Reproduce “ParaCodex: A Profiling-Guided Autonomous Coding Agent for Reliable Parallel Code Generation and Translation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Erel Kaplan, Tomer Bitan, Lian Ghrayeb, Le Chen, Tom Yotam, Niranjan Hasabnis, Gal Oren
PARACODEX introduces a profiling-guided autonomous coding agent that generates and translates parallel code for OpenMP GPU offload, achieving both functional correctness and high performance. The agent employs an iterative, multi-stage workflow integrating large language models with development tools to outperform zero-shot LLM baselines and often expert-written reference implementations.
Key insights
- A structured, artifact-driven agentic workflow with explicit planning and iterative refinement is crucial for reliable and high-performance parallel code generation, overcoming limitations of single-shot LLM outputs.
- Integrating hotspot analysis and explicit data strategy planning before code modification is essential to prevent common performance pitfalls like inefficient data transfers in OpenMP GPU offload.
- Profile-guided optimization, with a mechanism for reverting performance regressions, effectively closes the loop on performance tuning, enabling the agent to match or exceed expert-tuned reference implementations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.