int64_t x_row_dim = x_shape[x_shape.size() - 2];
  int64_t x_col_dim = x_shape[x_shape.size() - 1];
  int64_t y_row_dim = y_shape[y_shape.size() - 2];
  int64_t y_col_dim = y_shape[y_shape.size() - 1];
  int64_t out_row_dim = output_shape[output_shape.size() - 2];
  int64_t out_col_dim = output_shape[output_shape.size() - 1];

  int64_t expected_out_row_dim = op.getAdjX() ? x_col_dim : x_row_dim;

namespace {

std::vector<int64_t> TensorShapeAsVector(const tensorflow::TensorHandle& handle,
                                         tensorflow::Status* status) {
  std::vector<int64_t> shape;
  int rank = -1;
  *status = handle.NumDims(&rank);
  if (!status->ok()) {
    return shape;
  }
  shape.reserve(rank);
  for (int i = 0; i < rank; ++i) {
    int64_t dim;
    *status = handle.Dim(i, &dim);

// 'shape' is computed from the original shape and the broadcast dimensions to
// match result shape.
int64_t GetElementIndex(llvm::SmallVectorImpl<int64_t> &shape,
                        llvm::SmallVectorImpl<int64_t> &current_index) {
  int64_t ind = 0;
  int64_t mul = 1;
  for (int i = shape.size() - 1; i >= 0; --i) {
    ind += (current_index[i] % shape[i]) * mul;
    mul *= shape[i];
  }

        /*inputSpatialDimensions=*/SmallVector<int64_t>{1, 2},
        /*kernelInputFeatureDimension=*/2, /*kernelOutputFeatureDimension=*/3,
        /*kernelSpatialDimensions=*/SmallVector<int64_t>{0, 1},
        /*outputBatchDimension=*/0, /*outputFeatureDimension=*/3,
        /*outputSpatialDimensions=*/SmallVector<int64_t>{1, 2});

    // Determine the shape of the output tensor: [b, f, 0, 1] => [b, 0, 1, f]

constexpr float kRequantCost = 2.0;

// TODO(renjieliu): Ideally this should consider different kinds of SOCs as
// well.

// Get total bytes transferred.
int64_t GetTransferredTensorBytes(func::CallOp from_graph,
                                  func::CallOp to_graph) {
  int64_t total_size_transferred = 0;
  for (auto input : to_graph.getOperands()) {
    Operation* input_op = input.getDefiningOp();

  return th;
}

TFE_TensorHandle* DoubleTestMatrixTensorHandle(TFE_Context* ctx) {
  int64_t dims[] = {2, 2};
  double data[] = {1.0, 2.0, 3.0, 4.0};
  TF_Status* status = TF_NewStatus();
  TF_Tensor* t = TFE_AllocateHostTensor(ctx, TF_DOUBLE, &dims[0],
                                        sizeof(dims) / sizeof(int64_t), status);
  memcpy(TF_TensorData(t), &data[0], TF_TensorByteSize(t));

}
Status TapeOperation::SetAttrIntList(const char* attr_name,
                                     const int64_t* values, int num_values) {
  forward_op_.attrs.Set(
      attr_name, gtl::ArraySlice<const int64_t>(
                     reinterpret_cast<const int64_t*>(values), num_values));
  return parent_op_->SetAttrIntList(attr_name, values, num_values);
}

  if (num_bits >= 8) {
    return type;
  }
  int64_t qmin = QType::getDefaultMinimumForInteger(is_signed, num_bits);
  int64_t qmax = QType::getDefaultMaximumForInteger(is_signed, num_bits);
  if (narrow_range) {
    qmin += 1;
  }
  const int64_t storage_type_min = type.getStorageTypeMin();
  const int64_t storage_type_max = type.getStorageTypeMax();
  const double rate =

      RankedTensorType::get(static_cast<int64_t>(attr.size()),
                            IntegerType::get(attr.getContext(), 64));
  return DenseIntElementsAttr::get(ty, attr.getValue());
}

DenseIntElementsAttr GetI64ElementsAttr(ArrayRef<int64_t> values,
                                        Builder* builder) {
  RankedTensorType ty = RankedTensorType::get(
      {static_cast<int64_t>(values.size())}, builder->getIntegerType(64));

      const int64_t output_size = (lhs_size_dilated + stride - 1) / stride;
      const int64_t total_padding = std::max(
          (output_size - 1) * stride + rhs_size_dilated - lhs_size_dilated,
          static_cast<int64_t>(0));
      const int64_t padding_begin = total_padding / 2;
      const int64_t padding_end = total_padding - padding_begin;