Memory partition for SIMD in streaming dataflow architectures
Xiaowei Shen, Xiaochun Ye, Xu Tan, Da Wang, Zhimin Zhang, Zhimin Tang, Dongrui Fan · 2016
The high parallelism feature of scientific applications makes SIMD very suitable for streaming dataflow architectures. However, the splitting of SIMD memory requests increases the messages in on-chip networks and decreases the efficiency of streaming dataflow architectures. To process SIMD memory requests without splitting, a memory partition mechanism is proposed for SIMD in streaming dataflow architectures. The mechanism partitions input data of scientific applications into N (SIMD width) independent sub-spaces equally and groups operations on different data with same locations in N sub-spaces into SIMD operation. Moreover, the mechanism merges data with the same location in different sub-spaces to SIMD data and stores SIMD data in same SPMs to process SIMD memory requests as a whole. Simulation experimental results show that the proposed mechanism improves the performance per watt of streaming dataflow architectures by 2.5× on average on typical scientific applications.