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
|
# Owner(s): ["module: mkldnn"]
import itertools
import unittest
import torch
from torch import nn
from torch.testing._internal.common_utils import run_tests
from torch.testing._internal.jit_utils import JitTestCase
from test_tensorexpr import warmup_and_run_forward
FUSION_GROUP = 'prim::TensorExprGroup'
@unittest.skipIf(not torch._C.has_mkldnn, "MKL-DNN build is disabled")
class TestMkldnnFusion(JitTestCase):
def assertFused(self, graph, fused_patterns):
for pat in fused_patterns:
self.assertGraphContainsExactly(graph, pat, 0)
def _check_model(self, m, x, trace=False):
old_fusion_inlining = torch._C._debug_get_fusion_group_inlining()
torch._C._debug_set_fusion_group_inlining(False)
old_cpu_fuser_state = torch._C._jit_can_fuse_on_cpu()
torch._C._jit_override_can_fuse_on_cpu(True)
old_te_must_use_llvm_cpu = torch._C._jit_get_te_must_use_llvm_cpu()
torch._C._jit_set_te_must_use_llvm_cpu(False)
m.eval()
with torch.no_grad():
if trace:
script = torch.jit.trace(m, x)
else:
script = torch.jit.script(m)
script = torch.jit.freeze(script)
with torch.no_grad():
y = warmup_and_run_forward(script, x)
y = script(x)
y_ref = m(x)
graph = script.graph_for(*x)
self.assertEqual(y, y_ref)
torch._C._debug_set_fusion_group_inlining(old_fusion_inlining)
torch._C._jit_override_can_fuse_on_cpu(old_cpu_fuser_state)
torch._C._jit_set_te_must_use_llvm_cpu(old_te_must_use_llvm_cpu)
return graph
def test_single_conv(self):
class M(nn.Module):
def __init__(self, in_channels, out_channels, bias, **kwargs):
super(M, self).__init__()
self.conv = torch.nn.Conv2d(in_channels, out_channels, bias=bias, **kwargs)
def forward(self, x):
res = self.conv(x)
return res
for memory_format, enabled in [
[torch.contiguous_format, False],
[torch.channels_last, True],
]:
for trace in [True, False]:
input_size = 224
batch_size = 1
kernel_size = 3
options = itertools.product([True, False], [1, 2], [1, 4])
for bias, dilation, groups in options:
iC = 3 * groups
oC = 10 * groups
m = M(iC,
oC,
bias,
kernel_size=(kernel_size, kernel_size),
stride=2,
padding=1,
dilation=dilation,
groups=groups).to(memory_format=memory_format)
x = torch.randn(batch_size, iC, input_size, input_size).to(memory_format=memory_format)
graph = self._check_model(m, x, trace)
conv_node_name = 'aten::_convolution' if trace else 'aten::conv2d'
if enabled:
self.assertFused(graph, [conv_node_name])
self.assertGraphContainsExactly(graph, FUSION_GROUP, 1)
else:
self.assertGraphContains(graph, kind=conv_node_name)
def test_conv_eltwise(self):
class M(nn.Module):
def __init__(self, eltwise_fn, in_channels, out_channels, bias, **kwargs):
super(M, self).__init__()
self.conv = torch.nn.Conv2d(in_channels, out_channels, bias=bias, **kwargs)
self.eltwise = eltwise_fn
def forward(self, x):
x = self.conv(x)
x = self.eltwise(x)
return x
for memory_format, enabled in [
[torch.contiguous_format, False],
[torch.channels_last, True],
]:
for eltwise_fn in [torch.relu]:
for bias in [True, False]:
for oC in [1, 10]:
m = M(eltwise_fn, 3, oC, bias, kernel_size=(3, 3)).to(memory_format=memory_format)
x = torch.randn(1, 3, 224, 224).to(memory_format=memory_format)
graph = self._check_model(m, x)
if enabled:
self.assertFused(graph, ['aten::conv2d', 'aten::' + eltwise_fn.__name__])
self.assertGraphContainsExactly(graph, FUSION_GROUP, 1)
else:
self.assertGraphContains(graph, kind='aten::conv2d')
def test_unsupported_conv(self):
class M(nn.Module):
def __init__(self, m, in_channels, out_channels, bias, **kwargs):
super(M, self).__init__()
self.conv = m(in_channels, out_channels, bias=bias, **kwargs)
def forward(self, x):
res = self.conv(x)
return res
for module, dim, memory_format in [
[nn.Conv3d, 3, torch.contiguous_format],
[nn.Conv3d, 3, torch.channels_last_3d],
[nn.ConvTranspose2d, 2, torch.contiguous_format],
[nn.ConvTranspose2d, 2, torch.channels_last],
]:
trace = True
input_size = 224
batch_size = 1
kernel_size = 3
groups = 2
bias = True
iC = 3 * groups
oC = 10 * groups
dilation = 2
m = M(module,
iC,
oC,
bias,
kernel_size=kernel_size,
stride=2,
padding=1,
dilation=dilation,
groups=groups).to(memory_format=memory_format)
input_sizes = [batch_size, iC, input_size, input_size]
if dim == 3:
input_sizes.append(input_size)
x = torch.randn(input_sizes).to(memory_format=memory_format)
graph = self._check_model(m, x, trace)
self.assertGraphContains(graph, kind='aten::_convolution')
if __name__ == "__main__":
run_tests()
|
