Supersonic: Learning to Generate Source Code Optimizations in C/C++

TL;DR

Supersonic is a neural network-based tool that automates source code optimization in C/C++, outperforming large language models while being significantly smaller and more efficient.

Contribution

The paper introduces Supersonic, a seq2seq model that automates source code optimization, demonstrating superior performance and efficiency over large language models.

Findings

Supersonic outperforms GPT-3.5-Turbo and GPT-4 on code optimization tasks.

Supersonic is more than 600x smaller than GPT-3.5-Turbo and 3700x smaller than GPT-4.

Supersonic effectively minimizes code changes during optimization.

Abstract

Software optimization refines programs for resource efficiency while preserving functionality. Traditionally, it is a process done by developers and compilers. This paper introduces a third option, automated optimization at the source code level. We present Supersonic, a neural approach targeting minor source code modifications for optimization. Using a seq2seq model, Supersonic is trained on C/C++ program pairs ($x_{t}$, $x_{t+1}$), where $x_{t+1}$ is an optimized version of $x_{t}$, and outputs a diff. Supersonic's performance is benchmarked against OpenAI's GPT-3.5-Turbo and GPT-4 on competitive programming tasks. The experiments show that Supersonic not only outperforms both models on the code optimization task but also minimizes the extent of the change with a model more than 600x smaller than GPT-3.5-Turbo and 3700x smaller than GPT-4.