBegin() to fetch values from roots and update atomically by applying operation using my leaf value 1662a1729e3fSJunchao Zhang 1663a1729e3fSJunchao Zhang Collective 1664a1729e3fSJunchao Zhang 16654165533cSJose E. Roman Input Parameters: 1666a1729e3fSJunchao Zhang + sf - star forest 1667a1729e3fSJunchao Zhang . unit - data type 1668a1729e3fSJunchao Zhang . leafdata - leaf values to use in reduction 1669a1729e3fSJunchao Zhang - op - operation to use for reduction 1670a1729e3fSJunchao Zhang 16714165533cSJose E. Roman Output Parameters: 1672a1729e3fSJunchao Zhang + rootdata - root values to be updated, input state is seen by first process to perform an update 1673a1729e3fSJunchao Zhang - leafupdate - state at each leaf's respective root immediately prior to my atomic update 1674a1729e3fSJunchao Zhang 1675a1729e3fSJunchao Zhang Level: advanced 1676a1729e3fSJunchao Zhang 1677cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeEnd()`, `PetscSFReduceEnd()`, `PetscSFSetGraph()` 1678a1729e3fSJunchao Zhang @*/ 1679d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFFetchAndOpEnd(PetscSF sf, MPI_Datatype unit, void *rootdata, const void *leafdata, void *leafupdate, MPI_Op op) 1680d71ae5a4SJacob Faibussowitsch { 1681a1729e3fSJunchao Zhang PetscFunctionBegin; 1682a1729e3fSJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 16839566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_FetchAndOpEnd, sf, 0, 0, 0)); 1684dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, FetchAndOpEnd, unit, rootdata, leafdata, leafupdate, op); 16859566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_FetchAndOpEnd, sf, 0, 0, 0)); 1686a1729e3fSJunchao Zhang PetscFunctionReturn(0); 1687a1729e3fSJunchao Zhang } 1688a1729e3fSJunchao Zhang 1689a1729e3fSJunchao Zhang /*@C 1690cab54364SBarry Smith PetscSFComputeDegreeBegin - begin computation of degree for each root vertex, to be completed with `PetscSFComputeDegreeEnd()` 169195fce210SBarry Smith 169295fce210SBarry Smith Collective 169395fce210SBarry Smith 16944165533cSJose E. Roman Input Parameter: 169595fce210SBarry Smith . sf - star forest 169695fce210SBarry Smith 16974165533cSJose E. Roman Output Parameter: 169895fce210SBarry Smith . degree - degree of each root vertex 169995fce210SBarry Smith 170095fce210SBarry Smith Level: advanced 170195fce210SBarry Smith 1702cab54364SBarry Smith Note: 1703cab54364SBarry Smith The returned array is owned by `PetscSF` and automatically freed by `PetscSFDestroy()`. Hence no need to call `PetscFree()` on it. 1704ffe67aa5SVáclav Hapla 1705cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGatherBegin()`, `PetscSFComputeDegreeEnd()` 170695fce210SBarry Smith @*/ 1707d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComputeDegreeBegin(PetscSF sf, const PetscInt **degree) 1708d71ae5a4SJacob Faibussowitsch { 170995fce210SBarry Smith PetscFunctionBegin; 171095fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 171195fce210SBarry Smith PetscSFCheckGraphSet(sf, 1); 171295fce210SBarry Smith PetscValidPointer(degree, 2); 1713803bd9e8SMatthew G. Knepley if (!sf->degreeknown) { 17145b0d146aSStefano Zampini PetscInt i, nroots = sf->nroots, maxlocal; 171528b400f6SJacob Faibussowitsch PetscCheck(!sf->degree, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Calls to PetscSFComputeDegreeBegin() cannot be nested."); 17165b0d146aSStefano Zampini maxlocal = sf->maxleaf - sf->minleaf + 1; 17179566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nroots, &sf->degree)); 17189566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(PetscMax(maxlocal, 1), &sf->degreetmp)); /* allocate at least one entry, see check in PetscSFComputeDegreeEnd() */ 171929046d53SLisandro Dalcin for (i = 0; i < nroots; i++) sf->degree[i] = 0; 17209837ea96SMatthew G. Knepley for (i = 0; i < maxlocal; i++) sf->degreetmp[i] = 1; 17219566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(sf, MPIU_INT, sf->degreetmp - sf->minleaf, sf->degree, MPI_SUM)); 172295fce210SBarry Smith } 172395fce210SBarry Smith *degree = NULL; 172495fce210SBarry Smith PetscFunctionReturn(0); 172595fce210SBarry Smith } 172695fce210SBarry Smith 172795fce210SBarry Smith /*@C 1728cab54364SBarry Smith PetscSFComputeDegreeEnd - complete computation of degree for each root vertex, started with `PetscSFComputeDegreeBegin()` 172995fce210SBarry Smith 173095fce210SBarry Smith Collective 173195fce210SBarry Smith 17324165533cSJose E. Roman Input Parameter: 173395fce210SBarry Smith . sf - star forest 173495fce210SBarry Smith 17354165533cSJose E. Roman Output Parameter: 173695fce210SBarry Smith . degree - degree of each root vertex 173795fce210SBarry Smith 173895fce210SBarry Smith Level: developer 173995fce210SBarry Smith 1740cab54364SBarry Smith Note: 1741cab54364SBarry Smith The returned array is owned by `PetscSF` and automatically freed by `PetscSFDestroy()`. Hence no need to call `PetscFree()` on it. 1742ffe67aa5SVáclav Hapla 1743cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFGatherBegin()`, `PetscSFComputeDegreeBegin()` 174495fce210SBarry Smith @*/ 1745d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComputeDegreeEnd(PetscSF sf, const PetscInt **degree) 1746d71ae5a4SJacob Faibussowitsch { 174795fce210SBarry Smith PetscFunctionBegin; 174895fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 174995fce210SBarry Smith PetscSFCheckGraphSet(sf, 1); 175029046d53SLisandro Dalcin PetscValidPointer(degree, 2); 175195fce210SBarry Smith if (!sf->degreeknown) { 175228b400f6SJacob Faibussowitsch PetscCheck(sf->degreetmp, PETSC_COMM_SELF, PETSC_ERR_ARG_WRONGSTATE, "Must call PetscSFComputeDegreeBegin() before PetscSFComputeDegreeEnd()"); 17539566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(sf, MPIU_INT, sf->degreetmp - sf->minleaf, sf->degree, MPI_SUM)); 17549566063dSJacob Faibussowitsch PetscCall(PetscFree(sf->degreetmp)); 175595fce210SBarry Smith sf->degreeknown = PETSC_TRUE; 175695fce210SBarry Smith } 175795fce210SBarry Smith *degree = sf->degree; 175895fce210SBarry Smith PetscFunctionReturn(0); 175995fce210SBarry Smith } 176095fce210SBarry Smith 1761673100f5SVaclav Hapla /*@C 1762cab54364SBarry Smith PetscSFComputeMultiRootOriginalNumbering - Returns original numbering of multi-roots (roots of multi-SF returned by `PetscSFGetMultiSF()`). 176366dfcd1aSVaclav Hapla Each multi-root is assigned index of the corresponding original root. 1764673100f5SVaclav Hapla 1765673100f5SVaclav Hapla Collective 1766673100f5SVaclav Hapla 17674165533cSJose E. Roman Input Parameters: 1768673100f5SVaclav Hapla + sf - star forest 1769cab54364SBarry Smith - degree - degree of each root vertex, computed with `PetscSFComputeDegreeBegin()`/`PetscSFComputeDegreeEnd()` 1770673100f5SVaclav Hapla 17714165533cSJose E. Roman Output Parameters: 177266dfcd1aSVaclav Hapla + nMultiRoots - (optional) number of multi-roots (roots of multi-SF) 177366dfcd1aSVaclav Hapla - multiRootsOrigNumbering - original indices of multi-roots; length of this array is nMultiRoots 1774673100f5SVaclav Hapla 1775673100f5SVaclav Hapla Level: developer 1776673100f5SVaclav Hapla 1777cab54364SBarry Smith Note: 1778cab54364SBarry Smith The returned array multiRootsOrigNumbering is newly allocated and should be destroyed with `PetscFree()` when no longer needed. 