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
|
// This module implements a common subexpression elimination pass. We
// process the nodes in the graph in reverse postorder
// (i.e. inputs before their downstream dependencies). The rough algorithm is
// as follows:
//
// std::unordered_map<size_t, Node*> available
// for each node n in forward topological order:
// h = NodeHash(n)
// if available[h] exists and Equivalent(available(h), h)
// redirect downstream uses of outputs of n to available[h]
// remove n from graph
// else
// if available[h] does not exist
// available[h] = n
//
// This is similar to the global value number algorithm describe in this
// paper:
//
// "Global code motion/global value numbering", Cliff Click, PLDI '95
// Proceedings of the ACM SIGPLAN 1995 conference on Programming
// language design and implementation, Pages 246-257
// http://dl.acm.org/citation.cfm?id=207154
#include "tensorflow/core/graph/optimizer_cse.h"
#include <unordered_map>
#include "tensorflow/core/graph/algorithm.h"
#include "tensorflow/core/lib/gtl/map_util.h"
#include "tensorflow/core/lib/hash/hash.h"
#include "tensorflow/core/platform/logging.h"
namespace tensorflow {
class OptimizerCSE {
public:
explicit OptimizerCSE(Graph* g) : g_(g) {}
void Optimize(std::function<bool(const Node*)> consider_fn);
private:
struct Scratch;
static size_t NodeHash(const Node* n);
static bool Equivalent(const Node* a, const Node* b, Scratch* s);
static bool EqualAttrs(const Node* a, const Node* b, Scratch* s);
Graph* g_;
};
static void FillInputs(const Node* n,
gtl::InlinedVector<Node*, 4>* control_edges,
gtl::InlinedVector<std::pair<Node*, int>, 4>* in) {
DCHECK_EQ(in->size(), n->num_inputs());
control_edges->clear();
for (const Edge* e : n->in_edges()) {
if (e->IsControlEdge()) {
control_edges->push_back(e->src());
} else {
(*in)[e->dst_input()] = std::make_pair(e->src(), e->src_output());
}
}
std::sort(control_edges->begin(), control_edges->end());
if (n->op_def().is_commutative()) {
// For commutative inputs, we sort the input by the input Node*
// to get a canonical ordering (so that add(a,b) and add(b, a) will
// hash to the same value if is_commutative is true for 'add').
std::sort(in->begin(), in->end());
}
}
static size_t kIllegalNodeHash = 0;
size_t OptimizerCSE::NodeHash(const Node* n) {
const DataTypeVector& out = n->output_types();
string str_to_hash = strings::StrCat(n->type_string(), out.size());
for (DataType dt : out) {
strings::StrAppend(&str_to_hash, dt);
}
const int N_in = n->num_inputs();
strings::StrAppend(&str_to_hash, N_in);
gtl::InlinedVector<Node*, 4> control_edges;
gtl::InlinedVector<std::pair<Node*, int>, 4> in(N_in);
FillInputs(n, &control_edges, &in);
for (const auto& edge : in) {
strings::StrAppend(&str_to_hash, edge.first->id(), edge.second);
}
size_t h = Hash64(str_to_hash);
#if !defined(__ANDROID__) && !defined(ANDROID)
// Hash the attrs. For example, this makes sure different constants
// end up in different hash buckets.
string tmp;
for (const auto& attr : n->def().attr()) {
tmp = attr.first;
attr.second.AppendToString(&tmp);
// Add hashes of attrs, so the order of attrs doesn't matter.
h += Hash32(tmp.data(), tmp.size(), 0x87341245);
}
#endif
if (h == kIllegalNodeHash) h = kIllegalNodeHash + 1;
return h;
}
struct OptimizerCSE::Scratch {
// For EqualAttrs():
string a;
string b;
};
bool OptimizerCSE::EqualAttrs(const Node* a, const Node* b, Scratch* scratch) {
if (a->def().attr_size() != b->def().attr_size()) return false;
for (const auto& attr : b->def().attr()) {
auto iter = a->def().attr().find(attr.first);
if (iter == a->def().attr().end()) return false;
// Note: it should be safe to compare proto serializations of the attr
// values since at most one field should be set in each (indeed, it
// should be the same field).
iter->second.SerializeToString(&scratch->a);
attr.second.SerializeToString(&scratch->b);
if (scratch->a != scratch->b) return false;
}
return true;
}
static bool HasRefInput(const Node* n) {
for (auto dt : n->input_types()) {
if (IsRefType(dt)) return true;
}
return false;
}
bool OptimizerCSE::Equivalent(const Node* a, const Node* b, Scratch* scratch) {
// Different op names are different
if (a->type_string() != b->type_string()) return false;
// Never consider stateful nodes (such as non-const inputs) equivalent.
if (a->op_def().is_stateful()) return false;
// For now, we consider any node that takes a ref input to not be
// equivalent to any other node.
if (HasRefInput(a) || HasRefInput(b)) return false;
// Compare attrs. Note that equal attrs implies equal input and
// output types.
if (!EqualAttrs(a, b, scratch)) return false;
// Compare input sources
if (a->num_inputs() != b->num_inputs()) return false;
const int N_in = a->num_inputs();
gtl::InlinedVector<Node*, 4> a_control_edges;
gtl::InlinedVector<Node*, 4> b_control_edges;
gtl::InlinedVector<std::pair<Node*, int>, 4> a_in(N_in);
gtl::InlinedVector<std::pair<Node*, int>, 4> b_in(N_in);
FillInputs(a, &a_control_edges, &a_in);
FillInputs(b, &b_control_edges, &b_in);
if (a_in != b_in) return false;
if (a_control_edges != b_control_edges) return false;
return true;
}
void OptimizerCSE::Optimize(std::function<bool(const Node*)> consider_fn) {
// This very simple implementation works if the whole graph is one
// giant basic block (because we just traverse nodes in a
// topological order). We'll need to do something more
// sophisticated when we have control flow/loops/etc.
// TODO(jeff): We need to handle Update nodes specially, but dealing
// with more general control flow will also solve this issue, and for
// now, our updates are almost always the most downstream nodes in
// the graph.
std::vector<Node*> order;
GetReversePostOrder(*g_, &order);
// Our value is just a single Node*, meaning we keep just a single
// candidate for a given node hash value. This may cause us to
// (rarely) lose some optimization opportunities if there are
// hash collisions, but it allows us to avoid having the value
// be a set<Node*> (or equivalent).
std::unordered_map<size_t, Node*> available;
// Scratch space for Equivalent calls. Allocated here and passed in to
// Equivalent to avoid allocation inside the loop below.
Scratch scratch;
for (Node* n : order) {
if (!n->IsOp()) continue;
// See if we should consider this node at all
if (consider_fn != nullptr && !consider_fn(n)) continue;
size_t h = NodeHash(n);
Node** candidate = &available[h];
if (*candidate == nullptr) {
// No existing match: insert "n" into the hash table under "h"
*candidate = n;
} else if (Equivalent(*candidate, n, &scratch)) {
VLOG(1) << "CSE: equivalent: " << (*candidate)->name() << " and "
<< n->name();
// *candidate and n are equivalent. Therefore, we can replace
// n with *candidate by fixing up outgoing edges from "n" to instead
// come from "*candidate", and then delete n from the graph
for (const Edge* e : n->out_edges()) {
g_->AddEdge(*candidate, e->src_output(), e->dst(), e->dst_input());
}
g_->RemoveNode(n);
}
}
}
void OptimizeCSE(Graph* g, std::function<bool(const Node*)> consider_fn) {
OptimizerCSE opt(g);
opt.Optimize(consider_fn);
}
} // namespace tensorflow
|
