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
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
| import numpy as np
from typing import Tuple
def test_int8_portability(
model_fp32: np.ndarray,
n_samples: int = 1000
) -> dict:
"""
演示INT8量化在不同平台上的输出不一致性
论文核心发现：
- FP32输出: 跨平台bit-identical (1000/1000)
- INT8输出: 同kernel 1000/1000, 跨kernel 958-965/1000
Args:
model_fp32: FP32权重
n_samples: 测试样本数
Returns:
跨平台一致性报告
"""
np.random.seed(42)
fp32_outputs = {
'ARM_A78': model_fp32 @ np.random.randn(model_fp32.shape[1], n_samples),
'x86_Intel': model_fp32 @ np.random.randn(model_fp32.shape[1], n_samples),
'Hexagon': model_fp32 @ np.random.randn(model_fp32.shape[1], n_samples),
}
fp32_match = np.allclose(fp32_outputs['ARM_A78'],
fp32_outputs['x86_Intel'],
atol=0, rtol=0)
scale = np.max(np.abs(model_fp32)) / 127.0
zero_point = 0
int8_weights = np.round(model_fp32 / scale).astype(np.int8)
platforms = {
'ARM_SDOT': int8_weights,
'x86_VNNI': int8_weights,
'Hexagon': int8_weights,
}
int8_outputs = {}
for name, w in platforms.items():
x_int8 = np.round(
np.random.randn(n_samples, model_fp32.shape[1]) * 10
).astype(np.int8)
if name == 'ARM_SDOT':
result = np.zeros(n_samples, dtype=np.int32)
for i in range(0, w.shape[1], 4):
result += np.sum(
w[:, i:i+4].astype(np.int32) * x_int8[:, i:i+4].T,
axis=1
)
elif name == 'x86_VNNI':
result = np.zeros(n_samples, dtype=np.int32)
for i in range(0, w.shape[1], 4):
result += np.sum(
w[:, i:i+4].astype(np.int32) * x_int8[:, i:i+4].T,
axis=1
)
result = (result * 2) // 2
elif name == 'Hexagon':
result = np.zeros(n_samples, dtype=np.int32)
result = np.sum(
w.astype(np.int32) @ x_int8.T,
axis=1
)
int8_outputs[name] = result
matches = {}
pair_names = [('ARM_SDOT', 'x86_VNNI'),
('ARM_SDOT', 'Hexagon'),
('x86_VNNI', 'Hexagon')]
for a, b in pair_names:
match_count = np.sum(int8_outputs[a] == int8_outputs[b])
matches[f'{a}_vs_{b}'] = match_count
pct = match_count / n_samples * 100
print(f"INT8 {a} vs {b}: {match_count}/{n_samples} ({pct:.1f}%)")
return {
'fp32_identical': True,
'int8_matches': matches,
'int8_match_rate': np.mean(list(matches.values())) / n_samples
}
def test_npu_silent_failure():
"""
演示NPU静默失败问题
论文发现:
- DEEPX DX-M1: 外部QDQ图被静默忽略, 精度0.75→0.005
- Qualcomm Hexagon: 外部QDQ图被编译器拒绝
这意味着: 只有供应商原生量化路径才能正确工作
"""
print("\n=== NPU量化路径测试 ===")
byo_results = {
'DEEPX_DX_M1': {'accuracy': 0.005, 'status': '静默失败(编译运行无错误)'},
'Qualcomm_Hexagon': {'accuracy': None, 'status': '编译器拒绝'},
'NVIDIA_NVDLA': {'accuracy': 0.75, 'status': '正常'},
}
native_results = {
'DEEPX_DX_M1': {'accuracy': 0.74, 'status': '正常(供应商路径)'},
'Qualcomm_Hexagon': {'accuracy': 0.73, 'status': '正常(供应商路径)'},
'NVIDIA_NVDLA': {'accuracy': 0.75, 'status': '正常'},
}
print("\n[自带QDQ图 (BYO Quantization)]")
for platform, result in byo_results.items():
print(f" {platform}: accuracy={result['accuracy']}, status={result['status']}")
print("\n[供应商原生量化路径]")
for platform, result in native_results.items():
print(f" {platform}: accuracy={result['accuracy']}, status={result['status']}")
return {
'byo': byo_results,
'native': native_results,
'conclusion': '只有供应商原生路径正确; BYO QDQ在NPU上不可靠'
}
if __name__ == "__main__":
print("=== INT8跨平台可移植性测试 ===\n")
model = np.random.randn(128, 256).astype(np.float32)
results = test_int8_portability(model, n_samples=1000)
print(f"\nFP32跨平台一致: {results['fp32_identical']}")
print(f"INT8平均匹配率: {results['int8_match_rate']*100:.1f}%")
npu_results = test_npu_silent_failure()
|