NeonConvertFp16ToFp32Workload Class Reference

#include <NeonConvertFp16ToFp32Workload.hpp>

Inheritance diagram for NeonConvertFp16ToFp32Workload:

Public Member Functions
	NeonConvertFp16ToFp32Workload (const ConvertFp16ToFp32QueueDescriptor &descriptor, const WorkloadInfo &info)
virtual void	Execute () const override
void	ReplaceInputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
void	ReplaceOutputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
Public Member Functions inherited from MultiTypedWorkload< QueueDescriptor, InputDataType, OutputDataType >
	MultiTypedWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
Public Member Functions inherited from BaseWorkload< QueueDescriptor >
	BaseWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
void	ExecuteAsync (ExecutionData &executionData) 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 Attributes inherited from BaseWorkload< QueueDescriptor >
QueueDescriptor	m_Data
const arm::pipe::ProfilingGuid	m_Guid

Detailed Description

Definition at line 20 of file NeonConvertFp16ToFp32Workload.hpp.

Constructor & Destructor Documentation

NeonConvertFp16ToFp32Workload()

NeonConvertFp16ToFp32Workload	(	const ConvertFp16ToFp32QueueDescriptor &	descriptor,
		const WorkloadInfo &	info
	)

Definition at line 29 of file NeonConvertFp16ToFp32Workload.cpp.

     : Float16ToFloat32Workload<ConvertFp16ToFp32QueueDescriptor>(descriptor, info)
{
    this->m_Data.ValidateInputsOutputs("NeonConvertFp16ToFp32Workload", 1, 1);
 
    arm_compute::ITensor& input  = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
    arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
    if (arm_compute::NECast::validate(input.info(), output.info(), g_AclConvertPolicy))
    {
        // Use NECast if supported (needs hardware support for FP16)
        m_Cast.reset(new arm_compute::NECast());
        m_Cast->configure(&input, &output, g_AclConvertPolicy);
    }
    else
    {
        // Else use software implementation using Half.hpp
        GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
    }
}

References armnn::GatherTensorHandlePairs(), 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 51 of file NeonConvertFp16ToFp32Workload.cpp.

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvertFp16ToFp32Workload_Execute", this->GetGuid());
 
    if (m_Cast)
    {
        // Use NECast if supported and initialised
        m_Cast->run();
    }
    else
    {
        // Else use softare implementation using Half.hpp
        auto convertFunc = [](uint8_t* dst, const uint8_t* src, size_t size)
        {
            auto input = reinterpret_cast<const Half*>(src);
            auto output = reinterpret_cast<float*>(dst);
            size_t numElements = size/2; // 2 bytes per fp16
            armnnUtils::FloatingPointConverter::ConvertFloat16To32(input, numElements, output);
        };
 
        for (const auto& pair : m_TensorHandlePairs)
        {
            CopyTensorContentsGeneric(pair.first, pair.second, convertFunc);
        }
    }
}

References ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID, FloatingPointConverter::ConvertFloat16To32(), armnn::CopyTensorContentsGeneric(), and BaseWorkload< QueueDescriptor >::GetGuid().

ReplaceInputTensorHandle()

void ReplaceInputTensorHandle ( ITensorHandle *  tensorHandle, unsigned int  slot  )

overridevirtual

Reimplemented from BaseWorkload< QueueDescriptor >.

Definition at line 78 of file NeonConvertFp16ToFp32Workload.cpp.

{
    ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
    this->m_Data.m_Inputs[slot] = tensorHandle;
    try
    {
        Reconfigure();
    }
    catch(armnn::UnimplementedException& e)
    {
        // Cannot reconfigure, revert the slot back and throw the exception.
        this->m_Data.m_Inputs[slot] = backupHandle;
        throw e;
    }
}

References BaseWorkload< QueueDescriptor >::m_Data, and QueueDescriptor::m_Inputs.

ReplaceOutputTensorHandle()

void ReplaceOutputTensorHandle ( ITensorHandle *  tensorHandle, unsigned int  slot  )

overridevirtual

Reimplemented from BaseWorkload< QueueDescriptor >.

Definition at line 95 of file NeonConvertFp16ToFp32Workload.cpp.

{
    ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
    this->m_Data.m_Inputs[slot] = tensorHandle;
    try
    {
        Reconfigure();
    }
    catch(armnn::UnimplementedException& e)
    {
        // Cannot reconfigure, revert the slot back and throw the exception.
        this->m_Data.m_Inputs[slot] = backupHandle;
        throw e;
    }
}

References BaseWorkload< QueueDescriptor >::m_Data, and QueueDescriptor::m_Inputs.

---

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

• src/backends/neon/workloads/NeonConvertFp16ToFp32Workload.hpp
• src/backends/neon/workloads/NeonConvertFp16ToFp32Workload.cpp