Scholarlink

<p>Real-time embedded systems need to meet timing and energy constraints to avoid potential disasters. Compiler-level ScratchPad Memory (SPM) allocation can be used to optimize a program’s Worst-Case Execution Time (WCET) and energy consumption. However, static allocation is limited by SPM size constraints. Dynamic SPM allocation resolves this by allocating code to SPM during runtime, but copying code using the CPU increases WCET and energy consumption. To address this, we integrate a Direct Memory Access (DMA) model and DMA analysis at the compiler level and propose a single-objective DMA Call Placement Optimization (<i>DCPO</i>). In this paper, we consider functions and loops as dynamic allocation candidates. <i>DCPO</i> finds appropriate places within the code to place DMA transfer calls such that the DMA controller and the CPU run parallelly—minimizing the total execution time required by the DMA controller for dynamic allocation of functions and loops during runtime. Additionally, we propose a compiler-level DMA-aware multi-objective dynamic SPM allocation that uses <i>DCPO</i> and simultaneously minimizes WCET and energy objectives, yielding Pareto optimal solutions. Comparative evaluations demonstrate the superiority of our approach over state-of-the-art multi- and single-objective optimizations.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Compiler-level DMA-aware multi-objective dynamic SPM allocation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年3月1日

查看全文
  • Shashank Jadhav,
  • Heiko Falk

摘要

Real-time embedded systems need to meet timing and energy constraints to avoid potential disasters. Compiler-level ScratchPad Memory (SPM) allocation can be used to optimize a program’s Worst-Case Execution Time (WCET) and energy consumption. However, static allocation is limited by SPM size constraints. Dynamic SPM allocation resolves this by allocating code to SPM during runtime, but copying code using the CPU increases WCET and energy consumption. To address this, we integrate a Direct Memory Access (DMA) model and DMA analysis at the compiler level and propose a single-objective DMA Call Placement Optimization (DCPO). In this paper, we consider functions and loops as dynamic allocation candidates. DCPO finds appropriate places within the code to place DMA transfer calls such that the DMA controller and the CPU run parallelly—minimizing the total execution time required by the DMA controller for dynamic allocation of functions and loops during runtime. Additionally, we propose a compiler-level DMA-aware multi-objective dynamic SPM allocation that uses DCPO and simultaneously minimizes WCET and energy objectives, yielding Pareto optimal solutions. Comparative evaluations demonstrate the superiority of our approach over state-of-the-art multi- and single-objective optimizations.