#include <ClConvolution2dWorkload.hpp>

Inheritance diagram for ClConvolution2dWorkload:

Public Member Functions
	ClConvolution2dWorkload (const Convolution2dQueueDescriptor &descriptor, const WorkloadInfo &info, std::shared_ptr< arm_compute::MemoryManagerOnDemand > &memoryManager, const arm_compute::CLCompileContext &clCompileContext, const bool isFastMathEnabled=false)

void	Execute () const override

arm_compute::ConvolutionMethod	GetConvolutionMethod () const

bool	SupportsTensorHandleReplacement () const override

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

void	ExecuteAsync (WorkingMemDescriptor &workingMemDescriptor) override

void	PostAllocationConfigure () override

const Convolution2dQueueDescriptor &	GetData () const

profiling::ProfilingGuid	GetGuid () const final

Public Member Functions inherited from IWorkload
virtual	~IWorkload ()

virtual void	RegisterDebugCallback (const DebugCallbackFunction &)

Protected Member Functions
void	Reconfigure () override

Additional Inherited Members
Protected Attributes inherited from BaseWorkload< Convolution2dQueueDescriptor >
Convolution2dQueueDescriptor	m_Data

const profiling::ProfilingGuid	m_Guid

Detailed Description

Definition at line 31 of file ClConvolution2dWorkload.hpp.

Constructor & Destructor Documentation

◆ ClConvolution2dWorkload()

ClConvolution2dWorkload	(	const Convolution2dQueueDescriptor &	descriptor,
		const WorkloadInfo &	info,
		std::shared_ptr< arm_compute::MemoryManagerOnDemand > &	memoryManager,
		const arm_compute::CLCompileContext &	clCompileContext,
		const bool	isFastMathEnabled = `false`
	)

Definition at line 65 of file ClConvolution2dWorkload.cpp.

References ARMNN_SCOPED_PROFILING_EVENT, ConstTensorHandle::GetTensorInfo(), BaseWorkload< Convolution2dQueueDescriptor >::m_Data, Convolution2dQueueDescriptor::m_Weight, and armnn::Undefined.

     : ClBaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
     , m_ConvolutionLayer(memoryManager)
 {
     ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClConvolution2dWorkload");
     const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
 
     m_KernelTensor = std::make_unique<arm_compute::CLTensor>();
     BuildArmComputeTensor(*m_KernelTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
 
     const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(m_Data.m_Parameters.m_DilationX,
                                                                       m_Data.m_Parameters.m_DilationY);
 
     if (m_Data.m_Parameters.m_BiasEnabled)
     {
         m_BiasTensor = std::make_unique<arm_compute::CLTensor>();
         BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
     }
 
     m_Data.ValidateInputsOutputs("ClConvolution2dWorkload", 1, 1);
 
     arm_compute::ICLTensor& input  = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
     // Create Proxy tensor and set the initial tensor handle to it
     m_InputProxy = std::make_unique<ICLTensorProxy>(&input);
     m_OutputProxy = std::make_unique<ICLTensorProxy>(&output);
 
     arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
     input.info()->set_data_layout(aclDataLayout);
     output.info()->set_data_layout(aclDataLayout);
 
     arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
 
     const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
 
     {
         ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClConvolution2dWorkload_configure");
         m_ConvolutionLayer.configure(clCompileContext,
                                      m_InputProxy.get(),
                                      m_KernelTensor.get(),
                                      m_BiasTensor.get(),
                                      m_OutputProxy.get(),
                                      padStrideInfo,
                                      arm_compute::WeightsInfo(),
                                      aclDilationInfo,
                                      activationInfo,
                                      isFastMathEnabled);
     }
 
     m_ConvolutionMethod =
         m_ConvolutionLayer.get_convolution_method(input.info(),
                                                   m_KernelTensor->info(),
                                                   output.info(),
                                                   padStrideInfo,
                                                   arm_compute::WeightsInfo(),
                                                   activationInfo,
                                                   arm_compute::CLScheduler::get().target(),
                                                   aclDilationInfo,
                                                   isFastMathEnabled);
 
      // 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());
     detailsInfo.m_ConvolutionMethod = armnn::Optional<std::string>(GetConvolutionMethodString(m_ConvolutionMethod));
     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("ClConvolution2dWorkload_Construct",
                                          descriptor.m_Parameters,
                                          detailsInfo,
                                          this->GetGuid());
 
     InitializeArmComputeClTensorData(*m_KernelTensor, m_Data.m_Weight);
 
     if (m_BiasTensor)
     {
         InitializeArmComputeClTensorData(*m_BiasTensor, m_Data.m_Bias);
     }
 
     // Force Compute Library to perform the necessary copying and reshaping, after which
     // delete all the input tensors that will no longer be needed
     {
         ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClConvolution2dWorkload_prepare");
         m_ConvolutionLayer.prepare();
     }
     FreeUnusedTensors();
 }

Member Function Documentation

◆ Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 164 of file ClConvolution2dWorkload.cpp.

References ARMNN_SCOPED_PROFILING_EVENT_CL_GUID, CHECK_LOCATION, BaseWorkload< Convolution2dQueueDescriptor >::GetGuid(), and armnn::RunClFunction().

 {
     ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClConvolution2dWorkload_Execute", this->GetGuid());
     RunClFunction(m_ConvolutionLayer, CHECK_LOCATION());
 }

◆ GetConvolutionMethod()

arm_compute::ConvolutionMethod GetConvolutionMethod ( ) const

Definition at line 170 of file ClConvolution2dWorkload.cpp.

 {
     return m_ConvolutionMethod;
 }

◆ Reconfigure()

void Reconfigure ( )

overrideprotectedvirtual

Reimplemented from ClBaseWorkload< Convolution2dQueueDescriptor >.

Definition at line 181 of file ClConvolution2dWorkload.cpp.

References BaseWorkload< Convolution2dQueueDescriptor >::m_Data, QueueDescriptor::m_Inputs, and QueueDescriptor::m_Outputs.

Referenced by ClConvolution2dWorkload::SupportsTensorHandleReplacement().

 {
     arm_compute::ICLTensor& input  = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
     arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
     m_InputProxy->set(&input);
     m_OutputProxy->set(&output);
 }