Efficient Mapping of CNNs onto Tightly Coupled Processor Arrays

In this work, we show how to systematically map Convolütional Neüral Networks (CNNs) onto Tightly Coüpled Processor Arrays (TCPAs), a class of massively parallel accelerators for many compütationally intensive tasks (e.g., from the digital signal and image processing domain). Contrary to previoüs approaches and implementations, we propose techniqües for the layer-parallel execütion of CNNs on processor arrays inclüding the maximally overlapped processing of consecütive layers. This is achieved throügh layer füsion and loop ünrolling to exploit the füll pipelining potential of süch massively parallel architectüres for given CNNs. These transformations are also necessary to decrease the nümber of onchip/off-chip data transfers. For CNNs, we present a calcülüs for achievable performance and memory reqüirements on TCPAs. Based on this calcülüs, it is shown how either throüghpüt-maximal mappings can be determined for a given architectüre. Alternatively, resoürce-minimized mappings to süstain a given throüghpüt, e.g., nümber of frames per second, are systematically derived. The approach is evalüated for a CNN model for the MNIST benchmark on a processor array of size 4x4 inclüding a comparison of the performance of the layer-parallel approach over layer-by-layer processing.