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
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
|
/*
* Copyright (c) 2019 Samsung Electronics Co., Ltd. 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 "passes/optimizations/FuseArithmeticOps.h"
#include "passes/optimizations/OptimizationUtils.h"
#include "mir/ops/AddOp.h"
#include "mir/ops/ConstantOp.h"
#include "mir/ops/Conv2DOp.h"
#include "mir/ops/MulOp.h"
#include "mir/Graph.h"
#include "mir/Tensor.h"
#include "mir/Index.h"
#include "mir/TensorVariant.h"
#include "mir/ShapeRange.h"
#include <algorithm>
namespace nnc
{
namespace
{
using namespace mir;
using namespace std;
using namespace opt_util;
using OpType = Operation::Type;
using Edge = pair<Operation *, Operation *>;
/**
* This function used to get 'ConstantOp' with weights of 'AddOp', 'MulOp' or 'Conv2DOp'
* For each of these ops weights stored in second input node
*/
ops::ConstantOp *getSecondInputAsConst(Operation *op)
{
assert(op->getType() == OpType::add || op->getType() == OpType::mul ||
op->getType() == OpType::conv2D);
return dynamic_cast<ops::ConstantOp *>(op->getInput(1)->getNode());
}
// This function finds successive operations of given types, with ConstantOp as second input
vector<Edge> findSuccessiveOpsWithConstWeights(Graph *g, OpType first_op_type,
OpType second_op_type)
{
vector<Edge> matches;
unordered_set<Operation *> matched_nodes;
for (auto *first_op : g->getNodes())
{
if (first_op->getType() == first_op_type && getSecondInputAsConst(first_op))
{
for (auto &out : first_op->getOutputs())
{
for (Operation::Use use : out.getUses())
{
Operation *second_op = use.getNode();
if (second_op->getType() == second_op_type && getSecondInputAsConst(second_op))
{
/**
* Don't match already matched nodes, so for op1->op2->op3 this function
* will return {{f1, f2}} and not {{f1, f2}, {f2, f3}}
*/
if (matched_nodes.find(first_op) == matched_nodes.end() &&
matched_nodes.find(second_op) == matched_nodes.end())
{
matched_nodes.emplace(first_op);
matched_nodes.emplace(second_op);
matches.emplace_back(first_op, second_op);
}
}
}
}
}
}
return matches;
}
/**
* This function merges two ConstantOp into new one, by elementwise multiplication or addition
* If first ConstantOp rank > 1, second one broadcasting to first by axis=0
*/
Operation *mergeConstantOps(Graph *g, const ops::ConstantOp *const1_op,
const ops::ConstantOp *const2_op, OpType merge_type)
{
const auto &const1_val = const1_op->getValue();
const auto &const2_val = const2_op->getValue();
assert(const1_val.getShape().rank() >= const2_val.getShape().rank());
assert(const2_val.getShape().rank() == 1);
assert(const1_val.getShape().dim(0) == const2_val.getShape().dim(0));
// Create and fill TensorVariant for new ConstantOp
TensorVariant new_const_val(DataType::FLOAT32, const1_val.getShape());
Tensor<float> const1_accessor(const1_val);
Tensor<float> const2_accessor(const2_val);
Tensor<float> new_const_accessor(new_const_val);
ShapeRange const1_range(const1_val.getShape());
for (auto &idx : const1_range)
{
float operand1 = const1_accessor.at(idx);
/**
* Broadcast second ConstantOp to first one:
* idx of second constant always has rank 1 and equals to first dimension of first constant idx
*/
float operand2 = const2_accessor.at(Index{idx.at(0)});
switch (merge_type)
{
case OpType::mul:
new_const_accessor.at(idx) = operand1 * operand2;
break;
case OpType::add:
new_const_accessor.at(idx) = operand1 + operand2;
break;
default:
assert(false && "only 'mul' and 'add' constants merge types supported");
}
}
return g->create<ops::ConstantOp>(new_const_val);
}
// TODO: support 'DepthwiseConv'->'Mul'
/**
