HW-Flow-Fusion: Inter-Layer Scheduling for Convolutional Neural Network Accelerators with Dataflow Architectures

Energy and throughput efficient acceleration of convolutional neural networks (CNN) on devices with a strict power budget is achieved by leveraging different scheduling techniques to minimize data movement maximize reuse. Several dataflow mapping frameworks have been developed explore the optimal CNN layers reconfigurable accelerators. However, previous works usually optimize each layer singularly, without reuse between CNNs. In this work, we present an analytical model achieve searching for communication computation across layers. We call inter-layer framework HW-Flow-Fusion, as fused map-space multiple sharing available resources same accelerator, investigating constraints trade-offs execution workloads dependencies. propose memory-efficient model, tiling, resource partitioning strategies fuse recomputation. Compared standard single-layer scheduling, can reduce volume 51% 53% selected VGG16-E ResNet18 spatial array latency 39% 34% respectively, while also increasing ratio which improves memory bandwidth efficiency.