Overview

OpFlow is a header-only C++ embedded domain specific language (EDSL) designed for numerical simulation. It adopts the meta programming facilities provided by modern C++ heavily to provide an expressive front-end user interface while preserving the original expression context for backend optimization. The aims of developing OpFlow include:

• Enable users to use intuitive mathematical symbols & syntax to write numerical algorithms. Meanwhile, enable users to control the discretized method in a declarative manner instead of writing the detailed scheme repeatedly.

• Decouple the numerical algorithm with the underlying data structures. Users can switch between different implementations for the numerical expressions, trying out different data containers for the expressions for better performance while left the implementation of the algorithm untouched.

• Release the burden for the users to write tailored code for fully utilization of modern hardware. Users need and only need to care about the implementation of the algorithm in the highest level. Auto-parallelization is taken care by the language.

• Provide a nice platform for numerical method researchers to share & test methods from others easily and efficiently. The platform itself should also be as extensible as possible to support more situations.

Currently, OpFlow is in its very early stage of development. It currently supports finite difference method (FDM) on Cartesian grid, as well as an adaptive version based on the structured adaptive mesh refinement (SAMR) method. It uses CPU as the backend target and uses OpenMP & MPI for shared memory and distributed memory parallelization. We plan to introduce finer controlled thread models, as long as explicit vectorization for better performance. Flux and inner product semantics for finite volume method (FVM) and finite element method (FEM) are postponed to the later future.

The source code is available under the MPL v2.0 license at https://github.com/OpFlow-dev/OpFlow.