SparseFunction

class devito.types.SparseFunction(*args)[source]

Bases: AbstractSparseFunction

Tensor symbol representing a sparse array in symbolic equations.

A SparseFunction carries multi-dimensional data that are not aligned with the computational grid. As such, each data value is associated some coordinates. A SparseFunction provides symbolic interpolation routines to convert between Functions and sparse data points. These are based upon standard [bi,tri]linear interpolation.

Parameters:

name (str) – Name of the symbol.
npoint (int) – Number of sparse points.
grid (Grid) – The computational domain from which the sparse points are sampled.
coordinates (np.ndarray, optional) – The coordinates of each sparse point.
space_order (int, optional) – Discretisation order for space derivatives. Defaults to 0.
shape (tuple of ints, optional) – Shape of the object. Defaults to (npoint,).
dimensions (tuple of Dimension, optional) – Dimensions associated with the object. Only necessary if the SparseFunction defines a multi-dimensional tensor.
dtype (data-type, optional) – Any object that can be interpreted as a numpy data type. Defaults to np.float32.
initializer (callable or any object exposing the buffer interface, optional) – Data initializer. If a callable is provided, data is allocated lazily.
allocator (MemoryAllocator, optional) – Controller for memory allocation. To be used, for example, when one wants to take advantage of the memory hierarchy in a NUMA architecture. Refer to default_allocator.__doc__ for more information.

Examples

Creation

>>> from devito import Grid, SparseFunction
>>> grid = Grid(shape=(4, 4))
>>> sf = SparseFunction(name='sf', grid=grid, npoint=2)
>>> sf
sf(p_sf)

Inspection

>>> sf.data
Data([0., 0.], dtype=float32)
>>> sf.coordinates
sf_coords(p_sf, d)
>>> sf.coordinates_data
array([[0., 0.],
       [0., 0.]], dtype=float32)

Symbolic interpolation routines

>>> from devito import Function
>>> f = Function(name='f', grid=grid)
>>> exprs0 = sf.interpolate(f)
>>> exprs1 = sf.inject(f, sf)

Notes

The parameters must always be given as keyword arguments, since SymPy uses *args to (re-)create the Dimension arguments of the symbolic object. About SparseFunction and MPI. There is a clear difference between:

Where the sparse points physically live, i.e., on which MPI rank. This depends on the user code, particularly on how the data is set up.

and which MPI rank logically owns a given sparse point. The logical ownership depends on where the sparse point is located within self.grid.

Right before running an Operator (i.e., upon a call to op.apply), a SparseFunction “scatters” its physically owned sparse points so that each MPI rank gets temporary access to all of its logically owned sparse points. A “gather” operation, executed before returning control to user-land, updates the physically owned sparse points in self.data by collecting the values computed during op.apply from different MPI ranks.

property data

The domain data values, as a numpy.ndarray.

Elements are stored in row-major format.

Notes

With this accessor you are claiming that you will modify the values you get back. If you only need to look at the values, use data_ro() instead.

property data_domain

The domain data values.

Elements are stored in row-major format.

Notes

Alias to self.data.

With this accessor you are claiming that you will modify the values you get back. If you only need to look at the values, use data_ro_domain() instead.

property data_ro_domain: Read-only view of the domain data values.

property data_ro_with_halo: Read-only view of the domain+outhalo data values.

property data_with_halo

The domain+outhalo data values.

Elements are stored in row-major format.

Notes

With this accessor you are claiming that you will modify the values you get back. If you only need to look at the values, use data_ro_with_halo() instead.

property dimensions: Tuple of Dimensions representing the object indices.

property dtype

The data type of the object in the generated code, represented as a Python class:

numpy.dtype: basic data types. For example, np.float64 -> double.

ctypes: composite objects (e.g., structs), foreign types.

property grid: The Grid on which the discretization occurred.

guard(expr=None)

Generate guarded expressions, that is expressions that are evaluated by an Operator only if certain conditions are met. The introduced condition, here, is that all grid points in the support of a sparse value must fall within the grid domain (i.e., not on the halo).

Parameters:: expr (expr-like, optional) – Input expression, from which the guarded expression is derived. If not specified, defaults to self.

inject(*args, **kwargs): Implement an injection operation from a sparse point onto the grid

interpolate(*args, **kwargs): Implement an interpolation operation from the grid onto the given sparse points

property name: The name of the object.

shape

Shape of the domain region. The domain constitutes the area of the data written to by an Operator.

Notes

In an MPI context, this is the local domain region shape.

property space_order: The space order.