class IMS_NPU_Requirements:
    """
    IMS DMS/OMS NPU 选型需求矩阵
    
    核心约束：
    1. 车规认证（AEC-Q100 Grade 3/2）
    2. 功耗预算（通常 <5W 散热）
    3. 实时性（推理延迟 <33ms）
    4. 模型大小（量化后 <10MB）
    5. SDK 成熟度（量化/编译/运行时）
    """
    
    def __init__(self):
        self.requirements = {
            'dms_main': {
                'task': '驾驶员疲劳/分心检测',
                'model': 'YOLO-face + gaze estimation',
                'tops_required': '4-8 TOPS INT8',
                'latency': '<33ms (30fps)',
                'power': '<3W',
                'certification': 'AEC-Q100 Grade 2 (-40°C to 105°C)',
                'memory': '<5MB model size',
                'critical_features': [
                    '面部关键点检测',
                    'PERCLOS 计算',
                    '视线方向估计',
                    '头部姿态估计',
                    '手机使用检测'
                ]
            },
            'oms_main': {
                'task': '乘员检测/分类/姿态',
                'model': 'YOLO-pose + classification',
                'tops_required': '6-12 TOPS INT8',
                'latency': '<50ms (20fps)',
                'power': '<4W',
                'certification': 'AEC-Q100 Grade 3 (-40°C to 85°C)',
                'memory': '<8MB model size',
                'critical_features': [
                    '乘员检测与分类',
                    'OOP 异常姿态',
                    '安全带使用检测',
                    'CPD 儿童检测'
                ]
            },
            'cpd_radar': {
                'task': 'CPD 雷达信号处理',
                'model': 'CNN range-doppler',
                'tops_required': '1-2 TOPS',
                'latency': '<100ms (10fps)',
                'power': '<1W',
                'certification': 'AEC-Q100 Grade 3',
                'memory': '<1MB model size',
                'critical_features': [
                    '微动检测',
                    '呼吸/心跳识别',
                    '多人检测'
                ]
            }
        }
    
    def evaluate_chip(self, chip_name: str, tops: float, 
                      power: float, certification: str) -> dict:
        """评估芯片是否满足 IMS 需求"""
        results = {}
        for use_case, req in self.requirements.items():
            required_tops_str = req['tops_required']
            required_tops = float(required_tops_str.split('-')[0])
            required_power = float(req['power'].replace('<', '').replace('W', ''))
            
            meets_tops = tops >= required_tops
            meets_power = power <= required_power
            meets_cert = 'AEC-Q100' in certification
            
            results[use_case] = {
                'meets_tops': meets_tops,
                'meets_power': meets_power,
                'meets_cert': meets_cert,
                'overall': meets_tops and meets_power and meets_cert
            }
        
        return results

class SDKMaturityComparison:
    """NPU SDK 成熟度对比"""
    
    def __init__(self):
        self.vendors = {
            'Qualcomm': {
                'sdk': 'QAIRT/QNN/SNPE + AI Hub',
                'model_zoo': '完整（YOLO v3-v11）',
                'documentation': '公开 + 社区活跃',
                'quantization': 'INT8/INT16, PTQ + QAT',
                'simulator': '有',
                'deployment_artifact': '.dlc / QNN context binary',
                'lock_in': '中（专有格式，但 AI Hub 开放）',
                'ims_score': 9.5
            },
            'TI': {
                'sdk': 'TIDL + Processor SDK',
                'model_zoo': '完整（检测/分割/姿态）',
                'documentation': '公开 + GitHub',
                'quantization': 'INT8, PTQ',
                'simulator': '有',
                'deployment_artifact': '.tflite (量化)',
                'lock_in': '低（TFLite 标准）',
                'ims_score': 9.0
            },
            'NXP': {
                'sdk': 'eIQ + TIM-VX + Vela',
                'model_zoo': '部分（YOLOv4-tiny/v8）',
                'documentation': '混合（部分需账号）',
                'quantization': 'INT8',
                'simulator': '有',
                'deployment_artifact': '.vela',
                'lock_in': '中-高',
                'ims_score': 7.5
            },
            'Rockchip': {
                'sdk': 'RKNN-Toolkit2',
                'model_zoo': '完整（YOLO 全系列）',
                'documentation': 'GitHub 开放',
                'quantization': 'INT8/INT16/FP16',
                'simulator': '有',
                'deployment_artifact': '.rknn',
                'lock_in': '高（专有格式）',
                'ims_score': 6.5
            },
            'Amlogic': {
                'sdk': 'AMLNN Toolkit + libnnsdk.so',
                'model_zoo': '完整（GitHub playground）',
                'documentation': 'GitHub 开放',
                'quantization': 'INT8',
                'simulator': '有',
                'deployment_artifact': '.adla',
                'lock_in': '高',
                'ims_score': 6.0
            },
            'NVIDIA': {
                'sdk': 'CUDA + TensorRT',
                'model_zoo': '最完整',
                'documentation': '公开',
                'quantization': 'INT8/FP16/FP32',
                'simulator': '有',
                'deployment_artifact': '.engine',
                'lock_in': '中（TensorRT）',
                'ims_score': 8.5
            }
        }
    
    def rank_for_ims(self):
        """按 IMS 适配度排序"""
        ranked = sorted(self.vendors.items(), 
                        key=lambda x: x[1]['ims_score'], reverse=True)
        return ranked

class IMSDeploymentWorkflow:
    """IMS 模型量化部署工作流"""
    
    def __init__(self, vendor: str = 'qualcomm'):
        self.vendor = vendor
        self.workflow = {
            'qualcomm': [
                '1. PyTorch 模型训练',
                '2. 导出 ONNX',
                '3. QAIRT 量化（INT8 PTQ）',
                '4. 编译为 QNN context binary',
                '5. 推送到 QCS8550',
                '6. SNPE 运行时推理',
                '7. 性能分析（latency/memory/power）'
            ],
            'ti': [
                '1. PyTorch 模型训练',
                '2. 导出 TFLite',
                '3. TIDL 量化（INT8 PTQ）',
                '4. 编译为 TFLite 量化模型',
                '5. 推送到 TDA4VM',
                '6. TFLite Runtime 推理',
                '7. 性能分析'
            ],
            'rockchip': [
                '1. PyTorch 模型训练',
                '2. 导出 ONNX',
                '3. RKNN-Toolkit2 量化+编译',
                '4. 生成 .rknn',
                '5. 推送到 RK3588',
                '6. RKNN Runtime 推理',
                '7. 性能分析'
            ]
        }
    
    def estimate_deployment_time(self, model_complexity: str = 'medium'):
        """估算部署时间"""
        times = {
            'simple': {'training': 8, 'export': 1, 'quantize': 2, 
                      'compile': 2, 'deploy': 1, 'total': 14},
            'medium': {'training': 24, 'export': 2, 'quantize': 4,
                      'compile': 4, 'deploy': 2, 'total': 36},
            'complex': {'training': 48, 'export': 4, 'quantize': 8,
                        'compile': 8, 'deploy': 4, 'total': 72}
        }
        return times.get(model_complexity, times['medium'])

lock_in_risk = {
    'low': ['TI (TFLite 标准)'],
    'medium': ['Qualcomm (AI Hub 开放但专有格式)', 'NVIDIA (生态成熟)'],
    'high': ['NXP', 'Rockchip', 'Amlogic', 'STMicro', 'Renesas'],
    'critical': ['专用 SDK 账号要求 + NDA 限制']
}