tandem-fortran

Fortran bindings for Tandem8x32. The module tandem_rng wraps the reference C implementation and produces the stream of the specification bit for bit. The GPU modules write the same stream.

• API: the CPU module, device fills, kernels, OpenMP target and do concurrent fills, and parallel use.
• Design: the C mirror, bounded integers and normals.
• Tests: what the suite checks, where the fixtures come from, and what CI runs.
• Performance: CPU, A100 and offload figures.
• nvfortran defects: the nvfortran 25.3 defects the code works around.

Install

fpm build --profile release
fpm test

or add tandem_rng = { git = "https://github.com/tandem-rng/tandem-fortran" } to the dependencies in your fpm.toml. pixi run test supplies gfortran and fpm from conda-forge.

Fortran 2018, fpm or any build that compiles src/tandem_rng.f90 and src/c/tandem.c. Tested with LLVM flang 21 (with clang for the C), gfortran and ifx.

The normal fills use explicit fused multiply-adds, and tandem.c picks the hardware fma at run time on x86. Build it with --c-flag -ffp-contract=off in fpm, which keeps every other expression unfused, so all compilers give the same normals. pixi run test, the CI jobs and the Makefiles set it. With icx, add -fp-model=precise: its default fast model changes the Float32 exponentials.

The GPU module needs CUDA, which fpm cannot compile, so cuda/Makefile builds the whole stack with gfortran and nvcc. On a Linux host without a system CUDA install:

pixi run -e cuda test-cuda

The kernel module needs nvfortran from the NVIDIA HPC SDK, which installs without root from NVIDIA's tarball. With nvfortran and nvcc on the PATH:

make -f cuda/Makefile test-device

fpm install --profile release --prefix <prefix> installs the library and the module files. The packaging/ directory holds a Spack recipe (spack/package.py) and a conda-forge style recipe (conda/recipe.yaml) for the CPU module. Neither is submitted to Spack or conda-forge yet, and both build from the main branch.

Vendored sources

src/c/tandem.c and src/c/tandem.h come from tandem-c, src/c/tandem.cuh and src/c/tandem/core.hpp from tandem-cuda, unchanged. tools/sync_c.sh copies them and the dumps from the tandem-c and tandem-cuda commits pinned in it, and CI fails when any of them differ. test/conformance holds copies of the specification's conformance fixtures at tandem-spec 2a4bd08, and CI fails when they differ. tandem.c calls libm, so link with -lm; fpm does this through fpm.toml.

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.