import numpy as np

# 论文数据：INT8 vs FP32 加速比
platforms = {
    'ARM Cortex-A78 (SDOT)': 2.1,     # 有 dot-product 指令
    'ARM Cortex-A55 (无SDOT)': 0.59,   # 无 dot-product，反而变慢
    'x86 Intel (VNNI)': 1.8,          # 有 VNNI 指令
    'x86 AMD (无VNNI)': 0.59,         # 无 VNNI，变慢
    'Jetson Orin iGPU': 1.5,          # GPU 有 INT8 加速
    'NVDLA': 3.2,                      # 专用加速器
    'Qualcomm Hexagon HTP': 4.5,      # NPU 最快
    'DEEPX DX-M1': 3.8,               # NPU 很快
}

print("INT8 加速比 (vs FP32):")
print("-" * 45)
for platform, speedup in platforms.items():
    bar = "█" * int(speedup * 10)
    print(f"  {platform:30s} {speedup:4.1f}x {bar}")

print(f"\n最大差异: {max(platforms.values())/min(platforms.values()):.1f}x")
print(f"有ISA加速均值: {np.mean([2.1, 1.8, 1.5, 3.2, 4.5, 3.8]):.1f}x")
print(f"无ISA加速均值: {np.mean([0.59, 0.59]):.2f}x (变慢!)")

# NPU 量化兼容性测试
npu_results = {
    'Qualcomm Hexagon HTP': {
        'bring_your_own_qdq': {
            'status': 'silent_failure',  # 编译运行无报错
            'accuracy': 0.005,  # 但精度从 0.75 暴跌到 0.005
            'issue': '外部量化标度被忽略'
        },
        'vendor_native_path': {
            'status': 'success',
            'accuracy': 0.74,
        }
    },
    'DEEPX DX-M1': {
        'bring_your_own_qdq': {
            'status': 'compile_error',  # 编译器直接拒绝
            'accuracy': None,
            'issue': '不支持的 QDQ 图结构'
        },
        'vendor_native_path': {
            'status': 'success',
            'accuracy': 0.73,
        }
    }
}

for npu, results in npu_results.items():
    print(f"\n{npu}:")
    for method, result in results.items():
        print(f"  {method}: status={result['status']}, acc={result['accuracy']}")

class IMSQuantizationDeployer:
    """
    IMS 量化部署指南
    
    基于论文发现的实践建议
    """
    
    PLATFORM_PROFILES = {
        'qualcomm_8255': {
            'isa': 'hexagon_htp',
            'has_dotprod': True,
            'vendor_quant_required': True,
            'recommended_path': 'QNN SDK native',
            'fp32_baseline': 'onnxruntime',
        },
        'qualcomm_8295': {
            'isa': 'hexagon_htp',
            'has_dotprod': True,
            'vendor_quant_required': True,
            'recommended_path': 'QNN SDK native',
            'fp32_baseline': 'onnxruntime',
        },
        'jetson_orin': {
            'isa': 'nvdla',
            'has_dotprod': True,
            'vendor_quant_required': False,
            'recommended_path': 'TensorRT INT8',
            'fp32_baseline': 'TensorRT FP32',
        },
        'arm_cortex_a78': {
            'isa': 'arm_sdot',
            'has_dotprod': True,
            'vendor_quant_required': False,
            'recommended_path': 'ONNX Runtime INT8',
            'fp32_baseline': 'ONNX Runtime FP32',
        },
        'arm_cortex_a55': {
            'isa': 'arm_neon',
            'has_dotprod': False,  # 关键：无 SDOT
            'vendor_quant_required': False,
            'recommended_path': 'FP32 (INT8 反而更慢)',
            'fp32_baseline': 'ONNX Runtime FP32',
            'warning': 'INT8 在此平台会减速 1.7x！'
        },
    }
    
    def get_deployment_guide(self, model_name: str, target: str) -> dict:
        """
        获取量化部署指南
        
