Corpus ID: 235435936

ShortcutFusion: From Tensorflow to FPGA-based accelerator with reuse-aware memory allocation for shortcut data

@article{Nguyen2021ShortcutFusionFT,
  title={ShortcutFusion: From Tensorflow to FPGA-based accelerator with reuse-aware memory allocation for shortcut data},
  author={Duy-Thanh Nguyen and Hyeonseung Je and Tuan Nghia Nguyen and Soojung Ryu and Kyujung Lee and Hyuk-Jae Lee},
  journal={ArXiv},
  year={2021},
  volume={abs/2106.08167}
}
Residual block is a very common component in recent state-of-the art CNNs such as EfficientNet or EfficientDet. Shortcut data accounts for nearly 40% of feature-maps access in ResNet152 [8]. Most of the previous DNN compilers, accelerators ignore the shortcut data optimization. This paper presents ShortcutFusion, an optimization tool for FPGA-based accelerator with a reuse-aware static memory allocation for shortcut data, to maximize on-chip data reuse given resource constraints. From… Expand