namespace tensorflow {
namespace gradients {

/* Returns numerical grad inside `dtheta_approx` given `forward` model and
 * parameter specified by `input_index`.
 *
 * I.e. if y = <output of the forward model> and w = inputs[input_index],
 * this will calculate dy/dw numerically.
 *
 * `use_function` indicates whether to use graph mode(true) or eager(false).
 *

  // Returns the inputs of this op.
  virtual absl::Span<ImmediateExecutionTensorHandle* const> GetInputs()
      const = 0;
  virtual Status SetInput(size_t index,
                          ImmediateExecutionTensorHandle* input) = 0;

  virtual ImmediateExecutionContext* GetContext() const = 0;

  // Following two methods are used to support custom device.
  // Return true if the inputs contain custom device tensor handle. It means

// Return a shape op fetching the shape of `a`.
TFE_Op* ShapeOp(TFE_Context* ctx, TFE_TensorHandle* a);

// Return an allreduce op adding up input tensor `in` from `group_size` workers.
TFE_Op* AllReduceOp(TFE_Context* ctx, TFE_TensorHandle* in, int group_size);

// Return a SendOp op `op_name` with send input tensor `in` and attributes
// `send_device`, `recv_device`, and `send_device_incarnation` set.

// E.g. it could know whether we're in eager mode or graph mode, keeps track
// of gradient tapes, etc.
typedef struct TF_ExecutionContext TF_ExecutionContext;

// A TF_AbstractTensor is an input to an operation. E.g. it could be a union
// type of eager and graph tensors. It is also the result of executing an
// operation.
typedef struct TF_AbstractTensor TF_AbstractTensor;

  // The a handle for the resource-dtype tensor pointing to the variable's
  // buffer.
  TFE_TensorHandle* handle_;
  // The dtype of the variable's buffer (input dtype for assignments, output
  // dtype of read operations).
  TF_DataType type_;
};

// Creates a TFE_TensorHandle with value `v`.
TensorHandlePtr FloatTensorHandle(float v, TF_Status* status);

// parallel device creates a parallel tensor `x` with `a` on the first of
// `underlying_devices` and `b` on the second. Running `a_unpacked, b_unpacked =
// TPUReplicatedOutput(input=x, num_replicas=2)` un-packs the parallel tensor
// into its components.
//
// The filled `device` struct and the allocated `device_info` struct may be
// passed to TFE_RegisterCustomDevice. The `device_name` arguments must match.

// A compute function. This will never actually get called in this test, it's
// just nice to know that it compiles.
void compute(void* kernel, TF_OpKernelContext* ctx) {
  TF_Tensor* input;
  TF_Status* s = TF_NewStatus();
  TF_GetInput(ctx, 0, &input, s);
  TF_DeleteTensor(input);
  TF_DeleteStatus(s);
}

// Exercises tensorflow's C API.
int main(int argc, char** argv) {
  TF_InitMain(argv[0], &argc, &argv);

  // existing and given constraints will be performed.
  virtual Status SetDeviceName(const char* name) = 0;

  virtual Status AddInput(AbstractTensorHandle* input) = 0;
  virtual Status AddInputList(
      absl::Span<AbstractTensorHandle* const> inputs) = 0;
  virtual Status Execute(absl::Span<AbstractTensorHandle*> retvals,
                         int* num_retvals) = 0;

TF_Operation* RandomUniform(TF_Operation* shape, TF_DataType dtype,
                            TF_Graph* graph, TF_Status* s);

// Split `input` along the first dimension into 3 tensors
TF_Operation* Split3(TF_Operation* input, TF_Graph* graph, TF_Status* s,
                     const char* name = "split3");

bool IsPlaceholder(const tensorflow::NodeDef& node_def);

  }
};

// An abstract operation describes an operation by its type, name, and
// attributes. It can be "executed" by the context with some input tensors.
// It is allowed to reusing the same abstract operation for multiple execution
// on a given context, with the same or different input tensors.
class TracingOperation : public AbstractOperation {
 protected:
  explicit TracingOperation(AbstractOperationKind kind)