Enabling Manual-Controllable Compilation for Dataflow CGRAs
Felix Böseler, Jörg Walter, Verena Klös · 2025
Coarse Grained Reconfigurable Arrays (CGRAs) are particularly interesting accelerators because they uniquely combine domain-specific configurability with high energy-efficiency to execute computation-heavy kernels/algorithms. This domain specificity often requires extensive manual refinements of kernels from experts during compilation to fully exploit all accelerator advantages. However, there is currently little work that focuses on a systematic compilation approach which facilitates seamless manual controllability for experts during the kernel refinement process. To fill this gap, we contribute a compiler which revolves around a flexible yet understandable graph-based intermediate representation to enable the required manual controllability. We evaluate the compiler by compiling a neural network layer to a novel commercial dataflow CGRA.