// (as in the case of ConstOp kernel creation on GPU, which involves copying a
  // CPU tensor to GPU).
  // Setting a larger thread pool does not help with the Swift caller, as we use
  // a different TFE context for each thread of execution (for running graph
  // functions, and their send/recvs corountines).
  config.set_inter_op_parallelism_threads(1);

  TF_Buffer* ret = TF_NewBuffer();
  TF_CHECK_OK(MessageToBuffer(config, ret));

                       status.get()),
      variable_deleter);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  // Assign an initial value to the variable, mirroring it to each component
  // device.
  {
    TensorHandlePtr initial_value_cpu = FloatTensorHandle(20., status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

      if (operation_name == std::string("_EagerConst")) {
        // Non-parallel tensors from _EagerConst/tf.constant are implicitly
        // broadcast, i.e. set as the input to each parallel operation. This
        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit

                        const int in_flight_nodes_limit)
      : status_(TF_NewStatus()),
        // If the context's default exector is set to async, re-using that in
        // 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.

    status_ = TF_NewStatus();
    filesystem_ = new TF_Filesystem;
    filesystem_->plugin_filesystem = nullptr;
    // Because different tests requires different setup for filesystem. We
    // initialize filesystem in each testcase.
  }
  void TearDown() override {
    TF_DeleteStatus(status_);
    if (filesystem_->plugin_filesystem != nullptr)
      tf_gcs_filesystem::Cleanup(filesystem_);
    delete filesystem_;
  }

  TensorHandlePtr parallel_value = CreatePerDeviceValues(
      context.get(), components, device_name, status.get());
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Run a collective sum, so each component should now be the same.
  TensorHandlePtr reduced(
      CollectiveSum(context.get(), parallel_value.get(), 2, status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

        "memory regions");

  return OkStatus();
}

// Validates the operations supplied by the plugin.
//
// Uses the 4 simpler `ValidateHelper(const TF_...*)` to validate each
// individual function table and then checks that the function table for a
// specific file type exists if the plugin offers support for creating that
// type of files.

    // Get transpose attrs
    bool t_a;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("transpose_a", &t_a));

    bool t_b;
    TF_RETURN_IF_ERROR(forward_attrs_.Get("transpose_b", &t_b));

    // Conj each input
    AbstractTensorHandle* conj_output;
    std::string name = "Conj_A_MatMul_Grad";
    TF_RETURN_IF_ERROR(
        SafeConj(ctx, forward_inputs_[0], &conj_output, name.c_str()));

  // The cache has space for `block_count` blocks. The loop with i = 0 should
  // fill the cache, and the loop with i = 1 should be all cache hits. The
  // fetcher checks that it is called once and only once for each offset (to
  // fetch the corresponding block).
  for (int i = 0; i < 2; i++) {
    for (int j = 0; j < block_count; j++) {
      TF_EXPECT_OK(ReadCache(&cache, "", block_size * j, block_size, &out));
    }
  }

namespace tensorflow {
namespace {

using ::tensorflow::error::Code;

// As we need to test multiple URI schemes we need a parameterized test.
// Furthermore, since each test creates and deletes files, we will use the same
// fixture to create new directories in `SetUp`. Each directory will reside in
// `::testing::TempDir()`, will use a RNG component and the test name. This
// ensures that two consecutive runs are unlikely to clash.