MARCO: A High-performance Task <u>M</u> apping <u>a</u> nd <u>R</u> outing <u>Co</u> -optimization Framework for Point-to-Point NoC-based Heterogeneous Computing Systems

Heterogeneous computing systems (HCSs), which consist of various processing elements (PEs) that vary in their ability, are usually facilitated by the network-on-chip (NoC) to interconnect its components. The emerging point-to-point NoCs support single-cycle-multi-hop transmission, reduce or eliminate latency dependence on distance, addressing scalability concern raised high for long-distance transmission and enlarging design space routing algorithm search non-shortest paths. For such NoC-based HCSs, resource management strategies managed compilers, scheduler, controllers, e.g., mapping routing, complicated following reasons: (i) Due heterogeneity, need optimize computation communication concurrently (for homogeneous systems, only communication). (ii) Conducting consecutively cannot minimize schedule length most cases since PEs with ability may locate crowded area suffer from contention overhead. (iii) Since changing selection one task will reconstruct whole space, exploration is challenging. Therefore, this work, we propose MARCO, m apping a nd r outing co -optimization framework, decrease applications HCSs. Specifically, revise tabu explore evaluate quality routing. advanced reinforcement learning (RL)algorithm, i.e., advantage actor-critic, adopted efficiently compute We perform extensive experiments real applications, demonstrates MARCO achieves remarkable performance improvement terms (+44.94% ? +50.18%) when compared state-of-the-art co-optimization systems. also compare different combinations approaches.