MARS: Enabling Efficient DNN Mapping for Spatial Accelerators in Ruby Space

Spatial accelerators can significantly enhance the computational performance of deep neural networks (DNNs) by mapping each layer across spatial and temporal dimensions. However, state-of-the-art mapping tools predominantly rely on perfect factorization mapping (PFM), which restricts tile sizes to exact divisors of tensor dimensions and may overlook high-quality mappings in the broader design space. To address this limitation, we propose MARS, the first efficient exploration framework designed for Ruby space. Ruby space is defined as a generalized mapping space that relaxes the factor-only constraint by allowing any integer tile size within the valid range of a dimension. MARS first models the coupling of partitioning, scheduling, and hardware resource constraints in Ruby space as a constrained optimization problem. It then relaxes the discrete mapping space into a continuous domain and employs a combination of linear programming and a genetic algorithm to rapidly obtain high-quality candidate mappings. Experimental results on representative DNN models show that MARS achieves a reduction of \(15\% \) to \(65\% \) in energy-delay product compared to state-of-the-art methods on average. Furthermore, MARS demonstrates superior exploration efficiency and consistently better mapping quality.