UniPar: A Unified LLM-Based Framework for Parallel and Accelerated Code Translation in HPC
Introduces UniPar, an evaluation framework for assessing how large language models translate between parallel programming languages, achieving 69% compilation success and 33% functional correctness through fine-tuning and compiler-guided repair.
Abstract
This research introduces UniPar, an evaluation framework for assessing how large language models translate between parallel programming languages. The team targeted translations between serial code, CUDA, and OpenMP, testing models like GPT-4o-mini and LLaMA-3.3-70B-Instruct. Their approach combined fine-tuning, hyperparameter tuning, and compiler-guided repair, improving performance from 46% to 69% compilation success and 15% to 33% functional correctness. They also created the PARATRANS dataset covering serial-to-parallel and cross-paradigm code transformations, with both code and data made publicly available.
Related Research
Workflows vs Agents for Code Translation
Compares structured workflows versus agentic approaches for MATLAB-to-HDL translation, showing that agentic methods with the Model Context Protocol increase simulation reach rates by over 20 percentage points on mid-sized models.
Counting Without Running: Evaluating LLMs' Reasoning About Code Complexity
Introduces gpuFLOPBench, a benchmark containing 577 CUDA kernels to evaluate whether language models can predict floating-point operation counts without execution, revealing limitations in understanding hardware-specific performance details.