Convolutional Neural Networks (CNNs) have found widespread applications in artificial intelligence fields such as computer vision and edge computing. However, as input data dimensionality and convolutional model depth continue to increase, deploying CNNs on edge and embedded devices faces significant challenges, including high computational demands, excessive hardware resource consumption, and prolonged computation times. In contrast, the DecomposableWinograd Method (DWM), which decomposes large-size or largestride kernels into smaller kernels, provides a more efficient solution for inference acceleration in resource-constrained environments. This work proposes an approach employing the layerto- layer unified input transformation based on the DecomposableWinograd Method. This reduces computational complexity in the feature transformation unit through system-level parallel pipelining and operation reuse. Additionally, we introduce a reconfigurable, columnindexed Winograd computation unit design to minimize hardware resource consumption. We also design flexible data access patterns to support efficient computation. Finally, we propose a preprocessing shift network system that enables low-latency data access and dynamic selection of theWinograd computation unit. Experimental evaluations on VGG-16 and ResNet-18 networks demonstrate that our accelerator, deployed on the Xilinx XC7Z045 platform, achieves an average throughput of 683.26 GOPS. Compared to existing approaches, the design improves DSP efficiency (GOPS/DSPs) by 5.8×.
Loading....