ArmNN
 24.08
AsyncExecutionSample.cpp

Yet another variant of the SimpleSample application. In this little sample app you will be shown how to run a network multiple times asynchronously.

Note
This is currently an experimental interface
//
// Copyright © 2021 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include <armnn/Utils.hpp>
#include <iostream>
#include <thread>
/// A simple example of using the ArmNN SDK API to run a network multiple times with different inputs in an asynchronous
/// manner.
///
/// Background info: The usual runtime->EnqueueWorkload, which is used to trigger the execution of a network, is not
/// thread safe. Each workload has memory assigned to it which would be overwritten by each thread.
/// Before we added support for this you had to load a network multiple times to execute it at the
/// same time. Every time a network is loaded, it takes up memory on your device. Making the
/// execution thread safe helps to reduce the memory footprint for concurrent executions significantly.
/// This example shows you how to execute a model concurrently (multiple threads) while still only
/// loading it once.
///
/// As in most of our simple samples, the network in this example will ask the user for a single input number for each
/// execution of the network.
/// The network consists of a single fully connected layer with a single neuron. The neurons weight is set to 1.0f
/// to produce an output number that is the same as the input.
int main()
{
using namespace armnn;
// The first part of this code is very similar to the SimpleSample.cpp you should check it out for comparison
// The interesting part starts when the graph is loaded into the runtime
std::vector<float> inputs;
float number1;
std::cout << "Please enter a number for the first iteration: " << std::endl;
std::cin >> number1;
float number2;
std::cout << "Please enter a number for the second iteration: " << std::endl;
std::cin >> number2;
// Turn on logging to standard output
// This is useful in this sample so that users can learn more about what is going on
// Construct ArmNN network
NetworkId networkIdentifier;
float weightsData[] = {1.0f}; // Identity
TensorInfo weightsInfo(TensorShape({1, 1}), DataType::Float32, 0.0f, 0, true);
weightsInfo.SetConstant();
ConstTensor weights(weightsInfo, weightsData);
// Constant layer that now holds weights data for FullyConnected
IConnectableLayer* const constantWeightsLayer = myNetwork->AddConstantLayer(weights, "const weights");
FullyConnectedDescriptor fullyConnectedDesc;
IConnectableLayer* const fullyConnectedLayer = myNetwork->AddFullyConnectedLayer(fullyConnectedDesc,
"fully connected");
IConnectableLayer* InputLayer = myNetwork->AddInputLayer(0);
IConnectableLayer* OutputLayer = myNetwork->AddOutputLayer(0);
InputLayer->GetOutputSlot(0).Connect(fullyConnectedLayer->GetInputSlot(0));
constantWeightsLayer->GetOutputSlot(0).Connect(fullyConnectedLayer->GetInputSlot(1));
fullyConnectedLayer->GetOutputSlot(0).Connect(OutputLayer->GetInputSlot(0));
// Create ArmNN runtime
IRuntime::CreationOptions options; // default options
IRuntimePtr run = IRuntime::Create(options);
//Set the tensors in the network.
TensorInfo inputTensorInfo(TensorShape({1, 1}), DataType::Float32);
InputLayer->GetOutputSlot(0).SetTensorInfo(inputTensorInfo);
TensorInfo outputTensorInfo(TensorShape({1, 1}), DataType::Float32);
fullyConnectedLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
constantWeightsLayer->GetOutputSlot(0).SetTensorInfo(weightsInfo);
// Optimise ArmNN network
IOptimizedNetworkPtr optNet = Optimize(*myNetwork, {Compute::CpuRef}, run->GetDeviceSpec());
if (!optNet)
{
// This shouldn't happen for this simple sample, with reference backend.
// But in general usage Optimize could fail if the hardware at runtime cannot
// support the model that has been provided.
std::cerr << "Error: Failed to optimise the input network." << std::endl;
return 1;
}
// Load graph into runtime.
std::string errmsg; // To hold an eventual error message if loading the network fails
// Add network properties to enable async execution. The MemorySource::Undefined variables indicate
// that neither inputs nor outputs will be imported. Importing will be covered in another example.
run->LoadNetwork(networkIdentifier,
std::move(optNet),
errmsg,
networkProperties);
// Creates structures for inputs and outputs. A vector of float for each execution.
std::vector<std::vector<float>> inputData{{number1}, {number2}};
std::vector<std::vector<float>> outputData;
outputData.resize(2, std::vector<float>(1));
inputTensorInfo = run->GetInputTensorInfo(networkIdentifier, 0);
inputTensorInfo.SetConstant(true);
std::vector<InputTensors> inputTensors
{
{{0, armnn::ConstTensor(inputTensorInfo, inputData[0].data())}},
{{0, armnn::ConstTensor(inputTensorInfo, inputData[1].data())}}
};
std::vector<OutputTensors> outputTensors
{
{{0, armnn::Tensor(run->GetOutputTensorInfo(networkIdentifier, 0), outputData[0].data())}},
{{0, armnn::Tensor(run->GetOutputTensorInfo(networkIdentifier, 0), outputData[1].data())}}
};
// Lambda function to execute the network. We use it as thread function.
auto execute = [&](unsigned int executionIndex)
{
auto memHandle = run->CreateWorkingMemHandle(networkIdentifier);
run->Execute(*memHandle, inputTensors[executionIndex], outputTensors[executionIndex]);
};
// Prepare some threads and let each execute the network with a different input
std::vector<std::thread> threads;
for (unsigned int i = 0; i < inputTensors.size(); ++i)
{
threads.emplace_back(std::thread(execute, i));
}
// Wait for the threads to finish
for (std::thread& t : threads)
{
if(t.joinable())
{
t.join();
}
}
std::cout << "Your numbers were " << outputData[0][0] << " and " << outputData[1][0] << std::endl;
return 0;
}
armnn::InputLayer
A layer user-provided data can be bound to (e.g. inputs, outputs).
