namespace tensorflow {
namespace gradients {
namespace internal {

void CompareNumericalAndAutodiffGradients(
    Model model, Model grad_model, AbstractContext* ctx,
    absl::Span<AbstractTensorHandle* const> inputs, bool use_function,
    double abs_error = 1e-2);

void CheckTensorValue(AbstractTensorHandle* t, absl::Span<const float> manuals,
                      absl::Span<const int64_t> dims, double abs_error = 1e-2);

    if (!released_) f_();
    released_ = src.released_;
    f_ = src.release();
    return *this;
  }

  ~Cleanup() {
    if (!released_) f_();
  }

  // Releases the cleanup function instead of running it.
  // Hint: use c.release()() to run early.
  F release() {
    released_ = true;
    return std::move(f_);
  }

  bool is_released() const { return released_; }

 private:

namespace tensorflow {
namespace grappler {

// Plugin initialization function that a device plugin
// must define.
typedef void (*TFInitGraphPluginFn)(TP_OptimizerRegistrationParams* const,
                                    TF_Status* const);

// Registers Graph optimizers.
Status InitGraphPlugin(void* dso_handle);

// Allow registering a graph optimizer using a function (used for
// testing).

using Model = std::function<Status(AbstractContext*,
                                   absl::Span<AbstractTensorHandle* const>,
                                   absl::Span<AbstractTensorHandle*>)>;

// Runs `model` maybe wrapped in a function call op. This can be thought as
// being equivalent to the following python code.
//
// if use_function:
//   outputs = tf.function(model)(inputs)
// else:

#define TENSORFLOW_C_EAGER_C_API_REMOTE_TEST_UTIL_H_

// Run a function containing a MatMul op and check its output.
// If heavy_load_on_streaming_rpc is true, send some rpc requests before the one
// which creates a remote input, to simulate a scenario that the remote input
// is not ready when we start running an op or a function.
void TestRemoteExecuteSilentCopies(bool async, bool remote, bool func,

// here.
void TF_SetTracingImplementation(const char* name, TF_Status*);

// Creates a new TensorFlow function. A Function is an execution context, and as
// such it can trace operations through TF_ExecuteOperation. After completing
// tracing, a function can be obtained by TF_FinalizeFunction.
TF_ExecutionContext* TF_CreateFunction(const char* fn_name, TF_Status* status);

// `first_device` and `second_device`.
void BasicTestsForTwoDevices(TFE_Context* context, const char* first_device,
                             const char* second_device);

// Implementations of templated functions ******************************

template <std::size_t num_replicas>
TensorHandlePtr CreatePerDeviceValues(
    TFE_Context* context,
    const std::array<TFE_TensorHandle*, num_replicas>& components,

 * 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).
 *
 * `numerical_grad` is the pointer to the AbstractTensorHandle* which will
 * hold the numerical gradient data at the end of the function.
 */
Status CalcNumericalGrad(AbstractContext* ctx, Model forward,

// Struct for Optimizer. Plugin authors must provide an optimize function.
// Creation and deletion functions are optional.
typedef struct TP_Optimizer {
  size_t struct_size;
  void* ext;  // reserved for future use

  // [Optional]
  // Create function for optimizer.
  void* (*create_func)();

  // Optimizer function for optimizer. The first param is an optimizer created

// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.

#include <assert.h>
#include <stdlib.h>
#include <string.h>
#include "_cgo_export.h"

/* Test calling a Go function with multiple return values.  */

void
callMulti(void)
{
	struct multi_return result = multi();
	assert(strcmp(result.r0, "multi") == 0);
	assert(result.r1 == 0);
	free(result.r0);