Research And Implementation Of Transplant CUDA Program Based On Android

With the rapid evolution of mobile device, the change of traditional personal computers has become the fundamental development trend. While the GPU applications in image processing direction, it also provides a good general processing ability of computing platform. CUD A, which is short for Compute Unified Device Architecture, is a complete set of general purpose computation API for GPU. CUDA framework does not appear to solve the problem of parallel computing on mobile devices, although GPU evolved on mobile devices quickly. However, there are very limited in terms of computational efficiency and energy consumption relying solely on the on-chip GPU to boost computing power.This paper presents mobile CUDA environment to ensure the ability that CUDA program can be running on mobile devices and proposed a method to transplant CUDA program to Android platform. The proposed method realizes that CUDA program running on an Android based mobile device. After transplantation, the program can be achieved on an Android based mobile device to access the general purpose GPU computing resources located on high-performance servers. It has great significance to the high-performance computing located on mobile devices.Mobile CUDA runtime environment consists of two parts-server side and client side. The client side is deployed on the Android system based mobile devices, in order to fake the API function gets called with the function parameters CUDA program, send it to the server, the server to process the actual GPU operations. Because of the faked API, the original program does not require the introduction of other specific API to achieve compatibility characteristics.Transplant CUDA program for the Android based platform is relied on the detailed analysis of CUDA program compilation process. To change the original mixed compilation strategy, using the method of separation compile of different platforms. GPU side’s device code execution compiled dot fatbin format and standard C language code executed on the host side cross-platform using ARM compilation tools compiled into ARM platform target program.Experiments proofs that the CUDA transplanting method which presented in this paper has good computing performance and speedup of computationally intensive algorithms to improve the performance needed to run many times and the effect is significant.