The research on parallelized fast trilateral filter on GPU acceleration

This paper design and implement a parallel fast trilateral filter algorithm on GPU. The trilateral filter can be decomposed into two bilateral filters which are implemented by the parallelized bilateral filter on CUDA separately. The performance optimization strategies are intensively discussed. The occupancy of CUDA kernel for trilateral filter reaches 0.833 which is measured by Visual Profiler. The results indicate that more than two orders of magnitude speedup have been achieved in the GPU as compared to the CPU counterpart for trilateral filter (512×512 lena.bmp), which demonstrates the significance of parallelism.