* This function fuses some successive operations with constant weights into one:
* 'Add'->'Add' into 'Add'; 'Mul'->'Mul' into 'Mul'; 'Conv'->'Mul' into 'Conv';
* Before: | After:
* -------------------------|---------------------------
* [input] [Const1] | [input] [Const1*Const2]
* \\ // | \\ //
* [Mul] [Const2] | [Mul]
* \\ // |
* [Mul] |
* -------------------------|---------------------------
* [input] [Const1] | [input] [Const1+Const2]
* \\ // | \\ //
* [Add] [Const2] | [Add]
* \\ // |
* [Add] |
* -------------------------|---------------------------
* [input] [Const1] | [input] [Const1*Const2]
* \\ // | \\ //
* [Conv2D] [Const2] | [Conv2D]
* \\ // |
* [Mul] |
*/
bool fuseSuccessiveOps(Graph *g)
{
// Find all successive ops
vector<Edge> successive_ops;
auto mul_mul_vec = findSuccessiveOpsWithConstWeights(g, OpType::mul, OpType::mul);
successive_ops.insert(successive_ops.end(), mul_mul_vec.begin(), mul_mul_vec.end());
auto add_add_vec = findSuccessiveOpsWithConstWeights(g, OpType::add, OpType::add);
successive_ops.insert(successive_ops.end(), add_add_vec.begin(), add_add_vec.end());
auto conv_mul_vec = findSuccessiveOpsWithConstWeights(g, OpType::conv2D, OpType::mul);
successive_ops.insert(successive_ops.end(), conv_mul_vec.begin(), conv_mul_vec.end());
for (auto &edge : successive_ops)
{
auto const1_op = getSecondInputAsConst(edge.first);
auto const2_op = getSecondInputAsConst(edge.second);
assert(const1_op && const2_op);
// Create new constant operation and copy first successive operation
auto new_const_op = mergeConstantOps(g, const1_op, const2_op, edge.second->getType());
auto first_op_input = edge.first->getInput(0);
auto new_op = g->copyOpWithInputs(edge.first, {first_op_input, new_const_op->getOutput(0)});
// Replace second successive operation with new one and remove old nodes
g->replaceNode(edge.second, new_op);
removeNodeIfUnused(g, edge.first);
removeNodeIfUnused(g, const1_op);
removeNodeIfUnused(g, const2_op);
}
// If there is no successive operations to fuse - graph wasn't changed
return !successive_ops.empty();
}
/**
* This function sinks 'Add' through 'Mul'
* by multiplying 'Add' weights on 'Mul' weights
* Before: | After:
*--------------------------|--------------------------
* [input] [Const1] | [input] [Const2]
* \\ // | \\ //
* [Add] [Const2] | [Mul] [Const1*Const2]
* \\ // | \\ //
* [Mul] | [Add]
* |
*/
bool sinkAddThroughMul(Graph *g)
{
auto add_mul_edges = findSuccessiveOpsWithConstWeights(g, OpType::add, OpType::mul);
for (auto &edge : add_mul_edges)
{
auto old_add_op = edge.first;
auto old_mul_op = edge.second;
auto old_add_const_op = getSecondInputAsConst(old_add_op);
auto ols_mul_const_op = getSecondInputAsConst(old_mul_op);
assert(old_add_const_op && ols_mul_const_op);
// Create new operations
auto old_add_input = old_add_op->getInput(0);
auto new_mul_op =
g->copyOpWithInputs(old_mul_op, {old_add_input, ols_mul_const_op->getOutput(0)});
auto new_add_const_op = mergeConstantOps(g, old_add_const_op, ols_mul_const_op, OpType::mul);
auto new_add_op =
g->copyOpWithInputs(old_add_op, {new_mul_op->getOutput(0), new_add_const_op->getOutput(0)});
// Replace old mul with new add and remove old nodes
g->replaceNode(old_mul_op, new_add_op);
removeNodeIfUnused(g, old_add_op);
removeNodeIfUnused(g, old_add_const_op);
}
// If there is no add-mul edges - graph wasn't changed
return !add_mul_edges.empty();
}
} // unnamed namespace
nnc::PassData nnc::FuseArithmeticOps::run(nnc::PassData data)
{
auto g = static_cast<Graph *>(data);
bool graph_changed = true;
while (graph_changed)
{
graph_changed = false;
graph_changed |= fuseSuccessiveOps(g);
graph_changed |= sinkAddThroughMul(g);
}
return g;
}
} // namespace nnc
|
