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
|
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.dnnlowp_test_utils import check_quantized_results_close
from hypothesis import given, settings
dyndep.InitOpsLibrary("//caffe2/caffe2/quantization/server:dnnlowp_ops")
workspace.GlobalInit(["caffe2", "--caffe2_omp_num_threads=11"])
class DNNLowPMulOpTest(hu.HypothesisTestCase):
@given(
N=st.integers(32, 256),
is_empty=st.booleans(),
in_quantized=st.booleans(),
out_quantized=st.booleans(),
in_place=st.sampled_from([(False, False), (True, False), (False, True)]),
**hu.gcs_cpu_only
)
@settings(deadline=None)
def test_dnnlowp_elementwise_mul_int(
self, N, is_empty, in_quantized, out_quantized, in_place, gc, dc
):
if is_empty:
N = 0
# FIXME: DNNLOWP Mul doesn't support inplace operation and
# dequantize_output=1 at the same time
if in_place[0] or in_place[1]:
in_quantized = True
out_quantized = True
# All inputs have scale 1, so exactly represented after quantization
min_ = -100
max_ = min_ + 255
A = np.round(np.random.rand(N) * (max_ - min_) + min_)
A = A.astype(np.float32)
if N != 0:
A[0] = min_
A[1] = max_
B = np.round(np.random.rand(N) * 255 - 128).astype(np.float32)
if N != 0:
B[0] = -128
B[1] = 127
Output = collections.namedtuple("Output", ["Y", "engine"])
outputs = []
engine_list = ["", "DNNLOWP"]
for engine in 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_A = core.CreateOperator(
"Quantize", ["A"], ["A_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize_A])
quantize_B = core.CreateOperator(
"Quantize", ["B"], ["B_q"], engine=engine, device_option=gc
)
net.Proto().op.extend([quantize_B])
out = "Y"
if in_place[0]:
out = "A"
elif in_place[1]:
out = "B"
mul = core.CreateOperator(
"Mul",
["A_q", "B_q"] if do_quantize else ["A", "B"],
[(out + "_q") if do_dequantize else out],
dequantize_output=not do_dequantize,
engine=engine,
device_option=gc,
)
net.Proto().op.extend([mul])
if do_dequantize:
dequantize = core.CreateOperator(
"Dequantize", [out + "_q"], [out], engine=engine, device_option=gc
)
net.Proto().op.extend([dequantize])
self.ws.create_blob("A").feed(A, device_option=gc)
self.ws.create_blob("B").feed(B, device_option=gc)
self.ws.run(net)
outputs.append(Output(Y=self.ws.blobs[out].fetch(), engine=engine))
check_quantized_results_close(outputs)
@given(**hu.gcs_cpu_only)
@settings(deadline=None)
def test_dnnlowp_elementwise_mul_broadcast(self, gc, dc):
# Set broadcast and no axis, i.e. broadcasting last dimensions.
min_ = -100
max_ = min_ + 255
A = np.round(np.random.rand(2, 3, 4, 5) * (max_ - min_) + min_)
A = A.astype(np.float32)
A[0, 0, 0, 0] = min_
A[0, 0, 0, 1] = max_
B = np.round(np.random.rand(4, 5) * 255 - 128).astype(np.float32)
B[0, 0] = -128
B[0, 1] = 127
Output = collections.namedtuple("Output", ["Y", "engine"])
outputs = []
engine_list = ["", "DNNLOWP"]
for engine in engine_list:
net = core.Net("test_net")
mul = core.CreateOperator(
"Mul",
["A", "B"],
["Y"],
engine=engine,
device_option=gc,
broadcast=1,
dequantize_output=1,
)
net.Proto().op.extend([mul])
self.ws.create_blob("A").feed(A, device_option=gc)
self.ws.create_blob("B").feed(B, device_option=gc)
self.ws.run(net)
outputs.append(Output(Y=self.ws.blobs["Y"].fetch(), engine=engine))
check_quantized_results_close(outputs)
@given(**hu.gcs_cpu_only)
@settings(deadline=None)
def test_dnnlowp_elementwise_mul_broadcast_axis(self, gc, dc):
for bdim, axis in [
((3, 4), 1), # broadcasting intermediate dimensions
((2,), 0), # broadcasting the first dimension
((1, 4, 1), 1),
]:
# broadcasting with single elem dimensions at both ends
min_ = -100
max_ = min_ + 255
A = np.round(np.random.rand(2, 3, 4, 5) * (max_ - min_) + min_)
A = A.astype(np.float32)
B = np.round(np.random.rand(*bdim) * 255 - 128).astype(np.float32)
A.flat[0] = min_
A.flat[1] = max_
B.flat[0] = -128
B.flat[1] = 127
Output = collections.namedtuple("Output", ["Y", "engine"])
outputs = []
engine_list = ["", "DNNLOWP"]
for engine in engine_list:
net = core.Net("test_net")
mul = core.CreateOperator(
"Mul",
["A", "B"],
["Y"],
engine=engine,
device_option=gc,
broadcast=1,
axis=axis,
dequantize_output=1,
)
net.Proto().op.extend([mul])
self.ws.create_blob("A").feed(A, device_option=gc)
self.ws.create_blob("B").feed(B, device_option=gc)
self.ws.run(net)
outputs.append(Output(Y=self.ws.blobs["Y"].fetch(), engine=engine))
check_quantized_results_close(outputs)
|
