ArmNN
 21.11
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  std::vector<ImportedOutputId> ImportOutputs(NetworkId networkId, const OutputTensors& outputTensors);
60 
61  void ClearImportedInputs(NetworkId networkId, const std::vector<ImportedInputId> inputIds);
62  void ClearImportedOutputs(NetworkId networkId, const std::vector<ImportedOutputId> outputIds);
63 
64  // Evaluates network using input in inputTensors, outputs filled into outputTensors.
66  const InputTensors& inputTensors,
67  const OutputTensors& outputTensors);
68 
69  /// This is an experimental function.
70  /// Evaluates a network using input in inputTensors and outputs filled into outputTensors.
71  /// This function performs a thread safe execution of the network. Returns once execution is complete.
72  /// Will block until this and any other thread using the same workingMem object completes.
73  Status Execute(IWorkingMemHandle& workingMemHandle,
74  const InputTensors& inputTensors,
75  const OutputTensors& outputTensors,
76  std::vector<ImportedInputId> preImportedInputs,
77  std::vector<ImportedOutputId> preImportedOutputs);
78 
79  /// Unloads a network from the Runtime.
80  /// At the moment this only removes the network from the m_Impl->m_Network.
81  /// This might need more work in the future to be AndroidNN compliant.
82  /// @param [in] networkId Unique identifier for the network to be unloaded. Generated in LoadNetwork().
83  /// @return armnn::Status
84  Status UnloadNetwork(NetworkId networkId);
85 
86  const IDeviceSpec& GetDeviceSpec() const { return m_DeviceSpec; }
87 
88  /// Gets the profiler corresponding to the given network id.
89  /// @param networkId The id of the network for which to get the profile.
90  /// @return A pointer to the requested profiler, or nullptr if not found.
91  const std::shared_ptr<IProfiler> GetProfiler(NetworkId networkId) const;
92 
93  /// Create a new unique WorkingMemHandle object. Create multiple handles if you wish to have
94  /// overlapped Execution by calling this function from different threads.
95  std::unique_ptr<IWorkingMemHandle> CreateWorkingMemHandle(NetworkId networkId);
96 
97  /// Registers a callback function to debug layers performing custom computations on intermediate tensors.
98  /// @param networkId The id of the network to register the callback.
99  /// @param func callback function to pass to the debug layer.
100  void RegisterDebugCallback(NetworkId networkId, const DebugCallbackFunction& func);
101 
102  /// Creates a runtime for workload execution.
103  RuntimeImpl(const IRuntime::CreationOptions& options);
104 
105  ~RuntimeImpl();
106 
107  //NOTE: we won't need the profiling service reference but it is good to pass the service
108  // in this way to facilitate other implementations down the road
109  void ReportStructure();
110 
111 private:
112  friend void RuntimeLoadedNetworksReserve(RuntimeImpl* runtime); // See RuntimeTests.cpp
113 
114  friend profiling::ProfilingService& GetProfilingService(RuntimeImpl* runtime); // See RuntimeTests.cpp
115 
116  int GenerateNetworkId();
117 
118  LoadedNetwork* GetLoadedNetworkPtr(NetworkId networkId) const;
119 
120  template<typename Func>
121  void LoadedNetworkFuncSafe(NetworkId networkId, Func f)
122  {
123  std::lock_guard<std::mutex> lockGuard(m_Mutex);
124  auto iter = m_LoadedNetworks.find(networkId);
125  if (iter != m_LoadedNetworks.end())
126  {
127  f(iter->second.get());
128  }
129  }
130 
131  /// Loads any available/compatible dynamic backend in the runtime.
132  void LoadDynamicBackends(const std::string& overrideBackendPath);
133 
134  mutable std::mutex m_Mutex;
135 
136  /// Map of Loaded Networks with associated GUID as key
137  LoadedNetworks m_LoadedNetworks;
138 
139  std::unordered_map<BackendId, IBackendInternal::IBackendContextPtr> m_BackendContexts;
140 
141  int m_NetworkIdCounter;
142 
143  DeviceSpec m_DeviceSpec;
144 
145  /// List of dynamic backends loaded in the runtime
146  std::vector<DynamicBackendPtr> m_DynamicBackends;
147 
148  /// Profiling Service Instance
149  profiling::ProfilingService m_ProfilingService;
150 };
151 
152 } // namespace armnn
armnn::TensorInfo GetInputTensorInfo(NetworkId networkId, LayerBindingId layerId) const
Definition: Runtime.cpp:559
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:621
Status UnloadNetwork(NetworkId networkId)
Unloads a network from the Runtime.
Definition: Runtime.cpp:209
Status LoadNetwork(NetworkId &networkIdOut, IOptimizedNetworkPtr network)
Loads a complete network into the Runtime.
Definition: Runtime.cpp:145
std::vector< ImportedInputId > ImportInputs(NetworkId networkId, const InputTensors &inputTensors)
Definition: Runtime.cpp:569
std::vector< std::pair< LayerBindingId, class ConstTensor > > InputTensors
Definition: Tensor.hpp:392
armnn::TensorInfo GetOutputTensorInfo(NetworkId networkId, LayerBindingId layerId) const
Definition: Runtime.cpp:564
Copyright (c) 2021 ARM Limited and Contributors.
Status EnqueueWorkload(NetworkId networkId, const InputTensors &inputTensors, const OutputTensors &outputTensors)
Definition: Runtime.cpp:588
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:684
const std::shared_ptr< IProfiler > GetProfiler(NetworkId networkId) const
Gets the profiler corresponding to the given network id.
Definition: Runtime.cpp:264
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:357
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:276
const IDeviceSpec & GetDeviceSpec() const
Definition: Runtime.hpp:86
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:198
void ClearImportedInputs(NetworkId networkId, const std::vector< ImportedInputId > inputIds)
Definition: Runtime.cpp:579
Device specific knowledge to be passed to the optimizer.
Definition: Types.hpp:267
friend void RuntimeLoadedNetworksReserve(RuntimeImpl *runtime)
RuntimeImpl(const IRuntime::CreationOptions &options)
Creates a runtime for workload execution.
Definition: Runtime.cpp:291
std::vector< ImportedOutputId > ImportOutputs(NetworkId networkId, const OutputTensors &outputTensors)
Definition: Runtime.cpp:574
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:653
void ClearImportedOutputs(NetworkId networkId, const std::vector< ImportedOutputId > outputIds)
Definition: Runtime.cpp:583