Investigating performance variations of an optimized GPU-ported granulometry algorithm

In this article, we present an optimized GPU implementation of a granulometry algorithm which is used a lot in the study of material domain. The main contribution to this algorithm is the binarization of the input data which increases throughput while reducing data allocated memory space. Also, the optimized GPU implementation brings an order of magnitude speedup compared to a CPU multi-threaded implementation. Furthermore, we investigate the reasons why GPU performance drop for different input data dimensions. Three main factors are exposed: under-exploited threads, threadblocks and streaming multiprocessors. This study should help the reader understand the tight relation that exists between the CUDA programming paradigm and the gpu architecture as well as some main bottlenecks.