HPCTransCompile

Introduction

This repository contains the official implementation of HPCTransCompile: An AI Compiler Generated Dataset for High-Performance CUDA Transpilation and LLM Preliminary Exploration.

Framework Implementation

We propose a novel framework for generating high-performance CUDA and corresponding platform code pairs, leveraging AI compiler and automatic optimization technology. We further enhance the framework with a graph-based data augmentation method and introduce HPCTransEval, a benchmark for evaluating LLM performance on CUDA transpilation. We conduct experiments using CUDA-to-CPU transpilation as a case study on leading LLMs. The result demonstrates that our framework significantly improves CUDA transpilation, highlighting the potential of LLMs to address compatibility challenges within the CUDA ecosystem.

Setup

We provide two benchmarks, HPCTransEval and KernelBench_c. You can find them in the corresponding folders. In order to do the assessment correctly, you need to download a modified library tvm from https://github.com/hehesnail/tvm/tree/modify_style. Then, run:

pip install -r requirements.txt

1. To evaluate on HPCTransEval, run:

   bash EvalEngine/eval_HPCTransEval.sh
   

2. To evaluate on KernelBench_c, run:

   bash EvalEngine/eval_KernelBench_c.sh