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	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 >
const DepthwiseConvolution2dQueueDescriptor	m_Data
const profiling::ProfilingGuid	m_Guid

Detailed Description

Definition at line 24 of file NeonDepthwiseConvolutionWorkload.hpp.

Constructor & Destructor Documentation

NeonDepthwiseConvolutionWorkload()

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

Definition at line 72 of file NeonDepthwiseConvolutionWorkload.cpp.

References armnn::ConvertWeightTensorFromArmnnToAcl(), TensorInfo::GetNumBytes(), ConstCpuTensorHandle::GetTensorInfo(), BaseWorkload< DepthwiseConvolution2dQueueDescriptor >::m_Data, DepthwiseConvolution2dDescriptor::m_DataLayout, QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters, and DepthwiseConvolution2dQueueDescriptor::m_Weight.

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

    // Allocate a buffer for the swizzling of the weight tensor
    std::unique_ptr<unsigned char[]> permuteBuffer(new unsigned char[m_Data.m_Weight->GetTensorInfo().GetNumBytes()]);

    // Convert the weight format from ArmNN's [ M, I, H, W ] (does NOT depend on the data layout) to either
    // [ 1, H, W, I * M ] (if NHWC) or [ 1, I * M, H, W ] (if NCHW), as required by the compute library
    ConstTensor weightPermuted = ConvertWeightTensorFromArmnnToAcl(m_Data.m_Weight,
                                                                   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, weightPermuted.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);

    // Get the depth multiplier
    const unsigned int depthMultiplier = weightInfo.GetShape()[0];

    arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);

    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,
                                                        arm_compute::ActivationLayerInfo(),
                                                        aclDilationInfo);

    BOOST_ASSERT(m_pDepthwiseConvolutionLayer);

    ScopedCpuTensorHandle weightsPermutedHandle(weightPermuted);
    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 144 of file NeonDepthwiseConvolutionWorkload.cpp.

References ARMNN_SCOPED_PROFILING_EVENT_NEON.

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonDepthwiseConvolutionWorkload_Execute");
    BOOST_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