84 lines
3.1 KiB
Python
84 lines
3.1 KiB
Python
|
import sys
|
||
|
|
import os
|
||
|
|
|
||
|
|
# === 导入你的自定义模块 ===
|
||
|
|
from coverage_agent import CoverageAnalyzer # "眼"
|
||
|
|
from tb_injector import TestbenchInjector # "手"
|
||
|
|
|
||
|
|
# === 导入 CorrectBench 的核心模块 ===
|
||
|
|
# 假设你的 GPT_call.py 文件就在根目录，如果是在 utils 文件夹下，请改为 from utils.GPT_call import ...
|
||
|
|
try:
|
||
|
|
from GPT_call import llm_call, extract_code
|
||
|
|
except ImportError:
|
||
|
|
# 尝试从 utils 导入 (根据你的文件结构调整)
|
||
|
|
sys.path.append(os.path.join(os.path.dirname(__file__), 'utils'))
|
||
|
|
from GPT_call import llm_call, extract_code
|
||
|
|
|
||
|
|
def main(annotated_file, tb_file):
|
||
|
|
print(f"--- [CGA] Starting Coverage-Guided Agent Loop ---")
|
||
|
|
|
||
|
|
# 1. 初始化眼 (CoverageAnalyzer)
|
||
|
|
print(f"--- Step 1: Analyzing Coverage for {annotated_file} ---")
|
||
|
|
analyzer = CoverageAnalyzer(annotated_file)
|
||
|
|
score = analyzer.analyze()
|
||
|
|
print(f" Current Score: {score:.2f}%")
|
||
|
|
|
||
|
|
# 设定目标阈值
|
||
|
|
if score > 95.0:
|
||
|
|
print(" [Success] Coverage is already high. No action needed.")
|
||
|
|
return
|
||
|
|
|
||
|
|
# 2. 生成 Prompt (由覆盖率报告驱动)
|
||
|
|
prompt = analyzer.generate_prompt()
|
||
|
|
print(" [Info] Prompt generated successfully.")
|
||
|
|
|
||
|
|
# 3. 调用 CorrectBench 的大脑 (GPT_call)
|
||
|
|
print("--- Step 2: Asking LLM for Fix (via CorrectBench Infra) ---")
|
||
|
|
|
||
|
|
# 构建消息格式 (CorrectBench 标准格式)
|
||
|
|
messages = [{"role": "user", "content": prompt}]
|
||
|
|
|
||
|
|
# === 关键集成点 ===
|
||
|
|
# 使用 Qwen-max 或你配置里的任何模型
|
||
|
|
# 注意：确保你的 config/key_API.json 或环境变量配置正确
|
||
|
|
target_model = "qwen-max"
|
||
|
|
|
||
|
|
try:
|
||
|
|
response_text, other_infos = llm_call(
|
||
|
|
input_messages=messages,
|
||
|
|
model=target_model,
|
||
|
|
system_message="You are a Verilog expert specializing in testbench generation.",
|
||
|
|
temperature=0.2 # 保持低温度以获得稳定代码
|
||
|
|
)
|
||
|
|
print(f" [LLM] Response received. Tokens used: {other_infos['usage']['total_tokens']}")
|
||
|
|
|
||
|
|
except Exception as e:
|
||
|
|
print(f" [Error] LLM Call failed: {e}")
|
||
|
|
return
|
||
|
|
|
||
|
|
# 4. 提取代码
|
||
|
|
# 直接复用 GPT_call 里的 extract_code 函数
|
||
|
|
codes = extract_code(response_text, "verilog") # 或者是 "systemverilog"
|
||
|
|
|
||
|
|
if not codes:
|
||
|
|
print(" [Error] No code block found in LLM response.")
|
||
|
|
return
|
||
|
|
|
||
|
|
fix_code = codes[0] # 取第一个代码块
|
||
|
|
|
||
|
|
# 5. 注入代码 (手术)
|
||
|
|
print("--- Step 3: Injecting Code into Testbench ---")
|
||
|
|
injector = TestbenchInjector(tb_file)
|
||
|
|
try:
|
||
|
|
new_tb_path = injector.inject_task(fix_code, new_tb_path="enhanced_tb.v")
|
||
|
|
print(f" [Success] Enhanced Testbench saved to: {new_tb_path}")
|
||
|
|
print(" -> Now you can run simulation with 'enhanced_tb.v' to verify improvement.")
|
||
|
|
except Exception as e:
|
||
|
|
print(f" [Error] Injection failed: {e}")
|
||
|
|
|
||
|
|
if __name__ == "__main__":
|
||
|
|
if len(sys.argv) < 3:
|
||
|
|
print("Usage: python3 run_cga.py <annotated_file_path> <original_tb_path>")
|
||
|
|
print("Example: python3 run_cga.py logs/annotated/prob_lemmings4.v saves/lemmings4/final_TB.v")
|
||
|
|
else:
|
||
|
|
main(sys.argv[1], sys.argv[2])
|