1779ffe67aa5SVáclav Hapla 1780cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFComputeDegreeEnd()`, `PetscSFGetMultiSF()` 1781673100f5SVaclav Hapla @*/ 1782d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComputeMultiRootOriginalNumbering(PetscSF sf, const PetscInt degree[], PetscInt *nMultiRoots, PetscInt *multiRootsOrigNumbering[]) 1783d71ae5a4SJacob Faibussowitsch { 1784673100f5SVaclav Hapla PetscSF msf; 1785673100f5SVaclav Hapla PetscInt i, j, k, nroots, nmroots; 1786673100f5SVaclav Hapla 1787673100f5SVaclav Hapla PetscFunctionBegin; 1788673100f5SVaclav Hapla PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 17899566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, NULL, NULL, NULL)); 179029328920SVaclav Hapla if (nroots) PetscValidIntPointer(degree, 2); 179166dfcd1aSVaclav Hapla if (nMultiRoots) PetscValidIntPointer(nMultiRoots, 3); 179266dfcd1aSVaclav Hapla PetscValidPointer(multiRootsOrigNumbering, 4); 17939566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &msf)); 17949566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(msf, &nmroots, NULL, NULL, NULL)); 17959566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nmroots, multiRootsOrigNumbering)); 1796673100f5SVaclav Hapla for (i = 0, j = 0, k = 0; i < nroots; i++) { 1797673100f5SVaclav Hapla if (!degree[i]) continue; 1798ad540459SPierre Jolivet for (j = 0; j < degree[i]; j++, k++) (*multiRootsOrigNumbering)[k] = i; 1799673100f5SVaclav Hapla } 180008401ef6SPierre Jolivet PetscCheck(k == nmroots, PETSC_COMM_SELF, PETSC_ERR_PLIB, "sanity check fail"); 180166dfcd1aSVaclav Hapla if (nMultiRoots) *nMultiRoots = nmroots; 1802673100f5SVaclav Hapla PetscFunctionReturn(0); 1803673100f5SVaclav Hapla } 1804673100f5SVaclav Hapla 180595fce210SBarry Smith /*@C 1806cab54364SBarry Smith PetscSFGatherBegin - begin pointwise gather of all leaves into multi-roots, to be completed with `PetscSFGatherEnd()` 180795fce210SBarry Smith 180895fce210SBarry Smith Collective 180995fce210SBarry Smith 18104165533cSJose E. Roman Input Parameters: 181195fce210SBarry Smith + sf - star forest 181295fce210SBarry Smith . unit - data type 181395fce210SBarry Smith - leafdata - leaf data to gather to roots 181495fce210SBarry Smith 18154165533cSJose E. Roman Output Parameter: 181695fce210SBarry Smith . multirootdata - root buffer to gather into, amount of space per root is equal to its degree 181795fce210SBarry Smith 181895fce210SBarry Smith Level: intermediate 181995fce210SBarry Smith 1820cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFScatterBegin()` 182195fce210SBarry Smith @*/ 1822d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGatherBegin(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *multirootdata) 1823d71ae5a4SJacob Faibussowitsch { 1824a5526d50SJunchao Zhang PetscSF multi = NULL; 182595fce210SBarry Smith 182695fce210SBarry Smith PetscFunctionBegin; 182795fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 18289566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 18299566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &multi)); 18309566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(multi, unit, leafdata, multirootdata, MPI_REPLACE)); 183195fce210SBarry Smith PetscFunctionReturn(0); 183295fce210SBarry Smith } 183395fce210SBarry Smith 183495fce210SBarry Smith /*@C 1835cab54364SBarry Smith PetscSFGatherEnd - ends pointwise gather operation that was started with `PetscSFGatherBegin()` 183695fce210SBarry Smith 183795fce210SBarry Smith Collective 183895fce210SBarry Smith 18394165533cSJose E. Roman Input Parameters: 184095fce210SBarry Smith + sf - star forest 184195fce210SBarry Smith . unit - data type 184295fce210SBarry Smith - leafdata - leaf data to gather to roots 184395fce210SBarry Smith 18444165533cSJose E. Roman Output Parameter: 184595fce210SBarry Smith . multirootdata - root buffer to gather into, amount of space per root is equal to its degree 184695fce210SBarry Smith 184795fce210SBarry Smith Level: intermediate 184895fce210SBarry Smith 1849cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeEnd()`, `PetscSFScatterEnd()` 185095fce210SBarry Smith @*/ 1851d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFGatherEnd(PetscSF sf, MPI_Datatype unit, const void *leafdata, void *multirootdata) 1852d71ae5a4SJacob Faibussowitsch { 1853a5526d50SJunchao Zhang PetscSF multi = NULL; 185495fce210SBarry Smith 185595fce210SBarry Smith PetscFunctionBegin; 185695fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 18579566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &multi)); 18589566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(multi, unit, leafdata, multirootdata, MPI_REPLACE)); 185995fce210SBarry Smith PetscFunctionReturn(0); 186095fce210SBarry Smith } 186195fce210SBarry Smith 186295fce210SBarry Smith /*@C 1863cab54364SBarry Smith PetscSFScatterBegin - begin pointwise scatter operation from multi-roots to leaves, to be completed with `PetscSFScatterEnd()` 186495fce210SBarry Smith 186595fce210SBarry Smith Collective 186695fce210SBarry Smith 18674165533cSJose E. Roman Input Parameters: 186895fce210SBarry Smith + sf - star forest 186995fce210SBarry Smith . unit - data type 187095fce210SBarry Smith - multirootdata - root buffer to send to each leaf, one unit of data per leaf 187195fce210SBarry Smith 18724165533cSJose E. Roman Output Parameter: 187395fce210SBarry Smith . leafdata - leaf data to be update with personal data from each respective root 187495fce210SBarry Smith 187595fce210SBarry Smith Level: intermediate 187695fce210SBarry Smith 1877cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeBegin()`, `PetscSFScatterBegin()` 187895fce210SBarry Smith @*/ 1879d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFScatterBegin(PetscSF sf, MPI_Datatype unit, const void *multirootdata, void *leafdata) 1880d71ae5a4SJacob Faibussowitsch { 1881a5526d50SJunchao Zhang PetscSF multi = NULL; 188295fce210SBarry Smith 188395fce210SBarry Smith PetscFunctionBegin; 188495fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 18859566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 18869566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &multi)); 18879566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(multi, unit, multirootdata, leafdata, MPI_REPLACE)); 188895fce210SBarry Smith PetscFunctionReturn(0); 