ParBench: Benchmark for LLM Parallel Code Translation

ParBench, a novel benchmark, has been launched to assess large language models (LLMs) in the realm of parallel code translation tasks. As modern software demands shift across various accelerators, programming APIs, compiler stacks, and portability layers like CUDA, OpenMP, and OpenCL, the need for reliable migration solutions arises. Although LLMs and autonomous coding agents are being explored for these migrations, the industry lacks effective metrics to ensure they maintain essential low-level parallel semantics, such as thread indexing, synchronization, memory management, and host-device coordination. ParBench focuses on kernel-centric evaluation of LLM-driven parallel API translation under reproducible conditions, utilizing declarative benchmark specifications to standardize the build, run, and verification processes while concentrating solely on computational kernels. It incorporates several open-source HPC suites.