Automatic Loop Nest Parallelization for the Predictable Execution Model

Abstract

Currently, embedded real-time systems still widely use single-core processors. A major challenge in the adoption of multicore processors is the presence of shared hardware resources such as main memory. Contention between threads executing on different cores for access to such resources makes it difficult to tightly estimate the Worst-Case Execution Time (WCET) of applications. To safely employ multicore processors in real-time systems, previous work has introduced a PRedictable Execution Model (PREM) for embedded Multi-Processor Systems-on-a-Chip (MPSoCs). Under PREM, each thread is divided into memory phases, where the code and data required by the thread are moved from main memory to a local memory (cache or scratchpad) or vice versa, and execution phases, where the thread computes based on the code and data available in local memory. Memory phases are then scheduled by the Operating System (OS) to avoid contention among threads, thus resulting in tight WCET bounds. The main challenge in applying the model is to automatically generate optimized PREM-compliant code instead of rewriting programs manually. Note that many programs of interests, such as emerging AI and neural network kernels, comprise both compute-intensive and memory-intensive deeply nested loops. Hence, PREM code generation and optimization should be applicable to nested loop structures and consider whether performance is constrained by computation or memory transfers. In this thesis, we address the problem of automatically parallelizing and optimizing nested loop structure programs by presenting a workflow that automatically generates PREM-compliant optimized code. To correctly model the structure of nested loop programs, we leverage existing polyhedral compilation tools that analyze the original program and generate optimized executables. Two main techniques are adopted for optimization: loop tiling and parallelization. We build a timing model to estimate the length of execution and memory phases, and then construct a Directed Acyclic Graph (DAG) of program phases to estimate its makespan. During this process, our framework searches for the combination of tile sizes and thread numbers that minimize the makespan of the program; given the complexity of the optimization problem, we design a heuristic algorithm to find solutions close to the optimal. Finally, to show its usefulness, we evaluate our technique based on the Gem5 architectural simulator on computational kernels from the PolyBench-NN benchmark.