%4 = mhlo.multiply %3, %3 : tensor<?xi32>
      return %4 : tensor<?xi32>
    }
  })";
  CellReader<int64_t> legal_error(kFailedLegalizationStreamz);
  CellReader<int64_t> static_error(kNonStaticOpStreamz);
  CreateModule(kNonStaticFailure);

  auto result = Run();

  EXPECT_TRUE(result.failed());

  explicit TFE_MonitoringCounterReader(const char* name) {
    counter = std::make_unique<
        ::tensorflow::monitoring::testing::CellReader<int64_t>>(name);
  }
  template <typename... LabelType>
  int64_t Read(const LabelType&... labels);
  std::unique_ptr<::tensorflow::monitoring::testing::CellReader<int64_t>>
      counter;
};

    const DenseIntElementsAttr& dense_attr, PatternRewriter& builder) {
  std::vector<int32_t> ret(dense_attr.getNumElements());
  auto range = dense_attr.getValues<int64_t>();
  std::transform(range.begin(), range.end(), ret.begin(),
                 [](int64_t attr) { return static_cast<int32_t>(attr); });
  return DenseIntElementsAttr::get(
      RankedTensorType::get(ret.size(), builder.getIntegerType(32)), ret);
}

// nullptr will be returned.
TF_CAPI_EXPORT extern TF_Buffer* TF_CoordinationServiceGetKeyValue(
    const char* key, int64_t key_size, TF_CoordinationServiceAgent* agent,
    TF_Status* status);

TF_CAPI_EXPORT extern TF_Buffer* TF_CoordinationServiceGetKeyValueWithTimeout(
    const char* key, int64_t key_size, int64_t timeout_seconds,
    TF_CoordinationServiceAgent* agent, TF_Status* status);

  static constexpr int64_t kDefaultMicros = 3000000;

  int64_t max_iters = 0;   // Maximum iterations to run, ignored if <= 0.
  int64_t max_micros = 0;  // Maximum microseconds to run, ignored if <= 0.
};

// Stats holds statistics collected during benchmarking.
struct Stats {
  std::vector<int64_t> per_iter_us;  // Per-iteration deltas in us.
  int64_t total_us;                  // Total time in us.

  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;

  Value output_layer_norm_coefficients_;

  mlir::TFL::LSTMOp lstm_;

  Value none_;
  SmallVector<int64_t, 1> bias_slice_shape_;
  SmallVector<int64_t, 1> bias_size_values_;
  SmallVector<int64_t, 2> weight_slice_shape_;
  SmallVector<int64_t, 2> weight_slice_size_input_values_;
  SmallVector<int64_t, 2> weight_slice_size_recurrent_values_;
  OpBuilder builder_;
};

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();

  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 =

  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));