Skip to content

Why Python and C?

This project is a Python wrapper of C implementations.

Python

It's the original author's favourite language, so it seemed a reasonable choice (as good as any other).

C

The original ACM TOMS paper provides mathematical recipes for producing approximate random numbers, and the speed improvements derive from exploiting SIMD hardware. To go down to such a low-level brings us to the realms of the usual compiled HPC languages such as C, C++, Fortran, Rust, etc. The original author doesn't know Fortran or Rust very well, so that leaves us with C or C++.

Choosing between C and C++ was a hard decision, and either would be a good platform to implement the underlying routines. So why C and not C++:

  • The approximations use some type punning (between int and float), and this is easily done in C using unions (and is defined behaviour in C, including C99 onwards, cf. https://stackoverflow.com/a/25664954/5134817 and https://stackoverflow.com/a/25672839/5134817), whereas in C++ type punning is much trickier to do correctly (without producing undefined behaviour).
  • C has the keyword restrict which does not exist in C++ without compiler extensions (which we would prefer to avoid for portability), and this is something we take advantage of for optimisation.
  • C is the lingua franca for HPC and portability.
  • Python and NumPy have C APIs, and these are the initial targets for our wrappers.

Why not C++, Fortran, Rust, Julia, MATLAB, JavaScript, etc.

Creating interfaces for the various languages should be easy enough, and if the package proves popular we welcome contributions to help us widen the scope.

For the higher level languages we hope to wrap our C code similar to how we have for Python. For the lower level languages, either use wrappers, re-implement in the preferred language, or use langauge features, such as e.g. C++'s extern "C".