Lines Matching refs:PetscCallCUDA
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()
145 PetscCallCUDA(cudaMalloc((void **)&bas->leafbufdisp_d, nRemoteLeafRanks * sizeof(PetscInt))); in PetscSFSetUp_Basic_NVSHMEM()
146 PetscCallCUDA(cudaMalloc((void **)&bas->leafsigdisp_d, nRemoteLeafRanks * sizeof(PetscInt))); in PetscSFSetUp_Basic_NVSHMEM()
147 PetscCallCUDA(cudaMalloc((void **)&bas->iranks_d, nRemoteLeafRanks * sizeof(PetscMPIInt))); in PetscSFSetUp_Basic_NVSHMEM()
148 PetscCallCUDA(cudaMalloc((void **)&bas->ioffset_d, (nRemoteLeafRanks + 1) * sizeof(PetscInt))); in PetscSFSetUp_Basic_NVSHMEM()
150 …PetscCallCUDA(cudaMemcpyAsync(bas->leafbufdisp_d, bas->leafbufdisp, nRemoteLeafRanks * sizeof(Pets… in PetscSFSetUp_Basic_NVSHMEM()
151 …PetscCallCUDA(cudaMemcpyAsync(bas->leafsigdisp_d, bas->leafsigdisp, nRemoteLeafRanks * sizeof(Pets… in PetscSFSetUp_Basic_NVSHMEM()
152 …PetscCallCUDA(cudaMemcpyAsync(bas->iranks_d, bas->iranks + bas->ndiranks, nRemoteLeafRanks * sizeo… in PetscSFSetUp_Basic_NVSHMEM()
153 …PetscCallCUDA(cudaMemcpyAsync(bas->ioffset_d, bas->ioffset + bas->ndiranks, (nRemoteLeafRanks + 1)… in PetscSFSetUp_Basic_NVSHMEM()
222 PetscCallCUDA(cudaEventRecord(link->dataReady, link->stream)); in PetscSFLinkBuildDependenceBegin()
223 PetscCallCUDA(cudaStreamWaitEvent(link->remoteCommStream, link->dataReady, 0)); in PetscSFLinkBuildDependenceBegin()
238 PetscCallCUDA(cudaEventRecord(link->endRemoteComm, link->remoteCommStream)); in PetscSFLinkBuildDependenceEnd()
239 PetscCallCUDA(cudaStreamWaitEvent(link->stream, link->endRemoteComm, 0)); in PetscSFLinkBuildDependenceEnd()
321 PetscCallCUDA(cudaGetLastError()); in PetscSFLinkWaitSignalsOfCompletionOfGettingData_NVSHMEM()
397 PetscCallCUDA(cudaGetLastError()); in PetscSFLinkGetDataBegin_NVSHMEM()
403 PetscCallCUDA(cudaGetLastError()); in PetscSFLinkGetDataBegin_NVSHMEM()
416 PetscCallCUDA(cudaGetLastError()); in PetscSFLinkGetDataBegin_NVSHMEM()
458 PetscCallCUDA(cudaGetLastError()); in PetscSFLinkGetDataEnd_NVSHMEM()
460 PetscCallCUDA(cudaGetLastError()); in PetscSFLinkGetDataEnd_NVSHMEM()
566 PetscCallCUDA(cudaGetLastError()); in PetscSFLinkPutDataBegin_NVSHMEM()
633 PetscCallCUDA(cudaGetLastError()); in PetscSFLinkPutDataEnd_NVSHMEM()
662 PetscCallCUDA(cudaGetLastError()); in PetscSFLinkSendSignalsToAllowPuttingData_NVSHMEM()
673 PetscCallCUDA(cudaEventDestroy(link->dataReady)); in PetscSFLinkDestroy_NVSHMEM()
674 PetscCallCUDA(cudaEventDestroy(link->endRemoteComm)); in PetscSFLinkDestroy_NVSHMEM()
675 PetscCallCUDA(cudaStreamDestroy(link->remoteCommStream)); in PetscSFLinkDestroy_NVSHMEM()
761 PetscCallCUDA(cudaDeviceGetStreamPriorityRange(NULL, &greatestPriority)); in PetscSFLinkCreate_NVSHMEM()
762 …PetscCallCUDA(cudaStreamCreateWithPriority(&link->remoteCommStream, cudaStreamNonBlocking, greates… in PetscSFLinkCreate_NVSHMEM()
764 PetscCallCUDA(cudaEventCreateWithFlags(&link->dataReady, cudaEventDisableTiming)); in PetscSFLinkCreate_NVSHMEM()
765 PetscCallCUDA(cudaEventCreateWithFlags(&link->endRemoteComm, cudaEventDisableTiming)); in PetscSFLinkCreate_NVSHMEM()