Array Partitioning Method for Streaming Dataflow Optimization in High-level Synthesis

High-level synthesis (HLS) is a popular method that allows designers to describe the behavior-level functionality and automatically generates efficient register-transfer level (RTL) descriptions. In HLS, dataflow is the key micro-architecture to achieve high parallelism. However, strict conditions such as sequential access on the potential channels often limit the streaming dataflow. To settle this issue, this paper proposes an efficient array partitioning method for the streaming dataflow inference. The key is to explore the potential array partitioning mode that matches the sequential access requirements by streaming channels. An experimental case study is presented on the inference of the convolutional neural networks (CNN). It indicates that the proposed method can achieve about 28.6% performance improvements compared with the default dataflow, with the cost of 7.2% power increasement.