        Args:
            model_name: 模型名称
            target: 目标平台
            
        Returns:
            guide: 部署指南
        """
        profile = self.PLATFORM_PROFILES.get(target)
        if not profile:
            return {'error': f'Unknown platform: {target}'}
        
        guide = {
            'model': model_name,
            'target': target,
            'recommended_quant_path': profile['recommended_path'],
            'fp32_baseline': profile['fp32_baseline'],
            'has_dotprod': profile['has_dotprod'],
            'vendor_quant_required': profile.get('vendor_quant_required', False),
        }
        
        # 风险评估
        if not profile['has_dotprod']:
            guide['risk'] = 'HIGH'
            guide['warning'] = profile.get('warning', 'INT8 will be slower than FP32')
            guide['recommendation'] = 'Use FP32 instead of INT8 on this platform'
        elif profile.get('vendor_quant_required'):
            guide['risk'] = 'MEDIUM'
            guide['warning'] = 'Must use vendor SDK for quantization, QDQ graphs not portable'
            guide['recommendation'] = f"Use {profile['recommended_path']} for quantization"
        else:
            guide['risk'] = 'LOW'
            guide['recommendation'] = 'Standard INT8 PTQ is safe'
        
        return guide


# 测试
if __name__ == "__main__":
    deployer = IMSQuantizationDeployer()
    
    models = ['face_detection', 'eye_keypoints', 'fatigue_classifier', 'cpd_radar']
    platforms = ['qualcomm_8255', 'jetson_orin', 'arm_cortex_a78', 'arm_cortex_a55']
    
    print("IMS 量化部署指南:")
    print("=" * 80)
    
    for platform in platforms:
        print(f"\n{'─'*60}")
        print(f"平台: {platform}")
        print(f"{'─'*60}")
        
        for model in models:
            guide = deployer.get_deployment_guide(model, platform)
            risk_emoji = {'LOW': '🟢', 'MEDIUM': '🟡', 'HIGH': '🔴'}.get(guide.get('risk', ''), '⚪')
            print(f"  {risk_emoji} {model:20s} → {guide['recommended_quant_path']}")
            if 'warning' in guide:
                print(f"     ⚠️ {guide['warning']}")

def cross_platform_validation_workflow():
    """
    跨平台量化验证流程（基于论文发现）
    
    每个模型在部署前必须通过此流程
    """
    steps = [
        {
            'step': 1,
            'name': 'FP32 基线建立',
            'action': '在 ONNX Runtime FP32 上运行测试集',
            'output': 'FP32 预测结果 + 精度基线',
            'pass_criteria': 'Top-1 精度 ≥ 目标值',
        },
        {
            'step': 2,
            'name': 'INT8 量化（厂商路径）',
            'action': '使用目标平台厂商 SDK 量化',
            'output': '量化后模型 + 量化报告',
            'pass_criteria': '精度下降 < 1%',
        },
        {
            'step': 3,
            'name': '逐样本一致性验证',
            'action': '对比 INT8 与 FP32 的逐样本预测',
            'output': '一致率统计',
            'pass_criteria': '一致率 ≥ 99.5%（安全关键）',
            'critical': '论文发现 958-965/1000 不够安全！',
        },
        {
            'step': 4,
            'name': '延迟测量',
            'action': '在目标硬件上测量端到端延迟',
            'output': '延迟分布 P50/P95/P99',
            'pass_criteria': 'P99 ≤ 帧率要求（如 33ms@30fps）',
            'note': '注意输出传输延迟，NPU 可能受此限制',
        },
        {
            'step': 5,
            'name': '跨平台一致性回归',
            'action': '在所有目标平台上运行同一测试集',
            'output': '跨平台一致率矩阵',
            'pass_criteria': '所有平台对一致率 ≥ 99%',
            'critical': '论文核心发现：不同平台 INT8 输出不一致！',
        },
    ]
    
    for s in steps:
        print(f"Step {s['step']}: {s['name']}")
        print(f"  动作: {s['action']}")
        print(f"  通过标准: {s['pass_criteria']}")
        if s.get('critical'):
            print(f"  ⚠️ 关键: {s['critical']}")
        print()
    
    return steps

cross_platform_validation_workflow()