Automatic Data Layout Transformation for Heterogeneous Many-Core Systems

Applying appropriate data structures is critical to attain superior performance in heterogeneous many-core systems. A heterogeneous many-core system is comprised of a host for control flow management, and a device for massive parallel data processing. However, the host and device require different types of data structures. The host prefers Array-of-Structures (AoS) to ease the programming, while the device requires Structure-of-Arrays (SoA) for efficient data accesses. The conflicted preferences cost excessive effort for programmers to transform the data structures between two parts. The separately designed kernels with different coding styles also cause difficulty in maintaining programs. This paper addresses this issue by proposing a fully automated data layout transformation framework. Programmers can maintain the code in AoS style on the host, while the data layout is converted into SoA when being transferred to the device. The proposed framework streamlines the design flow and demonstrates up to 177% performance improvement.