Nowadays most of the applications of the stereo vision are related to the mobile devices, especially in the field of robotics. It is crucial to develop efficient and robust algorithms that would allow real time operation in a wide range of environments. This paper presents an efficient adaptive algorithm of stereo matching that was applied and optimized for the mobile Graphics Processing Unit. It is a well known problem that most of the stereo vision algorithms are based on the dense stereo matching methods that in most of the cases are the main factor for a demanding computation cost. The presented method introduces a novel approach to the stereo matching problem by adaptively combining the cost function that can be computed efficiently for small matching window and sparse accumulated windows similar to those applied in convolutional neural networks. Adaptability of this method is based on detecting edges in the processed images. Such a solution allows obtaining precise subpixel results on highly textured regions of the image as well as to obtain stable results on weakly matchable texture-less regions while sustaining high efficiency and not causing problems related to the memory bandwidth bottleneck.