Evaluation of OpenCL native math functions for image processing algorithms

Image enhancement plays an important role in different research fields such as medical image analysis. Since the same computations are usually performed on many image elements, those computations can be easily parallelized. Modern Graphics Processing Units (GPUs) are capable for doing many tasks in parallel. However, improving running times on GPUs usually leads to a loss of floating point precision. In this paper we evaluate the impact of GPU hardware implemented native functions on three GPUs, and one Central Processing Unit (CPU). As an example, the bilateral filter with built-in and native math functions was implemented and used for smoothing noisy brain Magnetic Resonance Images (MRI). For all experiments widely used error metrics were calculated. Experiments shows that native versions improve running times significantly (up to 155 times). As expected precision is lower for the measures which include a lot additions without normalization.