Abstract: As a general-purpose programming standard for multiple platforms, OpenCL has accelerated many applications. Due to the differences of different platforms in hardware and software environments, general optimization methods may not accelerate the application well for all. Taking the optimization of the mean shift algorithm on GPU and APU platforms as an example, the paper provided several insights on contributions of various optimization methods on different platforms. On one hand, we explored the architectures of different platforms. On the other hand, we compared the pros and cons of different optimization methods. Based on meticulous evaluations of the pros and cons, we looked for the optimal solution. Experimental results show that, on AMD 5850, Tesla C2050 and APU A6-3650 platforms, the optimized algorithm achieves 9.68 X，5.74 X and 1.27 X speedups, respectively, and 79.73 X，93.88 X and 2.22 X speedups comparcel to the serial version, respectively, and 1.27 X and 1.24 X speedups compared to the CUDA version OpenCV program for the first two platforms,respectively.

Key words: GPU, APU, OpenCL, Mean shift