HAIL: Hardware Automation through Intelligent Language Models
HAIL is a budget-aware agent workflow for repairing and optimizing AMD Vitis HLS C++ kernels. It uses C simulation, synthesis, and C/RTL co-simulation as feedback loops, then grades the resulting kernel for correctness and performance.
How it works
- Load the contract: each task supplies kernel code, immutable headers and testbenches, target FPGA part, clock, a credit budget, and an optional co-simulation requirement.
- Reach correctness: C simulation catches compile and functional errors; structural tasks (e.g. DATAFLOW stream deadlocks) additionally require C/RTL co-simulation. Focused log regions are fed back to the model after each failure.
- Optimize safely: correct candidates are synthesized, and optimizations are accepted only if they stay correct and improve measured latency.
- Grade independently: hidden tests, synthesis, and co-simulation run outside the agent budget, producing a correctness-gated score covering latency, resources, tool usage, time, and tokens.
HAIL supports a single-model OpenRouter backend and a CrewAI coder/reviewer backend, plus an optional Qdrant-based hybrid retrieval (RAG) layer that supplies relevant Vitis HLS guidance during optimization. Targets an AMD Alveo U55C with Vitis/Vivado 2025.2.
Team: Suraj Karki, Rabigh Khan (Hochschule Bielefeld), Sujan Sharma (TU Dortmund), Muhammad Shaheer (Paderborn University). Advisors: Qazi Arbab Ahmed and Thorsten Jungeblut (Hochschule Bielefeld).
Tech stack: Python · LLM Agents · CrewAI · OpenRouter · Qdrant (RAG) · AMD Vitis HLS · Vivado · Docker