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
|
/* 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.
==============================================================================*/
#include "tensorflow/core/grappler/costs/graph_memory.h"
#include "tensorflow/cc/ops/standard_ops.h"
#include "tensorflow/core/grappler/grappler_item.h"
#include "tensorflow/core/grappler/inputs/trivial_test_graph_input_yielder.h"
#include "tensorflow/core/platform/test.h"
namespace tensorflow {
namespace grappler {
namespace {
class GraphMemoryTest : public ::testing::Test {
protected:
std::unordered_map<string, DeviceProperties> devices_;
public:
GraphMemoryTest() {
devices_["/CPU:0"].set_type("CPU");
devices_["/CPU:0"].set_num_cores(1);
devices_["/CPU:0"].set_frequency(1);
devices_["/CPU:0"].set_bandwidth(1);
devices_["/GPU:0"].set_type("GPU");
devices_["/GPU:0"].set_num_cores(1);
devices_["/GPU:0"].set_frequency(1);
devices_["/CPU:0"].set_bandwidth(1);
(*devices_["/GPU:0"].mutable_environment())["architecture"] = "3";
}
};
TEST_F(GraphMemoryTest, Basic) {
TrivialTestGraphInputYielder fake_input(4, 1, 10, false, {"/CPU:0"});
GrapplerItem item;
CHECK(fake_input.NextItem(&item));
item.feed.clear();
GraphMemory memory(item);
Status s = memory.InferStatically(devices_);
TF_CHECK_OK(s);
const GraphMemory::MemoryUsage& mem_usage =
memory.GetPeakMemoryUsage("/CPU:0");
EXPECT_EQ(120, mem_usage.used_memory);
std::set<string> tensors;
for (const auto& t : mem_usage.live_tensors) {
tensors.insert(strings::StrCat(t.node, ":", t.output_id));
}
// When the execution of the 'Square' node completes, TF can start executing
// 'Square_1' and release the memory used by 'x'. Since we can't be sure of
// the order in which this takes place, in the worst case the 3 tensors are in
// memory.
std::set<string> expected;
expected.insert("Square:0");
expected.insert("Square_1:0");
expected.insert("x:0");
EXPECT_EQ(expected, tensors);
}
TEST_F(GraphMemoryTest, UnknownBatchSize) {
TrivialTestGraphInputYielder fake_input(4, 1, -1, false, {"/CPU:0"});
GrapplerItem item;
CHECK(fake_input.NextItem(&item));
item.feed.clear();
GraphMemory memory(item);
Status s = memory.InferStatically(devices_);
TF_CHECK_OK(s);
// Same maths as before, except that batch size is unknown and therefore
// assumed to be one.
const GraphMemory::MemoryUsage& mem_usage =
memory.GetPeakMemoryUsage("/CPU:0");
EXPECT_EQ(16, mem_usage.used_memory);
std::set<string> tensors;
for (const auto& t : mem_usage.live_tensors) {
tensors.insert(strings::StrCat(t.node, ":", t.output_id));
}
std::set<string> expected;
expected.insert("Const/Const:0");
expected.insert("Square:0");
expected.insert("x:0");
EXPECT_EQ(expected, tensors);
}
TEST_F(GraphMemoryTest, MultiDevice) {
TrivialTestGraphInputYielder fake_input(4, 2, 1024 * 1024, false,
{"/CPU:0", "/GPU:0"});
GrapplerItem item;
CHECK(fake_input.NextItem(&item));
item.feed.clear();
GraphMemory memory(item);
Status s = memory.InferStatically(devices_);
TF_CHECK_OK(s);
const GraphMemory::MemoryUsage& cpu_mem = memory.GetPeakMemoryUsage("/CPU:0");
EXPECT_EQ(16777216, cpu_mem.used_memory);
std::set<string> cpu_tensors;
for (const auto& t : cpu_mem.live_tensors) {
cpu_tensors.insert(strings::StrCat(t.node, ":", t.output_id));
}
std::set<string> cpu_expected;
cpu_expected.insert("Recv_Square_1_0_on_/CPU_0:0");
cpu_expected.insert("Square:0");
cpu_expected.insert("x:0");
cpu_expected.insert("AddN:0");
EXPECT_EQ(cpu_expected, cpu_tensors);
const GraphMemory::MemoryUsage& gpu_mem = memory.GetPeakMemoryUsage("/GPU:0");
EXPECT_EQ(16777216, gpu_mem.used_memory);
std::set<string> gpu_tensors;
for (const auto& t : gpu_mem.live_tensors) {
gpu_tensors.insert(strings::StrCat(t.node, ":", t.output_id));
}
std::set<string> gpu_expected;
gpu_expected.insert("Recv_AddN_0_on_/GPU_0:0");
gpu_expected.insert("Square_1:0");
gpu_expected.insert("AddN_1:0");
gpu_expected.insert("AddN_3:0");
EXPECT_EQ(gpu_expected, gpu_tensors);
}
TEST_F(GraphMemoryTest, GpuSwapping) {
TrivialTestGraphInputYielder fake_input(4, 2, 1024 * 1024, false, {"/GPU:0"});
GrapplerItem item;
CHECK(fake_input.NextItem(&item));
item.feed.clear();
{
// Estimate the max memory usage for the graph.
