ClDepthwiseConvolutionWorkload Class Reference

#include <ClDepthwiseConvolutionWorkload.hpp>

Inheritance diagram for ClDepthwiseConvolutionWorkload:

Public Member Functions
	ClDepthwiseConvolutionWorkload (const DepthwiseConvolution2dQueueDescriptor &descriptor, const WorkloadInfo &info, const arm_compute::CLCompileContext &clCompileContext)
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 &)

Protected Member Functions
void	FreeUnusedTensors ()

Protected Attributes
std::unique_ptr< arm_compute::IFunction >	m_DepthwiseConvolutionLayer
std::unique_ptr< arm_compute::CLTensor >	m_KernelTensor
std::unique_ptr< arm_compute::CLTensor >	m_BiasTensor
Protected Attributes inherited from BaseWorkload< DepthwiseConvolution2dQueueDescriptor >
DepthwiseConvolution2dQueueDescriptor	m_Data
const profiling::ProfilingGuid	m_Guid

Detailed Description

Definition at line 24 of file ClDepthwiseConvolutionWorkload.hpp.

Constructor & Destructor Documentation

ClDepthwiseConvolutionWorkload()

ClDepthwiseConvolutionWorkload	(	const DepthwiseConvolution2dQueueDescriptor &	descriptor,
		const WorkloadInfo &	info,
		const arm_compute::CLCompileContext &	clCompileContext
	)

Definition at line 75 of file ClDepthwiseConvolutionWorkload.cpp.

References ARMNN_REPORT_PROFILING_WORKLOAD_DESC, armnn::Convert1HWOTensorToAcl(), TensorInfo::GetNumBytes(), ConstTensorHandle::GetTensorInfo(), DepthwiseConvolution2dQueueDescriptor::m_Bias, DepthwiseConvolution2dDescriptor::m_BiasEnabled, WorkloadInfo::m_BiasTensorInfo, BaseWorkload< DepthwiseConvolution2dQueueDescriptor >::m_Data, DepthwiseConvolution2dDescriptor::m_DataLayout, WorkloadInfo::m_InputTensorInfos, ClDepthwiseConvolutionWorkload::m_KernelTensor, WorkloadInfo::m_OutputTensorInfos, QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters, DepthwiseConvolution2dQueueDescriptor::m_Weight, and WorkloadInfo::m_WeightsTensorInfo.

    : BaseWorkload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info)
{
    // 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("ClDepthwiseConvolutionWorkload_Construct",
                                         descriptor.m_Parameters,
                                         detailsInfo,
                                         this->GetGuid());

    // ArmNN's weight format is usually [ M, I, H, W ] but for depthwise its [ 1, H, W, I*M]
    // Permute to [ 1, I * M, H, W ] (if NCHW), as required by the compute library
    ConstTensor weightPermuted;
    unsigned int depthMultiplier;
    std::unique_ptr<unsigned char[]> permuteBuffer(new unsigned char[m_Data.m_Weight->GetTensorInfo().GetNumBytes()]);
    std::tie(weightPermuted, 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::CLTensor>();
    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::CLTensor>();
        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);


    std::string name = std::string("ClDepthwiseConvolutionWorkload");
    m_Data.ValidateInputsOutputs(name, 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();

    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_DepthwiseConvolutionLayer = std::make_unique<arm_compute::CLDepthwiseConvolutionLayer>();

    {
        ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClDepthwiseConvolutionWorkload_configure");
        static_cast<arm_compute::CLDepthwiseConvolutionLayer*>(m_DepthwiseConvolutionLayer.get())->configure(
                clCompileContext,
                &input,
                m_KernelTensor.get(),
                m_BiasTensor.get(),
                &output,
                padStrideInfo,
                depthMultiplier,
                activationInfo,
                aclDilationInfo);
    }
    ARMNN_ASSERT(m_DepthwiseConvolutionLayer);

    ScopedTensorHandle weightsPermutedHandle(weightPermuted);
    InitializeArmComputeClTensorData(*m_KernelTensor, &weightsPermutedHandle);

    if (m_BiasTensor)
    {
        InitializeArmComputeClTensorData(*m_BiasTensor, m_Data.m_Bias);
    }

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

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 172 of file ClDepthwiseConvolutionWorkload.cpp.

References ARMNN_ASSERT, ARMNN_SCOPED_PROFILING_EVENT_CL_GUID, CHECK_LOCATION, BaseWorkload< DepthwiseConvolution2dQueueDescriptor >::GetGuid(), ClDepthwiseConvolutionWorkload::m_DepthwiseConvolutionLayer, and armnn::RunClFunction().

{
    ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClDepthwiseConvolutionWorkload_Execute", this->GetGuid());
    ARMNN_ASSERT(m_DepthwiseConvolutionLayer);

    RunClFunction(*m_DepthwiseConvolutionLayer, CHECK_LOCATION());
}

FreeUnusedTensors()

void FreeUnusedTensors ( )

protected

Definition at line 166 of file ClDepthwiseConvolutionWorkload.cpp.

{
    FreeTensorIfUnused(m_KernelTensor);
    FreeTensorIfUnused(m_BiasTensor);
}

Member Data Documentation

m_BiasTensor

std::unique_ptr<arm_compute::CLTensor> m_BiasTensor

protected

Definition at line 39 of file ClDepthwiseConvolutionWorkload.hpp.

m_DepthwiseConvolutionLayer

std::unique_ptr<arm_compute::IFunction> m_DepthwiseConvolutionLayer

protected

Definition at line 36 of file ClDepthwiseConvolutionWorkload.hpp.

Referenced by ClDepthwiseConvolutionWorkload::Execute().

m_KernelTensor

std::unique_ptr<arm_compute::CLTensor> m_KernelTensor

protected

Definition at line 38 of file ClDepthwiseConvolutionWorkload.hpp.

Referenced by ClDepthwiseConvolutionWorkload::ClDepthwiseConvolutionWorkload().

---

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

• src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.hpp
• src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp