};
  const int block_size = 8;
  tf_gcs_filesystem::RamFileBlockCache cache(
      block_size, 2 * callers * block_size, 0, fetcher);
  std::vector<std::unique_ptr<Thread>> threads;
  threads.reserve(callers);
  for (int i = 0; i < callers; i++) {
    threads.emplace_back(
        Env::Default()->StartThread({}, "caller", [block_size, &cache, i]() {
          std::vector<char> out;
          TF_EXPECT_OK(

  StatusPtr first_bad_status(nullptr);
  for (int device_index = 0; device_index < underlying_devices_.size();
       ++device_index) {
    DeviceThread* device_thread = device_threads_[device_index].get();
    per_device_output_tensors.push_back(device_thread->Join(status));
    // We will run every Join even if there are bad statuses in case the user
    // wants to recover and continue running ops on the parallel device (which

      --enable-gnu-unique-object \
      --enable-initfini-array \
      --enable-languages="c,c++" \
      --enable-linker-build-id \
      --enable-plugin \
      --enable-shared \
      --enable-threads=posix \
      --with-default-libstdcxx-abi=${LIBSTDCXX_ABI} \
      --with-gcc-major-version-only \
      --with-linker-hash-style="gnu" \
      && \
      make -j$(nproc) && \
      make install

  // allocation/scheduling.
  //
  // TODO(allenl): Keep a map from outer thread to list of inner threads rather
  // than a single list of threads so aliased nested parallel devices don't
  // re-use a thread.
  std::vector<std::unique_ptr<DeviceThread>> device_threads_;
  // A cancellation manager to use if the caller does not provide one. When ops
  // are executed asynchronously this must outlive the queued op, so it can't be

// ops.
TF_CAPI_EXPORT extern void TFE_ContextClearCaches(TFE_Context* ctx);

// Sets a thread-local device placement policy. After this call, other calls to
// TFE_Execute in the same thread will use the device policy specified here
// instead of the device policy used to construct the context. This has no
// effect on the device policy used by other program threads.
TF_CAPI_EXPORT extern void TFE_ContextSetThreadLocalDevicePlacementPolicy(

                                  const std::shared_ptr<Block>& block,
                                  TF_Status* status) {
  absl::MutexLock lock(&mu_);
  if (block->timestamp == 0) {
    // The block was evicted from another thread. Allow it to remain evicted.
    return TF_SetStatus(status, TF_OK, "");
  }
  if (block->lru_iterator != lru_list_.begin()) {
    lru_list_.erase(block->lru_iterator);
    lru_list_.push_front(key);

  TFE_ContextOptionsSetDevicePlacementPolicy(opts, global_policy);
  TFE_Context* ctx = TFE_NewContext(opts, status.get());
  if (thread_policy != global_policy) {
    TFE_ContextSetThreadLocalDevicePlacementPolicy(ctx, thread_policy);
  }
  TFE_DeleteContextOptions(opts);
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  void (*cleanup)(TF_Filesystem* filesystem);

  /// Creates a new random access read-only file from given `path`.
  ///
  /// After this call `file` may be concurrently accessed by multiple threads.
  ///
  /// Plugins:
  ///   * Must set `status` to `TF_OK` if `file` was updated.
  ///   * Must set `status` to `TF_NOT_FOUND` if `path` doesn't point to an

  // 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));
  return ret;
}

      static_cast<tensorflow::ContextDevicePlacementPolicy>(policy));
}

// Note: this function looks up a thread local policy. So it should be called in
// the appropriate client thread. In particular, in async mode, it may not be
// safe to call this function from the async EagerExecutor threads.
extern TFE_ContextDevicePlacementPolicy TFE_ContextGetDevicePlacementPolicy(
    TFE_Context* ctx) {