NeonDepthwiseConvolutionWorkload Class Reference

#include <NeonDepthwiseConvolutionWorkload.hpp>

Inheritance diagram for NeonDepthwiseConvolutionWorkload:

Public Member Functions
	NeonDepthwiseConvolutionWorkload (const DepthwiseConvolution2dQueueDescriptor &descriptor, const WorkloadInfo &info)
virtual void	Execute () const override
Public Member Functions inherited from NeonBaseWorkload< DepthwiseConvolution2dQueueDescriptor >
	NeonBaseWorkload (const DepthwiseConvolution2dQueueDescriptor &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< DepthwiseConvolution2dQueueDescriptor >
	BaseWorkload (const DepthwiseConvolution2dQueueDescriptor &descriptor, const WorkloadInfo &info)
void	ExecuteAsync (ExecutionData &executionData) override
void	PostAllocationConfigure () override
const DepthwiseConvolution2dQueueDescriptor &	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< DepthwiseConvolution2dQueueDescriptor >
virtual void	Reconfigure ()
Protected Attributes inherited from BaseWorkload< DepthwiseConvolution2dQueueDescriptor >
DepthwiseConvolution2dQueueDescriptor	m_Data
const arm::pipe::ProfilingGuid	m_Guid

Detailed Description

Definition at line 26 of file NeonDepthwiseConvolutionWorkload.hpp.

Constructor & Destructor Documentation

NeonDepthwiseConvolutionWorkload()

NeonDepthwiseConvolutionWorkload	(	const DepthwiseConvolution2dQueueDescriptor &	descriptor,
		const WorkloadInfo &	info
	)

Definition at line 89 of file NeonDepthwiseConvolutionWorkload.cpp.

    : NeonBaseWorkload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info)
{
    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::ITensor& weights = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
    arm_compute::ITensor* biasesPtr = nullptr;
    weights.info()->set_are_values_constant(info.m_InputTensorInfos[1].IsConstant());
    if (m_Data.m_Parameters.m_BiasEnabled)
    {
        biasesPtr = &PolymorphicDowncast<IAclTensorHandle *>(m_Data.m_Inputs[2])->GetTensor();
        biasesPtr->info()->set_are_values_constant(info.m_InputTensorInfos[2].IsConstant());
        // We do not support dynamic bias
        ARMNN_ASSERT(info.m_InputTensorInfos[2].IsConstant() == true);
    }
 
    arm_compute::ITensorInfo* weightsInfo = weights.info();
    arm_compute::ITensorInfo* inputInfo = input.info();
    auto weightsShape = weightsInfo->tensor_shape();
    auto inputShape = inputInfo->tensor_shape();
 
    // The PermuteDepthwiseConv2dWeights backend optimization has been performed,
    // converting weights to have the same data layout as input.
    unsigned int depthMultiplier =
        ComputeDepthwiseConv2dDepthMultiplier(m_Data.m_Parameters.m_DataLayout, weightsShape, inputShape);
 
    const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
        m_Data.m_Parameters.m_DilationX, m_Data.m_Parameters.m_DilationY);
 
    uint32_t numInputs = m_Data.m_Parameters.m_BiasEnabled ? 3: 2;
    m_Data.ValidateInputsOutputs("NeonDepthwiseConvolutionWorkload", numInputs, 1);
 
    arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
    input.info()->set_data_layout(aclDataLayout);
    weights.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);
 
    m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
    static_cast<arm_compute::NEDepthwiseConvolutionLayer*>(
        m_pDepthwiseConvolutionLayer.get())->configure(&input,
                                                       &weights,
                                                       biasesPtr,
                                                       &output,
                                                       padStrideInfo,
                                                       depthMultiplier,
                                                       activationInfo,
                                                       aclDilationInfo);
 
    // 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("NeonDepthwiseConvolution2dWorkload_Construct",
                                         descriptor.m_Parameters,
                                         detailsInfo,
                                         GetGuid());
 
    ARMNN_ASSERT(m_pDepthwiseConvolutionLayer);
 
    m_pDepthwiseConvolutionLayer->prepare();
}

References ARMNN_ASSERT, armnn::info, DepthwiseConvolution2dDescriptor::m_BiasEnabled, BaseWorkload< DepthwiseConvolution2dQueueDescriptor >::m_Data, QueueDescriptor::m_Inputs, QueueDescriptor::m_Outputs, and QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters.

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 165 of file NeonDepthwiseConvolutionWorkload.cpp.

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonDepthwiseConvolutionWorkload_Execute", GetGuid());
    ARMNN_ASSERT(m_pDepthwiseConvolutionLayer);
 
    m_pDepthwiseConvolutionLayer->run();
}

References ARMNN_ASSERT, ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID, and BaseWorkload< DepthwiseConvolution2dQueueDescriptor >::GetGuid().

---

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

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