virtual std::string run() { return "Callback::run"; }
};

class Caller {
private:
	Callback *callback_;
public:
	Caller(): callback_(0) { }
	~Caller() { delCallback(); }
	void delCallback() { delete callback_; callback_ = 0; }
	void setCallback(Callback *cb) { delCallback(); callback_ = cb; }
	std::string call();

            device_value.release(), sizeof(DataType),
            [](void* data, size_t, void* arg) {
              delete reinterpret_cast<DataType*>(data);
            },
            nullptr),
        TF_DeleteTensor);
    // TODO(allenl): Here and when executing regular operations, we could hold
    // on to one TFE_Op per device and just call TFE_ResetOp to avoid parsing
    // device names repeatedly.

  // clients from directly destroying this object since it may manage its own
  // lifetime through ref counting. Thus clients MUST call Release() in order to
  // destroy an instance of this class.
  virtual void Release() = 0;

  // Creates an operation builder and ties it to this context.
  // The returned object can be used for setting operation's attributes,
  // adding inputs and finally executing (immediately or lazily as in tracing)

// passed to TFE_RegisterCustomDevice. The `device_name` arguments must match.
void AllocateParallelDevice(const char* device_name,
                            const char* const* underlying_devices,
                            int num_underlying_devices,
                            TFE_CustomDevice* device, void** device_info);

}  // namespace parallel_device
}  // namespace tensorflow

TF_CAPI_EXPORT extern TF_AttrBuilder* TF_NewAttrBuilder(const char* op_name);
TF_CAPI_EXPORT extern void TF_DeleteAttrBuilder(TF_AttrBuilder* builder);
TF_CAPI_EXPORT extern void TF_AttrBuilderSetType(TF_AttrBuilder* builder,
                                                 const char* attr_name,
                                                 TF_DataType value);
TF_CAPI_EXPORT extern void TF_AttrBuilderSetTypeList(TF_AttrBuilder* builder,

                           int64_t init_timeout_in_ms);

tensorflow::ServerDef ReplaceTaskInServerDef(
    const tensorflow::ServerDef& server_def, int task_index);

void ReplaceTaskInServerDef(tensorflow::ServerDef* server_def, int task_index,
                            const std::string& host, int port);

std::vector<std::string> ListDeviceNames(TFE_Context* ctx);

//go:build aix || darwin || dragonfly || freebsd || linux || netbsd || openbsd || solaris

#include <pthread.h>
#include "_cgo_export.h"

static void*
addThread(void *p)
{
	int i, max;
	
	max = *(int*)p;
	for(i=0; i<max; i++)
		Add(i);
	return 0;
}

void
doAdd(int max, int nthread)
{
	enum { MaxThread = 20 };
	int i;
	pthread_t thread_id[MaxThread];
	
	if(nthread > MaxThread)

// Converts eager tensor handle to DLPack (DLManagedTensor*), and return the
// void* for further PyCapsule construction.
TF_CAPI_EXPORT extern void* TFE_HandleToDLPack(TFE_TensorHandle* h,
                                               TF_Status* status);

// Converts DLPack (DLManagedTensor*) to eager tensor handle.
TF_CAPI_EXPORT extern TFE_TensorHandle* TFE_HandleFromDLPack(void* dlm,

  // Marks the following gradient as a result so it's not consumed by backward
  // functions.
  virtual void MarkAsResult(Gradient* gradient) const = 0;

  // Deletes the input tensor.
  virtual void DeleteGradient(Gradient* gradient) const = 0;
};

// Traces the execution of operations, doing eager garbage collection, and

#include "tensorflow/c/eager/gradients.h"

namespace tensorflow {
namespace gradients {
class TapeOperation : public AbstractOperation {
 public:
  explicit TapeOperation(AbstractOperation*, Tape*, const GradientRegistry&);
  void Release() override;
  Status Reset(const char* op, const char* raw_device_name) override;
  const string& Name() const override;
  const string& DeviceName() const override;
  Status SetDeviceName(const char* name) override;