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 (ExecutionData &executionData) override
void	PostAllocationConfigure () override
const FullyConnectedQueueDescriptor &	GetData () const
arm::pipe::ProfilingGuid	GetGuid () const final
virtual bool	SupportsTensorHandleReplacement () const override
Public Member Functions inherited from IWorkload
virtual	~IWorkload ()
virtual arm::pipe::ProfilingGuid	GetGuid () const =0
virtual bool	SupportsTensorHandleReplacement () const =0
virtual void	RegisterDebugCallback (const DebugCallbackFunction &)
virtual armnn::Optional< armnn::MemoryRequirements >	GetMemoryRequirements ()

Additional Inherited Members
Protected Member Functions inherited from NeonBaseWorkload< FullyConnectedQueueDescriptor >
virtual void	Reconfigure ()
Protected Attributes inherited from BaseWorkload< FullyConnectedQueueDescriptor >
FullyConnectedQueueDescriptor	m_Data
const arm::pipe::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 61 of file NeonFullyConnectedWorkload.cpp.

    : 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();
 
    // Copy the weights' tensor into arm_compute tensor.
    m_WeightsTensor = std::make_unique<arm_compute::Tensor>();
    m_WeightsTensorInfo = info.m_InputTensorInfos[1];
    BuildArmComputeTensor(*m_WeightsTensor, m_WeightsTensorInfo);
 
    if (m_Data.m_Parameters.m_BiasEnabled)
    {
        // Copy the biases tensor into arm_compute tensor.
        m_BiasesTensor = std::make_unique<arm_compute::Tensor>();
        m_BiasesTensorInfo = info.m_InputTensorInfos[2];
        BuildArmComputeTensor(*m_BiasesTensor, m_BiasesTensorInfo);
    }
 
    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());
 
    // 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>(info.m_InputTensorInfos[1]);
    if (descriptor.m_Parameters.m_BiasEnabled)
    {
        detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]);
    }
 
    // 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.
}

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

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 112 of file NeonFullyConnectedWorkload.cpp.

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonFullyConnectedWorkload_Execute", this->GetGuid());
    // The constant tensors may not be fully in place until the workload is Executed
    if (!prepared)
    {
        InitializeArmComputeTensorData(*m_WeightsTensor, m_WeightsTensorInfo, m_Data.m_Inputs[1]);
 
        if (m_Data.m_Parameters.m_BiasEnabled)
        {
            InitializeArmComputeTensorData(*m_BiasesTensor, m_BiasesTensorInfo, m_Data.m_Inputs[2]);
        }
        m_FullyConnectedLayer->prepare();
        FreeTensorIfUnused(m_WeightsTensor);
        FreeTensorIfUnused(m_BiasesTensor);
        prepared = true;
    }
    m_FullyConnectedLayer->run();
}

References ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID, BaseWorkload< FullyConnectedQueueDescriptor >::GetGuid(), armnn::InitializeArmComputeTensorData(), FullyConnectedDescriptor::m_BiasEnabled, BaseWorkload< FullyConnectedQueueDescriptor >::m_Data, QueueDescriptor::m_Inputs, and QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters.

---

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

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