Tutorial 13: Low-level API Poisson equation
Introduction and caveat
This tutorial is advanced and you only need to go through this if you want to know the internals of Gridap
and what it does under the hood. Even though you will likely want to use the high-level APIs in Gridap
, this tutorial will (hopefully) help if you want to become a Gridap
developer, not just a user. We also consider that this tutorial shows how powerful and expressive the Gridap
kernel is, and how mastering it you can implement new algorithms not currently provided by the library.
As any other Gridap tutorial, this tutorial is primarily designed to be executed in a Jupyter notebook environment. However, the usage of a Julia debugger (typically outside of a Jupyter notebook environment), such as, e.g., the Julia REPL-based Debugger.jl
package, or the one which comes along with the Visual Studio Code (VSCode) extension for the Julia programming language, may help the reader eager to understand the full detail of the explanations given. Some of the observations that come along with the code snippets are quite subtle/technical and may require a deeper exploration of the underlying code using a debugger.
Including Gridap's low-level API
Let us start including Gridap
and some of its submodules, to have access to a rich set of not so high-level methods. Note that the module Gridap
provides the high-level API, whereas the submodules such as, e.g., Gridap.FESpaces
, provide access to the different parts of the low-level API.
using Gridap
using Gridap.FESpaces
using Gridap.ReferenceFEs
using Gridap.Arrays
using Gridap.Geometry
using Gridap.Fields
using Gridap.CellData
using FillArrays
using Test
using InteractiveUtils
Discrete model and FE spaces set up using high-level API
We first create the geometry model and FE spaces using the high-level API. In this tutorial, we are not going to describe the geometrical machinery in detail, only what is relevant for the discussion. To simplify the analysis of the outputs, you can consider a 2D mesh, i.e., D=2
(everything below works for any spatial dimension without any extra complication). In order to make things slightly more interesting, i.e., having non-constant Jacobians, we have considered a mesh that is a stretching of an equal-sized structured mesh.
L = 2 # Domain length in each space dimension
D = 2 # Number of spatial dimensions
n = 4 # Partition (i.e., number of cells per space dimension)
function stretching(x::Point)
m = zeros(length(x))
m[1] = x[1]^2
for i in 2:D
m[i] = x[i]
end
Point(m)
end
pmin = Point(Fill(0,D))
pmax = Point(Fill(L,D))
partition = Tuple(Fill(n,D))
model = CartesianDiscreteModel(pmin,pmax,partition,map=stretching)
The next step is to build the global FE space of functions from which we are going to extract the unknown function of the differential problem at hand. This tutorial explores the Galerkin discretization of the scalar Poisson equation. Thus, we need to build H1-conforming global FE spaces. This can be achieved using $C^0$ continuous functions made of piece(cell)-wise polynomials. This is precisely the purpose of the following lines of code.
First, we build a scalar-valued (T = Float64
) Lagrangian reference FE of order order
atop a reference n-cube of dimension D
. To this end, we first need to create a Polytope
using an array of dimension D
with the parameter HEX_AXIS
, which encodes the reference representation of the cells in the mesh. Then, we create the Lagrangian reference FE using the reference geometry just created in the previous step. It is not the purpose of this tutorial to describe the (key) abstract concept of ReferenceFE
in Gridap.
T = Float64
order = 1
pol = Polytope(Fill(HEX_AXIS,D)...)
reffe = LagrangianRefFE(T,pol,order)
Second, we build the test (Vₕ) and trial (Uₕ) global finite element (FE) spaces out of model
and reffe
. At this point we also specify the notion of conformity that we are willing to satisfy, i.e., H1-conformity, and the region of the domain in which we want to (strongly) impose Dirichlet boundary conditions, the whole boundary of the box in this case.
Vₕ = FESpace(model,reffe;conformity=:H1,dirichlet_tags="boundary")
u(x) = x[1] # Analytical solution (for Dirichlet data)
Uₕ = TrialFESpace(Vₕ,u)
The CellDatum
abstract type and (some of) its subtypes
We also want to extract the triangulation out of the model and create a numerical quadrature. We use a quadrature rule with a higher number integration points than those strictly needed to integrate a mass matrix exactly, i.e., 4*order
, instead of 2*order
We do so in order to help the reader distinguish the axis used for quadrature points, and the one used for DoFs in multi-dimensional arrays, which contain the result of evaluating fields (or a differential operator acting on these) in a set of quadrature rule evaluation points.
Tₕ = Triangulation(model)
Qₕ = CellQuadrature(Tₕ,4*order)
Qₕ is an instance of type CellQuadrature
, a subtype of the CellDatum
abstract data type.
isa(Qₕ,CellDatum)
subtypes(CellDatum)
CellDatum
is the root of one out of three main type hierarchies in Gridap (along with the ones rooted at the abstract types Map
and Field
) on which the evaluation of variational methods in finite-dimensional spaces is grounded on. Any developer of Gridap should familiarize with these three hierarchies to some extent. Along this tutorial we will give some insight on the rationale underlying these, with some examples, but more effort in the form of self-research is expected from the reader as well.
Conceptually, an instance of a CellDatum
represents a collection of quantities (e.g., points in a reference system, or scalar-, vector- or tensor-valued fields, or arrays made of these), once per each cell of a triangulation. Using the get_data
generic function one can extract an array with such quantities. For example, in the case of Qₕ, we get an array of quadrature rules for numerical integration.
Qₕ_cell_data = get_data(Qₕ)
@test length(Qₕ_cell_data) == num_cells(Tₕ)
In this case we get the same quadrature rule in all cells (note that the returned array is of type Fill
). Gridap also supports different quadrature rules to be used in different cells. Exploring such feature is out of scope of the present tutorial.
Any CellDatum
has a trait, the so-called DomainStyle
trait. This information is consumed by Gridap
in different parts of the code. It specifies whether the quantities in it are either expressed in the reference (ReferenceDomain
) or the physical (PhysicalDomain
) domain. We can indeed check the DomainStyle
of a CellDatum
using the DomainStyle
generic function:
DomainStyle(Qₕ) == ReferenceDomain()
DomainStyle(Qₕ) == PhysicalDomain()
If we evaluate the two expressions above, we can see that the DomainStyle
trait of Qₕ is ReferenceDomain
. This means that the local FE space in the physical space in which our problem is posed is expressed in terms of the composition of a space in a reference FE in a parametric space (which is being shared by many or all FEs in the physical space) and the inverse of the geometrical map (from the parametric to the physical space).
In practise, the integration in the physical space is transformed into a numerical integration in the reference space (via a change of variables) using a quadrature. We can exploit this property for ReferenceDomain
FE spaces to reduce computations, i.e., to avoid applying the geometrical map to the quadrature points within Qₕ and its inverse at the shape functions in the physical space.
We note that, while finite elements may not be defined in this parametric space (it is though standard practice with Lagrangian FEs, and other FEs, because of performance reasons), finite element functions are always integrated in such a parametric space. However, for FE spaces that are genuinely defined in the physical space, i.e., the ones with the PhysicalDomain
trait, the transformation of quadrature points from the reference to the physical space is required.
