first_underlying_devices, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Create a second parallel device with the first parallel device and one
  // additional CPU.
  const char* second_device_name =
      "/job:localhost/replica:0/task:0/device:CUSTOM:1";
  std::array<const char*, 2> second_underlying_devices{

   * make it load its own version of FinalizableReferenceQueue. Then we need to interact with that
   * parallel version through reflection in order to exercise the parallel
   * FinalizableReferenceQueue, and then check that the parallel ClassLoader can be
   * garbage-collected after that.
   */

  public static class MyFinalizableWeakReference extends FinalizableWeakReference<Object> {

   * an edge connecting node A to node B if this {@link Network} has an edge connecting A to B.
   *
   * <p>If this network {@link #allowsParallelEdges() allows parallel edges}, parallel edges will be
   * treated as if collapsed into a single edge. For example, the {@link #degree(Object)} of a node
   * in the {@link Graph} view may be less than the degree of the same node in this {@link Network}.
   */

<img src="/img/async/parallel-burgers/parallel-burgers-01.png" class="illustration">

Then it's finally your turn, you place your order of 2 very fancy burgers for your crush and you.

You pay 💸.

<img src="/img/async/parallel-burgers/parallel-burgers-02.png" class="illustration">

The cashier goes to the kitchen.

    "groups" : [ {
      "testProject" : "excludeRuleMergingBuild",
      "coverage" : {
        "per_commit" : [ "linux" ]
      }
    } ]
  }, {
    "testId" : "org.gradle.performance.regression.corefeature.LargeDependencyGraphPerformanceTest.resolve large dependency graph (parallel = false, locking = true)",
    "groups" : [ {

            status, TF_INVALID_ARGUMENT,
            absl::StrCat(
                "Got a non-parallel tensor ",
                tensorflow::unwrap(absl::get<TFE_TensorHandle*>(input))
                    ->DebugString(),
                " as input to a parallel operation. First pack non-parallel "
                "tensors for each device into a parallel tensor explicitly.")
                .c_str());
        return absl::nullopt;
      }

      log += "three:run@${taskFaker.nanoTime} parallel=${taskFaker.isParallel}"
    }

    taskFaker.advanceUntil(0.µs)
    assertThat(log).isEmpty()

    taskFaker.advanceUntil(100.µs)
    assertThat(log).containsExactlyInAnyOrder(
      "one:run@100000 parallel=true",
      "two:run@100000 parallel=true",
      "three:run@100000 parallel=true",
    )

    taskFaker.assertNoMoreTasks()

        }
        try {
            final long start = System.currentTimeMillis();
            final Stream<QueryLog> stream = Stream.of(queryLogs);
            if (parallel) {
                stream.parallel();
            }
            final SuggestItem[] array = stream
                    .flatMap(queryLog -> contentsParser

  "durations" : [ {
    "testProject" : "excludeRuleMergingBuild",
    "linux" : 229
  } ]
}, {
  "scenario" : "org.gradle.performance.regression.corefeature.LargeDependencyGraphPerformanceTest.resolve large dependency graph (parallel = false, locking = true)",
  "durations" : [ {

        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.
        device_(device),
        executor_(
            TFE_NewExecutor(is_async, /*enable_streaming_enqueue=*/true,
                            /*in_flight_nodes_limit=*/in_flight_nodes_limit)),