4.5 Foundations

The two basic building blocks of all that is done in Nektar++ are the concepts of Points and of a Basis. The Point objects denote positions in space, either on compact domains (normally [-1,1]d where d is the dimension in a reference domain mapped to world-space) or periodic domains such as (0,2π] (i.e. in the case of points used in Fourier expansions). The Basis objects denote functions (e.g. polynomials) evaluated at a given set of points.

We rely heavily on the concept of “managers”, and this concept has its start right here in Foundations. When we started our development of Nektar++, we realized that a pure encapsulation strategy for Expansions would consist of every expansion holding a pointer to the basis (of some sort) that was used to specify it, and then correspondingly the basis pointing to a collection of points at which it was evaluated. From the computer science perspective, this is all very reasonable except for the fact that there are many cases in spectral/hp element expansions in which multiple elements use the same reference space basis functions, and correspondingly those basis functions are often evaluated at the same set of points (quadrature points). This led us to the conclusion that if we were going to have a code that could run over large number of elements (at the time tens of millions) without excessive memory usage, we would have to try to avoid as much of this duplication as possible. This led us to introduce to Nektar++ two important computer science tools: smart pointers and managers (i.e. a factory pattern).

Smart pointers were originally introduced as an add-on from the Boost library and later directly incorporated into Nektar++ when they were natively handled by C++11. With the tradition view of pointers, one has a variable that holds an address that points to a place in memory. Typically, in type-safe programming, the way this memory is to be interpreted is denoted by the type of pointer: a double pointer (i.e. double*) is a pointer to memory that should be interpreted as a double. The downside of this traditional view occurs when multiple pointers all point to the same place in memory. The question then arises: which pointer (or more properly which thread of control containing a pointer) is responsible for deleting the memory when it is no longer needed? Again, in the traditional view that each pointer points to unique memory and hence when the memory is not longer needed the memory can be released no longer is viable when multiple pointers all point to the same piece of memory. This probably was solved with the invention of the smart pointer, an object that is wrapped up to look like a pointer, and yet contains a reference counter within it. When a smart pointer is initiated and memory is originally assigned to it, the reference counter is incremented. As more and more references to the piece of memory are created, the increment counter continues to increase. As pointers go out of scope and are longer valid, the increment is decremented. Only when all the references to a particular piece of memory are removed can a free (i.e. delete) operation be accomplished. The introduction of smart pointers into C++ and into Nektar++ were particularly advantageous to us as they allowed us to have lots of different elements, expansions, etc. point to the same fundamental data structure in memory (for instance, a Basis object evaluated at a particular set of points) without having to worry about improper deallocation or memory leakage due to pointer mishandling.

The second important feature we exploited within the redesign of Nektar++ was the idea of managers (i.e. a factory pattern). You will see this use of the term manager throughout Nektar++. For us, a manager has the following characteristics: it expects as input a key, which gives the manager sufficient information to know the object for which it is receiving a request. If the manager already has one of those objects in storage, it passes back a (smart) pointer that that object. The manager itself continues to hold on to a pointer for the object (in case other requests are made of it for that object). If the manager does not have a copy of the requested object, then it has the ability to generate the object “on the fly” and store it for future use. This last component of the manager is the factory feature that we mentioned earlier – a manager has registered with it a collection of methods that know how to generate the various objects that might be requested of it.

Based upon these two concepts: two fundamental managers that exist at this level of the library are the PointsManager and the BasisManager. These are contained within ManagerAccess.h/cpp, and are defined as singletons and defined globally so that they can be accessed throughout Nektar++. Within these files you will find the RegisterCreator method calls that link particular Create methods with various keys. Only in rare cases will these every need to be modified, and only infrequently possibly added to (if additional Points or Basis information is given).

4.5.1 Points

One of the two primary data structures in this directory is Points. A points object consists of a PointsKey and then extra information needed to facilitate various operations at points. The basic layout of the data structure is shown in Figure 4.1.


pict

Figure 4.1: The basic Points class data object. It consists of a PointsKey (used by a PointsManager) and various other data members needed for point operations.


