Síntese de alto nível com modelos de linguagem de grande porte : injeção de diretivas automatizada por IA
Year: 2026
Type: dissertation
Abstract: This work investigates the use of Large Language Models (LLMs) for the automated selection and injection of optimization directives in High-Level Synthesis (HLS) projects. The manual choice of these directives directly impacts performance, area utilization, and power consumption in FPGA-based systems, but typically requires prior experience and specialized domain knowledge. An automated pipeline was developed to leverage commercial LLMs from the Google Gemini and Gemma families to analyze C/C++ source code and inject suitable pragmas to optimize loops, functions, and memory accesses, using a zero-shot prompt engineering strategy. The pipeline was evaluated against the MLDHLS benchmark suite using AMD Vitis HLS, and results were compared against an unoptimized baseline, the MLDHLS best-known solutions, and the HGBO state-of-the-art framework. The findings demonstrate that frontier LLMs can produce competitive directive configurations in a single inference pass, in some cases surpassing both the dataset reference and the iterative ML-based baseline, while also revealing systematic failure modes specific to zero-shot HLS optimization.
Author: Gomes, Sofia Popsin
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213483651