188995fce210SBarry Smith } 189095fce210SBarry Smith 189195fce210SBarry Smith /*@C 1892cab54364SBarry Smith PetscSFScatterEnd - ends pointwise scatter operation that was started with `PetscSFScatterBegin()` 189395fce210SBarry Smith 189495fce210SBarry Smith Collective 189595fce210SBarry Smith 18964165533cSJose E. Roman Input Parameters: 189795fce210SBarry Smith + sf - star forest 189895fce210SBarry Smith . unit - data type 189995fce210SBarry Smith - multirootdata - root buffer to send to each leaf, one unit of data per leaf 190095fce210SBarry Smith 19014165533cSJose E. Roman Output Parameter: 190295fce210SBarry Smith . leafdata - leaf data to be update with personal data from each respective root 190395fce210SBarry Smith 190495fce210SBarry Smith Level: intermediate 190595fce210SBarry Smith 1906cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFComputeDegreeEnd()`, `PetscSFScatterEnd()` 190795fce210SBarry Smith @*/ 1908d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFScatterEnd(PetscSF sf, MPI_Datatype unit, const void *multirootdata, void *leafdata) 1909d71ae5a4SJacob Faibussowitsch { 1910a5526d50SJunchao Zhang PetscSF multi = NULL; 191195fce210SBarry Smith 191295fce210SBarry Smith PetscFunctionBegin; 191395fce210SBarry Smith PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 19149566063dSJacob Faibussowitsch PetscCall(PetscSFGetMultiSF(sf, &multi)); 19159566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(multi, unit, multirootdata, leafdata, MPI_REPLACE)); 191695fce210SBarry Smith PetscFunctionReturn(0); 191795fce210SBarry Smith } 1918a7b3aa13SAta Mesgarnejad 1919d71ae5a4SJacob Faibussowitsch static PetscErrorCode PetscSFCheckLeavesUnique_Private(PetscSF sf) 1920d71ae5a4SJacob Faibussowitsch { 1921a072220fSLawrence Mitchell PetscInt i, n, nleaves; 1922a072220fSLawrence Mitchell const PetscInt *ilocal = NULL; 1923a072220fSLawrence Mitchell PetscHSetI seen; 1924a072220fSLawrence Mitchell 1925a072220fSLawrence Mitchell PetscFunctionBegin; 1926b458e8f1SJose E. Roman if (PetscDefined(USE_DEBUG)) { 19279566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, NULL, &nleaves, &ilocal, NULL)); 19289566063dSJacob Faibussowitsch PetscCall(PetscHSetICreate(&seen)); 1929a072220fSLawrence Mitchell for (i = 0; i < nleaves; i++) { 1930a072220fSLawrence Mitchell const PetscInt leaf = ilocal ? ilocal[i] : i; 19319566063dSJacob Faibussowitsch PetscCall(PetscHSetIAdd(seen, leaf)); 1932a072220fSLawrence Mitchell } 19339566063dSJacob Faibussowitsch PetscCall(PetscHSetIGetSize(seen, &n)); 193408401ef6SPierre Jolivet PetscCheck(n == nleaves, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "Provided leaves have repeated values: all leaves must be unique"); 19359566063dSJacob Faibussowitsch PetscCall(PetscHSetIDestroy(&seen)); 1936b458e8f1SJose E. Roman } 1937a072220fSLawrence Mitchell PetscFunctionReturn(0); 1938a072220fSLawrence Mitchell } 193954729392SStefano Zampini 1940a7b3aa13SAta Mesgarnejad /*@ 1941cab54364SBarry Smith PetscSFCompose - Compose a new `PetscSF` by putting the second `PetscSF` under the first one in a top (roots) down (leaves) view 1942a7b3aa13SAta Mesgarnejad 1943a7b3aa13SAta Mesgarnejad Input Parameters: 1944cab54364SBarry Smith + sfA - The first `PetscSF` 1945cab54364SBarry Smith - sfB - The second `PetscSF` 1946a7b3aa13SAta Mesgarnejad 1947a7b3aa13SAta Mesgarnejad Output Parameters: 1948cab54364SBarry Smith . sfBA - The composite `PetscSF` 1949a7b3aa13SAta Mesgarnejad 1950a7b3aa13SAta Mesgarnejad Level: developer 1951a7b3aa13SAta Mesgarnejad 1952a072220fSLawrence Mitchell Notes: 1953cab54364SBarry Smith Currently, the two `PetscSF`s must be defined on congruent communicators and they must be true star 195454729392SStefano Zampini forests, i.e. the same leaf is not connected with different roots. 195554729392SStefano Zampini 195654729392SStefano Zampini sfA's leaf space and sfB's root space might be partially overlapped. The composition builds 195754729392SStefano Zampini a graph with sfA's roots and sfB's leaves only when there is a path between them. Unconnected 195854729392SStefano Zampini nodes (roots or leaves) are not in sfBA. Doing a Bcast on the new SF is equivalent to doing a 195954729392SStefano Zampini Bcast on sfA, then a Bcast on sfB, on connected nodes. 1960a072220fSLawrence Mitchell 1961db781477SPatrick Sanan .seealso: `PetscSF`, `PetscSFComposeInverse()`, `PetscSFGetGraph()`, `PetscSFSetGraph()` 1962a7b3aa13SAta Mesgarnejad @*/ 1963d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCompose(PetscSF sfA, PetscSF sfB, PetscSF *sfBA) 1964d71ae5a4SJacob Faibussowitsch { 1965a7b3aa13SAta Mesgarnejad const PetscSFNode *remotePointsA, *remotePointsB; 1966d41018fbSJunchao Zhang PetscSFNode *remotePointsBA = NULL, *reorderedRemotePointsA = NULL, *leafdataB; 196754729392SStefano Zampini const PetscInt *localPointsA, *localPointsB; 196854729392SStefano Zampini PetscInt *localPointsBA; 196954729392SStefano Zampini PetscInt i, numRootsA, numLeavesA, numRootsB, numLeavesB, minleaf, maxleaf, numLeavesBA; 197054729392SStefano Zampini PetscBool denseB; 1971a7b3aa13SAta Mesgarnejad 1972a7b3aa13SAta Mesgarnejad PetscFunctionBegin; 1973a7b3aa13SAta Mesgarnejad PetscValidHeaderSpecific(sfA, PETSCSF_CLASSID, 1); 197429046d53SLisandro Dalcin PetscSFCheckGraphSet(sfA, 1); 197529046d53SLisandro Dalcin PetscValidHeaderSpecific(sfB, PETSCSF_CLASSID, 2); 197629046d53SLisandro Dalcin PetscSFCheckGraphSet(sfB, 2); 197754729392SStefano Zampini PetscCheckSameComm(sfA, 1, sfB, 2); 197829046d53SLisandro Dalcin PetscValidPointer(sfBA, 3); 19799566063dSJacob Faibussowitsch PetscCall(PetscSFCheckLeavesUnique_Private(sfA)); 19809566063dSJacob Faibussowitsch PetscCall(PetscSFCheckLeavesUnique_Private(sfB)); 198154729392SStefano Zampini 19829566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfA, &numRootsA, &numLeavesA, &localPointsA, &remotePointsA)); 19839566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfB, &numRootsB, &numLeavesB, &localPointsB, &remotePointsB)); 19840ea77edaSksagiyam /* Make sure that PetscSFBcast{Begin, End}(sfB, ...) works with root data of size */ 19850ea77edaSksagiyam /* numRootsB; otherwise, garbage will be broadcasted. */ 19860ea77edaSksagiyam /* Example (comm size = 1): */ 19870ea77edaSksagiyam /* sfA: 0 <- (0, 0) */ 19880ea77edaSksagiyam /* sfB: 100 <- (0, 0) */ 19890ea77edaSksagiyam /* 101 <- (0, 1) */ 19900ea77edaSksagiyam /* Here, we have remotePointsA = [(0, 0)], but for remotePointsA to be a valid tartget */ 19910ea77edaSksagiyam /* of sfB, it has to be recasted as [(0, 0), (-1, -1)] so that points 100 and 101 would */ 19920ea77edaSksagiyam /* receive (0, 0) and (-1, -1), respectively, when PetscSFBcast(sfB, ...) is called on */ 19930ea77edaSksagiyam /* remotePointsA; if not recasted, point 101 would receive a