The PointsKey object has three basis data members: the number of points, an enum specifying the point distribution, and a optional scaling factor. In almost all of Nektar++, only the first two are needed to uniquely specify a collection of points. The latter variable was added as part of the extensions of Nektar++ to accommodate NekMesh. As stated above, if you pass a PointsKey to a PointsManager you will receive back a pointer to a Points object with its various data members populated, among them a PointsKey identifier for the object. As can be seen by the diagram, the Points object has a collection (array) of three pointers to point positions. These three array pointers allow Points objects to point to 1D, 2D or 3D point positions. Corresponding to the dimension and associated with each point position is a weight associated with integration. This was done to facilitate both non-tensor product and tensor product constructions. To demonstrate non-tensor product constructions, consider if you were to be dealing with a 2D point set that is natively 2D (that is, not based upon tensor product construction such as the Electrostatic Points), then the number of points in the PointsKey would be the size of each array associated with m_points, and only one of the m_weight arrays (associated with m_weight[0]) would be valid such at:

∫                     nump∑oints- 1
     f(x0,x1)dx0dx1 ≈           ωif(z0,i,z1,i)
 Ω2D                     i=0

where z0,i and z1,i denote the points associated with m_points[0][i] and m_points[1][i] respectively. The weights in this case are calculated by routines found in NodalUtils via forming a Vandermonde system, solving for Lagrange basis functions that pass through a particular set of points, and then integrating those basis functions to obtain weights. This operation is done sufficiently often that we will try to spell it out after the Fekete Point paragraph below.

If one is dealing with a tensor product constructed 2D nodal point set, one would have:

                                ⌊                         ⌋
∫                     numpo∑ints-1  nump∑oints- 1
     f(x0,x1)dx0dx1 ≈           ⌈           ωiωjf(z0,i,z1,j)⌉
 Ω2D                      i=0         j=0

where z0,i and z1,i denote the points associated with m_points[0][i] and m_points[1][i] respectively and ωi and ωj are the weights associated with the two coordinate directions respectively. The tensor product construction above was implemented in Nektar++ assuming we would create derived nodal point sets for quadrilaterals and hexahedron; however, throughout much of Nektar++, we use the tensor product forms explicitly by specifying 1D point sets in the different directions as needed. This particular point will become more apparent when you get to reading the Chapter on StdRegions (Chapter 5). You will encounter, for instance, that a Standard Quadrilateral Expansion (StdQuadExp) requires two 1D Point objects denoting the two coordinates for integration. In this case, the number of points and number of weights match up per direction.

Electrostatic Points: The Electrostatic (Nodal) Points on the triangle and tetrahedron are based upon the work of Hesthaven [38]. They are an attempt to find the minimizer of a potential energy function based upon an electrostatic point source analogy, and provide points that correspondingly have low Lebesgue constant. The point positions as given in the paper are stored in the files NodalTriElecData.h and NodalTetElecData.h. Hesthaven uses a condensed format for the nodal positions which rely on the symmetries of the points (if you count the points in the file, you will see that fewer points are given than are needed to span the polynomial space. For example, for degree one, a single point it given. This point along with its three-fold symmetries represent all three points needed to support the linear space). The routine CalculatePoints() expands the condensed format of the points to the full set based upon the one, three (a and b denoting two types of three-symmetries) and six symmetries. Note that the ordering of the nodes expanded from the file does not respect any particular geometric ordering, so we reorder the points using NodalPointReorder2d(), which reorganizes the points in vertex points, edge points, and then interior points (i.e. a geometric decomposition that aligns more naturally with the way we organize nodes/modes within Nektar++). In Figure 4.2 we show an example of this reordering done for the points needed for expressing a total degree three expansion over a triangle.


pict

Figure 4.2: Example of how nodes are reordered to meet a geometric ordering. On the left, we show the nodal positions that would be used for building a total degree three polynomial space over a triangle ordered in canonical form. On the right, we show the reordering of those nodes to follow a vertex, edge and interior ordering. Note that in the case of a tetrahedron, the ordering would be vertex, edge, face and then interior.


Fekete Points: The Fekete (Nodal) Points on a triangle are based upon the work of Taylor and Wingate [5657], and are an alternative nodal point distribution on both triangles and tetrahedra. As an alternative strategy to the one mentioned above, Taylor and Wingate explicitly attempt to find a point distribution that minimizes the Lebesgue constant. For their particular point set, the minimization if over the Lebesgue function itself (not the electrostatic potential function, which can be viewed as a proxy for this function). The point positions as given in the paper are stored in the file NodalTriFeketeData.h and NodalTetFeketeData.h. Taylor and Wingate use a condensed format for the nodal positions similar to that used by Hesthaven.

