Agent Factories for High Level Synthesis: How Far Can General-Purpose Coding Agents Go in Hardware Optimization?

1. 1
   Define Target Algorithm

   Choose a high-level algorithmic specification (e.g., C/C++ function) that is suitable for hardware implementation via High-Level Synthesis (HLS). This will serve as the initial input for your AI agents.
2. 2
   Configure General-Purpose LLM Agents

   Set up access to a powerful Large Language Model (LLM) and instantiate client connections. Define distinct roles and initial prompts for each agent (e.g., 'HLS Designer', 'Optimizer', 'Evaluator') to guide them without explicit hardware-specific training.
3. 3
   Design the Agent Factory Orchestration

   Implement a 'two-stage pipeline' or a multi-agent coordination mechanism. This 'agent factory' will manage the construction, communication, and task assignment among your agents, guiding them through iterative optimization cycles.
4. 4
   Execute Optimization Cycles

   Initiate the agents to generate and refine hardware descriptions (e.g., Verilog, VHDL, or optimized C/C++ for HLS). Allow agents to iterate, propose modifications, and apply optimization strategies based on their general problem-solving capabilities.
5. 5
   Evaluate Hardware Metrics

   Use HLS tools (e.g., Vitis HLS, Intel HLS) to synthesize and analyze the generated hardware designs. Evaluate key metrics such as performance (latency, throughput), resource utilization (area), and power consumption to assess the agents' optimization effectiveness.

from openai import OpenAI # Or any other LLM client

# Dummy client for demonstration without an API key
class DummyLLMClient:
    def chat(self, messages):
        import time
        time.sleep(0.1) # Simulate API latency
        return type('obj', (object,), {'choices': [{'message': {'content': 'Generated initial hardware design concept.'}}]})()

class BaseAgent:
    def __init__(self, name, role_description, llm_client):
        self.name = name
        self.role_description = role_description
        self.llm = llm_client
        self.messages = []

    def generate_response(self, prompt, context=""):
        current_messages = [{"role": "system", "content": self.role_description}]
        if context: current_messages.append({"role": "user", "content": f"Context: {context}"})
        current_messages.append({"role": "user", "content": prompt})
        
        # Replace with actual LLM API call if using a real client
        response = self.llm.chat(messages=current_messages)
        agent_response = response.choices[0].message.content
        self.messages.append({"role": "user", "content": prompt}) # Keep history
        self.messages.append({"role": "assistant", "content": agent_response})
        return agent_response

# --- Example Usage ---
# llm_client = OpenAI(api_key="YOUR_OPENAI_API_KEY") # Use your actual LLM client
llm_client = DummyLLMClient() # For local testing

hls_designer_agent = BaseAgent(
    name="HLS Designer",
    role_description="You are an expert in High-Level Synthesis, skilled at translating C/C++ algorithms into efficient hardware descriptions. Focus on clarity and HLS pragmas.",
    llm_client=llm_client
)

optimization_agent = BaseAgent(
    name="Hardware Optimizer",
    role_description="You are an expert in hardware optimization, focusing on improving performance, area, and power for HLS generated designs. Identify bottlenecks and suggest C/C++ code transformations or HLS pragmas.",
    llm_client=llm_client
)

# Initial task for the HLS Designer
initial_design_prompt = "Generate a C/C++ function for a 4-tap FIR filter, suitable for HLS. Include basic HLS pragmas for pipelining if possible."
print(f"HLS Designer says: {hls_designer_agent.generate_response(initial_design_prompt)}")

# Example of a follow-up task for the Optimizer
optimization_prompt = "Analyze the provided FIR filter C/C++ code and suggest specific optimizations to reduce latency, assuming a target clock period of 10ns."
context_from_designer = "// (imagine C/C++ code for FIR filter generated by HLS Designer)"
print(f"Optimizer says: {optimization_agent.generate_response(optimization_prompt, context=context_from_designer)}")