LLM-Aided Refactoring and Optimization of C/C++ Code for High-Level Synthesis

Introduction

In High-Level Synthesis (HLS), converting a standard C/C++ program into an HLS-compatible version (HLS-C) requires significant manual effort. Our LLM-driven program repair framework aims to automate this process by taking regular C/C++ code as input and generating its optimized HLS-C code ready for synthesis.

Setup

To utilize this framework, you need to add specific Python scripts to your directory containing the regular C/C++ program. These scripts facilitate the interaction between LLM and the Catapult HLS tool for refactoring and optimization.

Scripts Description

• pre.py and ext.py: Extract reference examples from the HLS tool manual.
• reasoning.py: Step-wise reasoning to find potential errors.
• emb.py: Matches repair templates in the repair library.
• compile.py: Use this script to run your program as usual.
• inter_compile.py: Automates the interaction for repairing regular C code using the LLM.
• run.py: Uses the LLM to attempt multiple repairs on the C code.
• pass.py: Monitors and reports the pass rate of the repairs across multiple iterations, providing metrics on the effectiveness of the repairs.
• bwa.py: Analyze and fine-tune the bit width of floating-point variables.

Usage and Configuration

Modify the following paths in the scripts according to your environment setup:

• tcl_script_path: Path to your TCL script for running Catapult HLS.
  Example: /path/to/your/run_catapult.tcl
• log_file_path: Path where the Catapult log will be stored.
  Example: /path/to/your/catapult.log
• base_dir: Base directory for different HLS design instances.
  Example: /path/to/your/E3_BFS
• command: Command to invoke your installed design compiler.
  Example: /path/to/synopsys/syn/J-2014.09-SP3/bin/dc_shell -f /path/to/your/tcl_script
• catapult_bin: Directory path to your installed Catapult HLS tool.
  Example: /path/to/catapult/2022.2-1008433/Mgc_home/bin

Ensure these paths are correctly set to match your local environment.

Conclusion

The proposed LLM-aided C/C++ code refactoring and optimization framework for HLS covers runtime profiling and C/C++ code refactoring for hardware generation and performance optimization with minimal human effort.