#include <NeonConvolution2dWorkload.hpp>

Inheritance diagram for NeonConvolution2dWorkload:

Collaboration diagram for NeonConvolution2dWorkload:

Public Member Functions
	NeonConvolution2dWorkload (const Convolution2dQueueDescriptor &descriptor, const WorkloadInfo &info, std::shared_ptr< arm_compute::MemoryManagerOnDemand > &memoryManager, const bool isFastMathENabled=false)

void	Execute () const override

arm_compute::ConvolutionMethod	GetConvolutionMethod () const

Public Member Functions inherited from NeonBaseWorkload< Convolution2dQueueDescriptor >
	NeonBaseWorkload (const Convolution2dQueueDescriptor &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< QueueDescriptor >
	BaseWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)

virtual const std::string &	GetName () const override

void	PostAllocationConfigure () override

const QueueDescriptor &	GetData () const

arm::pipe::ProfilingGuid	GetGuid () const final

virtual bool	SupportsTensorHandleReplacement () const override

Public Member Functions inherited from IWorkload
virtual	~IWorkload ()

virtual void	RegisterDebugCallback (const DebugCallbackFunction &)

virtual armnn::Optional< armnn::MemoryRequirements >	GetMemoryRequirements ()

Additional Inherited Members
Protected Member Functions inherited from NeonBaseWorkload< Convolution2dQueueDescriptor >
virtual void	Reconfigure ()

Protected Attributes inherited from BaseWorkload< QueueDescriptor >
QueueDescriptor	m_Data

const arm::pipe::ProfilingGuid	m_Guid

const std::string	m_Name

Detailed Description

Definition at line 27 of file NeonConvolution2dWorkload.hpp.

Constructor & Destructor Documentation

◆ NeonConvolution2dWorkload()

NeonConvolution2dWorkload	(	const Convolution2dQueueDescriptor &	descriptor,
		const WorkloadInfo &	info,
		std::shared_ptr< arm_compute::MemoryManagerOnDemand > &	memoryManager,
		const bool	isFastMathENabled = `false`
	)

Definition at line 71 of file NeonConvolution2dWorkload.cpp.

     : NeonBaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
 {
     using arm_compute::NEConvolutionLayer;
  
     uint32_t numInputs = m_Data.m_Parameters.m_BiasEnabled ? 3: 2;
     m_Data.ValidateInputsOutputs("NeonConvolution2dWorkload", numInputs, 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();
  
     arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
     input.info()->set_data_layout(aclDataLayout);
     output.info()->set_data_layout(aclDataLayout);
  
     m_KernelTensor = std::make_unique<arm_compute::Tensor>();
     BuildArmComputeTensor(*m_KernelTensor, info.m_InputTensorInfos[1], m_Data.m_Parameters.m_DataLayout);
     m_KernelTensor->info()->set_are_values_constant(info.m_InputTensorInfos[1].IsConstant());
     if (m_Data.m_Parameters.m_BiasEnabled)
     {
         m_BiasTensor = std::make_unique<arm_compute::Tensor>();
         BuildArmComputeTensor(*m_BiasTensor, info.m_InputTensorInfos[2], m_Data.m_Parameters.m_DataLayout);
         m_BiasTensor->info()->set_are_values_constant(info.m_InputTensorInfos[2].IsConstant());
     }
  
     arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
  
     const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(m_Data.m_Parameters.m_DilationX,
                                                                       m_Data.m_Parameters.m_DilationY);
  
     const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
  
     auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(memoryManager);
     convolutionLayer->configure(&input,
                                 m_KernelTensor.get(),
                                 m_BiasTensor.get(),
                                 &output,
                                 padStrideInfo,
                                 arm_compute::WeightsInfo(),
                                 aclDilationInfo,
                                 activationInfo,
                                 isFastMathEnabled);
  
     m_ConvolutionMethod =
         convolutionLayer->get_convolution_method(input.info(),
                                                  m_KernelTensor->info(),
                                                  output.info(),
                                                  padStrideInfo,
                                                  arm_compute::WeightsInfo(),
                                                  aclDilationInfo,
                                                  activationInfo,
                                                  isFastMathEnabled);
  
     // Add details for profiling output
     WorkloadInfo detailsInfo;
  
     detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
     detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
     detailsInfo.m_ConvolutionMethod = armnn::Optional<std::string>(GetConvolutionMethodString(m_ConvolutionMethod));
  
     // Report Profiling Details
     ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonConvolution2dWorkload_Construct",
                                          descriptor.m_Parameters,
                                          detailsInfo,
                                          GetGuid());
  
     m_ConvolutionLayer.reset(convolutionLayer.release());
     m_KernelTensorInfo = info.m_InputTensorInfos[1];
  
     if (m_Data.m_Parameters.m_BiasEnabled)
     {
         m_BiasTensorInfo = info.m_InputTensorInfos[2];
     }
 }

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

Member Function Documentation

◆ Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 150 of file NeonConvolution2dWorkload.cpp.

 {
     ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID("NeonConvolution2dWorkload_Execute");
     // The constant tensors may not be fully in place until the workload is Executed
     if (!prepared)
     {
         InitializeArmComputeTensorData(*m_KernelTensor, m_KernelTensorInfo, m_Data.m_Inputs[1]);
  
         if (m_Data.m_Parameters.m_BiasEnabled)
         {
             InitializeArmComputeTensorData(*m_BiasTensor, m_BiasTensorInfo, m_Data.m_Inputs[2]);
         }
         m_ConvolutionLayer->prepare();
         FreeTensorIfUnused(m_KernelTensor);
         FreeTensorIfUnused(m_BiasTensor);
         prepared = true;
     }
     m_ConvolutionLayer->run();
 }

References ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID, armnn::InitializeArmComputeTensorData(), BaseWorkload< QueueDescriptor >::m_Data, and QueueDescriptor::m_Inputs.

◆ GetConvolutionMethod()

arm_compute::ConvolutionMethod GetConvolutionMethod ( ) const

Definition at line 170 of file NeonConvolution2dWorkload.cpp.

 {
     return m_ConvolutionMethod;
 }

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

src/backends/neon/workloads/NeonConvolution2dWorkload.hpp
src/backends/neon/workloads/NeonConvolution2dWorkload.cpp