Building Lagrange Interpolating Polynomials: One common operation done in both the case of the Electrostatic Points and the Fekete Points is to build the Lagrange interpolating polynomial associated with a point set. This is needed, for instance, for computing the integration weights associated with a particular collection of points (as mentioned above). This operation is done sufficiently often that we will try to spell it out here. Forming the Lagrange basis functions for a particular set of points, in particular in 2D and 3D, is non-trivial. The most common way of arriving at a set of actionable basis functions is to use a known basis (for which we are able to explicitly write down its form) and find the linear combination of those basis functions that yield the Lagrange basis. For example, let us assume we are dealing with the set of Electrostatic Points on a triangle. The number of points, numpoints, will specify the number of “modes” (coefficients) we expect to find. The total degree of the space is given by N = (M+1)(M+2)-
     2 where N is the degree of the polynomials and M is the number of points. As an example: note that when M = 3, we get N = 1 – that is, with three points we can support a polynomial of total degree one over a triangle. Let ϕi denote a known non-interpolating basis function. Although not done in practice, imagine that this is, for instance, a monomial. In 1D, this would equate to ϕi(x) = xi, and in multiple dimensions this would equate to ϕi(x,y) = xi1yi2 with some index mapping function i = σ(i1,i2) that gives us an index ordering. The total number of basis functions to span our degree N space is given by M. We first form the Vandermonde matrix V as follows:

    ⌊                                                 ⌋
    |  ϕ0(x0,0,x1,0)   ϕ1(x0,0,x1,0)   ...   ϕN(x0,0,x1,0) |
    ||  ϕ0(x0,1,x1,1)   ϕ1(x0,1,x1,1)   ...   ϕN(x0,1,x1,1) ||
V = |⌈                      ...                          |⌉

      ϕ0 (x0,N,x1,N)  ϕ1(x0,N ,x1,N )  ...  ϕN (x0,N ,x1,N)

where each row i denotes the evaluation of our N basis functions at a given point (x0,i,x1,i). Since we have selected the number of basis functions to be what can be uniquely resolved by our N points, this matrix is square and invertible. The conditioning of the matrix is based upon our basis choice; hence we know that monomials are not a good choice beyond approximately cubics so in general we use our internal modified basis for this system. With V now available, we can form the linear system:

Vc = b

where c is the vector of coefficients and where b = bi is a binary vector where the ith entry is set to one when finding the coefficients for the ith Lagrange basis function associated with a particular set of points. For each Lagrange basis function definition we seek to find, we update the right-hand-side vector and solve the linear system for the set of coefficients that denote the combination of our known basis that yields a Lagrange basis function. Note that since this operation (of “inverting" this linear system) must be done for each Lagrange basis function, we can optimize our operations by accomplishing LU decomposition on V first, which can be done in O(N2) operations, and then each solution for the coefficient vector can be done in O(N) operations through backsolves.

Finding Integration Weights: When the points are related to Gaussian quadrature, the array m_weights will contain the appropriate Gaussian quadrature weights computed using point and weight routines found in Foundations (our polymath library functions). In the case of weights associated with points sets that do not lie at the zeros of Jacobi polynomials, we must compute the weights directly based upon Lagrange interpolation through those points. Although this can be done in general for any point distribution, we will focus here on evenly-spaced points (note that when one applies this approach to Chebyshev points, one arrives at Clenshaw-Curtis quadrature [58]). Within Nektar++, given evenly-spaced points on the interval [-1,1] – if only one point is given, then the weight is set to 2.0 denoting a midpoint integration rule; otherwise, the weights are given by the following expression:

     ∫ 1
wj =     ℓj(ξ)dξ
      - 1

where ℓj(ξ) is the Lagrange basis function defined over an evenly-spaced set of points ξk ∈ [-1,1]. For 2D and 2D nodal point sets (such as Electrostatic and Fekete Points), this same procedure is used. We first form the Lagrange interpolating functions via the algorithm given above (using the Vandermonde system), and then for each quadrature weight compute the integral of that Lagrange basis function using some known quadrature rule (such as tensor product quadrature via Duffy transformation; this point will be more easily understood after reading Chapter 5)

4.5.2 Basis

The second of the two primary data structures in this directory is Basis. A basis object consists of a BasisKey and the extra information needed to facilitate various operations on the basis. The basis layout of the data structure is shown in Figure 4.3.


pict

Figure 4.3: The basic Basis class data object. It consists of a BasisKey (used by a BasisManager) and various other data members needed for point operations.


The Basis object consists of a BasisKey (which in turn holds a PointsKey), a pointer to the positions at which the basis is evaluated and then arrays which hold the evaluation of the basis (m_bdata) and the evaluation of the derivates of the basis (m_dbdata). The storage layout for these data structures are shown in Figure 4.3. The number of rows is given by the number of modes associated with a particular expansion, and the number of columns is given by the number of points. These arrays are stored as contiguous memory to help avoid excessive memory hopping.


pict

Figure 4.4: Diagram of the basic memory layout of the arrays associated with the basis and its derivatives evaluated at points.