garbage value. */ 19949566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numRootsB, &reorderedRemotePointsA)); 199554729392SStefano Zampini for (i = 0; i < numRootsB; i++) { 199654729392SStefano Zampini reorderedRemotePointsA[i].rank = -1; 199754729392SStefano Zampini reorderedRemotePointsA[i].index = -1; 199854729392SStefano Zampini } 199954729392SStefano Zampini for (i = 0; i < numLeavesA; i++) { 20000ea77edaSksagiyam PetscInt localp = localPointsA ? localPointsA[i] : i; 20010ea77edaSksagiyam 20020ea77edaSksagiyam if (localp >= numRootsB) continue; 20030ea77edaSksagiyam reorderedRemotePointsA[localp] = remotePointsA[i]; 200454729392SStefano Zampini } 2005d41018fbSJunchao Zhang remotePointsA = reorderedRemotePointsA; 20069566063dSJacob Faibussowitsch PetscCall(PetscSFGetLeafRange(sfB, &minleaf, &maxleaf)); 20079566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(maxleaf - minleaf + 1, &leafdataB)); 20080ea77edaSksagiyam for (i = 0; i < maxleaf - minleaf + 1; i++) { 20090ea77edaSksagiyam leafdataB[i].rank = -1; 20100ea77edaSksagiyam leafdataB[i].index = -1; 20110ea77edaSksagiyam } 20129566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(sfB, MPIU_2INT, remotePointsA, leafdataB - minleaf, MPI_REPLACE)); 20139566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(sfB, MPIU_2INT, remotePointsA, leafdataB - minleaf, MPI_REPLACE)); 20149566063dSJacob Faibussowitsch PetscCall(PetscFree(reorderedRemotePointsA)); 2015d41018fbSJunchao Zhang 201654729392SStefano Zampini denseB = (PetscBool)!localPointsB; 201754729392SStefano Zampini for (i = 0, numLeavesBA = 0; i < numLeavesB; i++) { 201854729392SStefano Zampini if (leafdataB[localPointsB ? localPointsB[i] - minleaf : i].rank == -1) denseB = PETSC_FALSE; 201954729392SStefano Zampini else numLeavesBA++; 202054729392SStefano Zampini } 202154729392SStefano Zampini if (denseB) { 2022d41018fbSJunchao Zhang localPointsBA = NULL; 2023d41018fbSJunchao Zhang remotePointsBA = leafdataB; 2024d41018fbSJunchao Zhang } else { 20259566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numLeavesBA, &localPointsBA)); 20269566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numLeavesBA, &remotePointsBA)); 202754729392SStefano Zampini for (i = 0, numLeavesBA = 0; i < numLeavesB; i++) { 202854729392SStefano Zampini const PetscInt l = localPointsB ? localPointsB[i] : i; 202954729392SStefano Zampini 203054729392SStefano Zampini if (leafdataB[l - minleaf].rank == -1) continue; 203154729392SStefano Zampini remotePointsBA[numLeavesBA] = leafdataB[l - minleaf]; 203254729392SStefano Zampini localPointsBA[numLeavesBA] = l; 203354729392SStefano Zampini numLeavesBA++; 203454729392SStefano Zampini } 20359566063dSJacob Faibussowitsch PetscCall(PetscFree(leafdataB)); 2036d41018fbSJunchao Zhang } 20379566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sfA), sfBA)); 20389566063dSJacob Faibussowitsch PetscCall(PetscSFSetFromOptions(*sfBA)); 20399566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*sfBA, numRootsA, numLeavesBA, localPointsBA, PETSC_OWN_POINTER, remotePointsBA, PETSC_OWN_POINTER)); 2040a7b3aa13SAta Mesgarnejad PetscFunctionReturn(0); 2041a7b3aa13SAta Mesgarnejad } 20421c6ba672SJunchao Zhang 204304c0ada0SJunchao Zhang /*@ 2044cab54364SBarry Smith PetscSFComposeInverse - Compose a new `PetscSF` by putting the inverse of the second `PetscSF` under the first one 204504c0ada0SJunchao Zhang 204604c0ada0SJunchao Zhang Input Parameters: 2047cab54364SBarry Smith + sfA - The first `PetscSF` 2048cab54364SBarry Smith - sfB - The second `PetscSF` 204904c0ada0SJunchao Zhang 205004c0ada0SJunchao Zhang Output Parameters: 2051cab54364SBarry Smith . sfBA - The composite `PetscSF`. 205204c0ada0SJunchao Zhang 205304c0ada0SJunchao Zhang Level: developer 205404c0ada0SJunchao Zhang 205554729392SStefano Zampini Notes: 205654729392SStefano Zampini Currently, the two SFs must be defined on congruent communicators and they must be true star 205754729392SStefano Zampini forests, i.e. the same leaf is not connected with different roots. Even more, all roots of the 205854729392SStefano Zampini second SF must have a degree of 1, i.e., no roots have more than one leaf connected. 205954729392SStefano Zampini 206054729392SStefano Zampini sfA's leaf space and sfB's leaf space might be partially overlapped. The composition builds 206154729392SStefano Zampini a graph with sfA's roots and sfB's roots only when there is a path between them. Unconnected 206254729392SStefano Zampini roots are not in sfBA. Doing a Bcast on the new SF is equivalent to doing a Bcast on sfA, then 206354729392SStefano Zampini a Reduce on sfB, on connected roots. 206454729392SStefano Zampini 2065db781477SPatrick Sanan .seealso: `PetscSF`, `PetscSFCompose()`, `PetscSFGetGraph()`, `PetscSFSetGraph()`, `PetscSFCreateInverseSF()` 206604c0ada0SJunchao Zhang @*/ 2067d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFComposeInverse(PetscSF sfA, PetscSF sfB, PetscSF *sfBA) 2068d71ae5a4SJacob Faibussowitsch { 206904c0ada0SJunchao Zhang const PetscSFNode *remotePointsA, *remotePointsB; 207004c0ada0SJunchao Zhang PetscSFNode *remotePointsBA; 207104c0ada0SJunchao Zhang const PetscInt *localPointsA, *localPointsB; 207254729392SStefano Zampini PetscSFNode *reorderedRemotePointsA = NULL; 207354729392SStefano Zampini PetscInt i, numRootsA, numLeavesA, numLeavesBA, numRootsB, numLeavesB, minleaf, maxleaf, *localPointsBA; 20745b0d146aSStefano Zampini MPI_Op op; 20755b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES) 20765b0d146aSStefano Zampini PetscBool iswin; 20775b0d146aSStefano Zampini #endif 207804c0ada0SJunchao Zhang 207904c0ada0SJunchao Zhang PetscFunctionBegin; 208004c0ada0SJunchao Zhang PetscValidHeaderSpecific(sfA, PETSCSF_CLASSID, 1); 208104c0ada0SJunchao Zhang PetscSFCheckGraphSet(sfA, 1); 208204c0ada0SJunchao Zhang PetscValidHeaderSpecific(sfB, PETSCSF_CLASSID, 2); 208304c0ada0SJunchao Zhang PetscSFCheckGraphSet(sfB, 2); 208454729392SStefano Zampini PetscCheckSameComm(sfA, 1, sfB, 2); 208504c0ada0SJunchao Zhang PetscValidPointer(sfBA, 3); 20869566063dSJacob Faibussowitsch PetscCall(PetscSFCheckLeavesUnique_Private(sfA)); 20879566063dSJacob Faibussowitsch PetscCall(PetscSFCheckLeavesUnique_Private(sfB)); 208854729392SStefano Zampini 20899566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfA, &numRootsA, &numLeavesA, &localPointsA, &remotePointsA)); 20909566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfB, &numRootsB, &numLeavesB, &localPointsB, &remotePointsB)); 20915b0d146aSStefano Zampini 20925b0d146aSStefano Zampini /* TODO: Check roots of sfB have degree of 1 */ 20935b0d146aSStefano Zampini /* Once we implement it, we can replace the MPI_MAXLOC 209483df288dSJunchao Zhang with MPI_REPLACE. In that case, MPI_MAXLOC and MPI_REPLACE have the same effect. 