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
|
/*
* 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 "Interpreter.h"
#include <stack>
#include <unordered_set>
#include "Registration.h"
#include "model/OperandIndexMap.h"
#include "util/logging.h"
#include "model/OperationVisitor.h"
namespace neurun
{
namespace exec
{
namespace interp
{
// TODO more structured execution kernel implementation
// TODO use cker for execution
// TODO divide tensor prepare and execution
// TODO introduce memory manager (buffer allocate and free)
class OperationExecutor : model::OperationVisitor
{
public:
#define OP(InternalName, IsNnApi) InternalName,
enum class NodeName
{
#include "model/Operations.lst"
};
#undef OP
public:
OperationExecutor(ExecEnv *env) : _env{env}
{
_kernels[NodeName::AddNode] = getAddNode();
_kernels[NodeName::Conv2DNode] = getConv2DNode();
_kernels[NodeName::MaxPool2DNode] = getMaxPool2DNode();
_kernels[NodeName::ConcatNode] = getConcatNode();
_kernels[NodeName::AvgPool2DNode] = getAvgPool2DNode();
_kernels[NodeName::FullyConnectedNode] = getFullyConnectedNode();
_kernels[NodeName::SoftmaxNode] = getSoftMaxNode();
_kernels[NodeName::ReshapeNode] = getReshapeNode();
_kernels[NodeName::DepthwiseConv2DNode] = getDepthwiseConvNode();
}
void execute(const model::OperationIndex &idx)
{
const auto nodeName = _env->model().operations.at(idx).getName();
VERBOSE(INTERPRETER) << "Prepare output operands and execute " << nodeName
<< " operation (id: " << idx.value() << ")" << std::endl;
_env->model().operations.at(idx).accept(*this);
}
private:
#define OP(InternalName, IsNnApi) \
virtual void visit(const model::operation::InternalName &node) override \
{ \
if (_kernels[NodeName::InternalName]->prepare != nullptr) \
{ \
_kernels[NodeName::InternalName]->prepare(_env, node); \
} \
_kernels[NodeName::InternalName]->invoke(_env, node); \
}
#include "model/Operations.lst"
#undef OP
private:
ExecEnv *_env;
std::unordered_map<NodeName, OpKernel *> _kernels;
};
void Interpreter::run()
{
VERBOSE(INTERPRETER) << "Interpreter is invoked " << std::endl;
// operand_stack: save operands prepared to use
std::stack<model::OperandIndex> operand_stack;
// Note: We should push input first, then constant.
// We use use-def for find operators ready to execution,
// but Use-Def cannot handle parameters (maybe constant, but not always)
// Note: If all model inputs are constant, it may not work (depend on tensors' order).
// But that scenario may not exist
for (auto ind : _env->model().inputs)
{
VERBOSE(INTERPRETER) << "Input: Push to operand stack " << ind.value() << std::endl;
operand_stack.push(ind);
}
_env->model().operands.iterate([&](const model::OperandIndex &ind, const model::Operand &obj) {
if (obj.isConstant())
{
VERBOSE(INTERPRETER) << "Constant: Push to operand stack " << ind.value() << std::endl;
operand_stack.push(ind);
}
});
// Execution
std::unordered_set<model::OperandIndex> ready_check;
std::unordered_set<model::OperationIndex> executed;
OperationExecutor executor{_env.get()};
while (!operand_stack.empty())
{
const auto current_operand_index = operand_stack.top();
operand_stack.pop();
VERBOSE(INTERPRETER) << "Poped operand " << current_operand_index.value()
<< " is checked ready to use" << std::endl;
assert(ready_check.find(current_operand_index) == ready_check.end());
ready_check.insert(current_operand_index);
// Find prepared operations by scan use of current operand
std::stack<model::OperationIndex> operation_stack;
const auto use_operators = _env->model().operands.at(current_operand_index).getUses();
for (auto use_operator : use_operators.list())
{
// Assumption: all parameters are ready to use
bool operator_ready = true;
for (auto input_index : _env->model().operations.at(use_operator).getInputs())
{
if (ready_check.find(input_index) == ready_check.end())
{
operator_ready = false;
break;
}
}
if (operator_ready)
{
VERBOSE(INTERPRETER) << "Ready to execute operation " << use_operator.value() << std::endl;
operation_stack.push(use_operator);
}
}
while (!operation_stack.empty())
{
const auto current_operation_index = operation_stack.top();
operation_stack.pop();
VERBOSE(INTERPRETER) << "Poped operation: " << current_operation_index.value() << "("
<< _env->model().operations.at(current_operation_index).getName() << ")"
<< std::endl;
// execution
// 1. Prepare output tensor
// 2. Call operation kernel
executor.execute(current_operation_index);
executed.insert(current_operation_index);
// 3. Push each output into operand stack
const auto def_operands = _env->model().operations.at(current_operation_index).getOutputs();
for (auto def_operand : def_operands)
{
VERBOSE(INTERPRETER) << "Buffer: Push to operand stack " << def_operand.value()
<< std::endl;
operand_stack.push(def_operand);
}
// 4. Free if lifetime of buffer operands used by input is finished
for (auto input_index : _env->model().operations.at(current_operation_index).getInputs())
{
const auto use_operators = _env->model().operands.at(input_index).getUses();
bool dead_buffer = true;
for (auto use_operator : use_operators.list())
{
if (executed.find(use_operator) == executed.end())
{
dead_buffer = false;
break;
}
}
if (dead_buffer)
{
_env->freeIfAllocated(input_index);
}
}
}
}
}
} // namespace interp
} // namespace exec
} // namespace neurun
|
