Announcement about new repository
We are migrating the repository to a new one (https://github.com/georgia-tech-synergy-lab/MAERI). The new reposiotry contains latest version of MAERI source code, which is used in a tutorial at HPCA 2019 (Feb 16, 2019).
MAERI (Multiply-Accumulate Engine with Reconfigurable Interconnect)
MAERI is a deep learning inference accelerator that enables fine-grained compute resource allocation at run time pusblished in ASPLOS 2018 (Hyoukjun Kwon, Ananda Samjadr, and Tushar Krishna, "MAERI: Enabling Flexible Dataflow Mapping over DNN Accelerators via Reconfigurable Interconnects." ASPLOS, 2018). ASPLOS 2018 paper, Sysml 2018 paper MAERI is written in Bluespec System Verilog. To run simulation and compile Verilog, you need to obtain Bluespec Compiler license. Please note that Bluespec provides free academic license to universities (http://bluespec.com/university/).
How to use the code
Please use script "Maeri" at the top level. Available options:(-c: compile simulation, -r: run simulation, -v: compile Verilog, -clean: clean up repo) For details, please refer to our tutorial slides in ./slides directory.
How to modify the configuration
Please change parameters in AcceleratorConfig.bsv. Please note that this code does not include compiler to generate multiplier/adder switch configurations so you will need to update ./testbench/testMAERI.bsv manually if you want to use settings other than included presets.