The Celerity High-level API: C++20 for Accelerator Clusters

Abstract Providing convenient APIs and notations for data parallelism which remain accessible programmers while still providing good performance has been a long-term goal of researchers as well language library designers. C++20 introduces ranges views, the composition operations on them using concise syntax, but efficient implementation these features is restricted to CPUs. We present Celerity High-level API, makes similarly mechanisms applicable GPUs accelerators, even distributed memory clusters GPUs. Crucially, we achieve this very high level abstraction without significant negative impact compared lower-level implementation, introducing any non-standard toolchain components or compilers, by implementing C++ infrastructure top system. This made possible two central API design strategies, form core our contribution. Firstly, gathering much information at compile-time metaprogramming techniques automatically fuse several distinctly formulated processing steps into single accelerator kernel invocation. And secondly, leveraging “Concepts” in order avoid type erasure, allowing highly code generation. have evaluated approach quantitatively comparison manual implementations benchmarks, demonstrating its low overhead. Additionally, investigated individual specific optimizations choices, illustrating advantages afforded Concepts-based approach.