The stencil skeleton is a particular data-parallel computation in which data dependence pattern is implemented by means of inter-worker communications.

A stencil is a gather in which all memory addresses used for reads are expressed as offsets relative to the location of a collection of outputs. Suppose we have an input array A, an output array B, and a set of integer offsets P. Then for every index k in the output array, we read from every input in A given by k + p for every p in P. Note that the index values in P may be multidimensional if the inputs and outputs are multidimensional. A stencil accesses a regular neighborhood of the input for each output.

The stencil pattern appears in a number of applications, but most notably in image processing and simulation. A wide variety of linear and non-linear image processing operations are specified using stencils, including linear convolution and non-linear noise reduction. Explicit solutions to partial differential equations, used in simulation and seismic reconstruction as well as in image processing, use iterative applications of stencil operations.

The implementation of the skeleton is entirely in Python and MPI. I have used mpi4py package and numpy for the matrix representation. The tests were run on top of a cluster of multicores machines using MPICHv1. Please refer to the official page of mpi4py package for further information regarding how to set up the proper testing environment.