In fact, the DomainStyle
metadata of CellDatum
allows Gridap
to do the right thing (as soon as it is implemented) for all combinations of points and FE spaces (both either expressed in the reference or physical space). This is accomplished by the change_domain
function in the API of CellDatum
.
Using the array of quadrature rules Qₕ_cell_data
, we can access specific entries. The object retrieved provides an array of points (Point
data type in Gridap
) in the cell reference parametric space $[0,1]^d$ and their corresponding weights.
q = Qₕ_cell_data[rand(1:num_cells(Tₕ))]
p = get_coordinates(q)
w = get_weights(q)
However, there is a more convenient way (for reasons made clear above) to work with the evaluation points of quadratures rules in Gridap
. Namely, using the get_cell_points
function we can extract a CellPoint
object out of a CellQuadrature
.
Qₕ_cell_point = get_cell_points(Qₕ)
CellPoint
(just as CellQuadrature
) is a subtype of CellDatum
as well
@test isa(Qₕ_cell_point, CellDatum)
and thus we can ask for the value of its DomainStyle
trait, and get an array of quantities out of it using the get_data
generic function
@test DomainStyle(Qₕ_cell_point) == ReferenceDomain()
qₖ = get_data(Qₕ_cell_point)
Not surprisingly, the DomainStyle
trait of the CellPoint
object is ReferenceDomain
, and we get a (cell) array with an array of Point
s per each cell out of a CellPoint
. As seen in the sequel, CellPoint
s are relevant objects because they are the ones that one can use in order to evaluate the so-called CellField
objects on the set of points of a CellPoint
.
CellField
is an abstract type rooted at a hierarchy that plays a cornerstone role in the implementation of the finite element method in Gridap
. At this point, the reader should keep in mind that the finite element method works with global spaces of functions which are defined piece-wise on each cell of the triangulation. In a nutshell (more in the sections below), a CellField
, as it being a subtype of CellDatum
, might be understood as a collection of Field
s (or arrays made out them) per each triangulation cell. Field
represents a field, e.g., a scalar, vector, or tensor field. Thus, the domain of a Field
are points in the physical domain (represented by a type Point
in Gridap
, which is a VectorValue
with a dimension matching that of the environment space) and the range is a scalar, vector (represented by VectorValue
) or tensor (represented by TensorValue
).
Unlike a plain array of Field
s, a CellField
is associated to a triangulation and is specifically designed having in mind FEs. For example, a global finite element function, or the collection of shape basis functions in the local FE space of each cell are examples of CellField
objects. As commented above, these fields can be defined in the physical or a reference space (combined with a geometrical map provided by the triangulation object for each cell). Thus, CellField
(as a sub-type of CellDatum
) has the DomainStyle
metadata that is used, e.g., for point-wise evaluations (as indicated above) of the fields and their derivatives (by implementing the transformations when taking a differential operators, e.g., the pull-back of the gradients).
Exploring our first CellField
objects
Let us work with our first CellField
objects, namely FEBasis
objects, and its evaluation. In particular, let us extract out of the global test space, Vₕ, and trial space, Uₕ, a collection of local test and trial finite element shape basis functions, respectively.
dv = get_fe_basis(Vₕ)
du = get_trial_fe_basis(Uₕ)
The objects returned are of FEBasis
type, one of the subtypes of CellField
. Apart from DomainStyle
, FEBasis
objects also have an additional trait, BasisStyle
, which specifies whether the cell-local shape basis functions are either of test or trial type (in the Galerkin method). This information is consumed in different parts of the code.
@test Gridap.FESpaces.BasisStyle(dv) == Gridap.FESpaces.TestBasis()
@test Gridap.FESpaces.BasisStyle(du) == Gridap.FESpaces.TrialBasis()
As expected, dv
is made out of test shape functions, and du
, of trial shape functions. We can also confirm that both dv
and du
are CellField
and CellDatum
objects (i.e., recall that FEBasis
is a subtype of CellField
, and the latter is a subtype of CellDatum
).
@test isa(dv,CellField) && isa(dv,CellDatum)
@test isa(du,CellField) && isa(du,CellDatum)
Thus, one may check the value of their DomainStyle
trait.
@test DomainStyle(dv) == ReferenceDomain()
@test DomainStyle(du) == ReferenceDomain()
We can see that the DomainStyle
of both FEBasis
objects is ReferenceDomain
. In the case of CellField
objects, this specifies that the point coordinates on which we evaluate the cell-local shape basis functions should be provided in the parametric space of the reference cell (to avoid the need to use the inverse of the geometrical map). However, the output from evaluation, as usual in finite elements defined parametrically, is the cell-local shape function in the physical domain evaluated at the corresponding mapped point.
Recall from above that CellField
objects are designed to be evaluated at CellPoint
objects, and that we extracted a CellPoint
object, Qₕ_cell_point
, out of a CellQuadrature
, of ReferenceDomain
trait DomainStyle
. Thus, we can evaluate dv
and du
at the quadrature rule evaluation points, on all cells, straight away as:
dv_at_Qₕ = evaluate(dv,Qₕ_cell_point)
du_at_Qₕ = evaluate(du,Qₕ_cell_point)
There are a pair of worth noting observations on the result of the previous two instructions. First, both dv_at_Qₕ
and du_at_Qₕ
are arrays of type Fill
(i.e., a constant array that only stores the entry once) because we are using the same quadrature and reference FE for all cells. This (same entry) is justified by: (1) the local shape functions are evaluated at the same set of points in the reference cell parametric space for all cells (i.e., the quadrature rule points), and (2) the shape functions in physical space have these very same values at the corresponding mapped points in the physical space for all cells. Thus they provide the same entry for whatever index we provide.
dv_at_Qₕ[rand(1:num_cells(Tₕ))]
du_at_Qₕ[rand(1:num_cells(Tₕ))]
At this point, the reader may want to observe which object results from the evaluation of, e.g., dv_at_Qₕ
, at a different set points for each cell (e.g. by building its own array of arrays of Points
).
Going back to our example, any entry of dv_at_Qₕ
is a rank-2 array of size 9x4 that provides in position [i,j]
the j-th test shape function at the i-th quadrature rule evaluation point. On the other hand, any entry of du_at_Qₕ
is a rank-3 array of size 9x1x4
that provides in position [i,1,j]
the j-th trial shape function at the i-th quadrature point. The reader might be wondering why the rank of these two arrays are different. The rationale is that, by means of the Julia broadcasting of the *
operation on these two arrays, we get the 9x4x4 array where the [i,j,k]
entry stores the product of the j-th test and k-th trial functions, both evaluated at the i-th quadrature point. If we sum over the $i$-index, we obtain part of the data required to compute the cell-local matrix that we assemble into the global matrix in order to get a mass matrix. For those readers more used to traditional finite element codes, the broadcast followed by the sum over i, provides the data required in order to implement the following triple standard for-nested loop:
M[:,:]=0.0
Loop over quadrature points i
detJK_wi=det(JK)*w[i]
Loop over shape test functions j
Loop over shape trial functions k
M[j,k]+=shape_test[i,j]*shape_trial[i,k]*detJK_wi
where det(JK)
represents the determinant of the reference-physical mapping of the current cell, and w[i]
the quadrature rule weight corresponding to the i-th evaluation point. Using Julia built-in support for broadcasting, we can vectorize the full operation, and get much higher performance.
