1 #include <../src/vec/is/sf/impls/basic/allgatherv/sfallgatherv.h> 2 3 /* Reuse the type. The difference is some fields (i.e., displs, recvcounts) are not used in Allgather on rank != 0, which is not a big deal */ 4 typedef PetscSF_Allgatherv PetscSF_Allgather; 5 6 PETSC_INTERN PetscErrorCode PetscSFBcastAndOpBegin_Gather(PetscSF,MPI_Datatype,PetscMemType,const void*,PetscMemType,void*,MPI_Op); 7 8 PetscErrorCode PetscSFSetUp_Allgather(PetscSF sf) 9 { 10 PetscInt i; 11 PetscSF_Allgather *dat = (PetscSF_Allgather*)sf->data; 12 13 PetscFunctionBegin; 14 for (i=PETSCSF_LOCAL; i<=PETSCSF_REMOTE; i++) { 15 sf->leafbuflen[i] = 0; 16 sf->leafstart[i] = 0; 17 sf->leafcontig[i] = PETSC_TRUE; 18 sf->leafdups[i] = PETSC_FALSE; 19 dat->rootbuflen[i] = 0; 20 dat->rootstart[i] = 0; 21 dat->rootcontig[i] = PETSC_TRUE; 22 dat->rootdups[i] = PETSC_FALSE; 23 } 24 25 sf->leafbuflen[PETSCSF_REMOTE] = sf->nleaves; 26 dat->rootbuflen[PETSCSF_REMOTE] = sf->nroots; 27 sf->persistent = PETSC_FALSE; 28 sf->nleafreqs = 0; /* MPI collectives only need one request. We treat it as a root request. */ 29 dat->nrootreqs = 1; 30 PetscFunctionReturn(0); 31 } 32 33 static PetscErrorCode PetscSFBcastAndOpBegin_Allgather(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata,MPI_Op op) 34 { 35 PetscErrorCode ierr; 36 PetscSFLink link; 37 PetscMPIInt sendcount; 38 MPI_Comm comm; 39 void *rootbuf = NULL,*leafbuf = NULL; /* buffer seen by MPI */ 40 MPI_Request *req; 41 42 PetscFunctionBegin; 43 ierr = PetscSFLinkCreate(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op,PETSCSF_BCAST,&link);CHKERRQ(ierr); 44 ierr = PetscSFLinkPackRootData(sf,link,PETSCSF_REMOTE,rootdata);CHKERRQ(ierr); 45 ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr); 46 ierr = PetscMPIIntCast(sf->nroots,&sendcount);CHKERRQ(ierr); 47 ierr = PetscSFLinkGetMPIBuffersAndRequests(sf,link,PETSCSF_ROOT2LEAF,&rootbuf,&leafbuf,&req,NULL);CHKERRQ(ierr); 48 ierr = MPIU_Iallgather(rootbuf,sendcount,unit,leafbuf,sendcount,unit,comm,req);CHKERRQ(ierr); 49 PetscFunctionReturn(0); 50 } 51 52 static PetscErrorCode PetscSFReduceBegin_Allgather(PetscSF sf,MPI_Datatype unit,PetscMemType leafmtype,const void *leafdata,PetscMemType rootmtype,void *rootdata,MPI_Op op) 53 { 54 PetscErrorCode ierr; 55 PetscSFLink link; 56 PetscInt rstart; 57 MPI_Comm comm; 58 PetscMPIInt rank,count,recvcount; 59 void *rootbuf = NULL,*leafbuf = NULL; /* buffer seen by MPI */ 60 PetscSF_Allgather *dat = (PetscSF_Allgather*)sf->data; 61 MPI_Request *req; 62 63 PetscFunctionBegin; 64 ierr = PetscSFLinkCreate(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op,PETSCSF_REDUCE,&link);CHKERRQ(ierr); 65 if (op == MPIU_REPLACE) { 66 /* REPLACE is only meaningful when all processes have the same leafdata to reduce. Therefore copy from local leafdata is fine */ 67 ierr = PetscLayoutGetRange(sf->map,&rstart,NULL);CHKERRQ(ierr); 68 ierr = (*link->Memcpy)(link,rootmtype,rootdata,leafmtype,(const char*)leafdata+(size_t)rstart*link->unitbytes,(size_t)sf->nroots*link->unitbytes);CHKERRQ(ierr); 69 #if defined(PETSC_HAVE_DEVICE) 70 if (PetscMemTypeHost(rootmtype) && PetscMemTypeDevice(leafmtype)) {ierr = (*link->d_SyncStream)(link);CHKERRQ(ierr);} 71 #endif 72 } else { 73 ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr); 74 ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 75 ierr = PetscSFLinkPackLeafData(sf,link,PETSCSF_REMOTE,leafdata);CHKERRQ(ierr); 76 ierr = PetscSFLinkGetMPIBuffersAndRequests(sf,link,PETSCSF_LEAF2ROOT,&rootbuf,&leafbuf,&req,NULL);CHKERRQ(ierr); 77 ierr = PetscMPIIntCast(dat->rootbuflen[PETSCSF_REMOTE],&recvcount);CHKERRQ(ierr); 78 if (!rank && !link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi]) { 79 ierr = PetscSFMalloc(sf,link->leafmtype_mpi,sf->leafbuflen[PETSCSF_REMOTE]*link->unitbytes,(void**)&link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi]);CHKERRQ(ierr); 80 } 81 if (!rank && link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi] == leafbuf) leafbuf = MPI_IN_PLACE; 82 ierr = PetscMPIIntCast(sf->nleaves*link->bs,&count);CHKERRQ(ierr); 83 ierr = MPI_Reduce(leafbuf,link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi],count,link->basicunit,op,0,comm);CHKERRMPI(ierr); 84 ierr = MPIU_Iscatter(link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi],recvcount,unit,rootbuf,recvcount,unit,0/*rank 0*/,comm,req);CHKERRQ(ierr); 85 } 86 PetscFunctionReturn(0); 87 } 88 89 static PetscErrorCode PetscSFBcastToZero_Allgather(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata) 90 { 91 PetscErrorCode ierr; 92 PetscSFLink link; 93 PetscMPIInt rank; 94 95 PetscFunctionBegin; 96 ierr = PetscSFBcastAndOpBegin_Gather(sf,unit,rootmtype,rootdata,leafmtype,leafdata,MPIU_REPLACE);CHKERRQ(ierr); 97 ierr = PetscSFLinkGetInUse(sf,unit,rootdata,leafdata,PETSC_OWN_POINTER,&link);CHKERRQ(ierr); 98 ierr = PetscSFLinkMPIWaitall(sf,link,PETSCSF_ROOT2LEAF);CHKERRQ(ierr); 99 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRMPI(ierr); 100 if (!rank && leafmtype == PETSC_MEMTYPE_DEVICE && !sf->use_gpu_aware_mpi) { 101 ierr = (*link->Memcpy)(link,PETSC_MEMTYPE_DEVICE,leafdata,PETSC_MEMTYPE_HOST,link->leafbuf[PETSCSF_REMOTE][PETSC_MEMTYPE_HOST],sf->leafbuflen[PETSCSF_REMOTE]*link->unitbytes);CHKERRQ(ierr); 102 } 103 ierr = PetscSFLinkReclaim(sf,&link);CHKERRQ(ierr); 104 PetscFunctionReturn(0); 105 } 106 107 PETSC_INTERN PetscErrorCode PetscSFCreate_Allgather(PetscSF sf) 108 { 109 PetscErrorCode ierr; 110 PetscSF_Allgather *dat = (PetscSF_Allgather*)sf->data; 111 112 PetscFunctionBegin; 113 sf->ops->BcastAndOpEnd = PetscSFBcastAndOpEnd_Basic; 114 sf->ops->ReduceEnd = PetscSFReduceEnd_Basic; 115 116 /* Inherit from Allgatherv */ 117 sf->ops->Reset = PetscSFReset_Allgatherv; 118 sf->ops->Destroy = PetscSFDestroy_Allgatherv; 119 sf->ops->FetchAndOpBegin = PetscSFFetchAndOpBegin_Allgatherv; 120 sf->ops->FetchAndOpEnd = PetscSFFetchAndOpEnd_Allgatherv; 121 sf->ops->GetRootRanks = PetscSFGetRootRanks_Allgatherv; 122 sf->ops->CreateLocalSF = PetscSFCreateLocalSF_Allgatherv; 123 sf->ops->GetGraph = PetscSFGetGraph_Allgatherv; 124 sf->ops->GetLeafRanks = PetscSFGetLeafRanks_Allgatherv; 125 126 /* Allgather stuff */ 127 sf->ops->SetUp = PetscSFSetUp_Allgather; 128 sf->ops->BcastAndOpBegin = PetscSFBcastAndOpBegin_Allgather; 129 sf->ops->ReduceBegin = PetscSFReduceBegin_Allgather; 130 sf->ops->BcastToZero = PetscSFBcastToZero_Allgather; 131 132 ierr = PetscNewLog(sf,&dat);CHKERRQ(ierr); 133 sf->data = (void*)dat; 134 PetscFunctionReturn(0); 135 } 136