1dd5b3ca6SJunchao Zhang 2dd5b3ca6SJunchao Zhang #include <../src/vec/is/sf/impls/basic/gatherv/sfgatherv.h> 3dd5b3ca6SJunchao Zhang 4eb02082bSJunchao Zhang /* Reuse the type. The difference is some fields (displs, recvcounts) are only significant 5eb02082bSJunchao Zhang on rank 0 in Gatherv. On other ranks they are harmless NULL. 6eb02082bSJunchao Zhang */ 7eb02082bSJunchao Zhang typedef PetscSF_Allgatherv PetscSF_Gatherv; 8eb02082bSJunchao Zhang 9d71ae5a4SJacob Faibussowitsch PETSC_INTERN PetscErrorCode PetscSFBcastBegin_Gatherv(PetscSF sf, MPI_Datatype unit, PetscMemType rootmtype, const void *rootdata, PetscMemType leafmtype, void *leafdata, MPI_Op op) 10d71ae5a4SJacob Faibussowitsch { 11cd620004SJunchao Zhang PetscSFLink link; 12855db38dSJunchao Zhang PetscMPIInt sendcount; 13dd5b3ca6SJunchao Zhang MPI_Comm comm; 14dd5b3ca6SJunchao Zhang PetscSF_Gatherv *dat = (PetscSF_Gatherv *)sf->data; 15cd620004SJunchao Zhang void *rootbuf = NULL, *leafbuf = NULL; /* buffer seen by MPI */ 16cd620004SJunchao Zhang MPI_Request *req; 17dd5b3ca6SJunchao Zhang 18dd5b3ca6SJunchao Zhang PetscFunctionBegin; 199566063dSJacob Faibussowitsch PetscCall(PetscSFLinkCreate(sf, unit, rootmtype, rootdata, leafmtype, leafdata, op, PETSCSF_BCAST, &link)); 209566063dSJacob Faibussowitsch PetscCall(PetscSFLinkPackRootData(sf, link, PETSCSF_REMOTE, rootdata)); 219566063dSJacob Faibussowitsch PetscCall(PetscSFLinkCopyRootBufferInCaseNotUseGpuAwareMPI(sf, link, PETSC_TRUE /* device2host before sending */)); 229566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 239566063dSJacob Faibussowitsch PetscCall(PetscMPIIntCast(sf->nroots, &sendcount)); 249566063dSJacob Faibussowitsch PetscCall(PetscSFLinkGetMPIBuffersAndRequests(sf, link, PETSCSF_ROOT2LEAF, &rootbuf, &leafbuf, &req, NULL)); 259566063dSJacob Faibussowitsch PetscCall(PetscSFLinkSyncStreamBeforeCallMPI(sf, link, PETSCSF_ROOT2LEAF)); 269566063dSJacob Faibussowitsch PetscCallMPI(MPIU_Igatherv(rootbuf, sendcount, unit, leafbuf, dat->recvcounts, dat->displs, unit, 0 /*rank 0*/, comm, req)); 27*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 28dd5b3ca6SJunchao Zhang } 29dd5b3ca6SJunchao Zhang 30d71ae5a4SJacob Faibussowitsch static PetscErrorCode PetscSFReduceBegin_Gatherv(PetscSF sf, MPI_Datatype unit, PetscMemType leafmtype, const void *leafdata, PetscMemType rootmtype, void *rootdata, MPI_Op op) 31d71ae5a4SJacob Faibussowitsch { 32cd620004SJunchao Zhang PetscSFLink link; 33dd5b3ca6SJunchao Zhang PetscMPIInt recvcount; 34dd5b3ca6SJunchao Zhang MPI_Comm comm; 35dd5b3ca6SJunchao Zhang PetscSF_Gatherv *dat = (PetscSF_Gatherv *)sf->data; 36cd620004SJunchao Zhang void *rootbuf = NULL, *leafbuf = NULL; /* buffer seen by MPI */ 37cd620004SJunchao Zhang MPI_Request *req; 38dd5b3ca6SJunchao Zhang 39dd5b3ca6SJunchao Zhang PetscFunctionBegin; 409566063dSJacob Faibussowitsch PetscCall(PetscSFLinkCreate(sf, unit, rootmtype, rootdata, leafmtype, leafdata, op, PETSCSF_REDUCE, &link)); 419566063dSJacob Faibussowitsch PetscCall(PetscSFLinkPackLeafData(sf, link, PETSCSF_REMOTE, leafdata)); 429566063dSJacob Faibussowitsch PetscCall(PetscSFLinkCopyLeafBufferInCaseNotUseGpuAwareMPI(sf, link, PETSC_TRUE /* device2host before sending */)); 