}

  RankedTensorType x_ty = GetRankedTensorTypeForOperand(op.getX());
  RankedTensorType y_ty = GetRankedTensorTypeForOperand(op.getY());

  if (!x_ty || !y_ty) return success();

  ArrayRef<int64_t> x_shape = x_ty.getShape();
  ArrayRef<int64_t> y_shape = y_ty.getShape();

  llvm::SmallVector<int64_t, 4> result_batch_shape;

  // batch size in lhs and rhs must be broadcastable
  RankedTensorType x_ty = op.getX().getType().dyn_cast<RankedTensorType>();
  RankedTensorType y_ty = op.getY().getType().dyn_cast<RankedTensorType>();

  if (!x_ty || !y_ty) return success();
  ArrayRef<int64_t> x_shape = x_ty.getShape();
  ArrayRef<int64_t> y_shape = y_ty.getShape();

  llvm::SmallVector<int64_t, 4> result_batch_shape;

    Tensor x_data = MakeData<X_T>(x_shape);
    Output x, y;
    MakeOp<X_T>(op_type, x_data, {4, 6}, align_corners, half_pixel_centers, &x,
                &y);
    JAC_T max_error;
    TF_ASSERT_OK((ComputeGradientError<X_T, Y_T, JAC_T>(
        scope_, x, x_data, y, {1, 4, 6, 1}, &max_error)));
    EXPECT_LT(max_error, 1.5e-3);
  }

  template <typename X_T, typename Y_T, typename JAC_T>

/// <X_T, Y_T, JAC_T> should be <double, double, double>
///
/// if y = Square(x), where x (and so y) are DT_COMPLEX64,
/// <X_T, Y_T, JAC_T> should be <complex64, complex64, float>
/// Note that JAC_T is always real-valued, and should be an appropriate
/// precision to host the partial derivatives for dy/dx
///
/// if y = ComplexAbs(x) where x is DT_COMPLEX64 (so y is DT_FLOAT)
/// <X_T, Y_T, JAC_T> should be <complex64, float, float>

    // complex-valued, we perturb its real and complex parts separately.
    for (int r = 0; r < x_size; ++r) {
      int unit_dimension = 0;
      for (X_T unit : BaseUnitsForType<X_T>::values()) {
        X_T x_delta = unit * X_T{delta};
        // Store current value of 'x' at 'r'.
        X_T v = x_data_flat(r);
        // Evaluate at positive delta.
        x_data_flat(r) = v + x_delta;
        std::vector<Tensor> y_pos;

//  BOOST_VERSION / 100000 is the major version

#define BOOST_VERSION 105500

//
//  BOOST_LIB_VERSION must be defined to be the same as BOOST_VERSION
//  but as a *string* in the form "x_y[_z]" where x is the major version
//  number, y is the minor version number, and z is the patch level if not 0.
//  This is used by <config/auto_link.hpp> to select which library version to link to.

				t.Errorf("String() mismatch:\nhave %s\nwant %s", s, tt.out)
			}
		})
	}
}

var lexTests = []struct {
	in  string
	out string
}{
	{"", ""},
	{"x", "x"},
	{"x.y", "x.y"},
	{"x_y", "x_y"},
	{"αx", "αx"},
	{"αx²", "αx err: invalid syntax at ²"},
	{"go1.2", "go1.2"},
	{"x y", "x y"},
	{"x!y", "x ! y"},
	{"&&||!()xy yx ", "&& || ! ( ) xy yx"},
	{"x~", "x err: invalid syntax at ~"},

    NDTRI
  };

  template <typename X_T, typename Y_T>
  void TestCWiseGrad(UnaryOpType op_type, const std::function<X_T(int)>& x_fn) {
    TF_ASSERT_OK(scope_.status());
    DataType x_type = DataTypeToEnum<X_T>::v();
    TensorShape shape({2, 3, 2});
    auto x = Placeholder(scope_, x_type, Placeholder::Shape(shape));
    Tensor x_data(x_type, shape);
    auto x_data_flat = x_data.flat<X_T>();

		t.Skipf("skipping on %s/%s because key generation takes too long", runtime.GOOS, runtime.GOARCH)
	}

	// Set up some roots, intermediate CAs, and leaf certs with various algorithms.
	// X_Y is X signed by Y.
	R1 := boringCert(t, "R1", boringRSAKey(t, 2048), nil, boringCertCA|boringCertFIPSOK)
	R2 := boringCert(t, "R2", boringRSAKey(t, 512), nil, boringCertCA)

    // Start replacing.
    auto k = !values.use_empty() ? k_values : k_indices;
    // Build scalar tensor k.
    auto k_ty = mlir::RankedTensorType::get({}, rewriter.getIntegerType(32));
    Value k_cst = rewriter.create<TFL::ConstOp>(
        op.getLoc(), DenseElementsAttr::get(k_ty, k));
    // Compute new result types.
    auto values_ty = mlir::dyn_cast<ShapedType>(values.getType());