PNNL

Abstract

SODA (Software Defined Accelerators) is an automatic and agile framework to streamline the generation of hardware accelerators from their high-level programming environment (e.g., PyTorch) by leveraging Google's Multi-Level Intermediate Representation (MLIR) compiler infrastructure and High-Level Synthesis (HLS) tools. It enables efficient synthesis of deep neural network (DNN) models (in high-level abstractions) into hardware descriptions via a combination of compiler-level transformations and synthesis techniques. This research focuses on improving the design quality and turnaround time of the SODA framework via exploring optimizations such as loop transformations (e.g., permutation, tiling, and unrolling) and memory optimizations (i.e., temporary buffer allocation and promotion of buffers to stack) to enable practical DNN accelerator synthesis. In particular, we propose ML4SODA, a machine learning (ML) based Design Space Exploration (DSE) engine that systematically evaluates HLS optimizations across different DNN layer types, such as convolution, depth-wise Convolution, and fully connected layers. Our DSE incorporates with a Decision Tree, which is trained from prior HLS runs and at inference time, it directly generates the optimal parameters for SODA framework to achieve performance-optimal or energy-efficiency-optimal design points. We embedded our decision-tree based DSE into SODA's compilation pass and our results demonstrate faster synthesis and high quality designs on large DNNs.