Improving Memory Access Efficiency for Heterogeneous Multiprocessor Systems

Recent trends show that using mobile GPUs to support 3D graphics and H.264 decoding based on OpenGL ES on smart phones are prevalent. To enhance the performance of both CPU and GPU, Heterogeneous System Architecture (HSA) is suggested which is one of methods accessing different kinds of memory using virtual memory structures. In this study, we suggest a method called Fusion Processor Architecture for accessing memory from CPU and GPU without copying contexts between them. We introduce control mechanisms including Job Manager, Re-mapper, and Pre-fetcher to control and distribute data between the two processors. Using these methods help reducing message exchanges and page-table requests between CPU and GPU, which in turn increasing efficiency of the two devices. Evaluations compared to Compute Unified Device Architecture (CUDA), OpenMP, and OpenCL prove the suggested mechanism works 198% faster on a QEMU emulation environment.