HPCTransCompile
收藏资源简介:
HPCTransCompile是一个由人工智能编译器生成的数据集,旨在支持高性能CUDA代码的转译和学习语言模型(LLM)的初步探索。该数据集利用了人工智能编译器和自动优化技术,生成高性能CUDA和相应平台代码对。数据集的构建过程包括操作符选择、计算图构建、无用依赖移除、图标签和数据集生成。HPCTransCompile旨在解决高性能CUDA代码转译的挑战,并评估LLMs在CUDA转译任务上的性能。
HPCTransCompile is a dataset generated by AI compilers, aimed at supporting the transpilation of high-performance CUDA code and preliminary explorations of large language models (LLMs). This dataset utilizes AI compilers and automated optimization techniques to generate pairs of high-performance CUDA code and corresponding platform-specific code. The dataset construction process includes operator selection, computational graph construction, redundant dependency removal, graph labeling, and dataset generation. HPCTransCompile aims to address the challenges of high-performance CUDA code transpilation and evaluate the performance of LLMs on CUDA transpilation tasks.




