dMazeRunner: Optimization Infrastructure for Programmable Dataflow Accelerators for Deep Learning

Deep neural networks (DNNs) can be efficiently executed on dataflow accelerators. However, the vast space of executing convolutions on computational and memory resources of accelerators makes it difficult for the programmers to automatically and efficiently accelerate the convolutions and for architects to achieve efficient accelerator designs. We propose dMazeRunner framework, which allows users to optimize execution methods for efficiently accelerating DNN operators like convolutions and GEMMs on a given architecture and to explore dataflow accelerator designs for different layers of DNN models. dMazeRunner optimizes the execution of DNN operators with various dataflow mechanisms and achieves highly efficient execution methods for DNN models within a few seconds. Check out the first comprehensive optimization infrastructure at - https://github.com/MPSLab-ASU/dMazeRunner Keywords — Hardware accelerators, energy-efficiency, mapping, deep learning, design space exploration - Learn about recent advances in designing accelerators for machine learning in this comprehensive survey: https://arxiv.org/abs/2007.00864 [In Proceedings of the IEEE, 2021]