Abstract: This paper proposes a compilation framework for mapping data-parallel program onto heterogeneous multi-
core systems. Based on the distributed memory model, the framework performs data alignment, data distribution and computation distribution, and finally generates parallel codes for programs. In contrast to other work, the proposed framework focuses on managing the data layout of the program to improve data locality. The experimental results show that the proposed framework can efficiently map programs onto the heterogeneous multi-core processors.

Key words: heterogeneous multi-core processor, automatic code generation, data distribution, computation distribution

摘要： 提出了一个异构多核编译框架，该框架将数据并行程序直接映射到异构多核系统上。基于分布式存储模型，编译器执行数据对齐、数据分布、计算分布操作，并为应用生成并行代码。与其他工作不同，该框架通过管理程序的数据布局来提高数据局部性。实验结果表明，该编译框架能够高效地将应用映射到异构多核处理器上。

关键词: 异构多核处理器, 代码自动生成, 数据分布, 计算分布