Resource-constrained spatial multi-tasking for embedded GPU

For recent smart devices, the embedded GPU is an indispensable unit and multi-task scheduling on the GPU becomes an important issue. To prevent the performance degradation on foreground GPU task by competing background GPU tasks, we propose a novel spatial multi-tasking which limits the number of available GPU cores for each task based on its priority. The proposed algorithm improved the performance of high priority task by up to 14% over the temporal budget based multi-tasking.