1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
|
import json
class LayerAllocation:
"""
分析模型各层在 NPU/CPU 上的分配
典型情况：
- 标准卷积 → NPU（高效）
- 特殊算子 → CPU 回退（慢）
- 未知算子 → CPU 回退（极慢）
"""
HEXAGON_SUPPORTED = [
'Conv2d', 'DepthwiseConv2d', 'Conv2dTranspose',
'BatchNorm2d', 'Relu', 'Relu6', 'Sigmoid',
'MaxPool2d', 'AvgPool2d', 'AdaptiveAvgPool2d',
'Concat', 'Add', 'Mul', 'Flatten',
'FullyConnected', 'Reshape',
'Sigmoid', 'Swish', 'HardSwish',
'ResizeNearest', 'ResizeBilinear',
]
HEXAGON_UNSUPPORTED = [
'TransformerAttention',
'LSTM', 'GRU',
'Einsum',
'ScatterAdd',
'CustomOp',
]
def analyze_model(self, model_graph: list) -> dict:
"""
分析模型层分配
Args:
model_graph: 模型层列表
Returns:
allocation: NPU/CPU 分配结果
"""
npu_layers = 0
cpu_layers = 0
cpu_fallback = []
for layer in model_graph:
op_type = layer['type']
if op_type in self.HEXAGON_SUPPORTED:
npu_layers += 1
else:
cpu_layers += 1
cpu_fallback.append({
'name': layer['name'],
'type': op_type,
'params': layer.get('params', {}),
'estimated_ms': self._estimate_cpu_time(layer)
})
total = npu_layers + cpu_layers
npu_ratio = npu_layers / total if total > 0 else 0
npu_time = npu_layers * 0.1
cpu_time = sum(l['estimated_ms'] for l in cpu_fallback)
return {
'npu_layers': npu_layers,
'cpu_layers': cpu_layers,
'npu_ratio': f"{npu_ratio:.1%}",
'estimated_npu_ms': npu_time,
'estimated_cpu_ms': cpu_time,
'estimated_total_ms': npu_time + cpu_time,
'cpu_fallbacks': cpu_fallback
}
def _estimate_cpu_time(self, layer):
"""估算 CPU 执行时间"""
base = {'LSTM': 5.0, 'GRU': 4.0, 'TransformerAttention': 8.0}
return base.get(layer['type'], 2.0)
if __name__ == "__main__":
analyzer = LayerAllocation()
model_graph = [
{'name': 'conv1', 'type': 'Conv2d'},
{'name': 'bn1', 'type': 'BatchNorm2d'},
{'name': 'relu1', 'type': 'Relu'},
{'name': 'conv2', 'type': 'Conv2d'},
{'name': 'lstm', 'type': 'LSTM'},
{'name': 'attn', 'type': 'TransformerAttention'},
{'name': 'fc', 'type': 'FullyConnected'},
]
result = analyzer.analyze_model(model_graph)
print(f"NPU 层数: {result['npu_layers']}/{result['npu_layers']+result['cpu_layers']}")
print(f"NPU 占比: {result['npu_ratio']}")
print(f"估算总延迟: {result['estimated_total_ms']:.1f}ms")
print(f"CPU 回退层: {[l['name'] for l in result['cpu_fallbacks']]}")
|