The highest-level possible way of performing the aforementioned broadcasted *
is by building a "new" CellField
instance by multiplying the two FEBasis
objects, and then evaluating the resulting object at the points in Qₕ_cell_point
. This is something common in Gridap
. One can create new CellField
objects out of existing ones, e.g., by performing operations among them, or by applying a differential operator, such as the gradient.
dv_mult_du = du*dv
dv_mult_du_at_Qₕ = evaluate(dv_mult_du,Qₕ_cell_point)
We can check that any entry of the resulting Fill
array is the 9x4x4
array resulting from the broadcasted *
of the two aforementioned arrays. In order to do so, we can use the so-called Broadcasting(*)
Gridap
Map
(one of the cornerstones of Gridap
).
A Map
represents a (general) function (a.k.a. map or mapping) that takes elements in its domain and return elements in its range. A Field
is a sub-type of Map
for the particular domain and ranges of physical fields detailed above. Why do we need to define the Map
type in Gridap
instead of using the Julia Function
? Map
is essential for performance, as we will explain later on.
The Map
below is a map that broadcasts the *
operation. When applied to arrays of numbers, it essentially translates into the built-in Julia broadcast (check that below!). However, as we will see along the tutorial, such a Map
can also be applied to, e.g., (cell) arrays of Field
s (arrays of Field
s, resp.) to build new (cell) arrays of Fields
(arrays of Field
s, resp.). This becomes extremely useful to build and evaluate discrete variational forms.
m=Broadcasting(*)
A=evaluate(m,dv_at_Qₕ[rand(1:num_cells(Tₕ))],du_at_Qₕ[rand(1:num_cells(Tₕ))])
B=broadcast(*,dv_at_Qₕ[rand(1:num_cells(Tₕ))],du_at_Qₕ[rand(1:num_cells(Tₕ))])
@test all(A .≈ B)
@test all(A .≈ dv_mult_du_at_Qₕ[rand(1:num_cells(Tₕ))])
Recall from above that CellField
objects are also CellDatum
objects. Thus, one can use the get_data
generic function to extract, in an array, the collection of quantities, one per each cell of the triangulation, out of them. As one may expect, in the case of our FEBasis
objects dv
and du
at hand, get_data
returns a (cell) array of arrays of Field
objects, i.e., the cell-local shape basis functions:
dv_array = get_data(dv)
du_array = get_data(du)
@test isa(dv_array,AbstractVector{<:AbstractVector{<:Field}})
@test isa(du_array,AbstractVector{<:AbstractArray{<:Field,2}})
@test length(dv_array) == num_cells(Tₕ)
@test length(du_array) == num_cells(Tₕ)
As expected, both dv_array
and du_array
are (conceptually) vectors (i.e, rank-1 arrays) with as many entries as cells. The concrete type of each vector differs, though, i.e., Fill
and LazyArray
, resp. (We will come back to LazyArray
s below, as they play a fundamental role in the way in which the finite element method is implemented in Gridap
.) For each cell, we have arrays of Field
objects. Recall from above that Map
and Field
(with Field
a subtype of Map
), and CellDatum
and CellField
(with CellField
a subtype of CellDatum
) and the associated type hierarchies, are fundamental in Gridap
for the implementation of variational methods in finite-dimensional spaces. Field
conceptually represents a physical (scalar, vector, or tensor) field. Field
objects can be evaluated at single Point
objects (or at an array of them in one shot), and they return scalars (i.e., a sub-type of Julia Number
), VectorValue
, or TensorValue
objects (or an array of them, resp.)
In order to evaluate a Field
object at a Point
object, or at an array of Points
, we can use the evaluate
generic function in its API
. For example, the following statement
ϕ₃ = dv_array[1][3]
evaluate(ϕ₃,[Point(0,0),Point(1,0),Point(0,1),Point(1,1)])
evaluates the 3rd test shape function of the local space of the first cell at the 4 vertices of the cell (recall from above that, for the implementation of Lagrangian finite elements being used in this tutorial, shape functions are thought to be evaluated at point coordinates expressed in the parametric space of the reference cell). As expected, ϕ₃ evaluates to one at the 3rd vertex of the cell, and to zero at the rest of vertices, as ϕ₃ is the shape function associated to the Lagrangian node/ DOF located at the 3rd vertex. We can also evaluate all shape functions of the local space of the first cell (i.e., an array of Field
s) at once at an array of Points
ϕ = dv_array[1]
evaluate(ϕ,[Point(0,0),Point(1,0),Point(0,1),Point(1,1)])
As expected, we get the Identity matrix, as the shape functions of the local space have, by definition, the Kronecker delta property.
However, and here comes one of the main take-aways of this tutorial, in Gridap
, (cell-wise) arrays of Fields
(or arrays of Fields
) are definitely NOT conceived to be evaluated following the approach that we used in the previous examples, i.e., by manually extracting the Field
(array of Field
s) corresponding to a cell, and then evaluating it (them) at a given set of Point
s. Instead, one uses the lazy_map
generic function, which combined with the evaluate
function, represents the operation of walking over all cells, and evaluating the fields, cell by cell, as a whole. This is illustrated in the following piece of code:
dv_array_at_qₖ = lazy_map(evaluate,dv_array,qₖ)
du_array_at_qₖ = lazy_map(evaluate,du_array,qₖ)
We note that the results of these two expressions are equivalent to the ones of evaluate(dv, Qₕ_cell_point)
and evaluate(du,Qₕ_cell_point)
, resp. (check it!) In fact, these latter two expressions translate under the hood into the calls to lazy_map
above. These calls to lazy_map
return an array of the same length of the input arrays, with their i-th entry conceptually defined, e.g., as evaluate(du_array[i],qₖ[i])
in the case of the second array. To be "conceptually defined as" does not mean that they are actually computed as evaluate(du_array[i],qₖ[i])
. Indeed they don't, this would not be high performant.
You might now be wondering what the main point behind lazy_map
is. lazy_map
turns out to be a cornerstone in Gridap
. (At this point, you may execute methods(lazy_map)
to observe that a large amount of programming logic is devoted to it.) Let us try to answer it more abstractly now. However, this will be revisited along the tutorial with additional examples.
lazy_map
can be applied to a Map
and an array or a set of arrays, all with the same layout, that provide at every entry the arguments of the map. It conceptually returns the array that results from applying the Map
to the arguments in each index of the argument array(s). Usually, the resulting type is a LazyArray
.
