To Rip or not to Rip: A Reinforcement Learning-based Rip-up and Reroute Algorithm for Global Routing

Fangzhou WANG, Wanying ZHENG

Keywords: Routing, Reinforcement Learning, EDA

Abstract:

Routing, including global routing and detailed routing, has been a critical step in
the design of integrated circuits. Most of the existing global routers will firstly use
techniques like pattern routing and layer assignment to quickly generate a routing
solution and optimize total wirelength and via usage. Then rip-up and reroute
(RRR) scheme will be applied to iteratively reduce the number of overflows in the
whole design. However, compared with initial routing stage, RRR will be much
more time consuming. It will rip up all the nets that pass through overflowed area
and reroute them sequentially. Even if the routing resources in one routing cell
is overused by 1, the router will rip up all the nets that are routed on the routing
cell, as it does not know which net will be the best choice to rip up. In this way,
RRR may be doing a lot of redundant work. Besides, some initial routing solutions
that are optimal in terms of wirelength will also be wasted when they are ripped
up, causing a loss of routing quality. Therefore, in this project, we propose to use
reinforcement learning to help decide which nets to rip up in each RRR iterations.
An actor-critic based Proximal Policy Optimization (PPO) agent is trained for this
task. Experimental results show that the proposed approach can successfully reduce
the number of rerouted nets with little loss of routing quality on the ICCAD’19
global routing contest benchmarks, which demonstrate the effectiveness of our
model.