IEEE Project Abstract

There exists an urgent need for determining theright amount and type of specialization while making a heterogeneous system as programmable and flexible as possible.Therefore, in this paper, we pioneer a self-optimizing and selfprogramming computing system (SOSPCS) design frameworkthat achieves both programmability and flexibility and exploitscomputing heterogeneity [e.g., CPUs, GPUs, and hardware accelerators (HWAs)]. First, at compile time, we form a task poolconsisting of hybrid tasks with different processing element (PE)affinities according to target applications. Tasks preferred tobe executed on GPUs or accelerators are detected from targetapplications by neural networks. Tasks suitable to run on CPUsare formed by community detection to minimize data movement overhead. Next, a distributed reinforcement learning-basedapproach is used at runtime to allow agents to map the tasksonto the network-on-chip-based heterogeneous PEs by learningan optimal policy based on Q values in the environment. We haveconducted experiments on a heterogeneous platform consistingof CPUs, GPUs, and HWAs with deep learning algorithmssuch as matrix multiplication, ReLU, and sigmoid functions.We concluded that SOSPCS provides performance improvementup to 4.12× and energy reduction up to 3.24× compared to thestate-of-the-art approaches.