A CGRA Definition Framework for Dataflow Applications

An End-to-End Programming Model for AI Engine Architectures

“…Coarse-Grained Reconfigurable Architectures (CGRAs) are architectures that can selectively use or disuse various components, subsets of the architecture to gain much higher performance per Watt over a more diverse set of applications than conventional processors such as multi-core CPUs and GPUs (Choi and Kee, 2015). In particular, owing to the reconfigurability of the device topology, CGRAs are well-suited for dataflow programs, i.e., specifications for the connectivity of functional units that explicitly represent and correspond to the flow of data in a program (Charitopoulos and Pnevmatikatos, 2020). Archetypical in this class of programs are multi-layered Deep Neural Networks (DNNs), wherein individual layers potentially map to discrete subsets of the CGRA and with data flowing between them in the form of activations (Choi and Kee, 2015).…”

Creating Customized CGRAs for Scientific Applications

“…The naive approach considers these nodes as separate, leading to CGRA designs that are harder to route due to cell heterogeneity. To address this, we include in the CSFD phase Node Merging; an algorithm that is designed to find whether two nodes with the same functionality should be merged under the same bit-width and what the optimal bit-width for the current application is described in detail in [27]. We use two metrics for Node Merging: the bit-width difference between the two nodes and the Percentage Gain of merging these two nodes.…”