GraphMemory memory(item);
Status s = memory.InferStatically(devices_);
TF_CHECK_OK(s);
const GraphMemory::MemoryUsage& gpu_mem =
memory.GetPeakMemoryUsage("/GPU:0");
EXPECT_EQ(20971520, gpu_mem.used_memory);
std::set<string> gpu_tensors;
for (const auto& t : gpu_mem.live_tensors) {
gpu_tensors.insert(strings::StrCat(t.node, ":", t.output_id));
}
std::set<string> gpu_expected;
gpu_expected.insert("Square:0");
gpu_expected.insert("Square_1:0");
gpu_expected.insert("AddN:0");
gpu_expected.insert("AddN_1:0");
gpu_expected.insert("AddN_2:0");
EXPECT_EQ(gpu_expected, gpu_tensors);
}
{
// Swap the first input to node AddN_1: its fanin (the square nodes) should
// not appear in the max cut anymore.
for (auto& node : *item.graph.mutable_node()) {
if (node.name() == "AddN_1") {
(*node.mutable_attr())["_swap_to_host"].mutable_list()->add_i(0);
}
}
GraphMemory memory(item);
Status s = memory.InferStatically(devices_);
TF_CHECK_OK(s);
const GraphMemory::MemoryUsage& new_gpu_mem =
memory.GetPeakMemoryUsage("/GPU:0");
EXPECT_EQ(20971520, new_gpu_mem.used_memory);
std::set<string> new_gpu_tensors;
for (const auto& t : new_gpu_mem.live_tensors) {
new_gpu_tensors.insert(strings::StrCat(t.node, ":", t.output_id));
}
std::set<string> new_gpu_expected;
new_gpu_expected.insert("AddN:0");
new_gpu_expected.insert("AddN_1:0");
new_gpu_expected.insert("AddN_2:0");
new_gpu_expected.insert("AddN_3:0");
new_gpu_expected.insert("AddN_4:0");
EXPECT_EQ(new_gpu_expected, new_gpu_tensors);
}
}
TEST_F(GraphMemoryTest, CtrlDependencies) {
// Build a simple graph with a control dependency.
Scope s = Scope::NewRootScope();
Output a = ops::Const(s.WithOpName("a").WithDevice("/CPU:0"), 10.0f, {3});
Output v =
ops::Variable(s.WithOpName("v").WithDevice("/CPU:0"), {3}, DT_FLOAT);
Output assign =
ops::Assign(s.WithOpName("assign").WithDevice("/CPU:0"), v, a);
ops::NoOp init(
s.WithOpName("init").WithDevice("/CPU:0").WithControlDependencies(
assign));
GrapplerItem item;
item.fetch.push_back("init");
TF_CHECK_OK(s.ToGraphDef(&item.graph));
GraphMemory memory(item);
Status status = memory.InferStatically(devices_);
TF_CHECK_OK(status);
const GraphMemory::MemoryUsage& mem = memory.GetPeakMemoryUsage("/CPU:0");
EXPECT_EQ(36, mem.used_memory);
std::set<string> tensors;
for (const auto& t : mem.live_tensors) {
tensors.insert(strings::StrCat(t.node, ":", t.output_id));
}
std::set<string> expected;
expected.insert("a:0");
expected.insert("v:0");
expected.insert("assign:0");
EXPECT_EQ(expected, tensors);
}
} // namespace
} // namespace grappler
} // namespace tensorflow
|
