ArmNN
 22.02
Runtime.hpp
Go to the documentation of this file.
1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #pragma once
6 
7 #include "LoadedNetwork.hpp"
8 #include "DeviceSpec.hpp"
9 
10 #include <armnn/INetwork.hpp>
11 #include <armnn/IRuntime.hpp>
12 #include <armnn/Tensor.hpp>
13 #include <armnn/BackendId.hpp>
14 
16 
17 #include <ProfilingService.hpp>
18 
19 #include <IProfilingService.hpp>
20 #include <IReportStructure.hpp>
21 
22 #include <mutex>
23 #include <unordered_map>
24 
25 namespace armnn
26 {
27 using LoadedNetworks = std::unordered_map<NetworkId, std::unique_ptr<LoadedNetwork>>;
29 
30 struct RuntimeImpl final : public IReportStructure
31 {
32 public:
33  /// Loads a complete network into the Runtime.
34  /// @param [out] networkIdOut - Unique identifier for the network is returned in this reference.
35  /// @param [in] network - Complete network to load into the Runtime.
36  /// The runtime takes ownership of the network once passed in.
37  /// @return armnn::Status
38  Status LoadNetwork(NetworkId& networkIdOut, IOptimizedNetworkPtr network);
39 
40  /// Load a complete network into the IRuntime.
41  /// @param [out] networkIdOut Unique identifier for the network is returned in this reference.
42  /// @param [in] network Complete network to load into the IRuntime.
43  /// @param [out] errorMessage Error message if there were any errors.
44  /// The runtime takes ownership of the network once passed in.
45  /// @return armnn::Status
46  Status LoadNetwork(NetworkId& networkIdOut,
47  IOptimizedNetworkPtr network,
48  std::string& errorMessage);
49 
50  Status LoadNetwork(NetworkId& networkIdOut,
51  IOptimizedNetworkPtr network,
52  std::string& errorMessage,
53  const INetworkProperties& networkProperties);
54 
57 
58  std::vector<ImportedInputId> ImportInputs(NetworkId networkId, const InputTensors& inputTensors,
59  MemorySource forceImportMemorySource = MemorySource::Undefined);
60  std::vector<ImportedOutputId> ImportOutputs(NetworkId networkId, const OutputTensors& outputTensors,
61  MemorySource forceImportMemorySource = MemorySource::Undefined);
62 
63  void ClearImportedInputs(NetworkId networkId, const std::vector<ImportedInputId> inputIds);
64  void ClearImportedOutputs(NetworkId networkId, const std::vector<ImportedOutputId> outputIds);
65 
66  // Evaluates network using input in inputTensors, outputs filled into outputTensors.
68  const InputTensors& inputTensors,
69  const OutputTensors& outputTensors,
70  std::vector<ImportedInputId> preImportedInputIds = {},
71  std::vector<ImportedOutputId> preImportedOutputIds = {});
72 
73  /// This is an experimental function.
74  /// Evaluates a network using input in inputTensors and outputs filled into outputTensors.
75  /// This function performs a thread safe execution of the network. Returns once execution is complete.
76  /// Will block until this and any other thread using the same workingMem object completes.
77  Status Execute(IWorkingMemHandle& workingMemHandle,
78  const InputTensors& inputTensors,
79  const OutputTensors& outputTensors,
80  std::vector<ImportedInputId> preImportedInputs,
81  std::vector<ImportedOutputId> preImportedOutputs);
82 
83  /// Unloads a network from the Runtime.
84  /// At the moment this only removes the network from the m_Impl->m_Network.
85  /// This might need more work in the future to be AndroidNN compliant.
86  /// @param [in] networkId Unique identifier for the network to be unloaded. Generated in LoadNetwork().
87  /// @return armnn::Status
88  Status UnloadNetwork(NetworkId networkId);
89 
90  const IDeviceSpec& GetDeviceSpec() const { return m_DeviceSpec; }
91 
92  /// Gets the profiler corresponding to the given network id.
93  /// @param networkId The id of the network for which to get the profile.
94  /// @return A pointer to the requested profiler, or nullptr if not found.
95  const std::shared_ptr<IProfiler> GetProfiler(NetworkId networkId) const;
96 
97  /// Create a new unique WorkingMemHandle object. Create multiple handles if you wish to have
98  /// overlapped Execution by calling this function from different threads.
99  std::unique_ptr<IWorkingMemHandle> CreateWorkingMemHandle(NetworkId networkId);
100 
101  /// Registers a callback function to debug layers performing custom computations on intermediate tensors.
102  /// @param networkId The id of the network to register the callback.
103  /// @param func callback function to pass to the debug layer.
104  void RegisterDebugCallback(NetworkId networkId, const DebugCallbackFunction& func);
105 
106  /// Creates a runtime for workload execution.
107  RuntimeImpl(const IRuntime::CreationOptions& options);
108 
109  ~RuntimeImpl();
110 
111  //NOTE: we won't need the profiling service reference but it is good to pass the service
112  // in this way to facilitate other implementations down the road
113  void ReportStructure();
114 
115 private:
116  friend void RuntimeLoadedNetworksReserve(RuntimeImpl* runtime); // See RuntimeTests.cpp
117 
118  friend profiling::ProfilingService& GetProfilingService(RuntimeImpl* runtime); // See RuntimeTests.cpp
119 
120  int GenerateNetworkId();
121 
122  LoadedNetwork* GetLoadedNetworkPtr(NetworkId networkId) const;
123 
124  template<typename Func>
125  void LoadedNetworkFuncSafe(NetworkId networkId, Func f)
126  {
127  std::lock_guard<std::mutex> lockGuard(m_Mutex);
128  auto iter = m_LoadedNetworks.find(networkId);
129  if (iter != m_LoadedNetworks.end())
130  {
131  f(iter->second.get());
132  }
133  }
134 
135  /// Loads any available/compatible dynamic backend in the runtime.
