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
|
#include "caffe2/core/common.h"
#include "caffe2/opt/converter.h"
#include "caffe2/opt/device.h"
#include <gtest/gtest.h>
using namespace nom::repr;
#define ADD_ARG(_op, _name, _type, _val) \
{ \
caffe2::Argument* arg = _op->add_arg(); \
arg->set_name(_name); \
arg->set_##_type(_val); \
}
TEST(DeviceTest, InsertCopies) {
caffe2::NetDef net;
for (auto i = 0; i < 9; ++i) {
if (i % 3 == 0) {
caffe2::OperatorDef* def = net.add_op();
def->set_type("Conv");
def->add_input("X");
def->add_input("W" + c10::to_string(i));
def->add_input("b" + c10::to_string(i));
ADD_ARG(def, "kernel", i, 3);
ADD_ARG(def, "stride", i, 1);
ADD_ARG(def, "pad", i, 0);
ADD_ARG(def, "order", s, "NCHW");
def->add_output("X");
def->mutable_device_option()->set_device_type(caffe2::PROTO_CPU);
} else {
caffe2::OperatorDef* def = net.add_op();
def->set_type("Relu");
def->add_input("X");
def->add_output("X");
def->mutable_device_option()->set_device_type(caffe2::PROTO_CPU);
}
}
auto nn = caffe2::convertToNNModule(net);
for (auto node : nn.dataFlow.getMutableNodes()) {
if (nn::is<Relu>(node)) {
auto annot = nn::get<NeuralNetOperator>(node)->getMutableAnnotation();
auto c2_annot = dyn_cast<caffe2::Caffe2Annotation>(annot);
c2_annot->setDeviceType(caffe2::PROTO_OPENCL);
}
}
caffe2::opt::insertCopies(
&nn,
[](NNGraph::NodeRef node) {
// Ignore all tensors
if (!nn::is<NeuralNetOperator>(node)) {
return true;
}
auto annot = nn::get<NeuralNetOperator>(node)->getMutableAnnotation();
NOM_REQUIRE_OR_RET_FALSE(annot);
auto c2_annot = dyn_cast<caffe2::Caffe2Annotation>(annot);
NOM_REQUIRE_OR_RET_FALSE(c2_annot);
return c2_annot->getDeviceType() == caffe2::PROTO_OPENCL;
},
[](NNGraph& g) {
return g.createNode(std::make_unique<GenericOperator>());
},
[](NNGraph& g) {
return g.createNode(std::make_unique<GenericOperator>());
});
auto proto = caffe2::convertToCaffe2Proto(nn, net);
// Conv -> Relu -> Relu
// becomes
// Conv -> Generic -> Relu -> Relu -> Generic
// thus
// 9 ops of this pattern becomes 15
EXPECT_EQ(proto.op().size(), 15);
}
|
