NeonFullyConnectedWorkload Class Reference

#include <NeonFullyConnectedWorkload.hpp>

Inheritance diagram for NeonFullyConnectedWorkload:

Public Member Functions
	NeonFullyConnectedWorkload (const FullyConnectedQueueDescriptor &descriptor, const WorkloadInfo &info, std::shared_ptr< arm_compute::MemoryManagerOnDemand > &memoryManager)
virtual void	Execute () const override
Public Member Functions inherited from NeonBaseWorkload< FullyConnectedQueueDescriptor >
	NeonBaseWorkload (const FullyConnectedQueueDescriptor &descriptor, const WorkloadInfo &info)
void	ReplaceInputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
void	ReplaceOutputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
Public Member Functions inherited from BaseWorkload< FullyConnectedQueueDescriptor >
	BaseWorkload (const FullyConnectedQueueDescriptor &descriptor, const WorkloadInfo &info)
void	ExecuteAsync (WorkingMemDescriptor &workingMemDescriptor) override
void	PostAllocationConfigure () override
const FullyConnectedQueueDescriptor &	GetData () const
profiling::ProfilingGuid	GetGuid () const final
virtual bool	SupportsTensorHandleReplacement () const override
Public Member Functions inherited from IWorkload
virtual	~IWorkload ()
virtual void	RegisterDebugCallback (const DebugCallbackFunction &)

Additional Inherited Members
Protected Member Functions inherited from NeonBaseWorkload< FullyConnectedQueueDescriptor >
virtual void	Reconfigure ()
Protected Attributes inherited from BaseWorkload< FullyConnectedQueueDescriptor >
FullyConnectedQueueDescriptor	m_Data
const profiling::ProfilingGuid	m_Guid

Detailed Description

Definition at line 27 of file NeonFullyConnectedWorkload.hpp.

Constructor & Destructor Documentation

NeonFullyConnectedWorkload()

NeonFullyConnectedWorkload	(	const FullyConnectedQueueDescriptor &	descriptor,
		const WorkloadInfo &	info,
		std::shared_ptr< arm_compute::MemoryManagerOnDemand > &	memoryManager
	)

Definition at line 54 of file NeonFullyConnectedWorkload.cpp.

References BaseWorkload< FullyConnectedQueueDescriptor >::m_Data, QueueDescriptor::m_Inputs, QueueDescriptor::m_Outputs, and QueueDescriptor::ValidateInputsOutputs().

    : NeonBaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info)
{
    m_Data.ValidateInputsOutputs("NeonFullyConnectedWorkload", 1, 1);

    arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
    arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();

    m_WeightsTensor = std::make_unique<arm_compute::Tensor>();
    BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo());

    if (m_Data.m_Parameters.m_BiasEnabled)
    {
        m_BiasesTensor = std::make_unique<arm_compute::Tensor>();
        BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo());
    }

    const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);

    arm_compute::FullyConnectedLayerInfo fc_info =
        ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor.m_Parameters, activationInfo);

    auto layer = std::make_unique<arm_compute::NEFullyConnectedLayer>(memoryManager);
    layer->configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info);
    m_FullyConnectedLayer.reset(layer.release());

    // Allocate
    if (m_Data.m_Weight->GetTensorInfo().GetDataType() == DataType::QAsymmU8)
    {
        InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight);
    }
    else
    {
        InitializeArmComputeTensorData(*m_WeightsTensor, m_Data.m_Weight);
    }

    if (m_BiasesTensor)
    {
        if (m_Data.m_Bias->GetTensorInfo().GetDataType() == DataType::Signed32)
        {
            InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias);
        }
        else
        {
            InitializeArmComputeTensorData(*m_BiasesTensor, m_Data.m_Bias);
        }
    }

    // Add details for profiling output
    WorkloadInfo detailsInfo;

    detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
    detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
    detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Weight->GetTensorInfo());
    if (descriptor.m_Parameters.m_BiasEnabled)
    {
        detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(descriptor.m_Bias->GetTensorInfo());
    }

    // Report Profiling Details
    ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonFullyConnectedWorkload_Construct",
                                         descriptor.m_Parameters,
                                         detailsInfo,
                                         this->GetGuid());

    // Force Compute Library to perform the necessary copying and reshaping, after which
    // delete all the input tensors that will no longer be needed
    m_FullyConnectedLayer->prepare();
    FreeUnusedTensors();
}

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 127 of file NeonFullyConnectedWorkload.cpp.

References ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID, and BaseWorkload< FullyConnectedQueueDescriptor >::GetGuid().

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonFullyConnectedWorkload_Execute", this->GetGuid());
    m_FullyConnectedLayer->run();
}

---

The documentation for this class was generated from the following files:

• src/backends/neon/workloads/NeonFullyConnectedWorkload.hpp
• src/backends/neon/workloads/NeonFullyConnectedWorkload.cpp