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
|
/* 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/optimizers/static_schedule.h"
#include <deque>
#include "tensorflow/core/framework/attr_value.pb.h"
#include "tensorflow/core/grappler/costs/graph_properties.h"
#include "tensorflow/core/grappler/costs/op_level_cost_estimator.h"
#include "tensorflow/core/grappler/costs/virtual_placer.h"
#include "tensorflow/core/grappler/op_types.h"
#include "tensorflow/core/grappler/utils.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/strings/strcat.h"
namespace tensorflow {
namespace grappler {
static Costs::NanoSeconds PredictExecutionTime(
const GraphProperties& properties, const OpLevelCostEstimator& estimator,
const VirtualPlacer& placer, const NodeDef& node) {
OpContext op_context;
op_context.op_info.set_op(node.op());
*op_context.op_info.mutable_attr() = node.attr();
std::vector<OpInfo::TensorProperties> inputs =
properties.GetInputProperties(node.name());
for (auto& input : inputs) {
op_context.op_info.add_inputs()->Swap(&input);
}
std::vector<OpInfo::TensorProperties> outputs =
properties.GetOutputProperties(node.name());
for (auto& output : outputs) {
op_context.op_info.add_outputs()->Swap(&output);
}
DeviceProperties device = placer.get_device(node);
op_context.op_info.mutable_device()->Swap(&device);
Costs::NanoSeconds estimate =
estimator.PredictCosts(op_context).execution_time;
// Make sure our estimates are at least one nanosecond per node.
return std::max(estimate, Costs::NanoSeconds(1));
}
Status EstimateEarliestExecutionTimes(
const GrapplerItem& item, const Cluster* cluster,
std::unordered_map<const NodeDef*, Costs::NanoSeconds>* completion_times) {
std::unordered_map<string, const NodeDef*> name_map;
std::unordered_map<const NodeDef*, int> pending_inputs;
std::deque<const NodeDef*> ready_nodes;
for (const NodeDef& node : item.graph.node()) {
name_map[node.name()] = &node;
if (node.input_size() == 0) {
ready_nodes.push_back(&node);
(*completion_times)[&node] = 0;
} else if (IsMerge(node)) {
// Merge nodes are processed as soon as one of the input becomes
// available.
pending_inputs[&node] = 1;
} else {
pending_inputs[&node] = node.input_size();
}
}
std::unordered_map<const NodeDef*, std::vector<const NodeDef*>> fanouts;
for (const NodeDef& node : item.graph.node()) {
for (const string& input : node.input()) {
string node_name = NodeName(input);
auto it = name_map.find(node_name);
if (it == name_map.end()) {
return errors::InvalidArgument(
strings::StrCat("Unknown input node ", input));
}
const NodeDef* fanin = it->second;
fanouts[fanin].push_back(&node);
}
}
name_map.clear();
GraphProperties properties(item);
TF_RETURN_IF_ERROR(properties.InferStatically(true));
OpLevelCostEstimator estimator;
VirtualPlacer placer(cluster);
while (!ready_nodes.empty()) {
const NodeDef* node = ready_nodes.front();
ready_nodes.pop_front();
Costs::NanoSeconds execution_time =
PredictExecutionTime(properties, estimator, placer, *node);
Costs::NanoSeconds completion_time =
execution_time + (*completion_times)[node];
(*completion_times)[node] = completion_time;
for (const NodeDef* fanout : fanouts[node]) {
int pending = pending_inputs[fanout];
if (pending == 0) {
// Already processed. Avoid going through loops more than once.
continue;
} else if (pending == 1) {
ready_nodes.push_back(fanout);
}
pending_inputs[fanout]--;
Costs::NanoSeconds ready_time =
std::max(completion_time, (*completion_times)[fanout]);
(*completion_times)[fanout] = ready_time;
}
}
return Status::OK();
}
Status EstimateRequiredTimes(
const GrapplerItem& item, const Cluster* cluster,
const std::unordered_map<const NodeDef*, Costs::NanoSeconds>&
execution_times,
std::unordered_map<const NodeDef*, Costs::NanoSeconds>* required_times) {
std::unordered_map<string, const NodeDef*> name_map;
for (const NodeDef& node : item.graph.node()) {
name_map[node.name()] = &node;
(*required_times)[&node] = Costs::NanoSeconds::max();
}
std::unordered_map<const NodeDef*, int> pending_fanouts;
for (const NodeDef& node : item.graph.node()) {
for (const string& input : node.input()) {
string node_name = NodeName(input);
auto it = name_map.find(node_name);
if (it == name_map.end()) {
return errors::InvalidArgument(
strings::StrCat("Unknown input node ", input));
}
const NodeDef* fanin = it->second;
pending_fanouts[fanin] += 1;
}
}
std::deque<const NodeDef*> ready_nodes;
for (const NodeDef& node : item.graph.node()) {
if (pending_fanouts[&node] == 0) {
auto it = execution_times.find(&node);
if (it != execution_times.end()) {
(*required_times)[&node] = it->second;
}
ready_nodes.push_back(&node);
}
}
GraphProperties properties(item);
TF_RETURN_IF_ERROR(properties.InferStatically(true));
OpLevelCostEstimator estimator;
VirtualPlacer placer(cluster);
while (!ready_nodes.empty()) {
const NodeDef* node = ready_nodes.front();
ready_nodes.pop_front();
Costs::NanoSeconds execution_time =
PredictExecutionTime(properties, estimator, placer, *node);
Costs::NanoSeconds required_time = (*required_times)[node] - execution_time;
for (const string& fanin_name : node->input()) {
const NodeDef* fanin = name_map[NodeName(fanin_name)];
(*required_times)[fanin] =
std::min((*required_times)[fanin], required_time);
int pending = pending_fanouts[fanin];
if (pending == 0) {
// Already processed. Avoid going through loops more than once.
continue;
} else if (pending == 1) {
ready_nodes.push_back(fanin);
}
pending_fanouts[fanin]--;
}
}
return Status::OK();
}
} // end namespace grappler
} // end namespace tensorflow
|
