pystencils

Run blazingly fast stencil codes on numpy arrays.

pystencils uses sympy to define stencil operations, that can be executed on numpy arrays. Exploiting the stencil structure makes pystencils run faster than normal numpy code and even as Cython and numba, as demonstrated in this notebook.

Here is a code snippet that computes the average of neighboring cells:

import pystencils as ps
import numpy as np

f, g = ps.fields("f, g : [2D]")
stencil = ps.Assignment(g[0, 0],
                       (f[1, 0] + f[-1, 0] + f[0, 1] + f[0, -1]) / 4)
kernel = ps.create_kernel(stencil).compile()

f_arr = np.random.rand(1000, 1000)
g_arr = np.empty_like(f_arr)
kernel(f=f_arr, g=g_arr)

pystencils is mostly used for numerical simulations using finite difference or finite volume methods. It comes with automatic finite difference discretization for PDEs:

c, v = ps.fields("c, v(2): [2D]")
adv_diff_pde = ps.fd.transient(c) - ps.fd.diffusion(c, sp.symbols("D")) + ps.fd.advection(c, v)
discretize = ps.fd.Discretization2ndOrder(dx=1, dt=0.01)
discretization = discretize(adv_diff_pde)

Installation

pip install pystencils[interactive]

Without [interactive] you get a minimal version with very little dependencies.

All options:

• gpu: use this if an NVIDIA GPU is available and CUDA is installed
• opencl: basic OpenCL support (experimental)
• alltrafos: pulls in additional dependencies for loop simplification e.g. libisl
• bench_db: functionality to store benchmark result in object databases
• interactive: installs dependencies to work in Jupyter including image I/O, plotting etc.
• autodiff: enable derivation of adjoint kernels and generation of Torch/Tensorflow operations
• doc: packages to build documentation

Options can be combined e.g.

pip install pystencils[interactive,gpu,doc]

Documentation

Read the docs here and check out the Jupyter notebooks in doc/notebooks.