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
|
//
// Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "Network.hpp"
#include "LayerFwd.hpp"
#include "Profiling.hpp"
#include <armnn/Tensor.hpp>
#include <armnn/backends/IBackendInternal.hpp>
#include <armnn/backends/IMemoryOptimizerStrategy.hpp>
#include <backendsCommon/TensorHandleFactoryRegistry.hpp>
#include <armnn/backends/Workload.hpp>
#include <armnn/backends/WorkloadFactory.hpp>
#include <backendsCommon/DefaultAllocator.hpp>
#include <backendsCommon/MemoryManager.hpp>
#include <backendsCommon/memoryOptimizerStrategyLibrary/strategies/SingleAxisPriorityList.hpp>
#include <ProfilingService.hpp>
#include <TimelineUtilityMethods.hpp>
#include <common/include/LabelsAndEventClasses.hpp>
#include <mutex>
#include <condition_variable>
#include <unordered_map>
namespace cl
{
class Context;
class CommandQueue;
class Device;
}
namespace armnn
{
class LoadedNetwork
{
public:
using WorkloadQueue = std::vector<std::unique_ptr<IWorkload>>;
~LoadedNetwork()
{
FreeWorkingMemory();
}
/// Create a new unique WorkingMemHandle object. Create multiple handles if you wish to have
/// overlapped Execution by calling this function from different threads.
std::unique_ptr<IWorkingMemHandle> CreateWorkingMemHandle(NetworkId networkId);
TensorInfo GetInputTensorInfo(LayerBindingId layerId) const;
TensorInfo GetOutputTensorInfo(LayerBindingId layerId) const;
std::vector<ImportedInputId> ImportInputs(const InputTensors& inputTensors,
MemorySource forceImportMemorySource = MemorySource::Undefined);
std::vector<ImportedOutputId> ImportOutputs(const OutputTensors& outputTensors,
MemorySource forceImportMemorySource = MemorySource::Undefined);
void ClearImportedInputs(const std::vector<ImportedInputId> inputIds);
void ClearImportedOutputs(const std::vector<ImportedOutputId> outputIds);
/// Single thread execution of the loaded network
Status EnqueueWorkload(const InputTensors& inputTensors, const OutputTensors& outputTensors,
std::vector<ImportedInputId> preImportedInputIds = {},
std::vector<ImportedOutputId> preImportedOutputIds = {});
/// Thread safe execution of the loaded network
Status Execute(const InputTensors& inputTensors,
const OutputTensors& outputTensors,
IWorkingMemHandle& workingMemHandle,
std::vector<ImportedInputId> preImportedInputs = {},
std::vector<ImportedOutputId> preImportedOutputs = {});
static std::unique_ptr<LoadedNetwork> MakeLoadedNetwork(std::unique_ptr<IOptimizedNetwork> net,
std::string& errorMessage,
const INetworkProperties& networkProperties,
profiling::ProfilingService& profilingService);
// NOTE we return by reference as the purpose of this method is only to provide
// access to the private m_Profiler and in theory we should not need to increment
// the shared_ptr's reference counter
const std::shared_ptr<IProfiler>& GetProfiler() const { return m_OptimizedNetwork->GetProfiler(); }
void FreeWorkingMemory();
void RegisterDebugCallback(const DebugCallbackFunction& func);
void SendNetworkStructure();
bool IsAsyncEnabled()
{
return m_NetworkProperties.m_AsyncEnabled;
}
profiling::ProfilingGuid GetNetworkGuid();
private:
void AllocateWorkingMemory(std::lock_guard<std::mutex>& lock);
void AllocateAndExecuteConstantWorkloads();
void AllocateAndExecuteConstantWorkloadsAsync();
std::unordered_map<LayerGuid, std::unique_ptr<IWorkload>> m_ConstantWorkloads;
std::unordered_map<LayerGuid, ITensorHandle*> m_ConstantTensorHandles;
std::unique_ptr<IMemoryOptimizerStrategy> m_ConstantStrategy = std::make_unique<SingleAxisPriorityList>();
