using tensorflow::ops::Mul;
using tensorflow::ops::ReluGrad;

namespace tensorflow {
namespace gradients {
namespace {

class ReluGradientFunction : public GradientFunction {
 public:
  explicit ReluGradientFunction(vector<AbstractTensorHandle*> f_outputs)
      : forward_outputs_(f_outputs) {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

        // allows code like "tf.constant(1.)" or "tf.reduce_sum(..., axis=1)"
        // (where the value starts on the host), without allowing other implicit
        // copies/broadcasts. Other implicit copies may be supported eventually,
        // but need special handling for gradients (gradient of copy-on is not
        // just copy-off but includes a sum) and consideration of performance.
        //

  bool* arrived_flag;
  // Set to true whenever an operation is executed
  bool* executed_flag;
  // If true, only explicit op placements are accepted. If false, uses
  // type-based dispatch.
  bool strict_scope_placement;
};

struct LoggedTensor {
  TFE_TensorHandle* tensor;
  LoggedTensor() = delete;
  explicit LoggedTensor(TFE_TensorHandle* tensor) : tensor(tensor) {}
  ~LoggedTensor() { TFE_DeleteTensorHandle(tensor); }
};

  TensorHandlePtr cpu_value(FloatTensorHandle(3., status.get()));
  ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());

  // Copying on to a parallel device must be explicit.
  TensorHandlePtr failed_copy_on_result(TFE_TensorHandleCopyToDevice(
      cpu_value.get(), context.get(), device_name, status.get()));
  EXPECT_EQ(TF_GetCode(status.get()), TF_UNIMPLEMENTED);

// multiple DeviceThreads should always be accessed in the same order to avoid
// deadlocks.
class DeviceThread {
 public:
  // Starts a background thread waiting for `StartExecute`.
  explicit DeviceThread(const std::string& device, const bool is_async,
                        const int in_flight_nodes_limit)
      : status_(TF_NewStatus()),
        // If the context's default exector is set to async, re-using that in

  TFE_ContextOptionsSetAsync(opts, static_cast<unsigned char>(async));
  TFE_ContextOptionsSetDevicePlacementPolicy(opts,
                                             TFE_DEVICE_PLACEMENT_EXPLICIT);
  TFE_Context* ctx = TFE_NewContext(opts, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteContextOptions(opts);

  TFE_ContextOptionsSetAsync(opts, static_cast<unsigned char>(async));
  TFE_ContextOptionsSetDevicePlacementPolicy(opts,
                                             TFE_DEVICE_PLACEMENT_EXPLICIT);
  TFE_Context* ctx = TFE_NewContext(opts, status);
  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteContextOptions(opts);

}
TEST(CAPI, TensorHandleSilentCopyLocalPolicy) {
  TensorHandleSilentCopy(false, TFE_DEVICE_PLACEMENT_EXPLICIT,
                         TFE_DEVICE_PLACEMENT_SILENT, false);
}
TEST(CAPI, TensorHandleSilentCopyLocalPolicyAsync) {
  TensorHandleSilentCopy(true, TFE_DEVICE_PLACEMENT_EXPLICIT,
                         TFE_DEVICE_PLACEMENT_SILENT, false);
}

void SetAndGetOpDevices(bool async) {

  TF_DeleteBuffer(api_def_buf);
  TF_DeleteApiDefMap(api_def_map);
  TF_DeleteBuffer(op_list_buf);
}

class DummyKernel : public tensorflow::OpKernel {
 public:
  explicit DummyKernel(tensorflow::OpKernelConstruction* context)
      : OpKernel(context) {}
  void Compute(tensorflow::OpKernelContext* context) override {}
};

// Test we can query kernels

struct TfDlManagedTensorCtx {
  TensorReference reference;
  std::vector<int64_t> shape;
  std::vector<int64_t> strides;
  DLManagedTensor tensor;

  explicit TfDlManagedTensorCtx(const TensorReference& ref) : reference(ref) {}
};

// Gets tensor from eager tensor handle.
const Tensor* GetTensorFromHandle(TFE_TensorHandle* h, TF_Status* status) {
  if (h == nullptr) {