20955b0d146aSStefano Zampini We use MPI_MAXLOC only to have a deterministic output from this routine if 20965b0d146aSStefano Zampini the root condition is not meet. 20975b0d146aSStefano Zampini */ 20985b0d146aSStefano Zampini op = MPI_MAXLOC; 20995b0d146aSStefano Zampini #if defined(PETSC_USE_64BIT_INDICES) 21005b0d146aSStefano Zampini /* we accept a non-deterministic output (if any) with PETSCSFWINDOW, since MPI_MAXLOC cannot operate on MPIU_2INT with MPI_Accumulate */ 21019566063dSJacob Faibussowitsch PetscCall(PetscObjectTypeCompare((PetscObject)sfB, PETSCSFWINDOW, &iswin)); 210283df288dSJunchao Zhang if (iswin) op = MPI_REPLACE; 21035b0d146aSStefano Zampini #endif 21045b0d146aSStefano Zampini 21059566063dSJacob Faibussowitsch PetscCall(PetscSFGetLeafRange(sfB, &minleaf, &maxleaf)); 21069566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(maxleaf - minleaf + 1, &reorderedRemotePointsA)); 210754729392SStefano Zampini for (i = 0; i < maxleaf - minleaf + 1; i++) { 210854729392SStefano Zampini reorderedRemotePointsA[i].rank = -1; 210954729392SStefano Zampini reorderedRemotePointsA[i].index = -1; 211054729392SStefano Zampini } 211154729392SStefano Zampini if (localPointsA) { 211254729392SStefano Zampini for (i = 0; i < numLeavesA; i++) { 211354729392SStefano Zampini if (localPointsA[i] > maxleaf || localPointsA[i] < minleaf) continue; 211454729392SStefano Zampini reorderedRemotePointsA[localPointsA[i] - minleaf] = remotePointsA[i]; 211554729392SStefano Zampini } 211654729392SStefano Zampini } else { 211754729392SStefano Zampini for (i = 0; i < numLeavesA; i++) { 211854729392SStefano Zampini if (i > maxleaf || i < minleaf) continue; 211954729392SStefano Zampini reorderedRemotePointsA[i - minleaf] = remotePointsA[i]; 212054729392SStefano Zampini } 212154729392SStefano Zampini } 212254729392SStefano Zampini 21239566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numRootsB, &localPointsBA)); 21249566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(numRootsB, &remotePointsBA)); 212554729392SStefano Zampini for (i = 0; i < numRootsB; i++) { 212654729392SStefano Zampini remotePointsBA[i].rank = -1; 212754729392SStefano Zampini remotePointsBA[i].index = -1; 212854729392SStefano Zampini } 212954729392SStefano Zampini 21309566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(sfB, MPIU_2INT, reorderedRemotePointsA - minleaf, remotePointsBA, op)); 21319566063dSJacob Faibussowitsch PetscCall(PetscSFReduceEnd(sfB, MPIU_2INT, reorderedRemotePointsA - minleaf, remotePointsBA, op)); 21329566063dSJacob Faibussowitsch PetscCall(PetscFree(reorderedRemotePointsA)); 213354729392SStefano Zampini for (i = 0, numLeavesBA = 0; i < numRootsB; i++) { 213454729392SStefano Zampini if (remotePointsBA[i].rank == -1) continue; 213554729392SStefano Zampini remotePointsBA[numLeavesBA].rank = remotePointsBA[i].rank; 213654729392SStefano Zampini remotePointsBA[numLeavesBA].index = remotePointsBA[i].index; 213754729392SStefano Zampini localPointsBA[numLeavesBA] = i; 213854729392SStefano Zampini numLeavesBA++; 213954729392SStefano Zampini } 21409566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(PetscObjectComm((PetscObject)sfA), sfBA)); 21419566063dSJacob Faibussowitsch PetscCall(PetscSFSetFromOptions(*sfBA)); 21429566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*sfBA, numRootsA, numLeavesBA, localPointsBA, PETSC_OWN_POINTER, remotePointsBA, PETSC_OWN_POINTER)); 214304c0ada0SJunchao Zhang PetscFunctionReturn(0); 214404c0ada0SJunchao Zhang } 214504c0ada0SJunchao Zhang 21461c6ba672SJunchao Zhang /* 2147cab54364SBarry Smith PetscSFCreateLocalSF_Private - Creates a local `PetscSF` that only has intra-process edges of the global `PetscSF` 21481c6ba672SJunchao Zhang 21491c6ba672SJunchao Zhang Input Parameters: 2150cab54364SBarry Smith . sf - The global `PetscSF` 21511c6ba672SJunchao Zhang 21521c6ba672SJunchao Zhang Output Parameters: 2153cab54364SBarry Smith . out - The local `PetscSF` 2154cab54364SBarry Smith 2155cab54364SBarry Smith .seealso: `PetscSF`, `PetscSFCreate()` 21561c6ba672SJunchao Zhang */ 2157d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFCreateLocalSF_Private(PetscSF sf, PetscSF *out) 2158d71ae5a4SJacob Faibussowitsch { 21591c6ba672SJunchao Zhang MPI_Comm comm; 21601c6ba672SJunchao Zhang PetscMPIInt myrank; 21611c6ba672SJunchao Zhang const PetscInt *ilocal; 21621c6ba672SJunchao Zhang const PetscSFNode *iremote; 21631c6ba672SJunchao Zhang PetscInt i, j, nroots, nleaves, lnleaves, *lilocal; 21641c6ba672SJunchao Zhang PetscSFNode *liremote; 21651c6ba672SJunchao Zhang PetscSF lsf; 21661c6ba672SJunchao Zhang 21671c6ba672SJunchao Zhang PetscFunctionBegin; 21681c6ba672SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 2169dbbe0bcdSBarry Smith if (sf->ops->CreateLocalSF) PetscUseTypeMethod(sf, CreateLocalSF, out); 2170dbbe0bcdSBarry Smith else { 21711c6ba672SJunchao Zhang /* Could use PetscSFCreateEmbeddedLeafSF, but since we know the comm is PETSC_COMM_SELF, we can make it fast */ 21729566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 21739566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &myrank)); 21741c6ba672SJunchao Zhang 21751c6ba672SJunchao Zhang /* Find out local edges and build a local SF */ 21769566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sf, &nroots, &nleaves, &ilocal, &iremote)); 21779371c9d4SSatish Balay for (i = lnleaves = 0; i < nleaves; i++) { 21789371c9d4SSatish Balay if (iremote[i].rank == (PetscInt)myrank) lnleaves++; 21799371c9d4SSatish Balay } 21809566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(lnleaves, &lilocal)); 21819566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(lnleaves, &liremote)); 21821c6ba672SJunchao Zhang 21831c6ba672SJunchao Zhang for (i = j = 0; i < nleaves; i++) { 21841c6ba672SJunchao Zhang if (iremote[i].rank == (PetscInt)myrank) { 21851c6ba672SJunchao Zhang lilocal[j] = ilocal ? ilocal[i] : i; /* ilocal=NULL for contiguous storage */ 21861c6ba672SJunchao Zhang liremote[j].rank = 0; /* rank in PETSC_COMM_SELF */ 21871c6ba672SJunchao Zhang liremote[j].index = iremote[i].index; 21881c6ba672SJunchao Zhang j++; 21891c6ba672SJunchao Zhang } 21901c6ba672SJunchao Zhang } 21919566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(PETSC_COMM_SELF, &lsf)); 21929566063dSJacob Faibussowitsch PetscCall(PetscSFSetFromOptions(lsf)); 21939566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(lsf, nroots, lnleaves, lilocal, PETSC_OWN_POINTER, liremote, PETSC_OWN_POINTER)); 21949566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(lsf)); 21951c6ba672SJunchao Zhang *out = lsf; 21961c6ba672SJunchao Zhang } 21971c6ba672SJunchao Zhang PetscFunctionReturn(0); 21981c6ba672SJunchao Zhang } 2199dd5b3ca6SJunchao Zhang 2200dd5b3ca6SJunchao Zhang /* Similar to PetscSFBcast, but only Bcast to leaves