tandem-kokkos

Kokkos implementation of Tandem8x32. It produces the stream of the specification bit for bit on every Kokkos backend. Two headers, C++20, no compiled library.

• API: tandem::fill and the other fills, tandem::Rng draws and parallel use.
• Design: the kernels and the portable core.
• Tests: what the suite checks, where the fixtures come from, and what CI runs.
• Performance: M4 Pro and A100 figures against Kokkos::fill_random.

Install

Header-only. Clone with the submodule:

git clone --recurse-submodules https://github.com/tandem-rng/tandem-kokkos
git submodule update --init           # in a clone made without it

With CMake, as a subdirectory or through FetchContent next to an existing Kokkos:

add_subdirectory(tandem-kokkos)        # finds Kokkos unless the Kokkos::kokkos target exists
target_link_libraries(app PRIVATE tandem::kokkos)

or installed:

cmake -S . -B build -DKokkos_ROOT=<kokkos prefix> && cmake --install build --prefix <prefix>

find_package(tandem-kokkos REQUIRED)
target_link_libraries(app PRIVATE tandem::kokkos)

Without CMake, add include/ and external/tandem-cuda/include/ to the include path of a Kokkos build. The headers need C++20, as Kokkos 5 does. external/tandem-cuda is a submodule pinned at 76eddae.

pixi.toml provides Kokkos 5.2.1, the newest conda-forge build, with the Serial and OpenMP backends (pixi run test), built with clang. The latest release is 5.2.2. pixi run -e gcc test-gcc builds the same with GCC 14 on Linux. The conda-forge package has no CUDA backend, so on a GPU host the cuda environment builds Kokkos 5.2.2 with Serial, OpenMP and CUDA from source into build/, with CUDA 12.8, clang 19 as the nvcc host compiler (-ccbin) and GCC 14 for its libstdc++. CUDA 13 needs a newer driver than the 570 on the GPU host, so the toolkit stays at 12.8:

pixi run -e cuda kokkos               # KOKKOS_ARCH=AMPERE80 by default
pixi run -e cuda test-cuda

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

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.