1dd5b3ca6SJunchao Zhang #include <../src/vec/is/sf/impls/basic/allgatherv/sfallgatherv.h> 2dd5b3ca6SJunchao Zhang 3dd5b3ca6SJunchao Zhang /* 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 */ 4dd5b3ca6SJunchao Zhang typedef PetscSF_Allgatherv PetscSF_Allgather; 5dd5b3ca6SJunchao Zhang 6ad227feaSJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFBcastBegin_Gather(PetscSF,MPI_Datatype,PetscMemType,const void*,PetscMemType,void*,MPI_Op); 7dd5b3ca6SJunchao Zhang 8cd620004SJunchao Zhang PetscErrorCode PetscSFSetUp_Allgather(PetscSF sf) 9cd620004SJunchao Zhang { 10cd620004SJunchao Zhang PetscInt i; 11cd620004SJunchao Zhang PetscSF_Allgather *dat = (PetscSF_Allgather*)sf->data; 12cd620004SJunchao Zhang 13cd620004SJunchao Zhang PetscFunctionBegin; 14cd620004SJunchao Zhang for (i=PETSCSF_LOCAL; i<=PETSCSF_REMOTE; i++) { 15cd620004SJunchao Zhang sf->leafbuflen[i] = 0; 16cd620004SJunchao Zhang sf->leafstart[i] = 0; 17cd620004SJunchao Zhang sf->leafcontig[i] = PETSC_TRUE; 18cd620004SJunchao Zhang sf->leafdups[i] = PETSC_FALSE; 19cd620004SJunchao Zhang dat->rootbuflen[i] = 0; 20cd620004SJunchao Zhang dat->rootstart[i] = 0; 21cd620004SJunchao Zhang dat->rootcontig[i] = PETSC_TRUE; 22cd620004SJunchao Zhang dat->rootdups[i] = PETSC_FALSE; 23cd620004SJunchao Zhang } 24cd620004SJunchao Zhang 25cd620004SJunchao Zhang sf->leafbuflen[PETSCSF_REMOTE] = sf->nleaves; 26cd620004SJunchao Zhang dat->rootbuflen[PETSCSF_REMOTE] = sf->nroots; 27cd620004SJunchao Zhang sf->persistent = PETSC_FALSE; 28cd620004SJunchao Zhang sf->nleafreqs = 0; /* MPI collectives only need one request. We treat it as a root request. */ 29cd620004SJunchao Zhang dat->nrootreqs = 1; 30cd620004SJunchao Zhang PetscFunctionReturn(0); 31cd620004SJunchao Zhang } 32cd620004SJunchao Zhang 33ad227feaSJunchao Zhang static PetscErrorCode PetscSFBcastBegin_Allgather(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata,MPI_Op op) 34dd5b3ca6SJunchao Zhang { 35dd5b3ca6SJunchao Zhang PetscErrorCode ierr; 36cd620004SJunchao Zhang PetscSFLink link; 37dd5b3ca6SJunchao Zhang PetscMPIInt sendcount; 38dd5b3ca6SJunchao Zhang MPI_Comm comm; 39cd620004SJunchao Zhang void *rootbuf = NULL,*leafbuf = NULL; /* buffer seen by MPI */ 40cd620004SJunchao Zhang MPI_Request *req; 41dd5b3ca6SJunchao Zhang 42dd5b3ca6SJunchao Zhang PetscFunctionBegin; 43cd620004SJunchao Zhang ierr = PetscSFLinkCreate(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op,PETSCSF_BCAST,&link);CHKERRQ(ierr); 44cd620004SJunchao Zhang ierr = PetscSFLinkPackRootData(sf,link,PETSCSF_REMOTE,rootdata);CHKERRQ(ierr); 4571438e86SJunchao Zhang ierr = PetscSFLinkCopyRootBufferInCaseNotUseGpuAwareMPI(sf,link,PETSC_TRUE/* device2host before sending */);CHKERRQ(ierr); 46dd5b3ca6SJunchao Zhang ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr); 47dd5b3ca6SJunchao Zhang ierr = PetscMPIIntCast(sf->nroots,&sendcount);CHKERRQ(ierr); 48cd620004SJunchao Zhang ierr = PetscSFLinkGetMPIBuffersAndRequests(sf,link,PETSCSF_ROOT2LEAF,&rootbuf,&leafbuf,&req,NULL);CHKERRQ(ierr); 4971438e86SJunchao Zhang ierr = PetscSFLinkSyncStreamBeforeCallMPI(sf,link,PETSCSF_ROOT2LEAF);CHKERRQ(ierr); 50820f2d46SBarry Smith ierr = MPIU_Iallgather(rootbuf,sendcount,unit,leafbuf,sendcount,unit,comm,req);CHKERRMPI(ierr); 