on rank 0 */ 2201d71ae5a4SJacob Faibussowitsch PetscErrorCode PetscSFBcastToZero_Private(PetscSF sf, MPI_Datatype unit, const void *rootdata, void *leafdata) 2202d71ae5a4SJacob Faibussowitsch { 2203eb02082bSJunchao Zhang PetscMemType rootmtype, leafmtype; 2204dd5b3ca6SJunchao Zhang 2205dd5b3ca6SJunchao Zhang PetscFunctionBegin; 2206dd5b3ca6SJunchao Zhang PetscValidHeaderSpecific(sf, PETSCSF_CLASSID, 1); 22079566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(sf)); 22089566063dSJacob Faibussowitsch PetscCall(PetscLogEventBegin(PETSCSF_BcastBegin, sf, 0, 0, 0)); 22099566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(rootdata, &rootmtype)); 22109566063dSJacob Faibussowitsch PetscCall(PetscGetMemType(leafdata, &leafmtype)); 2211dbbe0bcdSBarry Smith PetscUseTypeMethod(sf, BcastToZero, unit, rootmtype, rootdata, leafmtype, leafdata); 22129566063dSJacob Faibussowitsch PetscCall(PetscLogEventEnd(PETSCSF_BcastBegin, sf, 0, 0, 0)); 2213dd5b3ca6SJunchao Zhang PetscFunctionReturn(0); 2214dd5b3ca6SJunchao Zhang } 2215dd5b3ca6SJunchao Zhang 2216157edd7aSVaclav Hapla /*@ 2217cab54364SBarry Smith PetscSFConcatenate - concatenate multiple `PetscSF` into one 2218157edd7aSVaclav Hapla 2219157edd7aSVaclav Hapla Input Parameters: 2220157edd7aSVaclav Hapla + comm - the communicator 2221cab54364SBarry Smith . nsfs - the number of input `PetscSF` 2222cab54364SBarry Smith . sfs - the array of input `PetscSF` 2223*1f40158dSVaclav Hapla . rootMode - the root mode specifying how roots are handled 2224cab54364SBarry Smith - leafOffsets - the array of local leaf offsets, one for each input `PetscSF`, or NULL for contiguous storage 2225157edd7aSVaclav Hapla 2226157edd7aSVaclav Hapla Output Parameters: 2227cab54364SBarry Smith . newsf - The resulting `PetscSF` 2228157edd7aSVaclav Hapla 2229*1f40158dSVaclav Hapla Level: advanced 2230157edd7aSVaclav Hapla 2231157edd7aSVaclav Hapla Notes: 2232157edd7aSVaclav Hapla The communicator of all SFs in sfs must be comm. 2233157edd7aSVaclav Hapla 2234*1f40158dSVaclav Hapla Leaves are always concatenated locally, keeping them ordered by the input SF index and original local order. 2235157edd7aSVaclav Hapla The offsets in leafOffsets are added to the original leaf indices. 2236157edd7aSVaclav Hapla If all input SFs use contiguous leaf storage (ilocal = NULL), leafOffsets can be passed as NULL as well. 2237157edd7aSVaclav Hapla In this case, NULL is also passed as ilocal to the resulting SF. 2238157edd7aSVaclav Hapla If any input SF has non-null ilocal, leafOffsets is needed to distinguish leaves from different input SFs. 2239157edd7aSVaclav Hapla In this case, user is responsible to provide correct offsets so that the resulting leaves are unique (otherwise an error occurs). 2240157edd7aSVaclav Hapla 2241*1f40158dSVaclav Hapla All root modes retain the essential connectivity condition: 2242*1f40158dSVaclav Hapla If two leaves of the same input SF are connected (sharing the same root), they are also connected in the output SF. 2243*1f40158dSVaclav Hapla Parameter rootMode controls how the input root spaces are combined. 2244*1f40158dSVaclav Hapla For `PETSCSF_CONCATENATE_ROOTMODE_SHARED`, the root space is considered the same for each input SF (checked in debug mode) and is also the same in the output SF. 2245*1f40158dSVaclav Hapla For `PETSCSF_CONCATENATE_ROOTMODE_LOCAL` and `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL`, the input root spaces are taken as separate and joined. 2246*1f40158dSVaclav Hapla `PETSCSF_CONCATENATE_ROOTMODE_LOCAL` joins the root spaces locally; 2247*1f40158dSVaclav Hapla roots of sfs[0], sfs[1], sfs[2], ... are joined on each rank separately, ordered by input SF and original local index, and renumbered contiguously. 2248*1f40158dSVaclav Hapla `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL` joins the root spaces globally; 2249*1f40158dSVaclav Hapla roots of sfs[0], sfs[1], sfs[2, ... are joined globally, ordered by input SF index and original global index, and renumbered contiguously; 2250*1f40158dSVaclav Hapla the original root ranks are ignored. 2251*1f40158dSVaclav Hapla For both `PETSCSF_CONCATENATE_ROOTMODE_LOCAL` and `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL`, 2252*1f40158dSVaclav Hapla the output SF's root layout is such that the local number of roots is a sum of the input SF's local numbers of roots on each rank to keep the load balancing. 2253*1f40158dSVaclav Hapla However, for `PETSCSF_CONCATENATE_ROOTMODE_GLOBAL`, that means roots can move to different ranks. 2254*1f40158dSVaclav Hapla 2255*1f40158dSVaclav Hapla Example: 2256*1f40158dSVaclav Hapla We can use src/vec/is/sf/tests/ex18.c to compare the root modes. By running 2257*1f40158dSVaclav Hapla $ make -C $PETSC_DIR/src/vec/is/sf/tests ex18 2258*1f40158dSVaclav Hapla $ for m in {local,global,shared}; do 2259*1f40158dSVaclav Hapla $ mpirun -n 2 $PETSC_DIR/src/vec/is/sf/tests/ex18 -nsfs 2 -n 2 -root_mode $m -sf_view 2260*1f40158dSVaclav Hapla $ done 2261*1f40158dSVaclav Hapla we generate two identical SFs sf_0 and sf_1, 2262*1f40158dSVaclav Hapla $ PetscSF Object: sf_0 2 MPI processes 2263*1f40158dSVaclav Hapla $ type: basic 2264*1f40158dSVaclav Hapla $ rank #leaves #roots 2265*1f40158dSVaclav Hapla $ [ 0] 4 2 2266*1f40158dSVaclav Hapla $ [ 1] 4 2 2267*1f40158dSVaclav Hapla $ leaves roots roots in global numbering 2268*1f40158dSVaclav Hapla $ ( 0, 0) <- ( 0, 0) = 0 2269*1f40158dSVaclav Hapla $ ( 0, 1) <- ( 0, 1) = 1 2270*1f40158dSVaclav Hapla $ ( 0, 2) <- ( 1, 0) = 2 2271*1f40158dSVaclav Hapla $ ( 0, 3) <- ( 1, 1) = 3 2272*1f40158dSVaclav Hapla $ ( 1, 0) <- ( 0, 0) = 0 2273*1f40158dSVaclav Hapla $ ( 1, 1) <- ( 0, 1) = 1 2274*1f40158dSVaclav Hapla $ ( 1, 2) <- ( 1, 0) = 2 2275*1f40158dSVaclav Hapla $ ( 1, 3) <- ( 1, 1) = 3 2276*1f40158dSVaclav Hapla and pass them to `PetscSFConcatenate()` along with different choices of rootMode, yielding different result_sf: 2277*1f40158dSVaclav Hapla $ rootMode = local: 2278*1f40158dSVaclav Hapla $ PetscSF Object: result_sf 2 MPI processes 2279*1f40158dSVaclav Hapla $ type: basic 2280*1f40158dSVaclav Hapla $ rank #leaves #roots 2281*1f40158dSVaclav Hapla $ [ 0] 8 4 2282*1f40158dSVaclav Hapla $ [ 1] 8 4 2283*1f40158dSVaclav Hapla $ leaves roots roots in global numbering 2284*1f40158dSVaclav Hapla $ ( 0, 0) <- ( 0, 0) = 0 2285*1f40158dSVaclav Hapla $ ( 0, 1) <- ( 0, 1) = 1 2286*1f40158dSVaclav Hapla $ ( 0, 2) <- ( 1, 0) = 4 2287*1f40158dSVaclav Hapla $ ( 0, 3) <- ( 1, 1) = 5 2288*1f40158dSVaclav Hapla $ ( 0, 4) <- ( 0, 2) = 2 2289*1f40158dSVaclav Hapla $ ( 0, 5) <- ( 0, 3) = 3 2290*1f40158dSVaclav Hapla $ ( 0, 6) <- ( 1, 2) = 6 2291*1f40158dSVaclav Hapla $ ( 0, 7) <- ( 1, 3) = 7 2292*1f40158dSVaclav Hapla $ ( 1, 0) <- ( 0, 0) = 0 2293*1f40158dSVaclav Hapla $ ( 1, 1) <- ( 0, 1) = 1 