136  void LoadDynamicBackends(const std::string& overrideBackendPath);
137 
138  mutable std::mutex m_Mutex;
139 
140  /// Map of Loaded Networks with associated GUID as key
141  LoadedNetworks m_LoadedNetworks;
142 
143  std::unordered_map<BackendId, IBackendInternal::IBackendContextPtr> m_BackendContexts;
144 
145  int m_NetworkIdCounter;
146 
147  DeviceSpec m_DeviceSpec;
148 
149  /// List of dynamic backends loaded in the runtime
150  std::vector<DynamicBackendPtr> m_DynamicBackends;
151 
152  /// Profiling Service Instance
153  profiling::ProfilingService m_ProfilingService;
154 };
155 
156 } // namespace armnn
armnn::TensorInfo GetInputTensorInfo(NetworkId networkId, LayerBindingId layerId) const
Definition: Runtime.cpp:565
Status Execute(IWorkingMemHandle &workingMemHandle, const InputTensors &inputTensors, const OutputTensors &outputTensors, std::vector< ImportedInputId > preImportedInputs, std::vector< ImportedOutputId > preImportedOutputs)
This is an experimental function.
Definition: Runtime.cpp:648
Status UnloadNetwork(NetworkId networkId)
Unloads a network from the Runtime.
Definition: Runtime.cpp:214
Status LoadNetwork(NetworkId &networkIdOut, IOptimizedNetworkPtr network)
Loads a complete network into the Runtime.
Definition: Runtime.cpp:150
std::vector< std::pair< LayerBindingId, class ConstTensor > > InputTensors
Definition: Tensor.hpp:392
armnn::TensorInfo GetOutputTensorInfo(NetworkId networkId, LayerBindingId layerId) const
Definition: Runtime.cpp:570
Copyright (c) 2021 ARM Limited and Contributors.
profiling::IReportStructure IReportStructure
Definition: Runtime.hpp:28
void RegisterDebugCallback(NetworkId networkId, const DebugCallbackFunction &func)
Registers a callback function to debug layers performing custom computations on intermediate tensors...
Definition: Runtime.cpp:718
const std::shared_ptr< IProfiler > GetProfiler(NetworkId networkId) const
Gets the profiler corresponding to the given network id.
Definition: Runtime.cpp:270
std::function< void(LayerGuid guid, unsigned int slotIndex, ITensorHandle *tensorHandle)> DebugCallbackFunction
Define the type of callback for the Debug layer to call.
Definition: Types.hpp:371
int LayerBindingId
Type of identifiers for bindable layers (inputs, outputs).
Definition: Types.hpp:277
friend profiling::ProfilingService & GetProfilingService(RuntimeImpl *runtime)
Definition: TestUtils.cpp:57
void ReportStructure()
Definition: Runtime.cpp:282
const IDeviceSpec & GetDeviceSpec() const
Definition: Runtime.hpp:90
int NetworkId
Definition: IRuntime.hpp:25
std::vector< std::pair< LayerBindingId, class Tensor > > OutputTensors
Definition: Tensor.hpp:393
Status
enumeration
Definition: Types.hpp:29
std::unique_ptr< IOptimizedNetwork, void(*)(IOptimizedNetwork *network)> IOptimizedNetworkPtr
Definition: INetwork.hpp:242
void ClearImportedInputs(NetworkId networkId, const std::vector< ImportedInputId > inputIds)
Definition: Runtime.cpp:587
Device specific knowledge to be passed to the optimizer.
Definition: Types.hpp:267
friend void RuntimeLoadedNetworksReserve(RuntimeImpl *runtime)
Status EnqueueWorkload(NetworkId networkId, const InputTensors &inputTensors, const OutputTensors &outputTensors, std::vector< ImportedInputId > preImportedInputIds={}, std::vector< ImportedOutputId > preImportedOutputIds={})
Definition: Runtime.cpp:596
MemorySource
Define the Memory Source to reduce copies.
Definition: Types.hpp:217
std::vector< ImportedInputId > ImportInputs(NetworkId networkId, const InputTensors &inputTensors, MemorySource forceImportMemorySource=MemorySource::Undefined)
Definition: Runtime.cpp:575
RuntimeImpl(const IRuntime::CreationOptions &options)
Creates a runtime for workload execution.
Definition: Runtime.cpp:297
std::unordered_map< NetworkId, std::unique_ptr< LoadedNetwork > > LoadedNetworks
Definition: Runtime.hpp:27
std::unique_ptr< IWorkingMemHandle > CreateWorkingMemHandle(NetworkId networkId)
Create a new unique WorkingMemHandle object.
Definition: Runtime.cpp:687
std::vector< ImportedOutputId > ImportOutputs(NetworkId networkId, const OutputTensors &outputTensors, MemorySource forceImportMemorySource=MemorySource::Undefined)
Definition: Runtime.cpp:581
void ClearImportedOutputs(NetworkId networkId, const std::vector< ImportedOutputId > outputIds)
Definition: Runtime.cpp:591