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
|
/* Copyright 2015 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/common_runtime/graph_optimizer.h"
#include "tensorflow/core/common_runtime/constant_folding.h"
#include "tensorflow/core/common_runtime/function.h"
#include "tensorflow/core/graph/algorithm.h"
#include "tensorflow/core/graph/graph_constructor.h"
#include "tensorflow/core/graph/node_builder.h"
#include "tensorflow/core/graph/optimizer_cse.h"
namespace tensorflow {
GraphOptimizer::GraphOptimizer(const OptimizerOptions& opts) : opts_(opts) {
if (opts_.opt_level() >= OptimizerOptions::L1) {
opts_.set_do_common_subexpression_elimination(true);
opts_.set_do_constant_folding(true);
}
}
GraphOptimizer::~GraphOptimizer() {}
void GraphOptimizer::Optimize(
FunctionLibraryRuntime* runtime, Env* env, Device* device,
std::unique_ptr<Graph>* graph,
const std::unordered_map<string, std::vector<PartialTensorShape>>*
shape_map,
const std::function<bool(const Node*)>& cse_consider_fn,
const std::function<bool(const Node*)>& cf_consider_fn) {
Graph* g = graph->get();
DumpGraph("Initial", g);
bool changed = true;
const int kMaxRounds = 10;
for (int rounds = 0; rounds < kMaxRounds; ++rounds) {
changed = false;
if (RemoveListArrayConverter(g)) {
DumpGraph("RemoveListArrayConverter", g);
changed = true;
}
if (opts_.do_function_inlining() && RemoveDeadNodes(g)) {
DumpGraph("RemoveDeadNodes", g);
changed = true;
}
if (opts_.do_function_inlining() && RemoveIdentityNodes(g)) {
DumpGraph("RemoveIdentityNodes", g);
changed = true;
}
if (opts_.do_constant_folding()) {
ConstantFoldingOptions cf_opts;
cf_opts.shape_map = shape_map;
cf_opts.consider = cf_consider_fn;
if (opts_.max_folded_constant_in_bytes() > 0) {
cf_opts.max_constant_size_in_bytes =
opts_.max_folded_constant_in_bytes();
}
bool was_mutated;
ConstantFold(cf_opts, runtime, env, device, g, &was_mutated)
.IgnoreError();
if (was_mutated) {
RemoveDeadNodes(g);
DumpGraph("ConstFolding", g);
changed = true;
}
}
if (opts_.do_function_inlining() && FixupSourceAndSinkEdges(g)) {
DumpGraph("FixupSourceAndSinkEdges", g);
changed = true;
}
if (opts_.do_common_subexpression_elimination() &&
OptimizeCSE(g, cse_consider_fn)) {
DumpGraph("OptimizeCSE", g);
changed = true;
}
if (opts_.do_function_inlining() && ExpandInlineFunctions(runtime, g)) {
DumpGraph("ExpandInlineFunctions", g);
changed = true;
}
if (!changed) break;
}
// Note that we use the Graph constructor that copies the input
// FunctionLibraryDefinition, since the original lib def will go out of scope.
std::unique_ptr<Graph> copy(new Graph(g->flib_def()));
CopyGraph(*g, copy.get());
graph->swap(copy);
DumpGraph("ReCopy", graph->get());
}
} // end namespace tensorflow
|
