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
|
/*
* 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 "ExecutorBase.h"
#include "util/logging.h"
namespace onert
{
namespace exec
{
ExecutorBase::ExecutorBase(std::unique_ptr<ir::LoweredGraph> &&lowered_graph,
const backend::TensorBuilderSet &tensor_builders)
: _lowered_graph{std::move(lowered_graph)}, _graph{_lowered_graph->graph()}, _mutex()
{
auto build_itensor_list = [&](const onert::ir::OperandIndexSequence &ind_seq) {
std::vector<std::shared_ptr<backend::ITensor>> list;
for (auto ind : ind_seq)
{
std::shared_ptr<backend::ITensor> tensor;
for (auto &tensor_builder : tensor_builders)
{
tensor = tensor_builder->tensorAt(ind);
if (tensor != nullptr)
break;
}
assert(tensor != nullptr);
list.push_back(tensor);
}
return list;
};
_input_tensors = build_itensor_list(_graph.getInputs());
_output_tensors = build_itensor_list(_graph.getOutputs());
// Prepare each TensorManager on each backend
for (auto &tensor_builder : tensor_builders)
{
auto tensor_manager = tensor_builder->releaseTensorManager();
assert(tensor_manager != nullptr);
_tensor_mgrs.insert(std::move(tensor_manager));
}
}
std::unique_ptr<ISource> ExecutorBase::source(const ir::IOIndex &index, const ir::TypeInfo &type,
const void *buffer, size_t length,
ir::Layout io_layout)
{
using ir::DataType;
switch (type.type())
{
case DataType::FLOAT32:
return source<float>(index, buffer, length, io_layout);
case DataType::INT32:
return source<int32_t>(index, buffer, length, io_layout);
case DataType::UINT32:
return source<uint32_t>(index, buffer, length, io_layout);
case DataType::BOOL8:
case DataType::QUANT8_ASYMM:
case DataType::UINT8:
return source<uint8_t>(index, buffer, length, io_layout);
case DataType::QUANT8_SYMM:
return source<int8_t>(index, buffer, length, io_layout);
default:
throw std::runtime_error("Not supported yet");
}
}
std::unique_ptr<ISink> ExecutorBase::sink(const ir::IOIndex &index, const ir::TypeInfo &type,
void *buffer, size_t length, ir::Layout io_layout)
{
using ir::DataType;
switch (type.type())
{
case DataType::FLOAT32:
return sink<float>(index, buffer, length, io_layout);
case DataType::INT32:
return sink<int32_t>(index, buffer, length, io_layout);
case DataType::UINT32:
return sink<uint32_t>(index, buffer, length, io_layout);
case DataType::BOOL8:
case DataType::QUANT8_ASYMM:
case DataType::UINT8:
return sink<uint8_t>(index, buffer, length, io_layout);
case DataType::QUANT8_SYMM:
return sink<int8_t>(index, buffer, length, io_layout);
default:
throw std::runtime_error("Not supported yet");
}
}
void ExecutorBase::execute(const IODescription &desc)
{
// For thread-safe, use mutex
// TODO: if all used backends on this executor are thread-safe,
// do not need to use mutex (otherwise, use mutex)
std::lock_guard<std::mutex> lock(_mutex);
std::vector<std::unique_ptr<ISource>> sources{_graph.getInputs().size()};
std::vector<std::unique_ptr<ISink>> sinks{_graph.getOutputs().size()};
// Set input(s)
for (uint32_t n = 0; n < _graph.getInputs().size(); ++n)
{
ir::IOIndex input_index{n};
ir::OperandIndex index{_graph.getInputs().at(input_index)};
if (desc.inputs.at(n) == nullptr)
{
// Optional input
continue;
}
const auto operand_li = _lowered_graph->getLowerInfo()->operand.at(index).get();
if (operand_li->def_factors().empty())
{
// This input is not used (i.e. constant, EX. reshape's axis)
continue;
}
const auto &input = *desc.inputs.at(n);
sources.at(n) =
source(input_index, input.info.typeInfo(), input.buffer, input.size, input.layout);
auto setter = [&](::onert::backend::ITensor &tensor) { sources.at(n)->push(tensor); };
_input_tensors[n]->access(setter);
}
executeImpl();
// Get output(s)
for (uint32_t n = 0; n < _graph.getOutputs().size(); ++n)
{
ir::IOIndex output_index{n};
// Optional output
if (desc.outputs.at(n) == nullptr)
{
continue;
}
const auto &output = *desc.outputs.at(n);
sinks.at(n) =
sink(output_index, output.info.typeInfo(), output.buffer, output.size, output.layout);
auto getter = [&](::onert::backend::ITensor &tensor) { sinks.at(n)->pull(tensor); };
_output_tensors[n]->access(getter);
}
}
} // namespace exec
} // namespace onert
|
