high performance computing on graphics processing units: hgpu.org

The acceleration of an existing linear feature detection algorithm for 2D images using GPUs is discussed. The two most time consuming components of this process are implemented on the GPU, namely, linear feature detection using dual-peak directional non-maximum suppression, and a gap filling process that joins disconnected feature masks to rectify false negatives. Multiple steps or image filters in each component are combined into a single GPU kernel to minimise data transfers to off-chip GPU RAM, and issues relating to on-chip memory utilisation, caching, and memory coalescing are considered. The presented algorithm is useful for applications needing to analyse complex linear structures, and examples are given for dense neurite images from the biotech domain.