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());

// Sets a custom Executor for the current thread. All nodes created by this
// thread will be added to this Executor. It will override the current executor.
TF_CAPI_EXPORT extern void TFE_ContextSetExecutorForThread(TFE_Context*,
                                                           TFE_Executor*);

// Returns the Executor for the current thread.
TF_CAPI_EXPORT extern TFE_Executor* TFE_ContextGetExecutorForThread(

  const std::vector<std::string> underlying_devices_;
  // A sequence of thread wrappers, one per device, for executing operations in
  // parallel.
  //
  // Conceptually this is a thread pool with one thread per device. It requires
  // less synchronization than a thread pool would for this task, since Execute
  // acquires each thread in order (and so only one Execute will schedule

  // (now inaccessible) corresponding JVPs, but ForwardAccumulator's destructor
  // will release remaining references.
  //
  // This method is not thread-safe (and in general ForwardAccumulator is not
  // thread-safe).
  Status Accumulate(
      const string& op_type, const std::vector<TapeTensor>& input_tensors,
      const std::vector<TapeTensor>& output_tensors,

# Contributing guidelines

## Pull Request Checklist

Before sending your pull requests, make sure you do the following:

-   Read the [contributing guidelines](CONTRIBUTING.md).
-   Read the [Code of Conduct](CODE_OF_CONDUCT.md).
-   Ensure you have signed the
    [Contributor License Agreement (CLA)](https://cla.developers.google.com/).
-   Check if your changes are consistent with the

  // 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) {

  ops->writable_file_ops_api = TF_WRITABLE_FILE_OPS_API;
  ops->writable_file_ops_size = TF_WRITABLE_FILE_OPS_SIZE;
  ops->read_only_memory_region_ops_abi = TF_READ_ONLY_MEMORY_REGION_OPS_ABI;
  ops->read_only_memory_region_ops_api = TF_READ_ONLY_MEMORY_REGION_OPS_API;
  ops->read_only_memory_region_ops_size = TF_READ_ONLY_MEMORY_REGION_OPS_SIZE;
}

/// Initializes a TensorFlow plugin.
///

      # Sharded tests have target names like this:
      # WindowOpsTest.test_tflite_convert0 (<function hann_window at
      #     0x7fc61728dd40>, 10, False, tf.float32)
      # Where 0x... is a thread ID (or something) that is not important for
      # debugging, but breaks this "number of failures" counter because it's
      # different for repetitions of the same test. We use re.sub(r"0x\w+")
      # to remove it.

#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