When the resulting LazyArray
entries are also Map
s, one could evaluate
the LazyArray
on the array(s) that provide the argument(s) (i.e., its domain) using again a lazy_map
. E.g., for the sub-type Field
, one can create an array of fields, e.g., cell shape function, apply a Map
over this array, e.g., a scaling of the shape functions, using lazy_map
. The resulting array (conceptually also an array of Field
s) can be evaluated in a set of points applying evaluate
using lazy_map
, as in the two code lines above.
These lazy objects are cornerstones of Gridap
for the following reasons:
To keep memory allocation (and consumption) at very low levels,
lazy_map
NEVER returns an array that stores the result at all cells at once. In the two examples above, this is achieved usingFill
arrays. However, this is only possible in very particular scenarios (see discussion above). In more general cases, the array resulting fromlazy_map
does not have the same entry in all cells. In such cases,lazy_map
returns aLazyArray
, which is another essential component ofGridap
. In a nutshell, aLazyArray
is an array that applies entry-wise arrays ofMap
s (functions, operations) over array(s) that provide theMap
s arguments. These operations are only computed when accessing the corresponding index, thus the name lazy. Besides, the entries of these are computed in an efficient way, using a set of mechanisms that will be illustrated below with examples (e.g., using cache to store the entry-wise data without the need to allocate memory each time we access theLazyArray
).Apart from
Function
objects, such asevaluate
,lazy_map
can also be used to applyMap
s to arguments. For example,Broadcasting(*)
presented above. AMap
(or its sub-typeField
) can be applied vialazy_map
to otherMap
s (or arrays ofMap
s) to build a newMap
(or array ofMap
s). Thus, the recursive application oflazy_map
lets us build complex operation trees among arrays ofMap
s as the ones required for the implementation of variational forms. While building these trees, by virtue of Julia support for multiple type dispatching, there are plenty of opportunities for optimization by changing the order in which the operations are performed. These optimizations typically come in the form of a significant saving of FLOPs by exploiting the particular properties of theMap
s at hand, but could also come from higher granularity for vectorized array operations when the expressions are actually evaluated. Indeed, the arrays that one usually obtains fromlazy_map
differ in some cases from the trivialLazyArray
s that one would expect from a naive combination of the arguments tolazy_map
Using
lazy_map
we are hiding thousands of cell loops across the code (as the one for the computation of the element matrices above). As a result,Gridap
is much more expressive for cell-wise implementations.
Exploring another type of CellField
objects
Let us now work with another type of CellField
objects, the ones that are used to represent an arbitrary element of a global FE space of functions, i.e., a FE function. A global FE function can be understood conceptually as a collection of Field
s, one per each cell of the triangulation. The Field
corresponding to a cell represents the restriction of the global FE function to the cell. (Recall that in finite elements, global functions are defined piece-wise on each cell.) As we did in the previous section, we will explore, at different levels, how FE functions are evaluated. However, we will dig deeper into this by illustrating some of the aforementioned mechanisms on which LazyArray
relies in order to efficiently implement the entry-wise application of an operation (or array of operations) to a set of input arrays.
Let us now build a FE function belonging to the global trial space of functions Uₕ, with rand free DOF values. Using Gridap
higher-level API, this can be achieved as follows
uₕ = FEFunction(Uₕ,rand(num_free_dofs(Uₕ)))
As expected from the discussion above, the returned object is a CellField
object:
@test isa(uₕ,CellField)
Thus, we can, e.g., query the value of its DomainStyle
trait, that turns out to be ReferenceDomain
@test DomainStyle(uₕ) == ReferenceDomain()
Thus, in order to evaluate the Field
object that represents the restriction of the FE function to a given cell, we have to provide Point
s in the parametric space of the reference cell, and we get the value of the FE function at the corresponding mapped Point
s in the physical domain. This should not come as a surprise as we have that: (1) the restriction of the FE function to a given cell is mathematically defined as a linear combination of the local shape functions of the cell (with coefficients given by the values of the DOFs at the cell). (2) As observed in the previous section, the shape functions are such that their value at Point
s that are mapped from the reference cell to the physical cell by the cell geometrical map can be simply be obtained by evaluating the corresponding shape function in the reference FE at the same Point
s in the parametric space (without the need to compute the geometrical map and its inverse, i.e., exploiting the fact that the combination of this two is the identity map). This property is thus transferred to the FE function.
As FE functions are CellField
objects, we can evaluate them at CellPoint
objects. Let us do it at the points within Qₕ_cell_point
(see above for a justification of why this is possible):
uₕ_at_Qₕ = evaluate(uₕ,Qₕ_cell_point)
We note that internally this is just the application of evaluate
via lazy_map
for the raw (i.e., without the CellDatum
metadata) cell arrays of uₕ
and Qₕ_cell_point
(obtained via get_data
). Internally, a change_domain
is invoked if required, i.e., the two CellDatum
do not have the same DomainStyle
trait value (not the case here). You can check it by getting into this call using the VSCode debugger (write @enter
at the beginning of the line and run it). In any case, we provide many more details below.
For the first time in this tutorial, we have obtained a cell array of type LazyArray
from evaluating a CellField
at a CellPoint
.
@test isa(uₕ_at_Qₕ,LazyArray)
This makes sense as a finite element function restricted to a cell is, in general, different in each cell, i.e., it evaluates to different values at the quadrature rule evaluation points. In other words, the Fill
array optimization that was performed for the evaluation of the cell-wise local shape functions dv
and du
does not apply here, and a LazyArray
has to be used instead.
Although it is hard to understand the full concrete type name of uₕ_at_Qₕ
at this time
print(typeof(uₕ_at_Qₕ))
we will dissect LazyArray
s in this section up to an extent that will allow us to have a better grasp of it. By now, the most important thing for you to keep in mind is that LazyArray
s objects encode a recipe to produce its entries just-in-time when they are accessed. They NEVER store all of its entries at once. Even if the expression uₕ_at_Qₕ
typed in the REPL, (or inline evaluations in your code editor) show the array with all of its entries at once, don't get confused. This is because the Julia REPL is evaluating the array at all indices and collecting the result just for printing purposes.
Just as we did with FEBasis
, we can extract an array of Field
objects out of uₕ_at_Qₕ
, as uₕ_at_Qₕ
is also a CellBasis
object.
uₕ_array = get_data(uₕ)
As expected, uₕ_array
is (conceptually) a vector (i.e., rank-1 array) of Field
objects.
