Domain-aware Genetic Algorithms for Hardware and Mapping Optimization for Efficient DNN Acceleration

Thumbnail Image
Kao, Sheng-Chun
Krishna, Tushar
Associated Organization(s)
Supplementary to
The proliferation of AI across a variety of domains (vision, language, speech, recommendations, games) has led to the rise of domain-specific accelerators for deep learning. At design-time, these accelerators carefully architect the on-chip dataflow to maximize data reuse (over space and time) and size the hardware resources (PEs and buffers) to maximize performance and energy-efficiency, while meeting the chip’s area and power targets. At compile-time, the target Deep Neural Network (DNN) model is mapped over the accelerator. The mapping refers to tiling the computation and data (i.e., tensors) and scheduling them over the PEs and scratchpad buffers respectively, while honoring the microarchitectural constraints (number of PEs, buffer sizes, and dataflow). The design-space of valid hardware resource assignments for a given dataflow and the valid mappings for a given hardware is extremely large (~O(10^24)) per layer for state-of-the-art DNN models today. This makes exhaustive searches infeasible. Unfortunately, there can be orders of magnitude performance and energy-efficiency differences between an optimal and sub-optimal choice, making these decisions a crucial part of the entire design process. Moreover, manual tuning by domain experts become unprecedentedly challenged due to increased irregularity (due to neural architecture search) and sparsity of DNN models. This necessitate the existence of Map Space Exploration (MSE). In this thesis, our goal is to deliver a deep analysis of the MSE for DNN accelerators, propose different techniques to improve MSE, and generalize the MSE framework to a wider landscape (from mapping to HW-mapping co-exploration, from single-accelerator to multi-accelerator scheduling). As part of it, we discuss the correlation between hardware flexibility and the formed map space, formalized the map space representation by four mapping axes: tile, order, parallelism, and shape. Next, we develop dedicated exploration operators for these axes and use genetic algorithm framework to converge the solution. Next, we develop "sparsity-aware" technique to enable sparsity consideration in MSE and a "warm-start" technique to solve the search speed challenge commonly seen across learning-based search algorithms. Finally, we extend out MSE to support hardware and map space co-exploration and multi-accelerator scheduling.
Date Issued
Resource Type
Resource Subtype
Rights Statement
Rights URI