void NCCL<Dtype>::InitSingleProcess(vector<NCCL<Dtype>*>* nccls) {
ncclComm_t* comms = new ncclComm_t[nccls->size()];
int* gpu_list = new int[nccls->size()];
for (int i = 0; i < nccls->size(); ++i) {
gpu_list[i] = (*nccls)[i]->solver_->param().device_id();
}
NCCL_CHECK(ncclCommInitAll(comms, static_cast<int>(nccls->size()), gpu_list));
for (int i = 0; i < nccls->size(); ++i) {
(*nccls)[i]->comm_ = comms[i];
}
}
NcclCommList(const std::vector<int>& devices)
: comms(new ncclComm_t[devices.size()]), ndevices(devices.size()) {
CHECK(ncclCommInitAll(comms.get(), devices.size(), devices.data()));
}
