Devito: Fast Stencil Computation from Symbolic Specification
Devito is a Python package to implement optimized stencil computation (e.g., finite differences, image processing, machine learning) from high-level symbolic problem definitions. Devito builds on SymPy and employs automated code generation and just-in-time compilation to execute optimized computational kernels on several computer platforms, including CPUs, GPUs, and clusters thereof.
Devito provides a functional language to implement sophisticated operators that can be made up of multiple stencil computations, boundary conditions, sparse operations (e.g., interpolation), and much more. A typical use case is explicit finite difference methods for approximating partial differential equations. For example, a 2D diffusion operator may be implemented with Devito as follows
>>> grid = Grid(shape=(10, 10)) >>> f = TimeFunction(name='f', grid=grid, space_order=2) >>> eqn = Eq(f.dt, 0.5 * f.laplace) >>> op = Operator(Eq(f.forward, solve(eqn, f.forward)))
Operator generates low-level code from an ordered collection of
example above being for a single equation). This code may also be compiled and
There is virtually no limit to the complexity of an
Operator -- the Devito
compiler will automatically analyze the input, detect and apply optimizations
(including single- and multi-node parallelism), and eventually generate code
with suitable loops and expressions.
Key features include:
- A functional language to express finite difference operators.
- Straightforward mechanisms to adjust the discretization.
- Constructs to express sparse operators (e.g., interpolation), classic linear operators (e.g., convolutions), and tensor contractions.
- Seamless support for boundary conditions and adjoint operators.
- A flexible API to define custom stencils, sub-domains, sub-sampling, and staggered grids.
- Generation of highly optimized parallel code (SIMD vectorization, CPU and GPU parallelism via OpenMP, multi-node parallelism via MPI, blocking, aggressive symbolic transformations for FLOP reduction, etc.).
- Distributed NumPy arrays over multi-node (MPI) domain decompositions.
- Inspection and customization of the generated code.
- Autotuning framework to ease performance tuning.
- Smooth integration with popular Python packages such as NumPy, SymPy, Dask, and SciPy, as well as machine learning frameworks such as TensorFlow and PyTorch.
The easiest way to try Devito is through Docker using the following commands:
# get the code git clone https://github.com/devitocodes/devito.git cd devito # start a jupyter notebook server on port 8888 docker-compose up devito
After running the last command above, the terminal will display a URL such as
https://127.0.0.1:8888/?token=XXX. Copy-paste this URL into a browser window
to start a Jupyter notebook session where you can go
through the tutorials
provided with Devito or create your own notebooks.
To learn how to use Devito, here is a good place to start, with lots of examples and tutorials.
The website also provides access to other information, including documentation and instructions for citing us.
Some FAQ are discussed here.
If you are interested in any of the following
- Generation of parallel code (CPU, GPU, multi-node via MPI);
- Performance tuning;
- Benchmarking operators;
then you should take a look at this README.
You may also be interested in TheMatrix -- a cross-architecture benchmarking framework showing the performance of several production-grade seismic operators implemented with Devito. This is now our flagship project towards neat, open, and reproducible science.
Get in touch
If you're using Devito, we would like to hear from you. Whether you are facing issues or just trying it out, join the conversation.
Interactive jupyter notebooks
The tutorial jupyter notebook are available interactively at the public binder jupyterhub.