51855db38dSJunchao Zhang PetscFunctionReturn(0); 52855db38dSJunchao Zhang } 53855db38dSJunchao Zhang 54855db38dSJunchao Zhang static PetscErrorCode PetscSFReduceBegin_Allgather(PetscSF sf,MPI_Datatype unit,PetscMemType leafmtype,const void *leafdata,PetscMemType rootmtype,void *rootdata,MPI_Op op) 55855db38dSJunchao Zhang { 56855db38dSJunchao Zhang PetscErrorCode ierr; 57cd620004SJunchao Zhang PetscSFLink link; 58855db38dSJunchao Zhang PetscInt rstart; 59855db38dSJunchao Zhang MPI_Comm comm; 60cd620004SJunchao Zhang PetscMPIInt rank,count,recvcount; 61cd620004SJunchao Zhang void *rootbuf = NULL,*leafbuf = NULL; /* buffer seen by MPI */ 62cd620004SJunchao Zhang PetscSF_Allgather *dat = (PetscSF_Allgather*)sf->data; 63cd620004SJunchao Zhang MPI_Request *req; 64855db38dSJunchao Zhang 65855db38dSJunchao Zhang PetscFunctionBegin; 66cd620004SJunchao Zhang ierr = PetscSFLinkCreate(sf,unit,rootmtype,rootdata,leafmtype,leafdata,op,PETSCSF_REDUCE,&link);CHKERRQ(ierr); 6783df288dSJunchao Zhang if (op == MPI_REPLACE) { 68855db38dSJunchao Zhang /* REPLACE is only meaningful when all processes have the same leafdata to reduce. Therefore copy from local leafdata is fine */ 69855db38dSJunchao Zhang ierr = PetscLayoutGetRange(sf->map,&rstart,NULL);CHKERRQ(ierr); 7020c24465SJunchao Zhang ierr = (*link->Memcpy)(link,rootmtype,rootdata,leafmtype,(const char*)leafdata+(size_t)rstart*link->unitbytes,(size_t)sf->nroots*link->unitbytes);CHKERRQ(ierr); 7171438e86SJunchao Zhang if (PetscMemTypeDevice(leafmtype) && PetscMemTypeHost(rootmtype)) {ierr = (*link->SyncStream)(link);CHKERRQ(ierr);} /* Sync the device to host memcpy */ 72dd5b3ca6SJunchao Zhang } else { 73cd620004SJunchao Zhang ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr); 74ffc4695bSBarry Smith ierr = MPI_Comm_rank(comm,&rank);CHKERRMPI(ierr); 75cd620004SJunchao Zhang ierr = PetscSFLinkPackLeafData(sf,link,PETSCSF_REMOTE,leafdata);CHKERRQ(ierr); 7671438e86SJunchao Zhang ierr = PetscSFLinkCopyLeafBufferInCaseNotUseGpuAwareMPI(sf,link,PETSC_TRUE/* device2host before sending */);CHKERRQ(ierr); 77cd620004SJunchao Zhang ierr = PetscSFLinkGetMPIBuffersAndRequests(sf,link,PETSCSF_LEAF2ROOT,&rootbuf,&leafbuf,&req,NULL);CHKERRQ(ierr); 78cd620004SJunchao Zhang ierr = PetscMPIIntCast(dat->rootbuflen[PETSCSF_REMOTE],&recvcount);CHKERRQ(ierr); 79dd400576SPatrick Sanan if (rank == 0 && !link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi]) { 8020c24465SJunchao Zhang ierr = PetscSFMalloc(sf,link->leafmtype_mpi,sf->leafbuflen[PETSCSF_REMOTE]*link->unitbytes,(void**)&link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi]);CHKERRQ(ierr); 81cd620004SJunchao Zhang } 82dd400576SPatrick Sanan if (rank == 0 && link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi] == leafbuf) leafbuf = MPI_IN_PLACE; 83cd620004SJunchao Zhang ierr = PetscMPIIntCast(sf->nleaves*link->bs,&count);CHKERRQ(ierr); 8471438e86SJunchao Zhang ierr = PetscSFLinkSyncStreamBeforeCallMPI(sf,link,PETSCSF_LEAF2ROOT);CHKERRQ(ierr); 8571438e86SJunchao Zhang ierr = MPI_Reduce(leafbuf,link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi],count,link->basicunit,op,0,comm);CHKERRMPI(ierr); /* Must do reduce with MPI builltin datatype basicunit */ 8671438e86SJunchao Zhang ierr = MPIU_Iscatter(link->leafbuf_alloc[PETSCSF_REMOTE][link->leafmtype_mpi],recvcount,unit,rootbuf,recvcount,unit,0/*rank 0*/,comm,req);CHKERRMPI(ierr); 87dd5b3ca6SJunchao Zhang } 88dd5b3ca6SJunchao Zhang PetscFunctionReturn(0); 89dd5b3ca6SJunchao Zhang } 90dd5b3ca6SJunchao Zhang 91eb02082bSJunchao Zhang static PetscErrorCode PetscSFBcastToZero_Allgather(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata) 92dd5b3ca6SJunchao Zhang { 93dd5b3ca6SJunchao Zhang PetscErrorCode ierr; 94cd620004SJunchao Zhang PetscSFLink link; 95855db38dSJunchao Zhang PetscMPIInt rank; 96dd5b3ca6SJunchao Zhang 97dd5b3ca6SJunchao Zhang PetscFunctionBegin; 98ad227feaSJunchao Zhang ierr = PetscSFBcastBegin_Gather(sf,unit,rootmtype,rootdata,leafmtype,leafdata,MPI_REPLACE);CHKERRQ(ierr); 99cd620004SJunchao Zhang ierr = PetscSFLinkGetInUse(sf,unit,rootdata,leafdata,PETSC_OWN_POINTER,&link);CHKERRQ(ierr); 10071438e86SJunchao Zhang ierr = PetscSFLinkFinishCommunication(sf,link,PETSCSF_ROOT2LEAF);CHKERRQ(ierr); 101ffc4695bSBarry Smith ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRMPI(ierr); 102dd400576SPatrick Sanan if (rank == 0 && PetscMemTypeDevice(leafmtype) && !sf->use_gpu_aware_mpi) { 10320c24465SJunchao Zhang 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); 104855db38dSJunchao Zhang } 105cd620004SJunchao Zhang ierr = PetscSFLinkReclaim(sf,&link);CHKERRQ(ierr); 106dd5b3ca6SJunchao Zhang PetscFunctionReturn(0); 107dd5b3ca6SJunchao Zhang } 108dd5b3ca6SJunchao Zhang 109dd5b3ca6SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFCreate_Allgather(PetscSF sf) 110dd5b3ca6SJunchao Zhang { 111dd5b3ca6SJunchao Zhang PetscErrorCode ierr; 112dd5b3ca6SJunchao Zhang PetscSF_Allgather *dat = (PetscSF_Allgather*)sf->data; 113dd5b3ca6SJunchao Zhang 114dd5b3ca6SJunchao Zhang PetscFunctionBegin; 115ad227feaSJunchao Zhang sf->ops->BcastEnd = PetscSFBcastEnd_Basic; 116*9319200aSJunchao Zhang sf->ops->ReduceEnd = PetscSFReduceEnd_Allgatherv; 117dd5b3ca6SJunchao Zhang 118dd5b3ca6SJunchao Zhang /* Inherit from Allgatherv */ 119dd5b3ca6SJunchao Zhang sf->ops->Reset = PetscSFReset_Allgatherv; 120dd5b3ca6SJunchao Zhang sf->ops->Destroy = PetscSFDestroy_Allgatherv; 121dd5b3ca6SJunchao Zhang sf->ops->FetchAndOpBegin = PetscSFFetchAndOpBegin_Allgatherv; 122dd5b3ca6SJunchao Zhang sf->ops->FetchAndOpEnd = PetscSFFetchAndOpEnd_Allgatherv; 123dd5b3ca6SJunchao Zhang sf->ops->GetRootRanks = PetscSFGetRootRanks_Allgatherv; 124dd5b3ca6SJunchao Zhang sf->ops->CreateLocalSF = PetscSFCreateLocalSF_Allgatherv; 125dd5b3ca6SJunchao Zhang sf->ops->GetGraph = PetscSFGetGraph_Allgatherv; 126dd5b3ca6SJunchao Zhang sf->ops->GetLeafRanks = PetscSFGetLeafRanks_Allgatherv; 127dd5b3ca6SJunchao Zhang 128dd5b3ca6SJunchao Zhang /* Allgather stuff */ 129cd620004SJunchao Zhang sf->ops->SetUp = PetscSFSetUp_Allgather; 130ad227feaSJunchao Zhang sf->ops->BcastBegin = PetscSFBcastBegin_Allgather; 131dd5b3ca6SJunchao Zhang sf->ops->ReduceBegin = PetscSFReduceBegin_Allgather; 132dd5b3ca6SJunchao Zhang sf->ops->BcastToZero = PetscSFBcastToZero_Allgather; 133dd5b3ca6SJunchao Zhang 134dd5b3ca6SJunchao Zhang ierr = PetscNewLog(sf,&dat);CHKERRQ(ierr); 135dd5b3ca6SJunchao Zhang sf->data = (void*)dat; 136dd5b3ca6SJunchao Zhang PetscFunctionReturn(0); 137dd5b3ca6SJunchao Zhang } 138