@test isa(uₕ_array,AbstractVector{<:Field})
Its concrete type is, though, LazyArray
@test isa(uₕ_array,Gridap.Fields.LazyArray)
with full name, as above, of a certain complexity (to say the least):
print(typeof(uₕ_array))
As mentioned above, lazy_map
returns LazyArray
s in the most general scenarios. Thus, it is reasonable to think that get_data(uₕ)
returns an array that has been built via lazy_map
. (We advance now that this is indeed the case.) On the other hand, as uₕ_array
is (conceptually) a vector (i.e., rank-1 array) of Field
objects, this also tells us that the lazy_map
/LazyArray
pair does not only play a fundamental role in the evaluation of (e.g., cell) arrays of Field
s on (e.g., cell) arrays of arrays of Point
s, but also in building new cell arrays of Field
s (i.e., the local restriction of a FE function to each cell) out of existing ones (i.e., the cell array with the local shape functions). In the words of the previous section, we can use lazy_map
to build complex operation trees among arrays of Field
s, as required by the computer implementation of variational methods.
The key question now is: what is the point behind get_data(uₕ)
returning a LazyArray
of Field
s, and not just a plain Julia array of Field
s? At the end of the day, Field
objects themselves have very low memory demands, they only need to hold the necessary information to encode their action (evaluation) on a Point
/array of Point
s. This is in contrast to the evaluation of (e.g., cell) arrays of Field
s (or arrays of Field
s) at an array of Point
s, which does consume a significantly allocation of memory (if all entries are to be stored at once in memory, and not by demand). The short answer is higher performance. Using LazyArray
s to encode operation trees among cell arrays of Field
s, we can apply optimizations when evaluating these operation trees that would not be possible if we just computed a plain array of Field
s. If all this sounds quite abstract, (most probably it does), we are going to dig into this further in the rest of the section.
As mentioned above, uₕ_array
can be conceptually seen as an array of Field
s. Thus, if we access to a particular entry of it, we should get a Field
object. (Although possible, this is not the way in which uₕ_array
is conceived to be used, as was also mentioned in the previous section.) This is indeed confirmed when accessing, e.g., the third entry of uₕ_array
:
uₕ³ = uₕ_array[3]
@test isa(uₕ³,Field)
The concrete type of uₕ³
is LinearCombinationField
. This type represents a Field
defined as a linear combination of an existing vector of Field
s. This sort of Field
s can be built using the linear_combination
generic function. Among its methods, there is one which takes (1) a vector of scalars (i.e., Julia Number
s) with the coefficients of the expansion and (2) a vector of Field
s as its two arguments, and returns a LinearCombinationField
object. As mentioned above, this is the exact mathematical definition of a FE function restricted to a cell.
Let us manually build uₕ³. In order to do so, we can first use the get_cell_dof_values
generic function, which extracts out of uₕ a cell array of arrays with the DOF values of uₕ restricted to all cells of the triangulation (defined from a conceptual point of view).
Uₖ = get_cell_dof_values(uₕ)
(The returned array turns to be of concrete type LazyArray
, again to keep memory allocation low, but let us skip this detail for the moment.) If we restrict Uₖ
and dv_array
to the third cell
Uₖ³ = Uₖ[3]
ϕₖ³ = dv_array[3]
we get the two arguments that we need to invoke linear_combination
in order to build our manually built version of uₕ³
manual_uₕ³ = linear_combination(Uₖ³,ϕₖ³)
We can double-check that uₕ³
and manual_uₕ³
are equivalent by evaluating them at the quadrature rule evaluation points, and comparing the result:
@test evaluate(uₕ³,qₖ[3]) ≈ evaluate(manual_uₕ³,qₖ[3])
Following this idea, we can go even further and manually build a plain Julia vector of LinearCombinationField
objects as follows:
manual_uₕ_array = [linear_combination(Uₖ[i],dv_array[i]) for i=1:num_cells(Tₕ)]
And we can (lazily) evaluate this manually-built array of Field
s at a cell array of arrays of Point
s (i.e., at qₖ
) using lazy_map
:
manual_uₕ_array_at_qₖ = lazy_map(evaluate,manual_uₕ_array,qₖ)
The entries of the resulting array are equivalent to those of the array that we obtained from Gridap
automatically, i.e., uₕ_at_Qₕ
@test all( uₕ_at_Qₕ .≈ manual_uₕ_array_at_qₖ )
However, and here it comes the key of the discussion, the concrete types of uₕ_at_Qₕ
and manual_uₕ_array_at_qₖ
do not match.
@test typeof(uₕ_at_Qₕ) != typeof(manual_uₕ_array_at_qₖ)
This is because evaluate(uₕ,Qₕ_cell_point)
does not follow the (naive) approach that we followed to build manual_uₕ_array_at_qₖ
, but it instead calls lazy_map
under the hood as follows
uₕ_array_at_qₖ = lazy_map(evaluate,uₕ_array,qₖ)
Now we can see that the types of uₕ_array_at_qₖ
and uₕ_at_Qₕ
match:
@test typeof(uₕ_array_at_qₖ) == typeof(uₕ_at_Qₕ)
Therefore, why Gridap
does not build manual_uₕ_array_at_qₖ
? what's wrong with it? Let us first try to answer this quantitatively. Let us assume that we want to sum all entries of a LazyArray
. In the case of LazyArray
s of arrays, this operation is only well-defined if the size of the arrays of all entries matches. This is the case of the uₕ_array_at_qₖ
and manual_uₕ_array_at_qₖ
arrays, as we have the same quadrature rule at all cells. We can write this function following the Gridap
internals' way.
function smart_sum(a::LazyArray)
cache=array_cache(a) #Create cache out of a
sum=copy(getindex!(cache,a,1)) #We have to copy the output
#from get_index! to avoid array aliasing
for i in 2:length(a)
ai = getindex!(cache,a,i) #Compute the i-th entry of a
#re-using work arrays in cache
sum .= sum .+ ai
end
sum
end
The function uses the so-called "cache" of a LazyArray
. In a nutshell, this cache can be thought as a place-holder of work arrays that can be re-used among different evaluations of the entries of the LazyArray
(e.g., the work array in which the result of the computation of an entry of the array is stored.) This way, the code is more performant, as the cache avoids that these work arrays are created repeatedly when traversing the LazyArray
and computing its entries. It turns out that LazyArray
s are not the only objects in Gridap
that (can) work with caches. Map
and Field
objects also provide caches for reusing temporary storage among their repeated evaluation on different arguments of the same types. (For the eager reader, the cache can be obtained out of a Map
/Field
with the return_cache
abstract method; see also return_type
, return_value
, and evaluate!
functions of the abstract API of Map
s). When a LazyArray
is created out of objects that in turn rely on caches (e.g., a LazyArray
with entries defined as the entry-wise application of a Map
to two LazyArrays
), the caches of the latter objects are also handled by the former object, so that this scheme naturally accommodates top-down recursion, as per-required in the evaluation of complex operation trees among arrays of Field
s, and their evaluation at a set of Point
s. We warn the reader this is a quite complex mechanism. The reader is encouraged to follow with a debugger, step by step, the execution of the smart_sum
function with the LazyArray
s built above in order to gain some familiarity with this mechanism.