2294*1f40158dSVaclav Hapla $ ( 1, 2) <- ( 1, 0) = 4 2295*1f40158dSVaclav Hapla $ ( 1, 3) <- ( 1, 1) = 5 2296*1f40158dSVaclav Hapla $ ( 1, 4) <- ( 0, 2) = 2 2297*1f40158dSVaclav Hapla $ ( 1, 5) <- ( 0, 3) = 3 2298*1f40158dSVaclav Hapla $ ( 1, 6) <- ( 1, 2) = 6 2299*1f40158dSVaclav Hapla $ ( 1, 7) <- ( 1, 3) = 7 2300*1f40158dSVaclav Hapla $ 2301*1f40158dSVaclav Hapla $ rootMode = global: 2302*1f40158dSVaclav Hapla $ PetscSF Object: result_sf 2 MPI processes 2303*1f40158dSVaclav Hapla $ type: basic 2304*1f40158dSVaclav Hapla $ rank #leaves #roots 2305*1f40158dSVaclav Hapla $ [ 0] 8 4 2306*1f40158dSVaclav Hapla $ [ 1] 8 4 2307*1f40158dSVaclav Hapla $ leaves roots roots in global numbering 2308*1f40158dSVaclav Hapla $ ( 0, 0) <- ( 0, 0) = 0 2309*1f40158dSVaclav Hapla $ ( 0, 1) <- ( 0, 1) = 1 2310*1f40158dSVaclav Hapla $ ( 0, 2) <- ( 0, 2) = 2 2311*1f40158dSVaclav Hapla $ ( 0, 3) <- ( 0, 3) = 3 2312*1f40158dSVaclav Hapla $ ( 0, 4) <- ( 1, 0) = 4 2313*1f40158dSVaclav Hapla $ ( 0, 5) <- ( 1, 1) = 5 2314*1f40158dSVaclav Hapla $ ( 0, 6) <- ( 1, 2) = 6 2315*1f40158dSVaclav Hapla $ ( 0, 7) <- ( 1, 3) = 7 2316*1f40158dSVaclav Hapla $ ( 1, 0) <- ( 0, 0) = 0 2317*1f40158dSVaclav Hapla $ ( 1, 1) <- ( 0, 1) = 1 2318*1f40158dSVaclav Hapla $ ( 1, 2) <- ( 0, 2) = 2 2319*1f40158dSVaclav Hapla $ ( 1, 3) <- ( 0, 3) = 3 2320*1f40158dSVaclav Hapla $ ( 1, 4) <- ( 1, 0) = 4 2321*1f40158dSVaclav Hapla $ ( 1, 5) <- ( 1, 1) = 5 2322*1f40158dSVaclav Hapla $ ( 1, 6) <- ( 1, 2) = 6 2323*1f40158dSVaclav Hapla $ ( 1, 7) <- ( 1, 3) = 7 2324*1f40158dSVaclav Hapla $ 2325*1f40158dSVaclav Hapla $ rootMode = shared: 2326*1f40158dSVaclav Hapla $ PetscSF Object: result_sf 2 MPI processes 2327*1f40158dSVaclav Hapla $ type: basic 2328*1f40158dSVaclav Hapla $ rank #leaves #roots 2329*1f40158dSVaclav Hapla $ [ 0] 8 2 2330*1f40158dSVaclav Hapla $ [ 1] 8 2 2331*1f40158dSVaclav Hapla $ leaves roots roots in global numbering 2332*1f40158dSVaclav Hapla $ ( 0, 0) <- ( 0, 0) = 0 2333*1f40158dSVaclav Hapla $ ( 0, 1) <- ( 0, 1) = 1 2334*1f40158dSVaclav Hapla $ ( 0, 2) <- ( 1, 0) = 2 2335*1f40158dSVaclav Hapla $ ( 0, 3) <- ( 1, 1) = 3 2336*1f40158dSVaclav Hapla $ ( 0, 4) <- ( 0, 0) = 0 2337*1f40158dSVaclav Hapla $ ( 0, 5) <- ( 0, 1) = 1 2338*1f40158dSVaclav Hapla $ ( 0, 6) <- ( 1, 0) = 2 2339*1f40158dSVaclav Hapla $ ( 0, 7) <- ( 1, 1) = 3 2340*1f40158dSVaclav Hapla $ ( 1, 0) <- ( 0, 0) = 0 2341*1f40158dSVaclav Hapla $ ( 1, 1) <- ( 0, 1) = 1 2342*1f40158dSVaclav Hapla $ ( 1, 2) <- ( 1, 0) = 2 2343*1f40158dSVaclav Hapla $ ( 1, 3) <- ( 1, 1) = 3 2344*1f40158dSVaclav Hapla $ ( 1, 4) <- ( 0, 0) = 0 2345*1f40158dSVaclav Hapla $ ( 1, 5) <- ( 0, 1) = 1 2346*1f40158dSVaclav Hapla $ ( 1, 6) <- ( 1, 0) = 2 2347*1f40158dSVaclav Hapla $ ( 1, 7) <- ( 1, 1) = 3 2348*1f40158dSVaclav Hapla 2349*1f40158dSVaclav Hapla .seealso: `PetscSF`, `PetscSFCompose()`, `PetscSFGetGraph()`, `PetscSFSetGraph()`, `PetscSFConcatenateRootMode` 2350157edd7aSVaclav Hapla @*/ 2351*1f40158dSVaclav Hapla PetscErrorCode PetscSFConcatenate(MPI_Comm comm, PetscInt nsfs, PetscSF sfs[], PetscSFConcatenateRootMode rootMode, PetscInt leafOffsets[], PetscSF *newsf) 2352d71ae5a4SJacob Faibussowitsch { 2353157edd7aSVaclav Hapla PetscInt i, s, nLeaves, nRoots; 2354157edd7aSVaclav Hapla PetscInt *leafArrayOffsets; 2355157edd7aSVaclav Hapla PetscInt *ilocal_new; 2356157edd7aSVaclav Hapla PetscSFNode *iremote_new; 2357157edd7aSVaclav Hapla PetscBool all_ilocal_null = PETSC_FALSE; 2358*1f40158dSVaclav Hapla PetscLayout glayout = NULL; 2359*1f40158dSVaclav Hapla PetscInt *gremote = NULL; 2360*1f40158dSVaclav Hapla PetscMPIInt rank, size; 2361157edd7aSVaclav Hapla 2362157edd7aSVaclav Hapla PetscFunctionBegin; 2363157edd7aSVaclav Hapla { 2364157edd7aSVaclav Hapla PetscSF dummy; /* just to have a PetscObject on comm for input validation */ 2365157edd7aSVaclav Hapla 23669566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, &dummy)); 2367157edd7aSVaclav Hapla PetscValidLogicalCollectiveInt(dummy, nsfs, 2); 2368157edd7aSVaclav Hapla PetscValidPointer(sfs, 3); 2369157edd7aSVaclav Hapla for (i = 0; i < nsfs; i++) { 2370157edd7aSVaclav Hapla PetscValidHeaderSpecific(sfs[i], PETSCSF_CLASSID, 3); 2371157edd7aSVaclav Hapla PetscCheckSameComm(dummy, 1, sfs[i], 3); 2372157edd7aSVaclav Hapla } 2373*1f40158dSVaclav Hapla PetscValidLogicalCollectiveEnum(dummy, rootMode, 4); 2374157edd7aSVaclav Hapla if (leafOffsets) PetscValidIntPointer(leafOffsets, 5); 2375157edd7aSVaclav Hapla PetscValidPointer(newsf, 6); 23769566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&dummy)); 2377157edd7aSVaclav Hapla } 2378157edd7aSVaclav Hapla if (!nsfs) { 23799566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, newsf)); 23809566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*newsf, 0, 0, NULL, PETSC_OWN_POINTER, NULL, PETSC_OWN_POINTER)); 2381157edd7aSVaclav Hapla PetscFunctionReturn(0); 2382157edd7aSVaclav Hapla } 23839566063dSJacob Faibussowitsch PetscCallMPI(MPI_Comm_rank(comm, &rank)); 2384*1f40158dSVaclav Hapla PetscCallMPI(MPI_Comm_size(comm, &size)); 2385157edd7aSVaclav Hapla 2386*1f40158dSVaclav Hapla /* Calculate leaf array offsets */ 23879566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nsfs + 1, &leafArrayOffsets)); 2388157edd7aSVaclav Hapla leafArrayOffsets[0] = 0; 2389157edd7aSVaclav Hapla for (s = 0; s < nsfs; s++) { 2390157edd7aSVaclav Hapla PetscInt nl; 2391157edd7aSVaclav Hapla 23929566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfs[s], NULL, &nl, NULL, NULL)); 2393157edd7aSVaclav Hapla leafArrayOffsets[s + 1] = leafArrayOffsets[s] + nl; 2394157edd7aSVaclav Hapla } 2395157edd7aSVaclav Hapla nLeaves = leafArrayOffsets[nsfs]; 2396157edd7aSVaclav Hapla 2397*1f40158dSVaclav Hapla /* Calculate number of roots */ 2398*1f40158dSVaclav Hapla switch (rootMode) { 2399*1f40158dSVaclav Hapla case PETSCSF_CONCATENATE_ROOTMODE_SHARED: { 2400*1f40158dSVaclav Hapla PetscCall(PetscSFGetGraph(sfs[0], &nRoots, NULL, NULL, NULL)); 2401*1f40158dSVaclav Hapla if (PetscDefined(USE_DEBUG)) { 2402*1f40158dSVaclav Hapla for (s = 1; s < nsfs; s++) { 2403*1f40158dSVaclav Hapla PetscInt nr; 2404*1f40158dSVaclav Hapla 2405*1f40158dSVaclav Hapla PetscCall(PetscSFGetGraph(sfs[s], &nr, NULL, NULL, NULL)); 2406*1f40158dSVaclav 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); 2407*1f40158dSVaclav Hapla } 2408*1f40158dSVaclav Hapla } 2409*1f40158dSVaclav Hapla } break; 2410*1f40158dSVaclav Hapla case PETSCSF_CONCATENATE_ROOTMODE_GLOBAL: { 2411*1f40158dSVaclav Hapla /* Calculate also global layout in this case */ 2412*1f40158dSVaclav Hapla PetscInt *nls; 2413*1f40158dSVaclav Hapla PetscLayout *lts; 2414*1f40158dSVaclav Hapla PetscInt **inds; 2415*1f40158dSVaclav Hapla PetscInt j; 2416*1f40158dSVaclav Hapla PetscInt rootOffset = 0; 2417*1f40158dSVaclav Hapla 2418*1f40158dSVaclav Hapla PetscCall(PetscCalloc3(nsfs, <s, nsfs, &nls, nsfs, &inds)); 2419*1f40158dSVaclav Hapla PetscCall(PetscLayoutCreate(comm, &glayout)); 2420*1f40158dSVaclav Hapla glayout->bs = 1; 2421*1f40158dSVaclav Hapla glayout->n = 0; 2422*1f40158dSVaclav Hapla glayout->N = 0; 2423*1f40158dSVaclav Hapla for (s = 0; s < nsfs; s++) { 2424*1f40158dSVaclav Hapla PetscCall(PetscSFGetGraphLayout(sfs[s], <s[s], &nls[s], NULL, &inds[s])); 2425*1f40158dSVaclav Hapla glayout->n += lts[s]->n; 2426*1f40158dSVaclav Hapla glayout->N += lts[s]->N; 2427*1f40158dSVaclav Hapla } 2428*1f40158dSVaclav Hapla PetscCall(PetscLayoutSetUp(glayout)); 2429*1f40158dSVaclav Hapla PetscCall(PetscMalloc1(nLeaves, &gremote)); 2430*1f40158dSVaclav Hapla for (s = 0, j = 0; s < nsfs; s++) { 2431*1f40158dSVaclav Hapla for (i = 0; i < nls[s]; i++, j++) gremote[j] = inds[s][i] + rootOffset; 2432*1f40158dSVaclav Hapla rootOffset += lts[s]->N; 2433*1f40158dSVaclav Hapla PetscCall(PetscLayoutDestroy(<s[s])); 2434*1f40158dSVaclav Hapla PetscCall(PetscFree(inds[s])); 2435*1f40158dSVaclav Hapla } 2436*1f40158dSVaclav Hapla PetscCall(PetscFree3(lts, nls, inds)); 2437*1f40158dSVaclav Hapla nRoots = glayout->N; 2438*1f40158dSVaclav Hapla } break; 2439*1f40158dSVaclav Hapla case PETSCSF_CONCATENATE_ROOTMODE_LOCAL: 2440*1f40158dSVaclav Hapla /* nRoots calculated later in this case */ 2441*1f40158dSVaclav Hapla break; 2442*1f40158dSVaclav Hapla default: 2443*1f40158dSVaclav Hapla SETERRQ(comm, PETSC_ERR_ARG_WRONG, "Invalid PetscSFConcatenateRootMode %d", rootMode); 2444*1f40158dSVaclav Hapla } 2445*1f40158dSVaclav Hapla 2446157edd7aSVaclav Hapla if (!leafOffsets) { 2447157edd7aSVaclav Hapla all_ilocal_null = PETSC_TRUE; 2448157edd7aSVaclav Hapla for (s = 0; s < nsfs; s++) { 2449157edd7aSVaclav Hapla const PetscInt *ilocal; 2450157edd7aSVaclav Hapla 24519566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfs[s], NULL, NULL, &ilocal, NULL)); 2452157edd7aSVaclav Hapla if (ilocal) { 2453157edd7aSVaclav Hapla all_ilocal_null = PETSC_FALSE; 2454157edd7aSVaclav Hapla break; 2455157edd7aSVaclav Hapla } 2456157edd7aSVaclav Hapla } 2457157edd7aSVaclav 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"); 2458157edd7aSVaclav Hapla } 2459157edd7aSVaclav Hapla 2460157edd7aSVaclav Hapla /* Renumber and concatenate local leaves */ 2461157edd7aSVaclav Hapla ilocal_new = NULL; 2462157edd7aSVaclav Hapla if (!all_ilocal_null) { 24639566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nLeaves, &ilocal_new)); 2464157edd7aSVaclav Hapla for (i = 0; i < nLeaves; i++) ilocal_new[i] = -1; 2465157edd7aSVaclav Hapla for (s = 0; s < nsfs; s++) { 2466157edd7aSVaclav Hapla const PetscInt *ilocal; 2467157edd7aSVaclav Hapla PetscInt *ilocal_l = &ilocal_new[leafArrayOffsets[s]]; 2468157edd7aSVaclav Hapla PetscInt i, nleaves_l; 2469157edd7aSVaclav Hapla 24709566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfs[s], NULL, &nleaves_l, &ilocal, NULL)); 2471157edd7aSVaclav Hapla for (i = 0; i < nleaves_l; i++) ilocal_l[i] = (ilocal ? ilocal[i] : i) + leafOffsets[s]; 2472157edd7aSVaclav Hapla } 2473157edd7aSVaclav Hapla } 2474157edd7aSVaclav Hapla 2475157edd7aSVaclav Hapla /* Renumber and concatenate remote roots */ 2476*1f40158dSVaclav Hapla if (rootMode == PETSCSF_CONCATENATE_ROOTMODE_LOCAL || rootMode == PETSCSF_CONCATENATE_ROOTMODE_SHARED) { 2477*1f40158dSVaclav Hapla PetscInt rootOffset = 0; 2478*1f40158dSVaclav Hapla 24799566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nLeaves, &iremote_new)); 2480157edd7aSVaclav Hapla for (i = 0; i < nLeaves; i++) { 2481157edd7aSVaclav Hapla iremote_new[i].rank = -1; 2482157edd7aSVaclav Hapla iremote_new[i].index = -1; 2483157edd7aSVaclav Hapla } 2484157edd7aSVaclav Hapla for (s = 0; s < nsfs; s++) { 2485157edd7aSVaclav Hapla PetscInt i, nl, nr; 2486157edd7aSVaclav Hapla PetscSF tmp_sf; 2487157edd7aSVaclav Hapla const PetscSFNode *iremote; 2488157edd7aSVaclav Hapla PetscSFNode *tmp_rootdata; 2489157edd7aSVaclav Hapla PetscSFNode *tmp_leafdata = &iremote_new[leafArrayOffsets[s]]; 2490157edd7aSVaclav Hapla 24919566063dSJacob Faibussowitsch PetscCall(PetscSFGetGraph(sfs[s], &nr, &nl, NULL, &iremote)); 24929566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, &tmp_sf)); 2493157edd7aSVaclav Hapla /* create helper SF with contiguous leaves */ 24949566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(tmp_sf, nr, nl, NULL, PETSC_USE_POINTER, (PetscSFNode *)iremote, PETSC_COPY_VALUES)); 24959566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(tmp_sf)); 24969566063dSJacob Faibussowitsch PetscCall(PetscMalloc1(nr, &tmp_rootdata)); 2497*1f40158dSVaclav Hapla if (rootMode == PETSCSF_CONCATENATE_ROOTMODE_LOCAL) { 2498157edd7aSVaclav Hapla for (i = 0; i < nr; i++) { 2499*1f40158dSVaclav Hapla tmp_rootdata[i].index = i + rootOffset; 2500157edd7aSVaclav Hapla tmp_rootdata[i].rank = (PetscInt)rank; 2501157edd7aSVaclav Hapla } 2502*1f40158dSVaclav Hapla rootOffset += nr; 2503*1f40158dSVaclav Hapla } else { 2504*1f40158dSVaclav Hapla for (i = 0; i < nr; i++) { 2505*1f40158dSVaclav Hapla tmp_rootdata[i].index = i; 2506*1f40158dSVaclav Hapla tmp_rootdata[i].rank = (PetscInt)rank; 2507*1f40158dSVaclav Hapla } 2508*1f40158dSVaclav Hapla } 25099566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(tmp_sf, MPIU_2INT, tmp_rootdata, tmp_leafdata, MPI_REPLACE)); 25109566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(tmp_sf, MPIU_2INT, tmp_rootdata, tmp_leafdata, MPI_REPLACE)); 25119566063dSJacob Faibussowitsch PetscCall(PetscSFDestroy(&tmp_sf)); 25129566063dSJacob Faibussowitsch PetscCall(PetscFree(tmp_rootdata)); 2513157edd7aSVaclav Hapla } 2514*1f40158dSVaclav Hapla if (rootMode == PETSCSF_CONCATENATE_ROOTMODE_LOCAL) { nRoots = rootOffset; } // else nRoots already calculated above 2515157edd7aSVaclav Hapla 2516157edd7aSVaclav Hapla /* Build the new SF */ 25179566063dSJacob Faibussowitsch PetscCall(PetscSFCreate(comm, newsf)); 25189566063dSJacob Faibussowitsch PetscCall(PetscSFSetGraph(*newsf, nRoots, nLeaves, ilocal_new, PETSC_OWN_POINTER, iremote_new, PETSC_OWN_POINTER)); 2519*1f40158dSVaclav Hapla } else { 2520*1f40158dSVaclav Hapla /* Build the new SF */ 2521*1f40158dSVaclav Hapla PetscCall(PetscSFCreate(comm, newsf)); 2522*1f40158dSVaclav Hapla PetscCall(PetscSFSetGraphLayout(*newsf, glayout, nLeaves, ilocal_new, PETSC_OWN_POINTER, gremote)); 2523*1f40158dSVaclav Hapla } 25249566063dSJacob Faibussowitsch PetscCall(PetscSFSetUp(*newsf)); 2525*1f40158dSVaclav Hapla PetscCall(PetscSFViewFromOptions(*newsf, NULL, "-sf_concat_view")); 2526*1f40158dSVaclav Hapla PetscCall(PetscLayoutDestroy(&glayout)); 2527*1f40158dSVaclav Hapla PetscCall(PetscFree(gremote)); 25289566063dSJacob Faibussowitsch PetscCall(PetscFree(leafArrayOffsets)); 2529157edd7aSVaclav Hapla PetscFunctionReturn(0); 2530157edd7aSVaclav Hapla } 2531