high performance computing on graphics processing units: hgpu.org

This article presents an overview of the implementation of linear image filters in CPU and GPU. The main goal is to present a self contained discussion of different implementations and their background using tools from digital signal processing. First, using signal processing tools, we discuss different algorithms and estimate their computational cost. Then, we discuss the implementation of these filters in CPU and GPU. It is very common to find in the literature that GPUs can easily reduce computational times in many algorithms (straightforward implementations). In this work we show that GPU implementations not always reduce the computational time but also not all algorithms are suited for GPUs. We beleive this is a review that can help researchers and students working in this area. Although the experimental results are not meant to show which is the best implementation (in terms of running time), the main results can be extrapolated to CPUs and GPUs of different capabilities.