MEDEA: A Multi-objective Evolutionary Approach to DNN Hardware Mapping
Enrico Russo1,a, Maurizio Palesi1,b, Salvatore Monteleone2, Davide Patti1,c, Giuseppe Ascia1,d and Vincenzo Catania1,e
1Department of Electrical, Electronic, and Computer Engineering (DIEEI), University of Catania, I-95125 Catania, Italy
aenrico.russo7@studium.unict.it
bmaurizio.palesi@dieei.unict.it
cdavide.patti@dieei.unict.it
dgiuseppe.ascia@dieei.unict.it
evincenzo.catania@dieei.unict.it
2Department of Engineering, Niccoló Cusano University, I-00166 Rome, Italy
salvatore.monteleone@unicusano.it
ABSTRACT
Deep Neural Networks (DNNs) embedded domainspecific accelerators enable inference on resource-constrained devices. Making optimal design choices and efficiently scheduling neural network algorithms on these specialized architectures is challenging. Many choices can be made to schedule computation spatially and temporally on the accelerator. Each choice influences the access pattern to the buffers of the architectural hierarchy, affecting the energy and latency of the inference. Each mapping also requires specific buffer capacities and a number of spatial components instances that translate in different chip area occupation. The space of possible combinations, the mapping space, is so large that automatic tools are needed for its rapid exploration and simulation. This work presents MEDEA, an opensource multi-objective evolutionary algorithm based approach to DNNs accelerator mapping space exploration. MEDEA leverages the Timeloop analytical cost model. Differently from the other schedulers that optimize towards a single objective, MEDEA allows deriving the Pareto set of mappings to optimize towards multiple, sometimes conflicting, objectives simultaneously. We found that solutions found by MEDEA dominates in most cases those found by state-of-the-art mappers.
Keywords: Neural Network, Accelerator, Compiler, Scheduling, Genetic Algorithm, Multi-Objective Optimization.
