Accelerating genetic programming using pycuda

Nvidia's CUDA parallel computation is a good way to reduce computational cost when applying a filter expressed by an equation to an image. In fact, programs need to be compiled to build GPU kernels. Over the past decade, various implementation methods for the image filter using Genetic Programming (GP) have been developed to enhance its performance. By using GP, an appropriate image filter structure can be obtained through learning algorithms based on test data. In this case, each solution must be compiled; therefore, the required computational effort grows significantly. In this paper, we propose a PyCuda-based GP framework to reduce the computational efforts for evaluations. We verify that the proposed method can implement GPU kernels easily based on a sequential GP algorithm, thereby reducing the computational cost significantly.