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
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
|
import collections
import caffe2.python.hypothesis_test_util as hu
import hypothesis.strategies as st
import numpy as np
from caffe2.python import core, dyndep, workspace
from caffe2.quantization.server import utils as dnnlowp_utils
from caffe2.quantization.server.dnnlowp_test_utils import (
check_quantized_results_close,
run_conv_or_fc
)
from hypothesis import given
dyndep.InitOpsLibrary("//caffe2/caffe2/quantization/server:dnnlowp_ops")
workspace.GlobalInit(["caffe2", "--caffe2_omp_num_threads=11"])
class DNNLowPFullyConnectedAcc16OpTest(hu.HypothesisTestCase):
# correctness test with no quantization error in inputs
# fbgemm currently only supports N a multiple of 64
@given(
input_channels=st.sampled_from([32, 64]),
output_channels=st.sampled_from([64, 128, 256]),
batch_size=st.sampled_from([0, 32, 64, 128, 256]),
in_quantized=st.booleans(),
out_quantized=st.booleans(),
**hu.gcs_cpu_only
)
def test_dnnlowp_fully_connected_acc16_int(
self,
input_channels,
output_channels,
batch_size,
in_quantized,
out_quantized,
gc,
dc,
):
# X and W have scale 1, so exactly represented after quantization
# This was made sure by having at least one 0 and one 255 for unsigned
# 8-bit tensors, and at least one -128 and one 127 for signed 8-bit
# tensors.
# Since fbgemm_acc16 accumulates to 16-bit, To avoid overflow, we use
# small numbers except for those 0, 255, -128, and 127, for this test
# We also make sure 255, -128, or 127 are not multiplied together by
# putting them in different input channels and the corresponding input
# channel in other matrix is 0.
# For example, we put 255 in input channel 1 in X, so we make the
# corresponding input channel in W all zeros.
X_min = -77
X_max = X_min + 255
X = np.round(np.random.rand(batch_size, input_channels) * 4 + X_min)
X = X.astype(np.float32)
X[:, 0] = X_min
if batch_size != 0:
X[0, 1] = X_max
W_min = -100
W_max = W_min + 255
W = np.round(
np.random.rand(output_channels, input_channels) * 4 - 2 + W_min + 128
)
W = W.astype(np.float32)
W[0, 0] = W_min
W[1, 0] = W_max
W[:, 1] = W_min + 128
# No input quantization error in bias
b = np.round(np.random.randn(output_channels)).astype(np.float32)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("FC", ""),
("FC", "DNNLOWP_ACC16"),
("Int8FC", "DNNLOWP_ACC16"),
]
for op_type, engine in op_engine_list:
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
if do_quantize:
quantize = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine="DNNLOWP", device_option=gc
)
net.Proto().op.extend([quantize])
fc = core.CreateOperator(
op_type,
["X_q" if do_quantize else "X", "W", "b"],
["Y_q" if do_dequantize else "Y"],
dequantize_output=(0 if do_dequantize else 1),
engine=engine,
device_option=gc,
)
net.Proto().op.extend([fc])
if do_dequantize:
dequantize = core.CreateOperator(
"Dequantize", ["Y_q"], ["Y"], engine="DNNLOWP", device_option=gc
)
net.Proto().op.extend([dequantize])
run_conv_or_fc(
self, None, net, X, W, b, op_type, engine, None, gc, outputs
)
check_quantized_results_close(outputs)
@given(
input_channels=st.sampled_from([2]),
output_channels=st.sampled_from([4]),
batch_size=st.sampled_from([0, 1]),
nbits_in_non_outlier=st.sampled_from([0, 6]),
in_quantized=st.booleans(),
out_quantized=st.booleans(),
prepack_weight=st.booleans(),
**hu.gcs_cpu_only
)
def test_dnnlowp_fully_connected_acc16_outlier(
self,
input_channels,
output_channels,
batch_size,
nbits_in_non_outlier,
in_quantized,
out_quantized,
prepack_weight,
gc,
dc,
):
# X and W have scale 1, so exactly represented after quantization
# This was made sure by having at least one 0 and one 255 for unsigned
# 8-bit tensors, and at least one -128 and one 127 for signed 8-bit
# tensors.
# Since fbgemm_acc16 accumulates to 16-bit, To avoid overflow, we use
# small numbers except for those 0, 255, -128, and 127, for this test
# We also make sure 255, -128, or 127 are not multiplied together by
# putting them in different input channels and the corresponding input
# channel in other matrix is 0.
# For example, we put 255 in input channel 1 in X, so we make the
# corresponding input channel in W all zeros.
X_min = -77
X_max = X_min + 255
X = np.round(np.random.rand(batch_size, input_channels) * 4 + X_min)
X = X.astype(np.float32)
X[:, 0] = X_min
if batch_size != 0:
X[0, 1] = X_max
W_min = -100
W_max = W_min + 255
W = np.round(
np.random.rand(output_channels, input_channels) * 4 - 2 + W_min + 128
)
W = W.astype(np.float32)
W[0, 0] = W_min
W[1, 0] = W_max
W[:, 1] = W_min + 128
# No input quantization error in bias
b = np.round(np.random.randn(output_channels)).astype(np.float32)
Output = collections.namedtuple("Output", ["Y", "op_type", "engine"])
outputs = []
op_engine_list = [
("FC", ""),
("FC", "DNNLOWP_ACC16"),
("Int8FC", "DNNLOWP_ACC16"),
]
for op_type, engine in op_engine_list:
init_net = core.Net("test_init_net")
net = core.Net("test_net")
do_quantize = "DNNLOWP" in engine and in_quantized
do_dequantize = "DNNLOWP" in engine and out_quantized
do_prepack_weight = engine == "DNNLOWP" and prepack_weight
if do_quantize:
quantize = core.CreateOperator(
"Quantize", ["X"], ["X_q"], engine="DNNLOWP", device_option=gc
)
net.Proto().op.extend([quantize])
X_min = 0 if X.size == 0 else X.min()
X_max = 0 if X.size == 0 else X.max()
x_q_param = dnnlowp_utils.choose_quantization_params(X_min, X_max)
if do_prepack_weight:
inputs = ["W"]
if do_dequantize:
inputs += ["b"]
pack = core.CreateOperator(
"Int8FCPackWeight",
inputs,
["W_packed"],
in_scale=x_q_param.scale,
engine=engine,
)
init_net.Proto().op.extend([pack])
fc = core.CreateOperator(
op_type,
[
"X_q" if do_quantize else "X",
"W_packed" if do_prepack_weight else "W",
"b",
],
["Y_q" if do_dequantize else "Y"],
dequantize_output=(0 if do_dequantize else 1),
engine=engine,
nbits_in_non_outlier=nbits_in_non_outlier,
device_option=gc,
)
net.Proto().op.extend([fc])
if do_dequantize:
dequantize = core.CreateOperator(
"Dequantize", ["Y_q"], ["Y"], engine="DNNLOWP", device_option=gc
)
net.Proto().op.extend([dequantize])
run_conv_or_fc(
self, init_net, net, X, W, b, op_type, engine, None, gc, outputs
)
check_quantized_results_close(outputs)
|
