Reproduce “NEURA: A Unified and Retargetable Compilation Framework for Coarse-Grained Reconfigurable Architectures” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shangkun Li, Jinming Ge, Diyuan Tao, Zeyu Li, Jiawei Liang, Linfeng Du, Jiang Xu, Wei Zhang and 1 more
NEURA presents a unified and retargetable compilation framework for Coarse-Grained Reconfigurable Architectures (CGRAs) that effectively bridges the control-dataflow semantic gap by transforming complex control flow into a pure dataflow representation. The framework demonstrates geometric mean speedups up to 2.42x over state-of-the-art baselines for high-performance CGRAs, with negligible hardware overhead.
Key insights
- Unifying complex control flow into a pure dataflow graph via a predicated type system eliminates serialization bottlenecks and enables comprehensive exploitation of intra- and inter-basic block parallelism.
- Directly attaching predicate bits to data values allows for systematic and native handling of hierarchical control dependencies, resolving a key limitation in previous predication techniques.
- An MLIR-based intermediate representation (IR) achieves retargetability and extensibility, decoupling the compiler from specific CGRA execution models and requiring minimal, uniform ISA extensions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.