E2EMap: End-to-End Reinforcement Learning for CGRA Compilation via Reverse Mapping

Coarse-Grained Reconfigurable Arrays (CGRAs) are a promising architecture to cope with the challenges of increasing demand for high performance and high energy efficiency. However, the actual achieved performance of CGRA is highly dependent on the mappers. Traditional mappers using heuristics or combinatorial optimization can hardly learn from past experience, suffering from poor quality and portability. Recently, machine learning has been introduced to partial components in CGRA compilers, leaving other components to traditional heuristics, which is also prone to a sub-optimum, To this end, this paper proposes an end-to-end learning framework, E2EMap, for CGRA mapping that can cover the full mapping process. To reduce the complexity of the learning model, a reverse mapping problem is formulated, where various routing strategies can be thoroughly explored. To solve the problem, policy gradient reinforcement learning is introduced to learn from scratch. Experimental results demonstrate that E2EMap can achieve up to 2.23 x mapping quality across different CGRA settings while consuming even less compilation time as compared to state-of-the-art works.