%0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
    func.return %0 : tensor<2x1xf32>
    // CHECK: %[[VAL_0:.*]] = arith.constant dense<[2, 1]> : tensor<2xi32>
    // CHECK: %[[CONCAT:.*]] = "tfl.concatenation"(%arg0, %arg1) <{axis = 0 : i32, fused_activation_function = "NONE"}> : (tensor<1xf32>, tensor<1xf32>) -> tensor<2xf32>

                const std::vector<Output>& grad_inputs,
                std::vector<Output>* grad_outputs) {
  auto shift = op.input(1);
  auto axis = op.input(2);
  auto grad_op = Roll(scope, grad_inputs[0], Neg(scope, shift), axis);
  grad_outputs->push_back(grad_op);
  grad_outputs->push_back(NoGradient());
  grad_outputs->push_back(NoGradient());
  return scope.status();
}

  }

  SmallVector<double, 4> scales;
  SmallVector<int64_t, 4> zeroPoints;
  scales.reserve(axisSize);
  zeroPoints.reserve(axisSize);
  for (size_t axis = 0; axis != axisSize; ++axis) {
    double rmin = rmins[axis];
    double rmax = rmaxs[axis];
    if (std::fabs(rmax - rmin) < std::numeric_limits<double>::epsilon()) {
      scales.push_back(1.0);
      zeroPoints.push_back(qmin);
      continue;
    }

  %2 = "tfl.add"(%arg0, %arg3) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %3 = "tfl.pack"(%1, %2) {axis = 0 : i32, values_count = 2 : i32} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  func.return %3 : tensor<2x1xf32>
}

// CHECK: %[[CST:.*]] = arith.constant dense<1> : tensor<4xi32>

    func.func @main(%arg0: tensor<3x3x!tf_type.qint8>, %arg1: tensor<3x3x!tf_type.qint8>) -> tensor<6x3x!tf_type.qint8> {
      %axis = "tf.Const"() { value = dense<0> : tensor<i64> } : () -> tensor<i64>
      %1 = "tf.ConcatV2"(%arg0, %arg1, %axis) : (tensor<3x3x!tf_type.qint8>, tensor<3x3x!tf_type.qint8>, tensor<i64>) -> tensor<6x3x!tf_type.qint8>
      func.return %1 : tensor<6x3x!tf_type.qint8>
    }
  })";

// -----

func.func @pack_CPU(%arg0: tensor<100xf32>, %arg1: tensor<100xf32>) -> tensor<2x100xf32> attributes {tac.device = "CPU", tac.interface_name = "func_2"} {
  // CHECK: tac.cost = 1.000000e+02
  %0 = "tfl.pack"(%arg0, %arg1) {axis = 0 : i32, tac.device = "CPU", values_count = 2 : i32} : (tensor<100xf32>, tensor<100xf32>) -> tensor<2x100xf32>
  func.return %0 : tensor<2x100xf32>
}

  xs.push_back(Placeholder(scope, DT_DOUBLE, Placeholder::Shape(shape)));
  xs.push_back(Placeholder(scope, DT_DOUBLE, Placeholder::Shape(shape)));
  auto tmp = Stack(scope, xs, Stack::Axis(0));
  auto y = Unstack(scope, tmp, 2, Unstack::Axis(0));
  double max_error;
  TF_ASSERT_OK((ComputeGradientError<double, double, double>(
      scope, xs, {shape, shape}, y.output, {shape, shape}, &max_error)));
  EXPECT_LT(max_error, 1e-10);

//
// Description:
//   Given a tensor `input`, this operation inserts a dimension of 1 at the
//   dimension index `axis` of `input`'s shape. The dimension index `axis`
//   starts at zero; if you specify a negative number for `axis` it is counted
//   backward from the end.
//
//   This operation is useful if you want to add a batch dimension to a single

  [(IsNotInLiftedFunc $res), (IsConstTensor $b)], [], (addBenefit 1)>;

def LiftGather : Pat<
  (TF_GatherV2Op:$res $params, $indices, $axis, $batch_dims),
  (LiftAsTFPartitionedCall<"composite_gather_fn">
    (ArgumentList $params, $indices, $axis),
    (ResultList $res),
    (NamedAttributeList
      (NamedAttr<"batch_dims"> $batch_dims))),
  [(IsNotInLiftedFunc $res), (IsConstTensor $params)], [], (addBenefit 1)>;

  // CHECK-SAME: (tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<2xi32>
  %axis = "tf.Const"() {value = dense<0> : tensor<i32>} : () -> tensor<i32>
  %0 = "tf.VariableV2"() {container = "", shape = #tf_type.shape<>, shared_name = "x"} : () -> tensor<!tf_type.int32ref>
  %1 = "tf.ConcatV2"(%0, %0, %axis) : (tensor<!tf_type.int32ref>, tensor<!tf_type.int32ref>, tensor<i32>) -> tensor<2xi32>
  func.return %1 : tensor<2xi32>