#!/usr/bin/env python3 """ CircuitQA-Tool V7 Dataset Generator Generates the question-answer corpus with: - All 6 tools: synthesis_tool, simulation_tool, graph_extractor, static_analyzer, optimization_advisor, cross_stage_mapper - All 6 question categories with balanced distribution (~16.7% each) - Backend outputs from Yosys and Icarus Verilog (see eda_tools.py) - Two-stage training support (Stage 1: tool selection, Stage 2: answer generation) Input: Verilog files from CircuitQA_datasets Output: V7 JSONL with full CircuitQA-Tool structure """ import os import sys import json import random import hashlib from pathlib import Path from typing import Dict, List, Any, Optional, Tuple from dataclasses import dataclass, asdict, field from collections import defaultdict import argparse import re from concurrent.futures import ProcessPoolExecutor, as_completed import traceback import subprocess # Get conda path from environment or use default CONDA_PATH = os.environ.get('CONDA_PREFIX', os.path.expanduser("~/miniconda3")) def check_eda_dependencies(): """Verify EDA tools are available in conda 'eda' environment.""" try: # Check Yosys result = subprocess.run( f"source {CONDA_PATH}/bin/activate eda && yosys -V", shell=True, capture_output=True, text=True, timeout=10 ) if result.returncode != 0: print("WARNING: Yosys not found in conda 'eda' environment", flush=True) return False print(f"✓ Yosys available: {result.stdout.strip()}", flush=True) # Check iverilog result = subprocess.run( f"source {CONDA_PATH}/bin/activate eda && iverilog -V", shell=True, capture_output=True, text=True, timeout=10 ) if result.returncode != 0: print("WARNING: Icarus Verilog not found", flush=True) else: print(f"✓ Icarus Verilog available", flush=True) return True except Exception as e: print(f"WARNING: Failed to check EDA dependencies: {e}", flush=True) return False # Add processing directory to path sys.path.insert(0, os.path.dirname(__file__)) from eda_tools import ( synthesis_tool, simulation_tool, graph_extractor, static_analyzer, optimization_advisor, cross_stage_mapper, TOOL_REGISTRY, extract_module_name ) # ============================================================================ # Question Templates by Category # ============================================================================ QUESTION_TEMPLATES = { # Category 1: Design Quality Prediction (PPA Metrics) "design_quality": [ ("What is the estimated area of module {module}?", "area"), ("What is the estimated delay of module {module}?", "delay"), ("What is the estimated power consumption of module {module}?", "power"), ("How many gates are in module {module} after synthesis?", "gate_count"), ("What is the critical path delay of module {module}?", "critical_path"), ("Compare the area and delay trade-offs for module {module}.", "ppa_tradeoff"), ], # Category 2: Functional Behavior & Assertions "functional_behavior": [ ("What is the toggle rate of signal {signal} in module {module}?", "toggle_rate"), ("How many simulation cycles does module {module} require for coverage?", "coverage"), ("What is the signal activity profile of module {module}?", "activity"), ("Describe the functional behavior of module {module}.", "behavior"), ("What are the input/output relationships in module {module}?", "io_relation"), ], # Category 3: Structural Analysis "structural_analysis": [ ("How many nodes are in the {graph_type} graph of module {module}?", "node_count"), ("What is the average degree of nodes in module {module}'s graph?", "avg_degree"), ("Describe the graph structure of module {module}.", "graph_structure"), ("What are the main cell types in module {module}?", "cell_types"), ("How many edges connect the nodes in module {module}'s {graph_type}?", "edge_count"), ], # Category 4: Code Understanding & Explanation "code_understanding": [ ("Explain what module {module} does.", "explanation"), ("How many inputs and outputs does module {module} have?", "port_count"), ("What is the complexity of module {module}?", "complexity"), ("Describe the always blocks in module {module}.", "always_blocks"), ("What signals are defined in module {module}?", "signals"), ("Are there any potential issues in module {module}?", "lint"), ], # Category 5: Design Optimization & Recommendations "optimization": [ ("What synthesis optimizations should be applied to module {module} for {target}?", "recipe"), ("How can I reduce the area of module {module}?", "area_opt"), ("How can I improve the timing of module {module}?", "delay_opt"), ("How can I reduce power consumption in module {module}?", "power_opt"), ("What are the trade-offs of optimizing module {module} for {target}?", "tradeoffs"), ], # Category 6: Cross-Stage Reasoning "cross_stage": [ ("How does module {module} map from RTL to netlist?", "mapping"), ("What is the alignment score between RTL and netlist for module {module}?", "alignment"), ("Which RTL signals are unmapped in the netlist for module {module}?", "unmapped"), ("How many netlist cells correspond to module {module}?", "netlist_cells"), ], } # Tool mapping for each category CATEGORY_TOOLS = { "design_quality": "synthesis_tool", "functional_behavior": "simulation_tool", "structural_analysis": "graph_extractor", "code_understanding": "static_analyzer", "optimization": "optimization_advisor", "cross_stage": "cross_stage_mapper", } # ============================================================================ # V7 Data Structures # ============================================================================ @dataclass class V7QAPair: """V7 QA pair with full tool-calling support for all 6 categories.""" # Core QA fields question: str answer: str # Verilog context source_file: str module_name: str verilog_code: str # Tool calling tool_name: str tool_parameters: Dict[str, Any] tool_result: Dict[str, Any] # Metadata category: str subcategory: str difficulty: str question_hash: str = "" # Two-stage training stage1_target: str = "" stage2_context: str = "" def __post_init__(self): if not self.question_hash: self.question_hash = hashlib.md5( f"{self.question}{self.module_name}".encode() ).hexdigest()[:12] # ============================================================================ # Answer Generation Functions # ============================================================================ def generate_design_quality_answer(question: str, subcategory: str, tool_result: Dict, module_name: str) -> str: """Generate answer for design quality questions.""" if tool_result.get('status') != 'success': return f"Unable to analyze module {module_name}: {tool_result.get('error', 'unknown error')}" if subcategory == "area": area = tool_result.get('area', 0) return f"The estimated area of module {module_name} is {area:.2f} area units. This is based on synthesis with {tool_result.get('gate_count', 0)} gates." elif subcategory == "delay": delay = tool_result.get('delay', 0) return f"The estimated delay of module {module_name} is {delay:.2f} ps. The critical path estimate is {tool_result.get('critical_path_estimate', 0):.2f} ps." elif subcategory == "power": power = tool_result.get('power', 0) return f"The estimated power consumption of module {module_name} is {power:.4f} mW, based on {tool_result.get('gate_count', 0)} gates." elif subcategory == "gate_count": gate_count = tool_result.get('gate_count', 0) breakdown = tool_result.get('cell_breakdown', {}) breakdown_str = ", ".join(f"{k}: {v}" for k, v in list(breakdown.items())[:5]) return f"Module {module_name} synthesizes to {gate_count} gates. Cell breakdown: {breakdown_str}." elif subcategory == "critical_path": cp = tool_result.get('critical_path_estimate', 0) return f"The critical path delay of module {module_name} is approximately {cp:.2f} ps." else: # ppa_tradeoff return f"Module {module_name} has area={tool_result.get('area', 0):.2f}, delay={tool_result.get('delay', 0):.2f}ps, power={tool_result.get('power', 0):.4f}mW. Trade-off: reducing area may increase delay." def generate_functional_behavior_answer(question: str, subcategory: str, tool_result: Dict, module_name: str, signal_name: str = None) -> str: """Generate answer for functional behavior questions.""" if tool_result.get('status') not in ['success', 'estimated']: return f"Unable to simulate module {module_name}: {tool_result.get('error', 'unknown error')}" if subcategory == "toggle_rate" and signal_name: toggle_rates = tool_result.get('toggle_rates', {}) rate = toggle_rates.get(signal_name, 0.0) return f"The toggle rate of signal {signal_name} in module {module_name} is {rate:.3f} (toggles per cycle)." elif subcategory == "coverage": cycles = tool_result.get('simulation_cycles', 0) coverage = tool_result.get('coverage_estimate', 0) return f"Module {module_name} requires approximately {cycles} simulation cycles for {coverage:.1f}% coverage." elif subcategory == "activity": activity = tool_result.get('signal_activity', {}) return f"Module {module_name} has {activity.get('inputs', 0)} inputs, {activity.get('outputs', 0)} outputs, {activity.get('wires', 0)} wires, and {activity.get('regs', 0)} registers." else: # behavior, io_relation activity = tool_result.get('signal_activity', {}) return f"Module {module_name} processes {activity.get('inputs', 0)} inputs to produce {activity.get('outputs', 0)} outputs through {activity.get('regs', 0)} sequential elements." def generate_structural_analysis_answer(question: str, subcategory: str, tool_result: Dict, module_name: str, graph_type: str = "CDFG") -> str: """Generate answer for structural analysis questions.""" if tool_result.get('status') != 'success': return f"Unable to extract graph for module {module_name}: {tool_result.get('error', 'unknown error')}" if subcategory == "node_count": count = tool_result.get('node_count', 0) return f"The {graph_type} graph of module {module_name} contains {count} nodes." elif subcategory == "edge_count": count = tool_result.get('edge_count', 0) return f"The {graph_type} graph of module {module_name} contains {count} edges connecting the nodes." elif subcategory == "avg_degree": props = tool_result.get('graph_properties', {}) avg_deg = props.get('avg_degree', 0) return f"The average node degree in module {module_name}'s graph is {avg_deg:.2f}." elif subcategory == "cell_types": props = tool_result.get('graph_properties', {}) types = props.get('node_types', {}) types_str = ", ".join(f"{k}: {v}" for k, v in list(types.items())[:5]) return f"Module {module_name} contains the following cell types: {types_str}." else: # graph_structure return f"Module {module_name}'s {graph_type} has {tool_result.get('node_count', 0)} nodes and {tool_result.get('edge_count', 0)} edges with average degree {tool_result.get('graph_properties', {}).get('avg_degree', 0):.2f}." def generate_code_understanding_answer(question: str, subcategory: str, tool_result: Dict, module_name: str) -> str: """Generate answer for code understanding questions.""" if tool_result.get('status') != 'success': return f"Unable to analyze module {module_name}: {tool_result.get('error', 'unknown error')}" if subcategory == "explanation": ports = tool_result.get('port_count', {}) complexity = tool_result.get('code_complexity', {}) return f"Module {module_name} is a Verilog module with {ports.get('inputs', 0)} inputs and {ports.get('outputs', 0)} outputs. It has {tool_result.get('always_blocks', 0)} always blocks and {tool_result.get('assign_statements', 0)} assign statements." elif subcategory == "port_count": ports = tool_result.get('port_count', {}) return f"Module {module_name} has {ports.get('inputs', 0)} inputs, {ports.get('outputs', 0)} outputs, and {ports.get('inouts', 0)} bidirectional ports (total: {ports.get('total', 0)})." elif subcategory == "complexity": complexity = tool_result.get('code_complexity', {}) return f"Module {module_name} has cyclomatic complexity of {complexity.get('cyclomatic_complexity', 0)}, {complexity.get('lines_of_code', 0)} lines of code, and nesting depth of {complexity.get('nesting_depth', 0)}." elif subcategory == "always_blocks": count = tool_result.get('always_blocks', 0) return f"Module {module_name} contains {count} always blocks for sequential and combinational logic." elif subcategory == "signals": signals = tool_result.get('signal_count', {}) return f"Module {module_name} defines {signals.get('wires', 0)} wires, {signals.get('regs', 0)} registers, and {signals.get('parameters', 0)} parameters." else: # lint warnings = tool_result.get('lint_warnings', []) if warnings: return f"Potential issues in module {module_name}: {'; '.join(warnings[:3])}." return f"No obvious issues detected in module {module_name}." def generate_optimization_answer(question: str, subcategory: str, tool_result: Dict, module_name: str, target: str = "area") -> str: """Generate answer for optimization questions.""" if tool_result.get('status') != 'success': return f"Unable to analyze module {module_name}: {tool_result.get('error', 'unknown error')}" recipes = tool_result.get('recommended_recipes', []) suggestions = tool_result.get('optimization_suggestions', []) improvement = tool_result.get('estimated_improvement', {}) trade_offs = tool_result.get('trade_offs', []) if subcategory == "recipe": recipe_str = "; ".join(recipes[:2]) if recipes else "default synthesis" return f"For {target} optimization of module {module_name}, apply: {recipe_str}" elif subcategory in ["area_opt", "delay_opt", "power_opt"]: if suggestions: sugg_str = "; ".join(s.get('description', '') for s in suggestions[:2]) return f"To optimize {module_name}: {sugg_str}" return f"Apply standard {target} optimization to module {module_name}." else: # tradeoffs if trade_offs: return f"Trade-offs for optimizing {module_name} for {target}: {'; '.join(trade_offs[:2])}" return f"Optimizing {module_name} for {target} may affect other metrics." def generate_cross_stage_answer(question: str, subcategory: str, tool_result: Dict, module_name: str) -> str: """Generate answer for cross-stage reasoning questions.""" if tool_result.get('status') != 'success': return f"Unable to map module {module_name}: {tool_result.get('error', 'unknown error')}" stats = tool_result.get('mapping_statistics', {}) if subcategory == "mapping": return f"Module {module_name} maps {stats.get('rtl_signals', 0)} RTL signals to {stats.get('netlist_wires', 0)} netlist wires through {stats.get('netlist_cells', 0)} cells." elif subcategory == "alignment": score = tool_result.get('alignment_score', 0) return f"The RTL-to-netlist alignment score for module {module_name} is {score:.1f}%." elif subcategory == "unmapped": unmapped = tool_result.get('unmapped_elements', []) if unmapped: return f"Unmapped RTL signals in module {module_name}: {', '.join(unmapped[:5])}." return f"All RTL signals in module {module_name} are mapped to the netlist." else: # netlist_cells return f"Module {module_name} synthesizes to {stats.get('netlist_cells', 0)} netlist cells." # ============================================================================ # V7 Generator Class # ============================================================================ class V7Generator: """Generate V7 dataset with all 6 tools and 6 categories.""" def __init__(self, base_dir: str, output_file: str): self.base_dir = Path(base_dir) self.output_file = Path(output_file) # Statistics self.stats = defaultdict(int) self.category_counts = defaultdict(int) self.tool_counts = defaultdict(int) # Cache for tool results self.tool_cache: Dict[str, Dict] = {} def find_verilog_files(self) -> List[Path]: """Find all Verilog files in the base directory.""" verilog_files = [] for ext in ['*.v', '*.sv']: verilog_files.extend(self.base_dir.rglob(ext)) return verilog_files def read_verilog_file(self, filepath: Path) -> Tuple[str, str]: """Read Verilog file and extract module name.""" try: with open(filepath, 'r', encoding='utf-8', errors='ignore') as f: content = f.read() module_name = extract_module_name(content) return content, module_name or filepath.stem except Exception as e: return "", "" def get_tool_result(self, tool_name: str, verilog_code: str, **kwargs) -> Dict[str, Any]: """Get tool result with caching.""" # Use longer hash (16 chars = 64 bits) to reduce collision probability cache_key = f"{tool_name}:{hashlib.md5(verilog_code.encode()).hexdigest()[:16]}" if cache_key in self.tool_cache: return self.tool_cache[cache_key] tool_func = TOOL_REGISTRY.get(tool_name) if not tool_func: return {'status': 'error', 'error': f'Unknown tool: {tool_name}'} try: if tool_name == 'cross_stage_mapper': result = tool_func(verilog_code, None) elif tool_name == 'graph_extractor': result = tool_func(verilog_code, kwargs.get('graph_type', 'CDFG')) elif tool_name == 'optimization_advisor': result = tool_func(verilog_code, kwargs.get('target_metric', 'area')) elif tool_name == 'simulation_tool': result = tool_func(verilog_code, kwargs.get('num_cycles', 100)) elif tool_name == 'synthesis_tool': result = tool_func(verilog_code, kwargs.get('optimization_level', 'medium')) else: result = tool_func(verilog_code) self.tool_cache[cache_key] = result return result except Exception as e: return {'status': 'error', 'error': str(e)} def generate_qa_for_module(self, verilog_code: str, module_name: str, source_file: str) -> List[Dict]: """Generate QA pairs for a single module across all categories.""" qa_pairs = [] # Get signals for functional behavior questions signals = re.findall(r'(?:input|output|wire|reg)\s+(?:\[\d+:\d+\])?