1dd5b3ca6SJunchao Zhang #include <../src/vec/is/sf/impls/basic/allgatherv/sfallgatherv.h> 2dd5b3ca6SJunchao Zhang 3dd5b3ca6SJunchao Zhang #define PetscSFPackGet_Allgather PetscSFPackGet_Allgatherv 4dd5b3ca6SJunchao Zhang 5dd5b3ca6SJunchao 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 */ 6dd5b3ca6SJunchao Zhang typedef PetscSF_Allgatherv PetscSF_Allgather; 7dd5b3ca6SJunchao Zhang 8eb02082bSJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFBcastAndOpBegin_Gather(PetscSF,MPI_Datatype,PetscMemType,const void*,PetscMemType,void*,MPI_Op); 9dd5b3ca6SJunchao Zhang 10eb02082bSJunchao Zhang static PetscErrorCode PetscSFBcastAndOpBegin_Allgather(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata,MPI_Op op) 11dd5b3ca6SJunchao Zhang { 12dd5b3ca6SJunchao Zhang PetscErrorCode ierr; 13eb02082bSJunchao Zhang PetscSFPack link; 14dd5b3ca6SJunchao Zhang PetscMPIInt sendcount; 15dd5b3ca6SJunchao Zhang MPI_Comm comm; 16*855db38dSJunchao Zhang const void *rootbuf_mpi; /* buffer used by MPI */ 17*855db38dSJunchao Zhang void *leafbuf_mpi; 18*855db38dSJunchao Zhang PetscMemType rootmtype_mpi,leafmtype_mpi; 19dd5b3ca6SJunchao Zhang 20dd5b3ca6SJunchao Zhang PetscFunctionBegin; 21eb02082bSJunchao Zhang ierr = PetscSFPackGet_Allgather(sf,unit,rootmtype,rootdata,leafmtype,leafdata,&link);CHKERRQ(ierr); 22dd5b3ca6SJunchao Zhang ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr); 23dd5b3ca6SJunchao Zhang ierr = PetscMPIIntCast(sf->nroots,&sendcount);CHKERRQ(ierr); 24*855db38dSJunchao Zhang ierr = PetscSFBcastPrepareMPIBuffers_Allgatherv(sf,link,op,&rootmtype_mpi,&rootbuf_mpi,&leafmtype_mpi,&leafbuf_mpi);CHKERRQ(ierr); 25*855db38dSJunchao Zhang ierr = MPIU_Iallgather(rootbuf_mpi,sendcount,unit,leafbuf_mpi,sendcount,unit,comm,link->rootreqs[PETSCSF_ROOT2LEAF_BCAST][rootmtype_mpi]);CHKERRQ(ierr); 26*855db38dSJunchao Zhang PetscFunctionReturn(0); 27*855db38dSJunchao Zhang } 28*855db38dSJunchao Zhang 29*855db38dSJunchao Zhang static PetscErrorCode PetscSFReduceBegin_Allgather(PetscSF sf,MPI_Datatype unit,PetscMemType leafmtype,const void *leafdata,PetscMemType rootmtype,void *rootdata,MPI_Op op) 30*855db38dSJunchao Zhang { 31*855db38dSJunchao Zhang PetscErrorCode ierr; 32*855db38dSJunchao Zhang PetscSFPack link; 33*855db38dSJunchao Zhang PetscMPIInt sendcount; 34*855db38dSJunchao Zhang PetscInt rstart; 35*855db38dSJunchao Zhang MPI_Comm comm; 36*855db38dSJunchao Zhang const void *leafbuf_mpi; 37*855db38dSJunchao Zhang void *rootbuf_mpi; 38*855db38dSJunchao Zhang PetscMemType leafmtype_mpi,rootmtype_mpi; 39*855db38dSJunchao Zhang 40*855db38dSJunchao Zhang PetscFunctionBegin; 41*855db38dSJunchao Zhang ierr = PetscSFPackGet_Allgather(sf,unit,rootmtype,rootdata,leafmtype,leafdata,&link);CHKERRQ(ierr); 42*855db38dSJunchao Zhang ierr = PetscObjectGetComm((PetscObject)sf,&comm);CHKERRQ(ierr); 43dd5b3ca6SJunchao Zhang 44dd5b3ca6SJunchao Zhang if (op == MPIU_REPLACE) { 45*855db38dSJunchao Zhang /* REPLACE is only meaningful when all processes have the same leafdata to reduce. Therefore copy from local leafdata is fine */ 46*855db38dSJunchao Zhang ierr = PetscLayoutGetRange(sf->map,&rstart,NULL);CHKERRQ(ierr); 47*855db38dSJunchao Zhang ierr = PetscMemcpyWithMemType(rootmtype,leafmtype,rootdata,(const char*)leafdata+(size_t)rstart*link->unitbytes,(size_t)sf->nroots*link->unitbytes);CHKERRQ(ierr); 48dd5b3ca6SJunchao Zhang } else { 49*855db38dSJunchao Zhang ierr = PetscMPIIntCast(sf->nroots,&sendcount);CHKERRQ(ierr); 50*855db38dSJunchao Zhang ierr = PetscSFReducePrepareMPIBuffers_Allgatherv(sf,link,op,&rootmtype_mpi,&rootbuf_mpi,&leafmtype_mpi,&leafbuf_mpi);CHKERRQ(ierr); 51*855db38dSJunchao Zhang ierr = MPIU_Iscatter(leafbuf_mpi,sendcount,unit,rootbuf_mpi,sendcount,unit,0/*rank 0*/,comm,link->rootreqs[PETSCSF_LEAF2ROOT_REDUCE][rootmtype_mpi]);CHKERRQ(ierr); 52dd5b3ca6SJunchao Zhang } 53dd5b3ca6SJunchao Zhang PetscFunctionReturn(0); 54dd5b3ca6SJunchao Zhang } 55dd5b3ca6SJunchao Zhang 56eb02082bSJunchao Zhang static PetscErrorCode PetscSFBcastToZero_Allgather(PetscSF sf,MPI_Datatype unit,PetscMemType rootmtype,const void *rootdata,PetscMemType leafmtype,void *leafdata) 57dd5b3ca6SJunchao Zhang { 58dd5b3ca6SJunchao Zhang PetscErrorCode ierr; 59eb02082bSJunchao Zhang PetscSFPack link; 60*855db38dSJunchao Zhang PetscMPIInt rank; 61dd5b3ca6SJunchao Zhang 62dd5b3ca6SJunchao Zhang PetscFunctionBegin; 63eb02082bSJunchao Zhang ierr = PetscSFBcastAndOpBegin_Gather(sf,unit,rootmtype,rootdata,leafmtype,leafdata,MPIU_REPLACE);CHKERRQ(ierr); 64eb02082bSJunchao Zhang ierr = PetscSFPackGetInUse(sf,unit,rootdata,leafdata,PETSC_OWN_POINTER,&link);CHKERRQ(ierr); 65*855db38dSJunchao Zhang ierr = PetscSFPackWaitall(link,PETSCSF_ROOT2LEAF_BCAST);CHKERRQ(ierr); 66*855db38dSJunchao Zhang ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)sf),&rank);CHKERRQ(ierr); 67*855db38dSJunchao Zhang if (!rank && leafmtype == PETSC_MEMTYPE_DEVICE && !use_gpu_aware_mpi) { 68*855db38dSJunchao Zhang ierr = PetscMemcpyWithMemType(PETSC_MEMTYPE_DEVICE,PETSC_MEMTYPE_HOST,leafdata,link->leafbuf[PETSC_MEMTYPE_HOST],link->leafbuflen*link->unitbytes);CHKERRQ(ierr); 69*855db38dSJunchao Zhang } 70eb02082bSJunchao Zhang ierr = PetscSFPackReclaim(sf,&link);CHKERRQ(ierr); 71dd5b3ca6SJunchao Zhang PetscFunctionReturn(0); 72dd5b3ca6SJunchao Zhang } 73dd5b3ca6SJunchao Zhang 74dd5b3ca6SJunchao Zhang PETSC_INTERN PetscErrorCode PetscSFCreate_Allgather(PetscSF sf) 75dd5b3ca6SJunchao Zhang { 76dd5b3ca6SJunchao Zhang PetscErrorCode ierr; 77dd5b3ca6SJunchao Zhang PetscSF_Allgather *dat = (PetscSF_Allgather*)sf->data; 78dd5b3ca6SJunchao Zhang 79dd5b3ca6SJunchao Zhang PetscFunctionBegin; 80dd5b3ca6SJunchao Zhang 81dd5b3ca6SJunchao Zhang /* Inherit from Allgatherv */ 82dd5b3ca6SJunchao Zhang sf->ops->Reset = PetscSFReset_Allgatherv; 83dd5b3ca6SJunchao Zhang sf->ops->Destroy = PetscSFDestroy_Allgatherv; 84dd5b3ca6SJunchao Zhang sf->ops->BcastAndOpEnd = PetscSFBcastAndOpEnd_Allgatherv; 85dd5b3ca6SJunchao Zhang sf->ops->ReduceEnd = PetscSFReduceEnd_Allgatherv; 86dd5b3ca6SJunchao Zhang sf->ops->FetchAndOpBegin = PetscSFFetchAndOpBegin_Allgatherv; 87dd5b3ca6SJunchao Zhang sf->ops->FetchAndOpEnd = PetscSFFetchAndOpEnd_Allgatherv; 88dd5b3ca6SJunchao Zhang sf->ops->GetRootRanks = PetscSFGetRootRanks_Allgatherv; 89dd5b3ca6SJunchao Zhang sf->ops->CreateLocalSF = PetscSFCreateLocalSF_Allgatherv; 90dd5b3ca6SJunchao Zhang sf->ops->GetGraph = PetscSFGetGraph_Allgatherv; 91dd5b3ca6SJunchao Zhang sf->ops->GetLeafRanks = PetscSFGetLeafRanks_Allgatherv; 92dd5b3ca6SJunchao Zhang 93dd5b3ca6SJunchao Zhang /* Allgather stuff */ 94dd5b3ca6SJunchao Zhang sf->ops->BcastAndOpBegin = PetscSFBcastAndOpBegin_Allgather; 95dd5b3ca6SJunchao Zhang sf->ops->ReduceBegin = PetscSFReduceBegin_Allgather; 96dd5b3ca6SJunchao Zhang sf->ops->BcastToZero = PetscSFBcastToZero_Allgather; 97dd5b3ca6SJunchao Zhang 98dd5b3ca6SJunchao Zhang ierr = PetscNewLog(sf,&dat);CHKERRQ(ierr); 99dd5b3ca6SJunchao Zhang sf->data = (void*)dat; 100dd5b3ca6SJunchao Zhang PetscFunctionReturn(0); 101dd5b3ca6SJunchao Zhang } 102