Comparative Performance of GPU, SIMD and OpenMP Systems for Raw Template Matching in Computer Vision

Template matching is a traditional technique of Computer Vision whose advantages and disadvantages are known. However, advances in computer hardware allow computing it effectively with the use of SIMD instruction set, GPUs or multi-core systems. The computation of that low-level primitive in sub millisecond scale would improve high theoretical methods if they are used with high efficient primitives. This paper presents the comparative results of basic template matching by using SIMD instructions, multi-core systems and multi-GPU implementations. The results of this study will show that the high-specialized instruction in modern releases of SIMD and the use of multi-core systems outperforms the implementations based on GPUs for small mask size due to memory transfer cost. However, for big mask size GPU and SIMD systems have similar performance