Created by: NHZlX

The Following code is not thread safe:

void TensorRTEngine::Execute(int batch_size, std::vector<void *> *buffers,
                             cudaStream_t stream) {
  freshDeviceId();
  const std::thread::id tid = std::this_thread::get_id();
  batch_size_ = batch_size;
  if (infer_context_.find(tid) == infer_context_.end()) {
    PADDLE_ENFORCE_NOT_NULL(
        infer_engine_,
        "You should build engine first and then set the context.");
    infer_context_[tid].reset(infer_engine_->createExecutionContext());
  }
  infer_context_[tid]->enqueue(batch_size, buffers->data(), stream, nullptr);
  cudaStreamSynchronize(stream);
  SetRuntimeBatch(batch_size);
}

So, add a lock before run infer_context_[tid].reset(infer_engine_->createExecutionContext());.