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 BaseWorkload< DepthwiseConvolution2dQueueDescriptor >
	BaseWorkload (const DepthwiseConvolution2dQueueDescriptor &descriptor, const WorkloadInfo &info)
void	ExecuteAsync (WorkingMemDescriptor &workingMemDescriptor) override
void	PostAllocationConfigure () override
const DepthwiseConvolution2dQueueDescriptor &	GetData () const
profiling::ProfilingGuid	GetGuid () const final
Public Member Functions inherited from IWorkload
virtual	~IWorkload ()
virtual void	RegisterDebugCallback (const DebugCallbackFunction &)

Additional Inherited Members
Protected Attributes inherited from BaseWorkload< DepthwiseConvolution2dQueueDescriptor >
DepthwiseConvolution2dQueueDescriptor	m_Data
const profiling::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 76 of file NeonDepthwiseConvolutionWorkload.cpp.

References armnn::Convert1HWOTensorToAcl(), ConstTensorHandle::GetTensorInfo(), BaseWorkload< DepthwiseConvolution2dQueueDescriptor >::m_Data, DepthwiseConvolution2dDescriptor::m_DataLayout, WorkloadInfo::m_InputTensorInfos, QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters, and DepthwiseConvolution2dQueueDescriptor::m_Weight.

    : BaseWorkload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info)
{
    // ArmNN's weight format for depthwise is [ 1, H, W, I*M ]
    auto& weightInfo = m_Data.m_Weight->GetTensorInfo();

    ConstTensor weightsPermuted;
    unsigned int depthMultiplier;
    std::unique_ptr<unsigned char[]> permuteBuffer(new unsigned char[weightInfo.GetNumBytes()]);
    std::tie(weightsPermuted, depthMultiplier) = Convert1HWOTensorToAcl(m_Data.m_Weight,
                                                                        info.m_InputTensorInfos[0],
                                                                        m_Data.m_Parameters.m_DataLayout,
                                                                        permuteBuffer.get());

    // Convert the weights into the compute library format
    m_KernelTensor = std::make_unique<arm_compute::Tensor>();
    BuildArmComputeTensor(*m_KernelTensor, weightsPermuted.GetInfo(), m_Data.m_Parameters.m_DataLayout);

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

    const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(
        m_Data.m_Parameters.m_DilationX, m_Data.m_Parameters.m_DilationY);

    m_Data.ValidateInputsOutputs("NeonDepthwiseConvolutionWorkload", 1, 1);

    IAclTensorHandle* inputTensorHandle = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0]);
    IAclTensorHandle* outputTensorHandle = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0]);

    arm_compute::ITensor& input = inputTensorHandle->GetTensor();
    arm_compute::ITensor& output = outputTensorHandle->GetTensor();

    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);

    m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
    static_cast<arm_compute::NEDepthwiseConvolutionLayer*>(
        m_pDepthwiseConvolutionLayer.get())->configure(&input,
                                                       m_KernelTensor.get(),
                                                       m_BiasTensor.get(),
                                                       &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>(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("NeonDepthwiseConvolution2dWorkload_Construct",
                                         descriptor.m_Parameters,
                                         detailsInfo,
                                         this->GetGuid());

    ARMNN_ASSERT(m_pDepthwiseConvolutionLayer);

    ScopedTensorHandle weightsPermutedHandle(weightsPermuted);
    InitializeArmComputeTensorData(*m_KernelTensor, &weightsPermutedHandle);

    if (m_Data.m_Parameters.m_BiasEnabled)
    {
        InitializeArmComputeTensorData(*m_BiasTensor, m_Data.m_Bias);
    }

    m_pDepthwiseConvolutionLayer->prepare();
    FreeUnusedTensors();
}

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 163 of file NeonDepthwiseConvolutionWorkload.cpp.

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

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

    m_pDepthwiseConvolutionLayer->run();
}

---

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

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