Two Mapping Strategies for Autonomous Driving Accelerators

Many-core architectures hold tremendous potential in autonomous driving accelerators due to their significant energy efficiency advantages. However, effectively utilizing the vast computational resources of such architectures remains a major challenge and an open problem. To address this challenge, we propose two optimization strategies: Gluing Parallel Layers Dataflow (GPLD) and the Ouroboros-like resource allocation strategy. GPLD focuses on optimizing parallel layers, a dimension that has not yet been explored. We design a simple and elegant dataflow to directly glue parallel layers, allowing them to compute as a single layer. This approach reduces energy costs and improves the allocation of chiplet resources. The Ouroboros-like resource allocation strategy is employed to utilize idle resources during segment switching, thereby increasing overall throughput. Overall, these mapping optimizations improve 2.67× performance on average and reduce energy consumption by 50.68% compared to the state-of-the-art Tangram mapping strategy, optimized by us, across a wide range of DNNs.