Optimizing convolution operations on GPUs using adaptive tiling

The research domain of Multimedia Content Analysis (MMCA) considers all aspects of the automated extraction of knowledge from multimedia data. High-performance computing techniques are necessary to satisfy the ever increasing computational demands of MMCA applications. The introduction of Graphics Processing Units (GPUs) in modern cluster systems presents application developers with a challenge. While GPUs are well known to be capable of providing significant performance improvements, the programming complexity vastly increases. To this end, we have extended a user transparent parallel programming model for MMCA, named Parallel-Horus, to allow the execution of compute intensive operations on the GPUs present in the cluster. The most important class of operations in the MMCA domain are convolutions, which are typically responsible for a large fraction of the execution time. Existing optimization approaches for CUDA kernels in general as well as those specific to convolution operations are too limited in both performance and flexibility. In this paper, we present a new optimization approach, called adaptive tiling, to implement a highly efficient, yet flexible, library-based convolution operation for modern GPUs. To the best of our knowledge, our implementation is the most optimized and best performing implementation of 2D convolution in the spatial domain available to date.