ConvFusion: A Model for Layer Fusion in Convolutional Neural Networks

The superior accuracy and appealing universality of convolutional neural networks (CNNs) as a generic algorithm for many classification tasks have made the design of energy efficient CNN accelerators an important topic in both academia and industry. Of particular interest in the design and use of CNN accelerators is the scheduling of the computational workload, which can have a major impact on the quality of the final design. The many inherently independent operations in CNNs result in a vast scheduling space however, rendering the selection of the optimal schedule(s) non-trivial. To aid in this complex task, this work introduces a generic mathematical cost model of the external memory accesses, internal memory footprint, and compute load for CNN execution schedules. The model enables fast exploration of the scheduling space, including loop tiling, loop reordering, explicit data transfer scheduling, recomputation, and, crucially, layer fusion, which recently has attracted interest as a method to reduce external memory accesses. An accompanying open source tool is released to perform schedule space exploration for CNNs using the introduced cost model. Leveraging the code generation capabilities of this tool the proposed model is validated on six real world networks, demonstrating that layer fusion can reduce the external memory accesses by more than two orders of magnitude compared to the best non-fused schedules. Confusing at first glance however, a high-level energy analysis shows that the practical benefits of layer fusion may be overestimated if other parts of the system are not tuned accordingly.