\s*(\w+)', verilog_code) signals = list(set(signals))[:5] # Limit to 5 signals graph_types = ['CDFG', 'AIG', 'DFG'] targets = ['area', 'delay', 'power'] # Target 2 QA pairs per category for balanced distribution target_per_category = 2 for category in QUESTION_TEMPLATES.keys(): tool_name = CATEGORY_TOOLS[category] templates = list(QUESTION_TEMPLATES[category]) random.shuffle(templates) category_pairs_generated = 0 for template, subcategory in templates: # Stop once we have enough pairs for this category if category_pairs_generated >= target_per_category: break try: # Skip signal-based questions if no signals found if '{signal}' in template and not signals: continue # Format question signal = None graph_type = 'CDFG' target = 'area' if '{signal}' in template: signal = random.choice(signals) question = template.format(module=module_name, signal=signal) elif '{graph_type}' in template: graph_type = random.choice(graph_types) question = template.format(module=module_name, graph_type=graph_type) elif '{target}' in template: target = random.choice(targets) question = template.format(module=module_name, target=target) else: question = template.format(module=module_name) # Get tool parameters tool_params = self._get_tool_params(tool_name, graph_type, target) # Execute tool tool_result = self.get_tool_result(tool_name, verilog_code, **tool_params) # Generate answer answer = self._generate_answer(category, subcategory, question, tool_result, module_name, signal, graph_type, target) # Create stage1 target (tool call JSON) stage1_target = json.dumps({ 'tool': tool_name, 'parameters': tool_params }) # Create improved stage2 context (keep essential fields, truncate logs) essential_result = {k: v for k, v in tool_result.items() if k not in ['synthesis_log', 'simulation_log', 'nodes', 'edges']} stage2_context = f"Question: {question}\n\nTool: {tool_name}\nResult: {json.dumps(essential_result)}" if len(stage2_context) > 1500: stage2_context = stage2_context[:1500] + "..." # Determine difficulty (easy/medium) difficulty = self._estimate_difficulty(subcategory, tool_result) qa_pair = { 'question': question, 'answer': answer, 'source_file': source_file, 'module_name': module_name, 'verilog_code': verilog_code[:3000], # Limit code size 'tool_name': tool_name, 'tool_parameters': tool_params, 'tool_result': tool_result, 'category': category, 'subcategory': subcategory, 'difficulty': difficulty, 'question_hash': hashlib.md5(f"{question}{module_name}".encode()).hexdigest()[:12], 'stage1_target': stage1_target, 'stage2_context': stage2_context } qa_pairs.append(qa_pair) self.category_counts[category] += 1 self.tool_counts[tool_name] += 1 category_pairs_generated += 1 except Exception as e: self.stats['errors'] += 1 continue return qa_pairs def _get_tool_params(self, tool_name: str, graph_type: str = 'CDFG', target: str = 'area') -> Dict: """Get tool parameters based on tool name.""" if tool_name == 'synthesis_tool': return {'optimization_level': 'medium'} elif tool_name == 'simulation_tool': return {'num_cycles': 100, 'simulation_type': 'functional'} elif tool_name == 'graph_extractor': return {'graph_type': graph_type} elif tool_name == 'static_analyzer': return {'analysis_type': 'full'} elif tool_name == 'optimization_advisor': return {'target_metric': target} elif tool_name == 'cross_stage_mapper': return {} return {} def _generate_answer(self, category: str, subcategory: str, question: str, tool_result: Dict, module_name: str, signal: str = None, graph_type: str = 'CDFG', target: str = 'area') -> str: """Generate answer based on category.""" if category == 'design_quality': return generate_design_quality_answer(question, subcategory, tool_result, module_name) elif category == 'functional_behavior': return