RACO: Agentic CGRA Hardware/Software Co-Design for FMCW Radar with CHIA

Paper source · Quick start · Results and logs · Live exploration

---

📸 Demo

Recorded run showing mapping, RTL verification, and synthesis. Layout is optional.

🚀 Quick start

On Linux x86-64, install Git, make, a C compiler, and Python 3.10+. Then check the saved results without a model or Docker:

git clone https://github.com/zsjiang99/RACO.git
cd RACO
make -C chia-raco artifact-check

🖥️ Open the browser demo

This requires Node.js 20.19+ or 22.12+ and npm in addition to Python 3.10+. From the repository root:

python3.10 -m venv .venv
.venv/bin/python -m pip install -r gui/requirements.txt
(cd gui/frontend && npm ci && npm run build)
bash gui/start.sh

Open http://127.0.0.1:8765/ to inspect the saved CGRA design in the workbench. The paper's 38.73M-cycle run is in the saved results. To change the port, set A3_GUI_PORT before starting the server. The server has no authentication and binds to loopback; do not expose it publicly.

🤖 Run a new exploration

First complete the browser-demo setup above. Live exploration additionally needs Docker Engine, an OpenAI-compatible model endpoint, and at least 25 GB of free Docker storage. Docker supplies LLVM and EDA tools; no local GPU is needed when the model endpoint is remote.

🧰 Install the mapper and RTL tools

mkdir -p external
git clone https://github.com/tancheng/CGRA-Mapper.git external/CGRA-Mapper
git -C external/CGRA-Mapper checkout 5f8393acb6b3a17146806ec93a57f76f875be232
docker build -t cgramapper:v1 -f chia-raco/docker/mapper.Dockerfile external/CGRA-Mapper
docker pull cgra/neura-flow:20260114

📦 Install CHIA, RACO, and the agent classes

git clone https://github.com/ucb-bar/chia.git external/chia
git -C external/chia checkout 16c35e92aaaf9511c6453bf94cd5cf589698f4e3
.venv/bin/python -m pip install ./external/chia
.venv/bin/python -m pip install -e 'chia-raco[test]'
git clone https://github.com/coredac/MACO.git external/upstream-agents
git -C external/upstream-agents checkout 31c02ce013838d89ef2a6d211acfdf639ecb178d

The agent checkout supplies upstream classes; the project and interface are named RACO. To check the installation:

RACO_AGENT_DIR="$PWD/external/upstream-agents/agent" make -C chia-raco test PYTHON=../.venv/bin/python
make -C chia-raco mapper-smoke PYTHON=../.venv/bin/python

🔑 Configure model access

The GUI has no login or API-key field. Set model access in the same terminal that starts the server; the key stays on the server, not in the browser or repository. Live exploration needs an OpenAI-compatible Chat Completions endpoint that accepts this project's temperature, seed, and max_tokens requests and permits at least 4096 completion tokens.

For the OpenAI API, create a key in the OpenAI dashboard, then enter it without putting the key in shell history:

read -rsp 'OpenAI API key: ' RACO_LLM_API_KEY; echo
export RACO_LLM_API_KEY
export RACO_LLM_BASE_URL=https://api.openai.com/v1
export RACO_LLM_MODEL=your-compatible-model-id

For a local OpenAI-compatible service instead, start the service and set RACO_LLM_BASE_URL to its reachable /v1 endpoint and RACO_LLM_MODEL to its served model ID. Set RACO_LLM_API_KEY only if that service requires one. OPENAI_API_KEY alone is not read by this workflow; use RACO_LLM_API_KEY. Do not commit credentials or paste them into the GUI.

▶️ Start the live workflow

export RACO_AGENT_DIR="$PWD/external/upstream-agents/agent"
bash gui/start.sh

Open http://127.0.0.1:8765/, choose Multi-agent, Hardware only, or Single agent, then click Run exploration. Run one exploration at a time. The workflow maps the FMCW kernels, generates and verifies RTL, and synthesizes the selected design. Layout is a separate, optional action.

📄 License and credit

Zesong Jiang, Cheng Tan, and Jeff Zhang · Arizona State University. New integration code is BSD-3-Clause licensed. The agent modules are fetched from a pinned upstream checkout; other dependencies retain their own terms. See the third-party notices.