If we @time the smart_sum
function with uₕ_array_at_qₖ
and manual_uₕ_array_at_qₖ
smart_sum(uₕ_array_at_qₖ) # Execute once before to neglect JIT-compilation time
smart_sum(manual_uₕ_array_at_qₖ) # Execute once before to neglect JIT-compilation time
@time begin
for i in 1:100_000
smart_sum(uₕ_array_at_qₖ)
end
end
@time begin
for i in 1:100_000
smart_sum(manual_uₕ_array_at_qₖ)
end
end
we can observe that the array returned by Gridap
can be summed in significantly less time, using significantly less allocations.
Let us try to answer the question now qualitatively. In order to do so, we can take a look at the structure of both LazyArray
s using the print_op_tree
function provided by Gridap
print_op_tree(uₕ_array_at_qₖ)
print_op_tree(manual_uₕ_array_at_qₖ)
We can observe from the output of these calls the following:
uₕ_array_at_qₖ
is aLazyArray
whose entries are defined as the result of applying aFill
array ofLinearCombinationMap{Colon}
Map
s to aLazyArray
and aFill
array. The first array provides the FE function DOF values restricted to each cell, and the second the local basis shape functions evaluated at the quadrature points. As the shape functions in physical space have the same values in all cells at the corresponding mapped points in physical space, there is no need to re-evaluate them at each cell, we can evaluate them only once. And this is what the secondFill
array stores as its unique entry, i.e., a matrixM[i,j]
defined as the value of the j-thField
(i.e., shape function) evaluated at the i-thPoint
. This is indeed the main optimization thatlazy_map
applies compared to our manual construction ofuₕ_array_at_qₖ
. It is worth noting that, ifv
denotes the linear combination coefficients, andM
the matrix resulting from the evaluation of an array ofFields
at a set ofPoints
, withM[i,j]
being the value of the j-thField
evaluated at the i-th point, the evaluation ofLinearCombinationMap{Colon}
atv
andM
returns a vectorw
withw[i]
defined asw[i]=sum_k v[k]*M[i,k]
, i.e., the FE function evaluated at the i-th point.uₕ_array_at_qₖ
handles the cache ofLinearCombinationMap{Colon}
(which holds internal storage forw
) and that of the firstLazyArray
, so that when it retrieves the DOF valuesv
of a given cell, and then appliesLinearCombinationMap{Colon}
tov
andM
, it does not have to allocate any temporary working arrays, but re-uses the ones stored in the different caches.manual_uₕ_array_at_qₖ
is also aLazyArray
, but structured rather differently touₕ_array_at_qₖ
. In particular, its entries are defined as the result of applying a plain array ofLinearCombinationField
s to aFill
array ofPoint
s that holds the coordinates of the quadrature rule evaluation points in the parametric space of the reference cell (which are equivalent for all cells, thus theFill
array). The evaluation of aLinearCombinationField
on a set ofPoint
s ultimately depends onLinearCombinationMap
. As seen in the previous point, the evaluation of thisMap
requires a vectorv
and a matrixM
.v
was built in-situ when building eachLinearCombinationField
, and stored within these instances. However, in contrast touₕ_array_at_qₖ
,M
is not part ofmanual_uₕ_array_at_qₖ
, and thus it has to be (re-)computed each time that we evaluate a newLinearCombinationField
instance on a set of points. This is the main source of difference on the computation times observed. By eagerly constructing our array ofLinearCombinationField
s instead of deferring it until (lazy) evaluation vialazy_map
, we lost optimization opportunities. We stress thatmanual_uₕ_array_at_qₖ
also handles the cache ofLinearCombinationField
(that in turn handles the one ofLinearCombinationMap
), so that we do not need to allocateM
at each cell, we re-use the space within the cache ofLinearCombinationField
.
To conclude the section, we expect the reader to be convinced of the negative consequences in performance that an eager (early) evaluation of the entries of the array returned by a lazy_map
call can have in performance. The leitmotif of Gridap
is laziness. When building new arrays of Field
s (or arrays of Field
s), out of existing ones, or when evaluating them at a set of Point
s, ALWAYS use lazy_map
. This may expand across several recursion levels when building complex operation trees among arrays of Field
s. The more we defer the actual computation of the entries of LazyArray
s, the more optimizations will be available at the Gridap
's disposal by re-arranging the order of operations via exploitation of the particular properties of the arrays at hand. And this is indeed what we are going to do in the rest of the tutorial, namely calling lazy_map
to build new cell arrays out of existing ones, to end in a lazy cell array whose entries are the cell matrices and cell vectors contributions to the global linear system.
Let us, e.g., build Uₖ manually using this idea. First, we extract out of uₕ and Uₕ two arrays with the free and fixed (due to strong Dirichlet boundary conditions) DOF values of uₕ
uₕ_free_dof_values = get_free_dof_values(uₕ)
uₕ_dirichlet_dof_values = get_dirichlet_dof_values(Uₕ)
So far these are plain arrays, nothing is lazy. Then we extract out of Uₕ the global indices of the DOFs in each cell, the well-known local-to-global map in FE methods.
σₖ = get_cell_dof_ids(Uₕ)
Finally, we call lazy_map to build a LazyArray
, whose entries, when computed, contain the global FE function DOFs restricted to each cell.
m = Broadcasting(PosNegReindex(uₕ_free_dof_values,uₕ_dirichlet_dof_values))
manual_Uₖ = lazy_map(m,σₖ)
PosNegReindex
is a Map
that is built out of two vectors. We evaluate it at indices of array entries. When we give it a positive index, it returns the entry of the first vector corresponding to this index, and when we give it a negative index, it returns the entry of the second vector corresponding to the flipped-sign index. We can check this with the following expressions
@test evaluate(PosNegReindex(uₕ_free_dof_values,uₕ_dirichlet_dof_values),3) == uₕ_free_dof_values[3]
@test evaluate(PosNegReindex(uₕ_free_dof_values,uₕ_dirichlet_dof_values),-7) == uₕ_dirichlet_dof_values[7]
The Broadcasting(op)
Map
lets us, in this particular example, broadcast the PosNegReindex(uₕ_free_dof_values,uₕ_dirichlet_dof_values)
Map
to an array a global DOF ids, to obtain the corresponding cell DOF values. As regular, Broadcasting(op)
provides a cache with the work array required to store its result. LazyArray
uses this cache to reduce the number of allocations while computing its entries just-in-time. Please note that in Gridap
we put negative labels to fixed DOFs and positive to free DOFs in σₖ, thus we use an array that combines σₖ with the two arrays of free and fixed DOF values accessing the right one depending on the index. But everything is lazy, only computed when accessing the array. As mentioned multiple times, laziness is one f the leitmotifs in Gridap, the other being immutability.
Immutability is a feature that comes from functional programming. An immutable object cannot be modified after created. Since objects cannot change, one does not require to track how they change, i.e., there is no need to design (and understand) state diagrams. A code that strictly sticks to this principle is much more readable. Due to laziness, Gridap
objects are light-weight, and the (lazy) modification of existing (lazy) objects is highly efficient. You can find this action many times in the code above, in which we use lazy_map
to perform actions over lazy objects (e.g., LazyArray
or Fill
arrays) to create new lazy objects. However, strictly conforming to immutability can be inefficient in some very specific scenarios. Gridap
departs from immutability in the linear algebra part, since we want to re-use the memory allocation as much as possible for global arrays or symbolic/numeric factorisations in linear solvers.
