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
|
/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/
#ifdef INTEL_MKL
#include "tensorflow/core/graph/mkl_layout_pass.h"
#include "tensorflow/core/util/mkl_util.h"
#include <vector>
#include "tensorflow/core/framework/op.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/graph/graph.h"
#include "tensorflow/core/graph/graph_constructor.h"
#include "tensorflow/core/graph/testlib.h"
#include "tensorflow/core/kernels/ops_util.h"
#include "tensorflow/core/lib/random/simple_philox.h"
#include "tensorflow/core/lib/strings/str_util.h"
#include "tensorflow/core/lib/strings/stringprintf.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/protobuf.h"
#include "tensorflow/core/platform/test.h"
#include "tensorflow/core/platform/test_benchmark.h"
namespace tensorflow {
namespace {
static void InitGraph(const string& s, Graph* graph) {
GraphDef graph_def;
auto parser = protobuf::TextFormat::Parser();
// parser.AllowRelaxedWhitespace(true);
CHECK(parser.MergeFromString(s, &graph_def)) << s;
GraphConstructorOptions opts;
TF_CHECK_OK(ConvertGraphDefToGraph(opts, graph_def, graph));
}
class MklLayoutPassTest : public ::testing::Test {
public:
MklLayoutPassTest() : graph_(OpRegistry::Global()) {}
void InitGraph(const string& s) {
::tensorflow::InitGraph(s, &graph_);
original_ = CanonicalGraphString(&graph_);
}
static bool IncludeNode(const Node* n) { return n->IsOp(); }
static string EdgeId(const Node* n, int index) {
if (index == 0) {
return n->name();
} else if (index == Graph::kControlSlot) {
return strings::StrCat(n->name(), ":control");
} else {
return strings::StrCat(n->name(), ":", index);
}
}
string CanonicalGraphString(Graph* g) {
std::vector<string> nodes;
std::vector<string> edges;
for (const Node* n : g->nodes()) {
if (IncludeNode(n)) {
nodes.push_back(strings::StrCat(n->name(), "(", n->type_string(), ")"));
}
}
for (const Edge* e : g->edges()) {
if (IncludeNode(e->src()) && IncludeNode(e->dst())) {
edges.push_back(strings::StrCat(EdgeId(e->src(), e->src_output()), "->",
EdgeId(e->dst(), e->dst_input())));
}
}
// Canonicalize
std::sort(nodes.begin(), nodes.end());
std::sort(edges.begin(), edges.end());
return strings::StrCat(str_util::Join(nodes, ";"), "|",
str_util::Join(edges, ";"));
}
string DoMklLayoutOptimizationPass() {
string before = CanonicalGraphString(&graph_);
LOG(ERROR) << "Before MKL layout rewrite pass: " << before;
std::unique_ptr<Graph>* ug = new std::unique_ptr<Graph>(&graph_);
RunMklLayoutRewritePass(ug);
string result = CanonicalGraphString(&graph_);
LOG(ERROR) << "After MKL layout rewrite pass: " << result;
return result;
}
const string& OriginalGraph() const { return original_; }
Graph graph_;
string original_;
};
REGISTER_OP("Input").Output("o: float").SetIsStateful();
// Single Conv2D Op; No Mkl layer on the input and on the output.
// We will generate dummy Mkl tensor as 2nd input of Conv2D.
TEST_F(MklLayoutPassTest, Conv2D_Basic) {
InitGraph(
"node { name: 'A' op: 'Input'}"
"node { name: 'B' op: 'Input'}"
"node { name: 'C' op: 'Conv2D'"
" attr { key: 'T' value { type: DT_FLOAT } }"
" attr { key: 'data_format' value { s: 'NCHW' } }"
" attr { key: 'use_cudnn_on_gpu' value { b: false } }"
" attr { key: 'strides' value { list: {i: 1, i:1, i:1, i:1} } }"
" attr { key: 'padding' value { s: 'SAME' } }"
" input: ['A', 'B']}"
"node { name: 'D' op: 'Mul' attr { key: 'T' value { type: DT_FLOAT } }"
" input: ['B', 'C'] }");
EXPECT_EQ(DoMklLayoutOptimizationPass(),
"A(Input);B(Input);C(MklConv2D);D(Mul);DMT/_0(Const);DMT/_1(Const)|"
"A->C;B->C:2;B->D;C->D:1;DMT/_0->C:1;DMT/_1->C:3");
}
// 2 Conv2D Ops in sequence. Both should get transformed and 1st Conv2D will
// have 2 outputs, both of which will be inputs to next Conv2D.
TEST_F(MklLayoutPassTest, Conv2D_Positive1) {
InitGraph(
"node { name: 'A' op: 'Input'}"
"node { name: 'B' op: 'Input'}"
"node { name: 'C' op: 'Conv2D'"
" attr { key: 'T' value { type: DT_FLOAT } }"
" attr { key: 'data_format' value { s: 'NCHW' } }"
" attr { key: 'use_cudnn_on_gpu' value { b: false } }"
" attr { key: 'strides' value { list: {i: 1, i:1, i:1, i:1} } }"
" attr { key: 'padding' value { s: 'SAME' } }"
" input: ['A', 'B']}"
"node { name: 'D' op: 'Conv2D'"
" attr { key: 'T' value { type: DT_FLOAT } }"
" attr { key: 'data_format' value { s: 'NCHW' } }"
" attr { key: 'use_cudnn_on_gpu' value { b: false } }"
" attr { key: 'strides' value { list: {i: 1, i:1, i:1, i:1} } }"
" attr { key: 'padding' value { s: 'SAME' } }"
" input: ['A', 'C']}"
"node { name: 'E' op: 'Mul' attr { key: 'T' value { type: DT_FLOAT } }"
" input: ['C', 'D'] }");
EXPECT_EQ(DoMklLayoutOptimizationPass(),
"A(Input);B(Input);C(MklConv2D);D(MklConv2D);DMT/_0(Const);"
"DMT/_1(Const);DMT/_2(Const);E(Mul)|A->C;A->D;B->C:2;C->D:2;C->E;"
"C:1->D:3;D->E:1;DMT/_0->C:1;DMT/_1->C:3;DMT/_2->D:1");
}
static void BM_MklLayoutRewritePass(int iters, int op_nodes) {
testing::StopTiming();
string s;
for (int in = 0; in < 10; in++) {
s += strings::Printf("node { name: 'in%04d' op: 'Input'}", in);
}
random::PhiloxRandom philox(301, 17);
random::SimplePhilox rnd(&philox);
for (int op = 0; op < op_nodes; op++) {
s += strings::Printf(
"node { name: 'op%04d' op: 'Mul' attr { key: 'T' value { "
"type: DT_FLOAT } } input: ['in%04d', 'in%04d' ] }",
op, rnd.Uniform(10), rnd.Uniform(10));
}
bool first = true;
while (iters > 0) {
Graph* graph = new Graph(OpRegistry::Global());
InitGraph(s, graph);
int N = graph->num_node_ids();
if (first) {
testing::SetLabel(strings::StrCat("Per graph node. Nodes: ", N));
first = false;
}
{
testing::StartTiming();
std::unique_ptr<Graph> ug(graph);
RunMklLayoutRewritePass(&ug);
testing::StopTiming();
}
iters -= N; // Our benchmark units are individual graph nodes,
// not whole graphs
// delete graph;
}
}
BENCHMARK(BM_MklLayoutRewritePass)->Arg(1000)->Arg(10000);
} // namespace
} // namespace tensorflow
#endif /* INTEL_MKL */
|