Definition: InputLayer.hpp:13
armnn::INetworkPtr
std::unique_ptr< INetwork, void(*)(INetwork *network)> INetworkPtr
Definition: INetwork.hpp:339
armnn::IOptimizedNetworkPtr
std::unique_ptr< IOptimizedNetwork, void(*)(IOptimizedNetwork *network)> IOptimizedNetworkPtr
Definition: INetwork.hpp:340
armnn::Tensor
A tensor defined by a TensorInfo (shape and data type) and a mutable backing store.
Definition: Tensor.hpp:321
IRuntime.hpp
armnn::FullyConnectedDescriptor
A FullyConnectedDescriptor for the FullyConnectedLayer.
Definition: Descriptors.hpp:507
Descriptors.hpp
armnn::Compute::CpuRef
@ CpuRef
CPU Execution: Reference C++ kernels.
armnn::OutputSlot::SetTensorInfo
void SetTensorInfo(const TensorInfo &tensorInfo) override
Definition: Layer.cpp:95
armnn::TensorInfo
Definition: Tensor.hpp:152
armnn::ConfigureLogging
void ConfigureLogging(bool printToStandardOutput, bool printToDebugOutput, LogSeverity severity)
Configures the logging behaviour of the ARMNN library.
Definition: Utils.cpp:20
armnn::DataType::Float32
@ Float32
armnn::Layer::GetOutputSlot
const OutputSlot & GetOutputSlot(unsigned int index=0) const override
Get the const output slot handle by slot index.
Definition: Layer.hpp:339
ARMNN_NO_DEPRECATE_WARN_BEGIN
#define ARMNN_NO_DEPRECATE_WARN_BEGIN
Definition: Deprecated.hpp:33
armnn::OutputSlot::Connect
int Connect(InputSlot &destination)
Definition: Layer.cpp:123
armnn::Layer::GetInputSlot
const InputSlot & GetInputSlot(unsigned int index) const override
Get a const input slot handle by slot index.
Definition: Layer.hpp:337
armnn::TensorShape
Definition: Tensor.hpp:20
armnn::NetworkId
int NetworkId
Definition: IRuntime.hpp:35
INetwork.hpp
armnn::IRuntimePtr
std::unique_ptr< IRuntime, void(*)(IRuntime *runtime)> IRuntimePtr
Definition: IRuntime.hpp:41
armnn::IOutputSlot::SetTensorInfo
virtual void SetTensorInfo(const TensorInfo &tensorInfo)=0
armnn::MemorySource::Undefined
@ Undefined
Utils.hpp
armnn::OutputLayer
A layer user-provided data can be bound to (e.g. inputs, outputs).
Definition: OutputLayer.hpp:13
armnn::INetworkProperties
Definition: IRuntime.hpp:43
main
int main(int argc, char *argv[])
Definition: ArmnnConverter.cpp:329
armnn::IOutputSlot::Connect
virtual int Connect(IInputSlot &destination)=0
armnn::IRuntime::CreationOptions
Definition: IRuntime.hpp:78
ARMNN_NO_DEPRECATE_WARN_END
#define ARMNN_NO_DEPRECATE_WARN_END
Definition: Deprecated.hpp:34
armnn::IRuntime::Create
static IRuntimePtr Create(const CreationOptions &options)
Definition: Runtime.cpp:52
armnn::LogSeverity::Warning
@ Warning
armnn::IConnectableLayer::GetOutputSlot
virtual const IOutputSlot & GetOutputSlot(unsigned int index) const =0
Get the const output slot handle by slot index.
armnn
Copyright (c) 2021 ARM Limited and Contributors.
Definition: 01_00_quick_start.dox:6
armnn::IConnectableLayer::GetInputSlot
virtual const IInputSlot & GetInputSlot(unsigned int index) const =0
Get a const input slot handle by slot index.
armnn::ConstTensor
A tensor defined by a TensorInfo (shape and data type) and an immutable backing store.
Definition: Tensor.hpp:329
armnn::IConnectableLayer
Interface for a layer that is connectable to other layers via InputSlots and OutputSlots.
Definition: INetwork.hpp:80
armnn::TensorInfo::SetConstant
void SetConstant(const bool IsConstant=true)
Marks the data corresponding to this tensor info as constant.
Definition: Tensor.cpp:518
armnn::Optimize
IOptimizedNetworkPtr Optimize(const INetwork &network, const std::vector< BackendId > &backendPreferences, const IDeviceSpec &deviceSpec, const OptimizerOptionsOpaque &options=OptimizerOptionsOpaque(), Optional< std::vector< std::string > & > messages=EmptyOptional())
Create an optimized version of the network.
Definition: Network.cpp:2145
armnn::INetwork::Create
static INetworkPtr Create(const NetworkOptions &networkOptions={})
Definition: Network.cpp:682