tandem-c

Reference C and C++ implementation of Tandem8x32. It produces the stream of the specification bit for bit, with SIMD fills on CPUs and OpenMP target fills on GPUs.

• API: the C and C++ interface, the counter-based header, OpenMP target offload and parallel use.
• Design: how the fills, bounded integers, normals and exponentials work.
• Tests: what the suite checks, where the fixtures come from, and what CI runs.
• Performance: CPU and offload figures, and the standard generators.

Install

make                              # libtandem.a, built with clang
make install PREFIX=<prefix>      # libtandem.a, tandem.h, tandem.hpp, tandem.pc

Or compile tandem.c into your project, with tandem_normal_tables.h next to it. DESTDIR stages the installed files.

The library needs C11 and C++17 and libm, and nothing newer, so projects that vendor it can keep their own flags, with one exception: the floating-point model must stay strict. The Float32 exponential carries its logarithm as a high and a low part, and a compiler that reassociates or contracts those sums changes the values. Build tandem.c with -ffp-contract=off and without fast-math. The Intel compiler icx defaults to -fp-model=fast, so pass -fp-model=precise -ffp-contract=off there. make builds it as C23 and C++23 with clang, the primary compiler. Set CC and CXX for another compiler.

The packaging/ directory holds a Spack recipe (spack/package.py) and a conda-forge style recipe (conda/recipe.yaml). Neither is submitted to Spack or conda-forge yet, and both build from the main branch.

AI assistance

This implementation 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.