The geometrical model
From the triangulation we can also extract the cell map, i.e., the geometrical map that takes points in the parametric space $[0,1]^D$ (the SEGMENT
, QUAD
, or HEX
in 1, 2, or 3D, resp.) and maps it to the cell in the physical space $\Omega$.
ξₖ = get_cell_map(Tₕ)
We note that this map is just a LazyArray
of Field
s. The metadata related to CellField
is not required here, the cell map can only go from the reference to physical space, and its domain can only be a reference cell. For this reason, it is not a CellField
.
The cell map takes at each cell points in the parametric space and returns the mapped points in the physical space. Even though this space does not need a global definition (nothing has to be solved here), it is continuous across interior faces.
As usual, this cell_map is a LazyArray
. At each cell, it provides the Field
that maps Point
s in the parametric space of the reference cell to Point
s in physical space.
The node coordinates can be extracted from the triangulation, returning a global array of Point
s. You can see that such array is stored using Cartesian indices instead of linear indices. It is more natural for Cartesian meshes.
X = get_node_coordinates(Tₕ)
You can also extract a cell-wise array that provides the node indices per cell
cell_node_ids = get_cell_node_ids(Tₕ)
or the cell-wise nodal coordinates, combining the previous two arrays
_Xₖ = get_cell_coordinates(Tₕ)
A low-level definition of the cell map
Now, let us create the geometrical map almost from scratch, using the concepts that we have learned so far. In this example, we consider that the geometry is represented with a bilinear map, and we thus use a first-order, scalar-valued FE space to represent the nodal coordinate values. To this end, as we did before with the global space of FE functions, we first need to create a Polytope using an array of dimension D
with the parameter HEX_AXIS
. Then, this is used to create the scalar first order Lagrangian reference FE.
pol = Polytope(Fill(HEX_AXIS,D)...)
reffe_g = LagrangianRefFE(Float64,pol,1)
Next, we extract the basis of shape functions out of this Reference FE, which is a set of Field
s, as many as shape functions. We note that these Field
s have as domain the parametric space $[0,1]^D$. Thus, they can readily be evaluated for points in the parametric space.
ϕrg = get_shapefuns(reffe_g)
Now, we create a global cell array that has the same reference FE basis for all cells.
ϕrgₖ = Fill(ϕrg,num_cells(Tₕ))
Next, we use lazy_map
to build a LazyArray
that provides the coordinates of the nodes of each cell in physical space. To this end, we use the Broadcasting(Reindex(X))
and apply it to cell_node_ids
.
Xₖ = lazy_map(Broadcasting(Reindex(X)),cell_node_ids)
Reindex
is a Map
that is built out of a single vector, X
in this case. We evaluate it at indices of array entries, and it just returns the entry of the vector from which it is built corresponding to this index. We can check this with the following expressions:
@test evaluate(Reindex(X),3) == X[3]
If we combine Broadcasting
and Reindex
, then we can evaluate efficiently the Reindex
Map
at arrays of node ids, i.e., at each of the entries of cell_node_ids
. lazy_map
is used for reasons hopefully clear at this point (low memory consumption, efficient computation of the entries via caches, further opportunities for optimizations when combined with other lazy_map
calls, etc.). Finally, we can check that Xₖ
is equivalent to the array returned by Gridap
, i.e, _Xₖ
@test Xₖ == _Xₖ == get_cell_coordinates(Tₕ) # check
Next, we can compute the geometrical map as the linear combination of these shape functions in the parametric space with the node coordinates (at each cell)
ψₖ = lazy_map(linear_combination,Xₖ,ϕrgₖ)
This is the mathematical definition of the geometrical map in FEs! (see above for a description of the linear_combination
generic function). As expected, the FE map that we have built manually is equivalent to the one internally built by Gridap
.
@test lazy_map(evaluate,ψₖ,qₖ) == lazy_map(evaluate,ξₖ,qₖ) # check
It is good to stress (if it was not fully grasped yet) that lazy_map(k,a,b)
, with k
being a callable Julia object, is semantically (conceptually) equivalent to map(k,a,b)
but, among others, with a lazy result instead of a plain Julia array. A Julia object is callable if it makes sense to pass arguments to it. For example, objects k
such that isa(k,Map)==true
are callable. For these objects, k(x...)
is equivalent to evaluate(k,x...)
.
Following the same ideas, we can compute the Jacobian of the geometrical map (cell-wise). The Jacobian of the transformation is simply its gradient. The gradient in the parametric space can be built using two equivalent approaches. On the one hand, we can apply the Broadcasting(∇)
Map
to the array of Fields
with the local shape basis functions (i.e., ϕrg
). This results in an array of Field
s with the gradients, (Recall that Map
s can be applied to array of Field
s in order to get new array of Field
s) that we use to build a Fill
array with the result. Finally, we build the lazy array with the cell-wise Jacobians of the map as the linear combination of the node coordinates and the gradients of the local cell shape basis functions:
∇ϕrg = Broadcasting(∇)(ϕrg)
∇ϕrgₖ = Fill(∇ϕrg,num_cells(model))
J = lazy_map(linear_combination,Xₖ,∇ϕrgₖ)
We note that lazy_map
is not required in the first expression, as we are not actually working with cell arrays. On the other hand, using lazy_map
, we can apply Broadcasting(∇)
to the cell array of Field
s with the geometrical map.
lazy_map(Broadcasting(∇),ψₖ)
As mentioned above, those two approaches are equivalent
@test typeof(J) == typeof(lazy_map(Broadcasting(∇),ψₖ))
@test lazy_map(evaluate,J,qₖ) == lazy_map(evaluate,lazy_map(Broadcasting(∇),ψₖ),qₖ)
Computing the gradients of the trial and test FE space bases
Another salient feature of Gridap is that we can directly take the gradient of finite element bases. (In general, of any CellField
object.) In the following code snippet, we do so for dv
and du
grad_dv = ∇(dv)
grad_du = ∇(du)
The result of this operation when applied to a FEBasis
object is a new FEBasis
object.
@test isa(grad_dv, Gridap.FESpaces.FEBasis)
@test isa(grad_du, Gridap.FESpaces.FEBasis)
We can also extract an array of arrays of Fields
, as we have done before with FEBasis
objects.
grad_dv_array = get_data(grad_dv)
grad_du_array = get_data(grad_du)
The resulting LazyArray
s encode the so-called pull back transformation of the gradients. We need this transformation in order to compute the gradients in physical space. The gradients in physical space are indeed the ones that we need to integrate in the finite element method, not the reference ones, even if we always evaluate the integrals in the parametric space of the reference cell. We can also check that the DomainStyle
trait of grad_dv
and grad_du
is ReferenceDomain
@test DomainStyle(grad_dv) == ReferenceDomain()
@test DomainStyle(grad_du) == ReferenceDomain()
This should not come as a surprise, as this is indeed the nature of the pull back transformation of the gradients. We provide Point
s in the parametric space of the reference cell, and we get back the gradients in physical space evaluated at the mapped Point
s in physical space.
