Mapping Streaming Applications to OpenCL

Graphic processing units (GPUs) have been gaining popularity in general purpose and high performance computing. A GPU is made up of a number of streaming multiprocessors (SM), each of which consists of many processing cores. A large number of general-purpose applications have been mapped onto GPUs efficiently. Stream processing applications, however, exhibit properties such as unfavorable data movement patterns and low computation-tocommunication ratio that might lead to a poor performance on a GPU. We describe the automated mapping framework developed earlier that maps most stream processing applications onto NVIDIA GPUs efficiently by taking into account its architectural characteristics. We then discuss the implementation details of porting the mapping framework to OpenCL running on AMD GPUs and evaluate the performance of the mapping framework by running several benchmarks. Performance between the generated CUDA and OpenCL code is compared based on different heuristics.