LoadedNetwork(std::unique_ptr<IOptimizedNetwork> net,
const INetworkProperties& networkProperties,
profiling::ProfilingService& profilingService);
void EnqueueInput(const BindableLayer& layer, ITensorHandle* tensorHandle, const TensorInfo& tensorInfo);
void EnqueueOutput(const BindableLayer& layer, ITensorHandle* tensorHandle, const TensorInfo& tensorInfo);
void EnqueueInput(const ConstTensor& inputTensor, ITensorHandle* inputTensorHandle);
void ImportOutputTensor(const Tensor& outputTensor, ITensorHandle* outputTensorHandle);
bool Execute(std::unique_ptr<profiling::TimelineUtilityMethods>& timelineUtils,
profiling::ProfilingGuid inferenceGuid);
const IWorkloadFactory& GetWorkloadFactory(const Layer& layer) const;
inline LayerBindingId ValidateImportedInputID(ImportedInputId id);
inline LayerBindingId ValidateImportedOutputID(ImportedOutputId id);
void CreateMemoryProfile();
void CreateMemoryProfileAsync();
std::unique_ptr<MemoryManager> CreateExternalMemoryManger(
std::vector<std::pair<std::shared_ptr<TensorMemory>, MemorySource>>& tensorMemory);
using BackendPtrMap = std::unordered_map<BackendId, IBackendInternalUniquePtr>;
BackendPtrMap m_Backends;
std::vector<IBackendInternal::IMemoryManagerSharedPtr> m_BackendMemoryMangers;
using WorkloadFactoryMap = std::unordered_map<BackendId, IBackendInternal::IWorkloadFactoryPtr>;
WorkloadFactoryMap m_WorkloadFactories;
std::unique_ptr<IOptimizedNetwork> m_OptimizedNetwork;
WorkloadQueue m_InputQueue;
WorkloadQueue m_WorkloadQueue;
WorkloadQueue m_OutputQueue;
mutable std::mutex m_WorkingMemMutex;
bool m_IsWorkingMemAllocated = false;
INetworkProperties m_NetworkProperties;
TensorHandleFactoryRegistry m_TensorHandleFactoryRegistry;
profiling::ProfilingService& m_ProfilingService;
struct ImportedTensorHandlePin
{
ImportedTensorHandlePin()
{}
ImportedTensorHandlePin(LayerBindingId layerBindingId,
std::unique_ptr<ITensorHandle> tensorHandle)
: m_LayerBindingId(layerBindingId)
, m_TensorHandle(std::move(tensorHandle))
{}
ImportedTensorHandlePin(ImportedTensorHandlePin&&) = default;
~ImportedTensorHandlePin()
{
if (m_TensorHandle)
{
m_TensorHandle->Unimport();
}
}
LayerBindingId m_LayerBindingId;
std::unique_ptr<ITensorHandle> m_TensorHandle;
};
std::vector<ImportedTensorHandlePin> m_PreImportedInputHandles;
std::vector<ImportedTensorHandlePin> m_PreImportedOutputHandles;
ImportedInputId m_CurImportedInputId = 0;
ImportedInputId m_CurImportedOutputId = 0;
std::unordered_map<BackendId, std::vector<MemBlock>> m_MemBlockMap;
std::unordered_map<BackendId, std::vector<MemBin>> m_MemBinMap;
std::vector<ITensorHandle*> m_Tensorhandles;
std::vector<std::pair<std::shared_ptr<TensorMemory>, MemorySource>> m_TensorMemory;
std::unique_ptr<MemoryManager> m_ExternalMemoryManager;
std::unordered_map<BackendId, bool> m_SupportsExternallyManagedMemory;
// A set of vectors to record the workload queue indexes and their corresponding Input/Output Slot indexes
// which are connected to Inputs and Outputs for the network.
struct WorkloadIndices
{
unsigned int m_WorkloadIndex;
unsigned int m_SlotIndex;
};
struct OutputWorkloadIndices
{
WorkloadIndices m_OutputSlotIndices;
std::vector<WorkloadIndices> m_InputSlotIndices;
};
std::unordered_map<LayerBindingId, std::vector<WorkloadIndices>> m_InputWorkloadSlotPairs;
std::unordered_map<LayerBindingId, OutputWorkloadIndices> m_OutputWorkloadSlotPairs;
std::vector<bool> m_IsInputImported;
std::vector<bool> m_IsOutputImported;
};
}
|
