TransMap: An Efficient CGRA Mapping Framework via Transformer and Deep Reinforcement Learning

Coarse-grained reconfigurable architectures (CGRAs) are garnering increasing attention as domain-specific accelerators owing to their high flexibility and energy efficiency. Mapping encompasses both placement and routing, constituting a crucial part of the CGRA toolchain. Achieving high mapping quality while minimizing mapping time has been a key objective. In this paper, we propose TransMap, an efficient CGRA mapping framework. TransMap enhances the graph representation by incorporating position and type embeddings of Data Flow Graphs (DFGs) and CGRAs. The transformer encoder serves as a graph learner, extracting attention relationships between nodes and edges in the graph representation. Concurrently, a deep reinforcement learning agent is developed to decode the global attention information embedded in the transformer output. Through the integrated learning of the transformer and deep reinforcement learning, experiments demonstrate that our approach exhibits superior performance in mapping quality and compilation time compared to state-of-the-art mapping algorithms.