Answers are grounded in this paper and cite their sources.
In this work, we show how to systematically map Convoltional Neral Networks (CNNs) onto Tightly Copled Processor Arrays (TCPAs), a class of massively parallel accelerators for many comptationally intensive tasks (e.g., from the digital signal and image processing domain). Contrary to previos approaches and implementations, we propose techniqes for the layer-parallel exection of CNNs on processor arrays inclding the maximally overlapped processing of consective layers. This is achieved throgh layer fsion and loop nrolling to exploit the fll pipelining potential of sch massively parallel architectres for given CNNs. These transformations are also necessary to decrease the nmber of onchip/off-chip data transfers. For CNNs, we present a calcls for achievable performance and memory reqirements on TCPAs. Based on this calcls, it is shown how either throghpt-maximal mappings can be determined for a given architectre. Alternatively, resorce-minimized mappings to sstain a given throghpt, e.g., nmber of frames per second, are systematically derived. The approach is evalated for a CNN model for the MNIST benchmark on a processor array of size 4x4 inclding a comparison of the performance of the layer-parallel approach over layer-by-layer processing.
No takes yet. Share an insight, caveat, or question.
Christian Heidorn (2019) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: