RepoOMP uses AI to speed up C++ code 8x via OpenMP
Auto-parallelizer RepoOMP cuts compile time by 47% while hitting 8.96× speedups...
Researchers from Tsinghua University and the Chinese Academy of Sciences today open-sourced RepoOMP, a repository-aware framework that automates OpenMP parallelization of performance-critical loops.
RepoOMP combines deterministic rule checks with an LLM agent to solve the dependency problem that stumps previous tools. It first builds a Multi-granularity Attributes Performance (MAP) graph to identify “hotspots,” then routes each hotspot either to a rule engine or an LLM agent based on compilation safety. A Structured Transformation Context (STC) keeps the agent focused on only the dependency facts it needs, cutting irrelevant repository text by up to 68%.
On standard benchmarks, RepoOMP hit average speedups of 8.23× on NPB and 8.96× on BOTS. Across nine real-world kernels, it delivered a cross-backbone mean speedup of 5.25× and improved the best baseline by 18–28%. For the 330 accepted real-world hotspots, median speedup was 2.25×. The code and benchmarks are open-source at the provided arXiv link.
- RepoOMP analyzes 951 hotspots across NPB, BOTS, FFmpeg, NCNN and GROMACS repositories.
- Average speedups of 8.23× on NPB and 8.96× on BOTS; median 2.25× for 330 real-world hotspots.
- Cuts agent-side token cost by 47–68% by feeding only dependency facts (STC) to the LLM.
Why It Matters
Developers can now auto-parallelize mature C/C++ codebases with near-linear speedups while burning fewer tokens and compile cycles.