LAUSR

Convolutions are the core operation of deep learning applications based on Convolutional Neural Networks (CNNs). Current GPU architectures are highly efficient for training and deploying deep CNNs, and are largely used in production. State–of–the–art implementations, however, present low efficiency for some commonly used network configurations. In this paper we propose a GPU-based implementation of the convolution operation for CNN inference that favors coalesced accesses, without requiring prior data transformations. Our experiments demonstrate that it yields notable performance improvements in a range of common CNN forward-propagation convolution configurations, with speedups of up to 2.29 × with respect to the best implementation in cuDNN, covering a relevant region in currently existing approaches. This improvement results in speedups of up to 7.4% for CNN online inference use cases.