Abstract: OOpenCL is a general-purpose programming framework forheterogeneous computing platforms, however, due to the differences in hardware architecture,how to achieve performance portability on different platforms based on the function portability is still to be studied. Currently most of the researches on algorithm optimization are aimed at a sin- gle hardware platform, and difficult to achieve the efficient running on different platforms. This paper analysed the differences between the underlying hardware architectures of GPU, and studied the effects of different GPU platforms using different optimization methods on performance from the access efficiency of global memory, full use of the(}PU compute resource, the constraints with hardware resource and other aspects. Based on this, the Laplace image enhance- ment algorithm based on OpenCL was implemented. Experimental results show that optimized algorithm gets 3. 7一136. 1 times and 56. 7 times on average speedup(without calculate the data transfer time) on both AMD and NVIDIA GPU, and the performance of the optimized kernel increases 12. 3%一346. 7 0 o and 143. 1 0 o on average than the CUDA ver- sion in NVIDIA NPP library,which verifies the effectiveness and cross-platform ability of optimization methods.

Key words: OpenCL, General-purpose computing, Laplace, Across platform