MARS: Exploiting Multi-Level Parallelism for DNN Workloads on Adaptive Multi-Accelerator Systems | IEEE Conference Publication | IEEE Xplore