Learning Performance-Improving Code Edits

TL;DR

This paper presents a framework that leverages large language models and a curated dataset to improve high-level program performance optimizations in C++, achieving significant speedups over human efforts.

Contribution

It introduces a novel dataset, evaluation environment, and adaptation strategies for LLMs to perform high-level code optimizations, surpassing human performance in speedup.

Findings

Achieved a mean speedup of 6.86 with eight generations.

Set a new upper limit on speedup at 9.64, surpassing human bests.

Developed environment based on gem5 for reliable performance measurement.

Abstract

With the decline of Moore's law, optimizing program performance has become a major focus of software research. However, high-level optimizations such as API and algorithm changes remain elusive due to the difficulty of understanding the semantics of code. Simultaneously, pretrained large language models (LLMs) have demonstrated strong capabilities at solving a wide range of programming tasks. To that end, we introduce a framework for adapting LLMs to high-level program optimization. First, we curate a dataset of performance-improving edits made by human programmers of over 77,000 competitive C++ programming submission pairs, accompanied by extensive unit tests. A major challenge is the significant variability of measuring performance on commodity hardware, which can lead to spurious "improvements." To isolate and reliably evaluate the impact of program optimizations, we design an environment based on the gem5 full system simulator, the de facto simulator used in academia and industry. Next, we propose a broad range of adaptation strategies for code optimization; for prompting, these include retrieval-based few-shot prompting and chain-of-thought, and for finetuning, these include performance-conditioned generation and synthetic data augmentation based on self-play. A combination of these techniques achieves a mean speedup of 6.86 with eight generations, higher than average optimizations from individual programmers (3.66). Using our model's fastest generations, we set a new upper limit on the fastest speedup possible for our dataset at 9.64 compared to using the fastest human submissions available (9.56).