Exact Scheduling to Minimize Off-Chip Data Movement for Deep Learning Accelerators

Specialized hardware accelerators are increasingly utilized to provide performance/power efficiency for Deep Neural Network (DNN) applications. However their benefits are limited by expensive off-chip data movement between host memory and the accelerator's on-chip scratchpad, which can consume significantly more energy than accelerator computation [13]. While application-level DNN operators can have arbitrary sizes, accelerators typically support fixed-sized operations due to constrained on-chip memory and micro-architectures. Consequently, mapping an application-level operator to an accelerator involves decomposing it into loops of smaller tiles. Different choices of tile sizes, loop orders and memory partition across tensors result in a vast design space with huge differences in off-chip data movement volume. To address this challenge, we introduce Shoehorn, a schedule optimization framework that jointly optimizes loop tiling, loop ordering, and memory partitioning for mapping application-level DNN operators to hardware accelerators. Shoehorn can generate optimal schedules in subseconds and outperforms state-of-the-art approaches, reducing up to 51% total off-chip memory traffic relative to competing schedulers for several widely-used DNN applications on three distinct hardware accelerator targets.