78   PetscInt       i, nRemoteRootRanks, nRemoteLeafRanks;  in PetscSFSetUp_Basic_NVSHMEM()  local
88   nRemoteRootRanks      = sf->nranks - sf->ndranks;  in PetscSFSetUp_Basic_NVSHMEM()
90   sf->nRemoteRootRanks  = nRemoteRootRanks;  in PetscSFSetUp_Basic_NVSHMEM()
93   PetscCall(PetscMalloc2(nRemoteLeafRanks, &rootreqs, nRemoteRootRanks, &leafreqs));  in PetscSFSetUp_Basic_NVSHMEM()
95   stmp[0] = nRemoteRootRanks;  in PetscSFSetUp_Basic_NVSHMEM()
110   PetscCall(PetscMalloc2(nRemoteRootRanks, &sf->rootsigdisp, nRemoteRootRanks, &sf->rootbufdisp));  in PetscSFSetUp_Basic_NVSHMEM()
111 …for (i = 0; i < nRemoteRootRanks; i++) PetscCallMPI(MPIU_Irecv(&sf->rootsigdisp[i], 1, MPIU_INT, s…  in PetscSFSetUp_Basic_NVSHMEM()
113   PetscCallMPI(MPI_Waitall(nRemoteRootRanks, leafreqs, MPI_STATUSES_IGNORE));  in PetscSFSetUp_Basic_NVSHMEM()
115 …for (i = 0; i < nRemoteRootRanks; i++) PetscCallMPI(MPIU_Irecv(&sf->rootbufdisp[i], 1, MPIU_INT, s…  in PetscSFSetUp_Basic_NVSHMEM()
120   PetscCallMPI(MPI_Waitall(nRemoteRootRanks, leafreqs, MPI_STATUSES_IGNORE));  in PetscSFSetUp_Basic_NVSHMEM()
122   PetscCallCUDA(cudaMalloc((void **)&sf->rootbufdisp_d, nRemoteRootRanks * sizeof(PetscInt)));  in PetscSFSetUp_Basic_NVSHMEM()
123   PetscCallCUDA(cudaMalloc((void **)&sf->rootsigdisp_d, nRemoteRootRanks * sizeof(PetscInt)));  in PetscSFSetUp_Basic_NVSHMEM()
124   PetscCallCUDA(cudaMalloc((void **)&sf->ranks_d, nRemoteRootRanks * sizeof(PetscMPIInt)));  in PetscSFSetUp_Basic_NVSHMEM()
125   PetscCallCUDA(cudaMalloc((void **)&sf->roffset_d, (nRemoteRootRanks + 1) * sizeof(PetscInt)));  in PetscSFSetUp_Basic_NVSHMEM()
127 …PetscCallCUDA(cudaMemcpyAsync(sf->rootbufdisp_d, sf->rootbufdisp, nRemoteRootRanks * sizeof(PetscI…  in PetscSFSetUp_Basic_NVSHMEM()
128 …PetscCallCUDA(cudaMemcpyAsync(sf->rootsigdisp_d, sf->rootsigdisp, nRemoteRootRanks * sizeof(PetscI…  in PetscSFSetUp_Basic_NVSHMEM()
129 …PetscCallCUDA(cudaMemcpyAsync(sf->ranks_d, sf->ranks + sf->ndranks, nRemoteRootRanks * sizeof(Pets…  in PetscSFSetUp_Basic_NVSHMEM()
130 …PetscCallCUDA(cudaMemcpyAsync(sf->roffset_d, sf->roffset + sf->ndranks, (nRemoteRootRanks + 1) * s…  in PetscSFSetUp_Basic_NVSHMEM()
135 …for (i = 0; i < nRemoteRootRanks; i++) PetscCallMPI(MPI_Send(&i, 1, MPIU_INT, sf->ranks[i + sf->nd…  in PetscSFSetUp_Basic_NVSHMEM()
139   for (i = 0; i < nRemoteRootRanks; i++) {  in PetscSFSetUp_Basic_NVSHMEM()
316     n   = sf->nRemoteRootRanks;  in PetscSFLinkWaitSignalsOfCompletionOfGettingData_NVSHMEM()
357     nsrcranks = sf->nRemoteRootRanks;  in PetscSFLinkGetDataBegin_NVSHMEM()
383     ndstranks = sf->nRemoteRootRanks;  in PetscSFLinkGetDataBegin_NVSHMEM()
445     nsrcranks  = sf->nRemoteRootRanks;  in PetscSFLinkGetDataEnd_NVSHMEM()
542     ndstranks = sf->nRemoteRootRanks;  in PetscSFLinkPutDataBegin_NVSHMEM()
616     nsrcranks = sf->nRemoteRootRanks;  in PetscSFLinkPutDataEnd_NVSHMEM()
625     ndstranks  = sf->nRemoteRootRanks;  in PetscSFLinkPutDataEnd_NVSHMEM()
649     nsrcranks    = sf->nRemoteRootRanks;  in PetscSFLinkSendSignalsToAllowPuttingData_NVSHMEM()