NeonFloorDivWorkload Class Reference

#include <NeonFloorDivWorkload.hpp>

Inheritance diagram for NeonFloorDivWorkload:

Collaboration diagram for NeonFloorDivWorkload:

Public Member Functions
	NeonFloorDivWorkload (const DivisionQueueDescriptor &descriptor, const WorkloadInfo &info)
virtual void	Execute () const override
Public Member Functions inherited from NeonBaseWorkload< DivisionQueueDescriptor >
	NeonBaseWorkload (const DivisionQueueDescriptor &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< QueueDescriptor >
	BaseWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
virtual const std::string &	GetName () const override
void	PostAllocationConfigure () override
const QueueDescriptor &	GetData () const
arm::pipe::ProfilingGuid	GetGuid () const final
virtual bool	SupportsTensorHandleReplacement () const override
Public Member Functions inherited from IWorkload
virtual	~IWorkload ()
virtual void	RegisterDebugCallback (const DebugCallbackFunction &)
virtual armnn::Optional< armnn::MemoryRequirements >	GetMemoryRequirements ()

Additional Inherited Members
Protected Member Functions inherited from NeonBaseWorkload< DivisionQueueDescriptor >
virtual void	Reconfigure ()
Protected Attributes inherited from BaseWorkload< QueueDescriptor >
QueueDescriptor	m_Data
const arm::pipe::ProfilingGuid	m_Guid
const std::string	m_Name

Detailed Description

Definition at line 26 of file NeonFloorDivWorkload.hpp.

Constructor & Destructor Documentation

NeonFloorDivWorkload()

NeonFloorDivWorkload	(	const DivisionQueueDescriptor &	descriptor,
		const WorkloadInfo &	info
	)

Definition at line 129 of file NeonFloorDivWorkload.cpp.

        : NeonBaseWorkload<DivisionQueueDescriptor>(descriptor, info)
{
    m_Data.ValidateInputsOutputs("NeonFloorDivWorkload", 2, 1);
 
    TensorInfo input0Info = info.m_InputTensorInfos[0];
    TensorInfo input1Info = info.m_InputTensorInfos[1];
    TensorInfo outputInfo = info.m_OutputTensorInfos[0];
 
    arm_compute::ITensor& input0   = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
    arm_compute::ITensor& input1 = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
    arm_compute::ITensor& output  = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
    // Get data type of input and output
    arm_compute::DataType inputDataType  = PolymorphicDowncast<NeonTensorHandle*>(m_Data.m_Inputs[0])->GetDataType();
    arm_compute::DataType outputDataType = PolymorphicDowncast<NeonTensorHandle*>(m_Data.m_Outputs[0])->GetDataType();
 
    const arm_compute::ActivationLayerInfo activationInfo =
        ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
 
    // If Tensors are Signed32 we need to Cast them to floats, this is to ensure we get the correct
    // output if the result is a negative number, as we should floor towards -(infinity)
    if(inputDataType == arm_compute::DataType::S32 && outputDataType == arm_compute::DataType::S32)
    {
        // Create new Cast layer pointers if type is S32
        m_CastLayer0.reset(new arm_compute::NECast());
        m_CastLayer1.reset(new arm_compute::NECast());
        m_CastLayer2.reset(new arm_compute::NECast());
 
        // Cast Input 0 to type float32
        TensorInfo outputCast0_Info = ConvertTensorToFloat32(input0Info);
 
        // Initialise output tensor based on Float32 type
        BuildArmComputeTensor(m_OutputCast0, outputCast0_Info);
        armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_OutputCast0);
 
        // Configure first Cast Layer
        m_CastLayer0->configure(&input0, &m_OutputCast0, arm_compute::ConvertPolicy::WRAP);
 
        // Cast Input 1 to type Float32
        TensorInfo outputCast1_Info = ConvertTensorToFloat32(input1Info);
 
        // Initialise Output tensor based on Float32 type
        BuildArmComputeTensor(m_OutputCast1, outputCast1_Info);
        armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_OutputCast1);
 
        // Configure second Cast Layer
        m_CastLayer1->configure(&input1, &m_OutputCast1, arm_compute::ConvertPolicy::WRAP);
 
        // Create Div output tensor
        TensorInfo outputDiv_Info = ConvertTensorToFloat32(outputInfo);
        BuildArmComputeTensor(m_OutputDiv, outputDiv_Info);
        armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_OutputDiv);
 
        // Configure Div Layer
        m_DivLayer.configure(&m_OutputCast0, &m_OutputCast1, &m_OutputDiv, activationInfo);
 
        // Create Floor output tensor
        BuildArmComputeTensor(m_OutputFloor, outputDiv_Info);
        armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_OutputFloor);
 
        // Configure Floor Layer
        m_FloorLayer.configure(&m_OutputDiv, &m_OutputFloor);
 
        // Configure third Cast Layer
        m_CastLayer2->configure(&m_OutputFloor, &output, arm_compute::ConvertPolicy::WRAP);
    }
    else
    {
        // Create Div output tensor
        BuildArmComputeTensor(m_OutputDiv, outputInfo);
        armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_OutputDiv);
 
        // Configure Div Layer
        m_DivLayer.configure(&input0, &input1, &m_OutputDiv, activationInfo);
 
        // Configure Floor Layer
        m_FloorLayer.configure(&m_OutputDiv, &output);
    }
}

References armnn::ConvertAdditionalInfoToAclActivationLayerInfo(), armnn::ConvertTensorToFloat32(), armnn::info, BaseWorkload< QueueDescriptor >::m_Data, QueueDescriptor::m_Inputs, QueueDescriptor::m_Outputs, and QueueDescriptor::ValidateInputsOutputs().

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 211 of file NeonFloorDivWorkload.cpp.

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID("NeonFloorDivWorkload_Execute");
    // Only run Cast Layers if needed. e.g. if it exists
    if(m_CastLayer0 && m_CastLayer1)
    {
        m_CastLayer0->run();
        m_CastLayer1->run();
 
        // Delete objects after running
        m_CastLayer0.reset();
        m_CastLayer1.reset();
    }
    m_DivLayer.run();
    m_FloorLayer.run();
    if(m_CastLayer2)
    {
        m_CastLayer2->run();
 
        // Delete object after running layer
        m_CastLayer2.reset();
    }
}

References ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID.

---

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

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