Formula Translation in Blitz++, NumPy and Modern Fortran: A Case Study of the Language Choice Tradeoffs

Abstract

Three object-oriented implementations of a prototype solver of the advection equation are introduced. The presented programs are based on Blitz++ (C++), NumPy (Python) and Fortran's built-in array containers. The solvers constitute implementations of the Multidimensional Positive-Definite Advective Transport Algorithm (MPDATA). The introduced codes serve as examples for how the application of object-oriented programming (OOP) techniques and new language constructs from C++11 and Fortran 2008 allow to reproduce the mathematical notation used in the literature within the program code. A discussion on the tradeoffs of the programming language choice is presented. The main angles of comparison are code brevity and syntax clarity (and hence maintainability and auditability) as well as performance. All performance tests are carried out using free and open-source compilers. In the case of Python, a significant performance gain is observed when switching from the standard interpreter (CPython) to the PyPy implementation of Python. Entire source code of all three implementations is embedded in the text and is licensed under the terms of the GNU GPL license.

Python bindings for libcloudph++

“…The Python bindings significantly facilitate the use of the library and add relatively little runtime overhead (particularly in the case of the resource-intensive particle-based scheme). Python has simpler syntax than C++, its philosophy emphasises succinct code (see [3] for a geoscience-relevant case study comparing Python, C++ and Fortran). Moreover, Python is widespread across the atmospheric science community [7].…”

libmpdata++ 0.1: a library of parallel MPDATA solvers for systems of generalised transport equations

“…The adopted design contributes to the readability, maintainability and conciseness of the code. The current development of libmpdata++ is an extension of the research on OOP implementation of the basic MPDATA scheme presented in Arabas et al (2014).…”

“…The adopted design contributes to the readability, maintainability and conciseness of the code. The current development of libmpdata++ is an extension of the research on OOP implementation of the basic MPDATA scheme presented in Arabas et al (2014). The goal of this article is twofold: first, to document the library interface by providing usage examples; and second, to validate the correctness of the implementation by verifying the results against published benchmarks.…”

PyMPDATA v1: Numba-accelerated implementation of MPDATA with examples in Python, Julia and Matlab

“…PyMPDATA is aimed to address several aspects which steepen the learning curve and limit collaborative usage and development of existing C++ (e.g., Jaruga et al, 2015) and Fortran (e.g., Kühnlein et al, 2019) implementations of MPDATA. Performance on par with compiled-language implementations is targeted by employing just-in-time (JIT) compilation using Numba (Lam et al, 2015), which translates Python code into fast machine code using the Low Level Virtual Machine (LLVM, https://llvm.org/) compiler infrastructure (for a discussion of another JIT implementation of MPDATA using PyPy, see Arabas et al, 2014). PyMPDATA is engineered aiming at both performance and usability, the latter encompassing research user's, developer's and maintainer's perspectives.…”

“…and based on Figure 5 from Arabas et al (2014). As of the current version, PyMPDATA supports homogeneous transport in one (1D), two (2D), and three dimensions (3D) using structured meshes, optionally generalised by coordinate transformation (Smolarkiewicz & Clark, 1986;Smolarkiewicz & Margolin, 1993).…”