In the adaptive optics(AO) system, imaging is a very important process. The imaging process of the detectors and Hartman-Shack wavefront sensor in the AO simulation system is calculated with two-dimensional discrete convolution, whose numerical algorithm can be implemented with fast Fourier transformation (FFT). With the increase of the matrix size, the computation of convolution increases dramatically, and becomes a bottleneck of the AO system. To solve the problem, convolution on graphic processing unit(GPU) was implemented, which is a powerful accelerator in modern high performance computers. By applying various optimizations, significant performance improvement is obtained. On the NVIDIA Tesla C2050 GPU, a speedup of 5 to 24 times compared to serial program on the CPU with different image sizes is achieved.