//
// DeviceThread itself is thread-safe, in that StartExecute will block if there
// is a pending execution. Since StartExecute is equivalent to grabbing a lock,
// multiple DeviceThreads should always be accessed in the same order to avoid
// deadlocks.
class DeviceThread {
 public:
  // Starts a background thread waiting for `StartExecute`.
  explicit DeviceThread(const std::string& device, const bool is_async,

  // A non-blocking version of `Execute`. After each call, `Join` must be called
  // before `StartExecute` is called again. Using `StartExecute` with `Join`
  // allows the caller to schedule computation on multiple ParallelDevices
  // without sequencing those operations (first call `StartExecute` on each
  // parallel device, then call `Join` on each; even if some of the `Join`s

                     status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  CancellationManager cancellation_manager;
  parallel_device.StartExecute(context.get(), std::vector<ParallelTensor*>(),
                               "VarHandleOp", TFE_OpGetAttrs(handle_op.get()),
                               /*expected_max_outputs=*/1,