EXPECT_EQ(6 * TF_DataTypeSize(TF_UINT64), TF_TensorByteSize(a));
  EXPECT_EQ(6 * TF_DataTypeSize(TF_UINT64), TF_TensorByteSize(b));

  // Check that a write to one tensor shows up in the other.
  *(static_cast<int64_t*>(TF_TensorData(a))) = 4;
  EXPECT_EQ(4, *(static_cast<int64_t*>(TF_TensorData(b))));
  *(static_cast<int64_t*>(TF_TensorData(b))) = 6;
  EXPECT_EQ(6, *(static_cast<int64_t*>(TF_TensorData(a))));

}

Status ModularRandomAccessFile::Read(uint64 offset, size_t n,
                                     StringPiece* result, char* scratch) const {
  if (ops_->read == nullptr)
    return errors::Unimplemented(
        tensorflow::strings::StrCat("Read() not implemented for ", filename_));

  UniquePtrTo_TF_Status plugin_status(TF_NewStatus(), TF_DeleteStatus);
  int64_t read =

  ASSERT_EQ(errors::OK, status_.code()) << status_.message();

  auto ReluGradModel = BuildGradModel(ReluModel, registry_);

  float X_vals[] = {1.0f, 2.0f, 3.0f, -5.0f, -4.0f, -3.0f, 2.0f, 10.0f, -1.0f};
  int64_t X_dims[] = {3, 3};
  AbstractTensorHandlePtr X;
  {
    AbstractTensorHandle* X_raw;
    status_ = TestTensorHandleWithDims<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), X_vals, X_dims, 2, &X_raw);

  virtual Status NumElements(int64_t* num_elements) const = 0;
  // Returns size of specified dimension
  //
  // -1 indicates an unknown axis length; this is unreachable for most standard
  // ImmediateExecutionTensorHandles, but comes up for example when computing
  // the shape of a parallel tensor with component shapes differing across
  // devices.
  virtual Status Dim(int dim_index, int64_t* dim) const = 0;

  EXPECT_EQ(
      1, metrics->point_set_map.at("test/counter")->points.at(0)->int64_value);

  TFE_MonitoringCounterCellIncrementBy(cell, 5);
  EXPECT_EQ(TFE_MonitoringCounterCellValue(cell), 6);
  metrics = collection_registry->CollectMetrics(options);
  EXPECT_EQ(
      6, metrics->point_set_map.at("test/counter")->points.at(0)->int64_value);

  TFE_MonitoringDeleteCounter0(counter);

// Return a tensor handle with given type, values and dimensions.
template <class T, TF_DataType datatype>
Status TestTensorHandleWithDims(AbstractContext* ctx, const T* data,
                                const int64_t* dims, int num_dims,
                                AbstractTensorHandle** tensor) {
  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  TFE_Context* eager_ctx =

  /* Want to test simple MatMul example:
    [[0,0],    *   [[0,0],    =   [[0,0],
     [0,0]]         [0,0]]         [0,0]]
  */

  // Build an abstract input tensor.
  int64_t dims[] = {2, 2};  // Matrices will be 2 x 2
  int num_dims = sizeof(dims) / sizeof(dims[0]);

  float vals[] = {0.0f, 0.0f, 0.0f, 0.0f};
  TFE_Context* eager_ctx = TF_ExecutionContextGetTFEContext(ctx, status.get());

void TF_GraphSetOutputHandleShapesAndTypes(TF_Graph* graph, TF_Output output,
                                           int num_shapes_and_types,
                                           const int64_t** shapes,
                                           const int* ranks,
                                           const TF_DataType* types,
                                           TF_Status* status);

  TFE_ContextOptions* opts = TFE_NewContextOptions();
  TFE_Context* ctx = TFE_NewContext(opts, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
  TFE_DeleteContextOptions(opts);

  std::vector<int64_t> dims(4, 1);
  TFE_Op* op = TFE_NewOp(ctx, "AvgPool", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  TF_Tensor* tensor =

}

// Checks whether the stride array matches the layout of compact, row-majored
// data.
bool IsValidStrideCompactRowMajorData(int64_t* shape_arr, int64_t* stride_arr,
                                      int ndim) {
  bool valid = true;
  int64_t expected_stride = 1;
  for (int i = ndim - 1; i >= 0; --i) {
    // Empty tensors are always compact regardless of strides.
    if (shape_arr[i] == 0) return true;