439566063dSJacob Faibussowitsch PetscCall(PetscObjectGetComm((PetscObject)sf, &comm)); 449566063dSJacob Faibussowitsch PetscCall(PetscMPIIntCast(sf->nroots, &recvcount)); 459566063dSJacob Faibussowitsch PetscCall(PetscSFLinkGetMPIBuffersAndRequests(sf, link, PETSCSF_LEAF2ROOT, &rootbuf, &leafbuf, &req, NULL)); 469566063dSJacob Faibussowitsch PetscCall(PetscSFLinkSyncStreamBeforeCallMPI(sf, link, PETSCSF_LEAF2ROOT)); 479566063dSJacob Faibussowitsch PetscCallMPI(MPIU_Iscatterv(leafbuf, dat->recvcounts, dat->displs, unit, rootbuf, recvcount, unit, 0, comm, req)); 48*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 49dd5b3ca6SJunchao Zhang } 50dd5b3ca6SJunchao Zhang 51d71ae5a4SJacob Faibussowitsch PETSC_INTERN PetscErrorCode PetscSFFetchAndOpBegin_Gatherv(PetscSF sf, MPI_Datatype unit, PetscMemType rootmtype, void *rootdata, PetscMemType leafmtype, const void *leafdata, void *leafupdate, MPI_Op op) 52d71ae5a4SJacob Faibussowitsch { 53dd5b3ca6SJunchao Zhang PetscFunctionBegin; 54dd5b3ca6SJunchao Zhang /* In Gatherv, each root only has one leaf. So we just need to bcast rootdata to leafupdate and then reduce leafdata to rootdata */ 559566063dSJacob Faibussowitsch PetscCall(PetscSFBcastBegin(sf, unit, rootdata, leafupdate, MPI_REPLACE)); 569566063dSJacob Faibussowitsch PetscCall(PetscSFBcastEnd(sf, unit, rootdata, leafupdate, MPI_REPLACE)); 579566063dSJacob Faibussowitsch PetscCall(PetscSFReduceBegin(sf, unit, leafdata, rootdata, op)); 58*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 59dd5b3ca6SJunchao Zhang } 60dd5b3ca6SJunchao Zhang 61d71ae5a4SJacob Faibussowitsch PETSC_INTERN PetscErrorCode PetscSFCreate_Gatherv(PetscSF sf) 62d71ae5a4SJacob Faibussowitsch { 63dd5b3ca6SJunchao Zhang PetscSF_Gatherv *dat = (PetscSF_Gatherv *)sf->data; 64dd5b3ca6SJunchao Zhang 65dd5b3ca6SJunchao Zhang PetscFunctionBegin; 66ad227feaSJunchao Zhang sf->ops->BcastEnd = PetscSFBcastEnd_Basic; 67cd620004SJunchao Zhang sf->ops->ReduceEnd = PetscSFReduceEnd_Basic; 68cd620004SJunchao Zhang 69dd5b3ca6SJunchao Zhang /* Inherit from Allgatherv */ 70dd5b3ca6SJunchao Zhang sf->ops->SetUp = PetscSFSetUp_Allgatherv; 71dd5b3ca6SJunchao Zhang sf->ops->Reset = PetscSFReset_Allgatherv; 72dd5b3ca6SJunchao Zhang sf->ops->Destroy = PetscSFDestroy_Allgatherv; 73dd5b3ca6SJunchao Zhang sf->ops->GetGraph = PetscSFGetGraph_Allgatherv; 74dd5b3ca6SJunchao Zhang sf->ops->GetLeafRanks = PetscSFGetLeafRanks_Allgatherv; 75dd5b3ca6SJunchao Zhang sf->ops->GetRootRanks = PetscSFGetRootRanks_Allgatherv; 76dd5b3ca6SJunchao Zhang sf->ops->FetchAndOpEnd = PetscSFFetchAndOpEnd_Allgatherv; 77dd5b3ca6SJunchao Zhang sf->ops->CreateLocalSF = PetscSFCreateLocalSF_Allgatherv; 78dd5b3ca6SJunchao Zhang 79dd5b3ca6SJunchao Zhang /* Gatherv stuff */ 80ad227feaSJunchao Zhang sf->ops->BcastBegin = PetscSFBcastBegin_Gatherv; 81dd5b3ca6SJunchao Zhang sf->ops->ReduceBegin = PetscSFReduceBegin_Gatherv; 82dd5b3ca6SJunchao Zhang sf->ops->FetchAndOpBegin = PetscSFFetchAndOpBegin_Gatherv; 83dd5b3ca6SJunchao Zhang 844dfa11a4SJacob Faibussowitsch PetscCall(PetscNew(&dat)); 85dd5b3ca6SJunchao Zhang sf->data = (void *)dat; 86*3ba16761SJacob Faibussowitsch PetscFunctionReturn(PETSC_SUCCESS); 87dd5b3ca6SJunchao Zhang } 88