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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "Runtime.hpp"
#include <armnn/Version.hpp>
#include <backendsCommon/BackendRegistry.hpp>
#include <iostream>
#ifdef ARMCOMPUTECL_ENABLED
#include <arm_compute/core/CL/OpenCL.h>
#include <arm_compute/core/CL/CLKernelLibrary.h>
#include <arm_compute/runtime/CL/CLScheduler.h>
#endif
#include <boost/log/trivial.hpp>
#include <boost/polymorphic_cast.hpp>
using namespace armnn;
using namespace std;
namespace armnn
{
IRuntime* IRuntime::CreateRaw(const CreationOptions& options)
{
return new Runtime(options);
}
IRuntimePtr IRuntime::Create(const CreationOptions& options)
{
return IRuntimePtr(CreateRaw(options), &IRuntime::Destroy);
}
void IRuntime::Destroy(IRuntime* runtime)
{
delete boost::polymorphic_downcast<Runtime*>(runtime);
}
int Runtime::GenerateNetworkId()
{
return m_NetworkIdCounter++;
}
Status Runtime::LoadNetwork(NetworkId& networkIdOut, IOptimizedNetworkPtr inNetwork)
{
std::string ignoredErrorMessage;
return LoadNetwork(networkIdOut, std::move(inNetwork), ignoredErrorMessage);
}
Status Runtime::LoadNetwork(NetworkId& networkIdOut,
IOptimizedNetworkPtr inNetwork,
std::string & errorMessage)
{
IOptimizedNetwork* rawNetwork = inNetwork.release();
unique_ptr<LoadedNetwork> loadedNetwork = LoadedNetwork::MakeLoadedNetwork(
std::unique_ptr<OptimizedNetwork>(boost::polymorphic_downcast<OptimizedNetwork*>(rawNetwork)),
errorMessage);
if (!loadedNetwork)
{
return Status::Failure;
}
networkIdOut = GenerateNetworkId();
{
std::lock_guard<std::mutex> lockGuard(m_Mutex);
// Stores the network
m_LoadedNetworks[networkIdOut] = std::move(loadedNetwork);
}
return Status::Success;
}
Status Runtime::UnloadNetwork(NetworkId networkId)
{
#ifdef ARMCOMPUTECL_ENABLED
if (arm_compute::CLScheduler::get().context()() != NULL)
{
// Waits for all queued CL requests to finish before unloading the network they may be using.
try
{
// Coverity fix: arm_compute::CLScheduler::sync() may throw an exception of type cl::Error.
arm_compute::CLScheduler::get().sync();
}
catch (const cl::Error&)
{
BOOST_LOG_TRIVIAL(warning) << "WARNING: Runtime::UnloadNetwork(): an error occurred while waiting for "
"the queued CL requests to finish";
return Status::Failure;
}
}
#endif
{
std::lock_guard<std::mutex> lockGuard(m_Mutex);
if (m_LoadedNetworks.erase(networkId) == 0)
{
BOOST_LOG_TRIVIAL(warning) << "WARNING: Runtime::UnloadNetwork(): " << networkId << " not found!";
return Status::Failure;
}
#ifdef ARMCOMPUTECL_ENABLED
if (arm_compute::CLScheduler::get().context()() != NULL && m_LoadedNetworks.empty())
{
// There are no loaded networks left, so clear the CL cache to free up memory
m_ClContextControl.ClearClCache();
}
#endif
}
BOOST_LOG_TRIVIAL(debug) << "Runtime::UnloadNetwork(): Unloaded network with ID: " << networkId;
return Status::Success;
}
const std::shared_ptr<IProfiler> Runtime::GetProfiler(NetworkId networkId) const
{
auto it = m_LoadedNetworks.find(networkId);
if (it != m_LoadedNetworks.end())
{
auto& loadedNetwork = it->second;
return loadedNetwork->GetProfiler();
}
return nullptr;
}
Runtime::Runtime(const CreationOptions& options)
: m_ClContextControl(options.m_GpuAccTunedParameters.get(),
options.m_EnableGpuProfiling)
, m_NetworkIdCounter(0)
, m_DeviceSpec{BackendRegistryInstance().GetBackendIds()}
{
BOOST_LOG_TRIVIAL(info) << "ArmNN v" << ARMNN_VERSION << "\n";
}
Runtime::~Runtime()
{
std::vector<int> networkIDs;
try
{
// Coverity fix: The following code may throw an exception of type std::length_error.
std::transform(m_LoadedNetworks.begin(), m_LoadedNetworks.end(),
std::back_inserter(networkIDs),
[](const auto &pair) { return pair.first; });
}
catch (const std::exception& e)
{
// Coverity fix: BOOST_LOG_TRIVIAL (typically used to report errors) may throw an
// exception of type std::length_error.
// Using stderr instead in this context as there is no point in nesting try-catch blocks here.
std::cerr << "WARNING: An error has occurred when getting the IDs of the networks to unload: " << e.what()
<< "\nSome of the loaded networks may not be unloaded" << std::endl;
}
// We then proceed to unload all the networks which IDs have been appended to the list
// up to the point the exception was thrown (if any).
for (auto networkID : networkIDs)
{
try
{
// Coverity fix: UnloadNetwork() may throw an exception of type std::length_error,
// boost::log::v2s_mt_posix::odr_violation or boost::log::v2s_mt_posix::system_error
UnloadNetwork(networkID);
}
catch (const std::exception& e)
{
// Coverity fix: BOOST_LOG_TRIVIAL (typically used to report errors) may throw an
// exception of type std::length_error.
// Using stderr instead in this context as there is no point in nesting try-catch blocks here.
std::cerr << "WARNING: An error has occurred when unloading network " << networkID << ": " << e.what()
<< std::endl;
}
}
}
LoadedNetwork* Runtime::GetLoadedNetworkPtr(NetworkId networkId) const
{
std::lock_guard<std::mutex> lockGuard(m_Mutex);
return m_LoadedNetworks.at(networkId).get();
}
TensorInfo Runtime::GetInputTensorInfo(NetworkId networkId, LayerBindingId layerId) const
{
return GetLoadedNetworkPtr(networkId)->GetInputTensorInfo(layerId);
}
TensorInfo Runtime::GetOutputTensorInfo(NetworkId networkId, LayerBindingId layerId) const
{
return GetLoadedNetworkPtr(networkId)->GetOutputTensorInfo(layerId);
}
Status Runtime::EnqueueWorkload(NetworkId networkId,
const InputTensors& inputTensors,
const OutputTensors& outputTensors)
{
LoadedNetwork* loadedNetwork = GetLoadedNetworkPtr(networkId);
static thread_local NetworkId lastId = networkId;
if (lastId != networkId)
{
LoadedNetworkFuncSafe(lastId, [](LoadedNetwork* network)
{
network->FreeWorkingMemory();
});
}
lastId=networkId;
return loadedNetwork->EnqueueWorkload(inputTensors, outputTensors);
}
}
|
