CNNFlow: Memory-driven Data Flow Optimization for Convolutional Neural Networks

Convolution Neural Networks (CNNs) are widely deployed in computer vision applications. The datasets large, and the data reuse across different parts is heavily interleaved. Given that memory access (SRAM especially DRAM) more expensive both performance energy than computation, maximizing to reduce movement hierarchy critical improving execution efficiency. This even important for common use case of CNNs on mobile devices where computing/memory resources limited. We propose CNNFlow, a memory-driven dataflow optimization framework automatically schedule CNN computation given architecture maximize at each level hierarchy. provide mathematical calculation reuses terms parameters including loop ordering, blocking, memory-bank allocation tensors CNN. then present series techniques help prune large search space cost exploration. provides, first time, an exact practical algorithm optimal solutions minimize efficacy demonstrated two used algorithms: AlexNet VGG16 with 5 13 convolution layers, respectively. CNNFlow finds solution layer within tens minutes compute time. Its requires about 20% fewer DRAM accesses 40%–80% SRAM compared state-of-the-art algorithms literature.