Catapult: An LLM-based Automatic Repair Tool for C/C++ High-Level Synthesis Programs

Catapult is an innovative open-source project that uses large language models to automate the repair of C/C++ programs in the high-level synthesis (HLS) workflow. It addresses the time-consuming and labor-intensive issues of traditional manual debugging, significantly improving the efficiency of hardware design verification. It provides an intelligent solution for HLS technology, which bridges software algorithms and hardware implementation.

High-level synthesis (HLS) technology is a key bridge connecting software algorithms and hardware implementation. However, converting C/C++ code into efficient hardware circuits faces several issues: syntax compatibility differences, timing constraint violations, abnormal resource utilization, and insufficient parallelism exploitation. The traditional manual repair process requires repeated iterations, and complex errors may take hours or even days to fix.

Catapult's core architecture consists of four key components: 1. Error Detection and Diagnosis Module: Extracts key error information from HLS logs through LLM semantic understanding; 2. Context-Aware Code Analysis: Considers the global perspective of function control flow, data dependencies, and HLS pragma directives; 3. Intelligent Repair Strategy Generation: Generates candidate solutions based on fine-tuned large language models (e.g., CodeLlama, StarCoder); 4. Repair Verification and Feedback Loop: Automatically verifies repair solutions and provides feedback to optimize subsequent strategies.

Catapult has three key technical features: 1. Multi-dimensional Error Classification: Covers syntax-level, semantic-level, timing-level, and resource-level errors; 2. Domain-Specific Prompt Engineering: Designs prompt templates that include HLS expertise; 3. Progressive Repair Strategy: First resolves basic errors, then gradually optimizes performance to avoid the risk of over-modification.

Catapult's application scenarios include: 1. Accelerating FPGA Prototype Verification: Shortens the cycle from algorithm to bitstream; 2. Legacy Code Modernization: Helps migrate mature algorithms to FPGA platforms; 3. Education and Training Assistance: Generates repair explanations to help students understand HLS principles.

Currently, Catapult is mainly optimized for Xilinx Vivado HLS, and support for other tools is pending development. The success rate of deep architecture refactoring optimization needs to be improved. Future directions: Expand multi-toolchain support, use reinforcement learning to optimize repair quality, develop an interactive repair interface, and build a pre-synthesis performance prediction model.

Catapult represents a significant progress in AI-assisted hardware design automation, lowering the threshold for FPGA programming and improving development efficiency. As large model technology evolves, such tools will play a more important role in the hardware design field, enabling more software developers to harness the computing power of programmable hardware.