Hardware/Software Co-Design, Friedrich-Alexander University Erlangen-Nü rnberg, Germany.

---

Abstract—In this work, we show how to systematically map Convolutional Neural Networks (CNNs) onto Tightly Coupled Processor Arrays (TCPAs), a class of massively parallel accelerators for many computationally intensive tasks (e.g., from the digital signal and image processing domain). Contrary to previous approaches and implementations, we propose techniques for the layer-parallel execution of CNNs on processor arrays including the maximally overlapped processing of consecutive layers. This is achieved through layer fusion and loop unrolling to exploit the full pipelining potential of such massively parallel architectures for given CNNs. These transformations are also necessary to decrease the number of on-chip/off-chip data transfers. For CNNs, we present a calculus for achievable performance and memory requirements on TCPAs. Based on this calculus, it is shown how either throughput-maximal mappings can be determined for a given architecture. Alternatively, resource-minimized mappings to süstain a given throughput, e.g., number of frames per second, are systematically derived. The approach is evaluated for a CNN model for the MNIST benchmark on a processor array of size 4x4 including a comparison of the performance of the layer-parallel approach over layer-by-layer processing.

Index Terms—CNN accelerator, TCPA.

[PDF]

Cite: Christian Heidorn, Michael Witterauf, Frank Hannig, Jürgen Teich, "Efficient Mapping of CNNs onto Tightly Coupled Processor Arrays," Journal of Computers vol. 14, no. 8, pp. 541-556, 2019.