//    "/job:localhost/replica:0/task:0/device:GPU:1"}
// Then executing an operation on CUSTOM:0 will execute it on GPU:0 and GPU:1.
//
// Implicit copies onto `device_name` are allowed, replicating the value once
// per device in `underlying_devices`. Implicit copies off of the device throw
// an error.
//
// All component tensors must have the same dtype. Currently they must also have

        // 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
        // copies/broadcasts. Other implicit copies may be supported eventually,

#include "tensorflow/core/platform/status.h"

namespace tensorflow {

// Attributes of an op.
class AbstractOpAttrs {
 protected:
  enum AbstractOpAttrsKind { kEager, kTfrt };
  explicit AbstractOpAttrs(AbstractOpAttrsKind kind) : kind_(kind) {}

 public:
  // Returns which subclass is this instance of.
  AbstractOpAttrsKind getKind() const { return kind_; }
  virtual ~AbstractOpAttrs() = default;

  void operator()(TF_Buffer* buf) const { TF_DeleteBuffer(buf); }
};
using OwnedTFBuffer = std::unique_ptr<TF_Buffer, TFBufferDeleter>;

class CGraphOptimizer : public CustomGraphOptimizer {
 public:
  explicit CGraphOptimizer(TP_Optimizer optimizer, const char* device_type)
      : optimizer_(optimizer), device_type_(device_type) {
    if (optimizer.create_func != nullptr) {
      c_optimizer_ = (*optimizer_.create_func)();

  // For LLVM style RTTI.
  static bool classof(const AbstractTensorHandle* ptr) {
    return ptr->getKind() == kEager || ptr->getKind() == kTfrt;
  }

 protected:
  explicit ImmediateExecutionTensorHandle(AbstractTensorHandleKind kind)
      : AbstractTensorHandle(kind) {}
  ~ImmediateExecutionTensorHandle() override {}
};

namespace internal {
struct ImmediateExecutionTensorHandleDeleter {

};

struct TF_DeviceList {
  std::vector<tensorflow::DeviceAttributes> response;
};

struct TF_Function {
  tensorflow::FunctionRecord* record;
};

struct TF_ApiDefMap {
  explicit TF_ApiDefMap(const tensorflow::OpList& op_list)
      :
#if !defined(IS_MOBILE_PLATFORM) && !defined(IS_SLIM_BUILD)
        api_def_map(op_list),
#endif  // !defined(IS_MOBILE_PLATFORM) && !defined(IS_SLIM_BUILD)

}
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

template <typename F>
class Cleanup {
 public:
  Cleanup() : released_(true), f_() {}

  template <typename G>
  explicit Cleanup(G&& f)          // NOLINT
      : f_(std::forward<G>(f)) {}  // NOLINT(build/c++11)

  Cleanup(Cleanup&& src)  // NOLINT
      : released_(src.is_released()), f_(src.release()) {}

  // Implicitly move-constructible from any compatible Cleanup<G>.
  // The source will be released as if src.release() were called.

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();
      }
    }
  }