Answers are grounded in this paper and cite their sources.
Demonstrates effective sparse CNN mapping in streaming CGRAs, highlighting improved efficiency.
Streaming coarse-grained reconfigurable array (CGRA), with its reconfigurability and efficient memory access, is a promising platform for the convolutional neural network (CNN) inference. While model sparsity is widely used to accelerate scaling CNNs, existing methods lack support for mapping sparse CNNs onto streaming CGRAs, mainly due to the substantial configuration overhead and complex I/O data scheduling at the processing element array (PEA) side. To bridge this gap, we propose SparseMap, the first mapping framework for sparse CNN processing on streaming CGRA. Firstly, SparseMap introduces an inter-operator configuration context sharing mechanism to reduce configuration overhead, along with a Multi-graph Maximum Common Subgraph (Multi-MCS)-based optimization method, which jointly fine-tunes sparsity patterns across operators to maximize the sharing benefits. Secondly, to handle the irregularity of sparse processing, SparseMap establishes a co-optimization approach that integrates PEA-side I/O data planning with operation scheduling&binding, coordinating I/O data and computation to ensure high-throughput operator mapping. In addition, SparseMap introduces a hierarchical mapping strategy to improve mapping efficiency. Experiments show that SparseMap effectively enables sparse CNN processing on streaming CGRAs. Compared to baseline methods, SparseMap reduces the configuration overhead by one to three orders of magnitude, effectively alleviating the configuration pressure, while achieving a speedup of 1.24 × to 2.49 × over dense computation.
No takes yet. Share an insight, caveat, or question.
Ni et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: