tandem-numpy

NumPy BitGenerator for Tandem8x32. It wraps tandem-c and produces the stream of the specification bit for bit.

• API: Tandem, TandemGenerator and parallel use.
• Design: the bounded integer, normal and exponential contracts.
• Tests: what the suite checks.
• Performance: fill and Generator figures on the Apple M4.

Install

git clone --recurse-submodules https://github.com/tandem-rng/tandem-numpy
pip install ./tandem-numpy

Needs Python 3.11, NumPy 2.0, a C compiler, and meson-python. The C code is the external/tandem-c submodule, pinned at tandem-c 1c75956. For development, pixi install then pixi run test.

The reference C implementation sits in the external/tandem-c git submodule. Clone with git clone --recurse-submodules, or run git submodule update --init in an existing clone. GitHub's ZIP download omits submodules and does not build.

The submodule is pinned at tandem-c 1c75956. The extension is built with -ffp-contract=off and no -mfma: the normal loop uses explicit fused multiply-adds, so its bits do not depend on the compiler, and on x86 the AVX2 and FMA copy is chosen at run time, also in a wheel built for a baseline x86-64.

The build uses meson-python and needs a C compiler. For development, pixi install creates an environment with the package installed editable, which rebuilds the extension on import when the sources change, and pixi run test runs the tests.

Normals are bit exact with tandem-c.

AI assistance

This port was written with the help of large language models under human direction. The design and the specification are human work, as is much of the Julia implementation. The code is tested bit for bit against every vector of the specification and against long stream dumps from the Julia implementation, and every value must match. The output does not depend on who or what wrote the code.