Vectors¶
Overview¶
Warning
The API of the classes probably will change in the future as currently parallelization is not supported.
NeoN implements several field classes:
Vector<ValueType>
the basic GPU capable container class supporting algebraic operationsBoundaryData<ValueType>
A GPU friendly datastructure storing boundary data.Field<ValueType>
The combination of an internal field and its corresponding boundary data.
Besides these container like field classes several finite volume specific field classes are implemented. The corresponding classes are:
DomainMixin<ValueType>
Mixin class combining aField
and the corresponding mesh.VolumeVector<ValueType>
Uses the DomainMixin and implements finite volume specific members, including the notion of concrete boundary condiditonsSurfaceVector<ValueType>
The surface field equivalent toVolumeVector
The Vector<ValueType> class¶
The Vector class is the basic container class and is the central component for implementing a platform portable CFD framework.
One of the key differences between accessing the elements of a Vector
and typical std
sequential data containers is the lack of subscript or direct element access operators.
This is to prevent accidental access to device memory from the host.
The correct procedure to access Vector
elements is indirectly through a view
, as shown below:
.. code-block:: cpp
// Host Vectors Vector<T> hostVector(Executor::CPUExecutor, size_); auto hostVectorView = hostVector.view(); hostVectorView[1] = 1; // assuming size_ > 2.
// Device Vectors Vector<T> deviceVector(Executor::GPUExecutor, size_); auto deviceVectorOnHost = deviceVector.copyToHost(); auto deviceVectorOnHostView = deviceVectorOnHost.view(); deviceVectorOnHostView[1] = 1; // assuming size_ > 2.
Vectors support basic algebraic operations such as binary operations like the addition or subtraction of two fields, or scalar operations like the multiplication of a field with a scalar. In the following, some implementation details of the field operations are detailed using the additions operator as an example. The block of code below shows an example implementation of the addition operator.
[[nodiscard]] Vector<T> operator+(const Vector<T>& rhs)
{
Vector<T> result(exec_, size_);
result = *this;
add(result, rhs);
return result;
}
Besides creating a temporary for the result it mainly calls the free standing add
function which is implemented in include/NeoN/field/fieldFreeFunctions.hpp
.
This, in turn, dispatches to the fieldBinaryOp
function, passing the actual kernel as lambda.
The fieldBinaryOp
is implemented using our parallelFor implementations which ultimately dispatch to the Kokkos::parallel_for
function, see Kokkos documentation for more details.
template<typename ValueType>
void add(Vector<ValueType>& a, const Vector<std::type_identity_t<ValueType>>& b)
{
detail::fieldBinaryOp(
a, b, KOKKOS_LAMBDA(ValueType va, ValueType vb) { return va + vb; }
);
}
A simplified version of the parallelFor
function is shown below.
The code snippet highlights another important aspect, the executor.
The executor defines the Kokkos::RangePolicy
, see Kokkos Programming Model.
Besides defining the RangePolicy, the executor also holds functions for allocating and deallocationg memory.
See our documentation for more details on the executor model.
Further Details.
Cell Centred Specific Vectors¶
Within in the finiteVolume/cellCentred
folder and the namespace
NeoN::finiteVolume::cellCentred
two specific field types, namely the VolumeVector
and the SurfaceVector
are implemented.
Both derive from the DomainMixin
a mixin class which handles that all derived fields contain geometric information via the mesh data member and field specific data via the Field
data member.
Field
acts as the fundamental data container within this structure, offering both read and write to the internalVector
and boundaryVectors
data structure holding actual boundary data.
The VolumeVector
and the SurfaceVector
hold a vector of boundary conditions implemented in finiteVolume/cellCentred/boundary
and a correctBoundaryConditions
member function that updates the field’s boundary condition.
Functionally, the VolumeVector
and the SurfaceVector
classes are comparable to OpenFOAM classes such as volScalarVector
, volVec3Vector
, and volTensorVector
or surfaceScalarVector
, surfaceVec3Vector
, and surfaceTensorVector
respectively.
A difference in the SurfaceVector implementation is that the internalVector
also contains the boundary values, so no branches (if) are required when iterating over all cell faces.
Thus the size of the internalVector
in NeoN differs from that of OpenFOAM.
Further details VolumeVector and ScalarVector.