We can manually build grad_dv_array
and grad_du_array
as follows
ϕr = get_shapefuns(reffe)
∇ϕr = Broadcasting(∇)(ϕr)
∇ϕrₖ = Fill(∇ϕr,num_cells(Tₕ))
manual_grad_dv_array = lazy_map(Broadcasting(push_∇),∇ϕrₖ,ξₖ)
∇ϕrᵀ = Broadcasting(∇)(transpose(ϕr))
∇ϕrₖᵀ = Fill(∇ϕrᵀ,num_cells(Tₕ))
manual_grad_du_array = lazy_map(Broadcasting(push_∇),∇ϕrₖᵀ,ξₖ)
We note the use of the Broadcasting(push_∇)
Map
at the last step. For Lagrangian FE spaces, this Map
represents the pull back of the gradients. This transformation requires the gradients of the shape functions in the reference space, and the (gradient of the) geometrical map. The last step, e.g., the construction of manual_grad_dv_array
, actually translates into the combination of the following calls to lazy_map
to build the final transformation:
Build array of Field
s with the Jacobian transposed at each cell
Jt = lazy_map(Broadcasting(∇),ξₖ)
Build array of Field
s with the inverse of the Jacobian transposed at each cell
inv_Jt = lazy_map(Operation(inv),Jt)
Build array of arrays of Field
s defined as the broadcasted single contraction of the Jacobian inverse transposed and the gradients of the shape functions in the reference space
low_level_manual_gradient_dv_array = lazy_map(Broadcasting(Operation(⋅)),inv_Jt,∇ϕrₖ)
As always, we check that all arrays built are are equivalent
@test typeof(grad_dv_array) == typeof(manual_grad_dv_array)
@test lazy_map(evaluate,grad_dv_array,qₖ) == lazy_map(evaluate,manual_grad_dv_array,qₖ)
@test lazy_map(evaluate,grad_dv_array,qₖ) == lazy_map(evaluate,low_level_manual_gradient_dv_array,qₖ)
@test lazy_map(evaluate,grad_dv_array,qₖ) == evaluate(grad_dv,Qₕ_cell_point)
With the lessons learned so far in this section, it is left as an exercise for the reader to manually build the array that get_data
returns when we call it with the CellField
object resulting from taking the gradient of uₕ as an argument, i.e., get_data(∇(uₕ))
.
A low-level implementation of the residual integration and assembly
Let us now create manually an array of Field
s uₖ that returns the FE function uₕ at each cell, and another array with its gradients, ∇uₖ. We hope that the next set of instructions can be already understood with the material covered so far
ϕrₖ = Fill(ϕr,num_cells(Tₕ))
∇ϕₖ = manual_grad_dv_array
uₖ = lazy_map(linear_combination,Uₖ,ϕrₖ)
∇uₖ = lazy_map(linear_combination,Uₖ,∇ϕₖ)
Let us consider now the integration of (bi)linear forms. The idea is to compute first the following residual for our random function uₕ
intg = ∇(uₕ)⋅∇(dv)
but we are going to do it using low-level methods instead.
First, we create an array that for each cell returns the dot product of the gradients
Iₖ = lazy_map(Broadcasting(Operation(⋅)),∇uₖ,∇ϕₖ)
This array is equivalent to the one within the intg
CellField
object
@test all(lazy_map(evaluate,Iₖ,qₖ) .≈ lazy_map(evaluate,get_data(intg),qₖ))
Now, we can finally compute the cell-wise residual array, which using the high-level integrate
function is
res = integrate(∇(uₕ)⋅∇(dv),Qₕ)
In a low-level, what we do is to apply (create a LazyArray
) the IntegrationMap
Map
over the integrand evaluated at the integration points, the quadrature rule weights, and the Jacobian evaluated at the integration points
Jq = lazy_map(evaluate,J,qₖ)
intq = lazy_map(evaluate,Iₖ,qₖ)
iwq = lazy_map(IntegrationMap(),intq,Qₕ.cell_weight,Jq)
@test all(res .≈ iwq)
The result is the cell-wise residual (previous to assembly). This is a lazy array but you could collect the element residuals into a plain Julia array if you want
collect(iwq)
Alternatively, we can use the following syntactic sugar
cellvals = ∫( ∇(dv)⋅∇(uₕ) )*Qₕ
and check that we get the same cell-wise residual as the one defined above
@test all(cellvals .≈ iwq)
Assembling a residual
Now, we need to assemble these cell-wise (lazy) residual contributions in a global (non-lazy) array. With all this, we can assemble our vector using the cell-wise residual contributions and the assembler. Let us create a standard assembler struct for the finite element spaces at hand. This will create a vector of size global number of DOFs, and a SparseMatrixCSC
, to which we can add contributions.
assem = SparseMatrixAssembler(Uₕ,Vₕ)
We create a tuple with 1-entry arrays with the cell vectors (i.e., iwq
) and cell-wise global DOF identifiers (i.e., σₖ
). If we had additional terms, we would have more entries in the array. You can take a look at the SparseMatrixAssembler
struct for more details.
rs = ([iwq],[σₖ])
b = allocate_vector(assem,rs)
assemble_vector!(b,assem,rs)
A low-level implementation of the Jacobian integration and assembly
After computing the residual, we use similar ideas for the Jacobian. The process is the same as above, so it does not require additional explanations
∇ϕₖᵀ = manual_grad_du_array
int = lazy_map(Broadcasting(Operation(⋅)),∇ϕₖ,∇ϕₖᵀ)
@test all(collect(lazy_map(evaluate,int,qₖ)) .==
collect(lazy_map(evaluate,get_data(∇(du)⋅∇(dv)),qₖ)))
intq = lazy_map(evaluate,int,qₖ)
Jq = lazy_map(evaluate,J,qₖ)
iwq = lazy_map(IntegrationMap(),intq,Qₕ.cell_weight,Jq)
jac = integrate(∇(dv)⋅∇(du),Qₕ)
@test collect(iwq) == collect(jac)
rs = ([iwq],[σₖ],[σₖ])
A = allocate_matrix(assem,rs)
A = assemble_matrix!(A,assem,rs)
Now we can obtain the free DOFs and add the solution to the initial guess
x = A \ b
uf = get_free_dof_values(uₕ) - x
ufₕ = FEFunction(Uₕ,uf)
@test sum(integrate((u-ufₕ)*(u-ufₕ),Qₕ)) <= 10^-8
or if you like Unicode symbols
@test ∑(∫(((u-ufₕ)*(u-ufₕ)))Qₕ) <= 10^-8
This page was generated using Literate.jl.