#include "tensorflow/core/platform/status.h"
#include "tensorflow/core/platform/str_util.h"
#include "tensorflow/core/platform/strcat.h"
#include "tensorflow/core/platform/stringpiece.h"
#include "tensorflow/core/platform/thread_annotations.h"
#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/public/session.h"
#include "tensorflow/core/public/version.h"

// The implementation below is at the top level instead of the

  // unnecessary allocations each Execute call, we keep one heap-allocated
  // version for the thread.
  StatusPtr status_ TF_GUARDED_BY(execution_mutex_);

  const std::string device_;
  ExecutorPtr executor_ TF_GUARDED_BY(execution_mutex_);
  mutable OpPtr op_ TF_GUARDED_BY(execution_mutex_);
  std::unique_ptr<Thread> thread_;
};

DeviceThread::~DeviceThread() {
  {
    tensorflow::mutex_lock l(execution_mutex_);

  // (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();

    notification.Notify();
    // Wait for other thread to issue read.
    Env::Default()->SleepForMicroseconds(100000);  // 0.1 secs
    TF_SetStatus(status, TF_OK, "");
    return n;
  };
  tf_gcs_filesystem::RamFileBlockCache cache(block_size, block_size, 0,
                                             fetcher);
  // Fork off thread for parallel read.
  std::unique_ptr<Thread> concurrent(

                                  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());

  cc_options.numa_node = options->numa_node;
  return reinterpret_cast<TF_Thread*>(::tensorflow::Env::Default()->StartThread(
      cc_options, thread_name, [=]() { (*work_func)(param); }));
}

void TF_JoinThread(TF_Thread* thread) {
  // ::tensorflow::Thread joins on destruction
  delete reinterpret_cast<::tensorflow::Thread*>(thread);
}

void* TF_LoadSharedLibrary(const char* library_filename, TF_Status* status) {

}

}  // namespace

TEST(TestEnv, TestThreads) {
  TF_ThreadOptions options;
  TF_DefaultThreadOptions(&options);
  SomeThreadData data;
  TF_Thread* thread =
      TF_StartThread(&options, "SomeThreadName", &SomeThreadFunc, &data);
  TF_JoinThread(thread);
  ::tensorflow::mutex_lock l(data.mu);
  ASSERT_TRUE(data.did_work);

  tensorflow::unwrap(ctx)->SetThreadLocalDevicePlacementPolicy(
      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(