generate_functional_behavior_answer(question, subcategory, tool_result, module_name, signal) elif category == 'structural_analysis': return generate_structural_analysis_answer(question, subcategory, tool_result, module_name, graph_type) elif category == 'code_understanding': return generate_code_understanding_answer(question, subcategory, tool_result, module_name) elif category == 'optimization': return generate_optimization_answer(question, subcategory, tool_result, module_name, target) elif category == 'cross_stage': return generate_cross_stage_answer(question, subcategory, tool_result, module_name) return f"Analysis of module {module_name} completed." def _estimate_difficulty(self, subcategory: str, tool_result: Dict) -> str: """Estimate question difficulty (easy or medium).""" if tool_result.get('status') == 'error': return 'medium' # Simple heuristics if subcategory in ['port_count', 'gate_count', 'node_count', 'alignment']: return 'easy' return 'medium' def generate(self, max_files: int = None, target_pairs: int = 100000): """Generate V7 dataset.""" import sys print(f"Finding Verilog files in {self.base_dir}...", flush=True) verilog_files = self.find_verilog_files() print(f"Found {len(verilog_files)} Verilog files", flush=True) if max_files: verilog_files = verilog_files[:max_files] all_qa_pairs = [] with open(self.output_file, 'w') as f: for i, filepath in enumerate(verilog_files): if len(all_qa_pairs) >= target_pairs: print(f"Reached target of {target_pairs} pairs", flush=True) break try: verilog_code, module_name = self.read_verilog_file(filepath) if not verilog_code or not module_name: self.stats['skipped'] += 1 continue # Skip very small or very large files if len(verilog_code) < 50 or len(verilog_code) > 50000: self.stats['skipped'] += 1 continue source_file = str(filepath.relative_to(self.base_dir)) qa_pairs = self.generate_qa_for_module(verilog_code, module_name, source_file) for qa in qa_pairs: f.write(json.dumps(qa) + '\n') all_qa_pairs.append(qa) f.flush() # Flush after each module self.stats['files_processed'] += 1 if (i + 1) % 50 == 0: print(f"Processed {i+1}/{len(verilog_files)} files, {len(all_qa_pairs)} QA pairs", flush=True) sys.stdout.flush() except Exception as e: self.stats['errors'] += 1 continue self.stats['total_pairs'] = len(all_qa_pairs) return all_qa_pairs def print_stats(self): """Print generation statistics.""" print("\n" + "=" * 60) print("V7 Dataset Generation Statistics") print("=" * 60) print(f"Files processed: {self.stats['files_processed']}") print(f"Files skipped: {self.stats['skipped']}") print(f"Errors: {self.stats['errors']}") print(f"Total QA pairs: {self.stats['total_pairs']}") print("\nCategory Distribution:") total = sum(self.category_counts.values()) for cat, count in sorted(self.category_counts.items()): pct = count / total * 100 if total > 0 else 0 print(f" {cat}: {count} ({pct:.1f}%)") print("\nTool Usage:") for tool, count in sorted(self.tool_counts.items()): pct = count / total * 100 if total > 0 else 0 print(f" {tool}: {count} ({pct:.1f}%)") # ============================================================================ # Main Function # ============================================================================ def main(): parser = argparse.ArgumentParser(description='Generate V7 CircuitQA-Tool dataset') parser.add_argument('--base-dir', type=str, default='.', help='Base directory containing Verilog files') parser.add_argument('--output', type=str, default='v7_qa_dataset.jsonl', help='Output JSONL file') parser.add_argument('--max-files', type=int, default=None, help='Maximum number of files to process') parser.add_argument('--target-pairs', type=int, default=100000, help='Target number of QA pairs') parser.add_argument('--skip-eda-check', action='store_true', help='Skip EDA tool dependency check') args = parser.parse_args() # Check EDA dependencies if not args.skip_eda_check: print("Checking EDA tool dependencies...", flush=True) if not check_eda_dependencies(): print("WARNING: Some EDA tools missing. Results may be estimated.", flush=True) print("", flush=True) generator = V7Generator(args.base_dir, args.output) generator.generate(max_files=args.max_files, target_pairs=args.target_pairs) generator.print_stats() if __name__ == '__main__': main()