ConvertFp32NetworkToFp16Impl Class Reference

#include <ConvertFp32NetworkToFp16.hpp>

Public Member Functions
void	Run (Graph &graph, Layer &layer) const

Protected Member Functions
	ConvertFp32NetworkToFp16Impl ()=default
	~ConvertFp32NetworkToFp16Impl ()=default

Detailed Description

Definition at line 15 of file ConvertFp32NetworkToFp16.hpp.

Constructor & Destructor Documentation

ConvertFp32NetworkToFp16Impl()

ConvertFp32NetworkToFp16Impl ( )

protecteddefault

Referenced by ConvertFp32NetworkToFp16Impl::Run().

~ConvertFp32NetworkToFp16Impl()

~ConvertFp32NetworkToFp16Impl ( )

protecteddefault

Referenced by ConvertFp32NetworkToFp16Impl::Run().

Member Function Documentation

Run()

void Run ( Graph &  graph, Layer &  layer  ) const

inline

Definition at line 18 of file ConvertFp32NetworkToFp16.hpp.

References Layer::BeginInputSlots(), Layer::BeginOutputSlots(), armnn::ConvertFp16ToFp32, ConvertFp32NetworkToFp16Impl::ConvertFp32NetworkToFp16Impl(), armnn::ConvertFp32ToFp16, Layer::EndInputSlots(), Layer::EndOutputSlots(), armnn::Float16, armnn::Float32, TensorInfo::GetDataType(), Layer::GetDataType(), Layer::GetType(), armnn::Input, armnn::InsertConvertFp16ToFp32LayersBefore(), armnn::InsertConvertFp32ToFp16LayersAfter(), armnn::Output, TensorInfo::SetDataType(), and ConvertFp32NetworkToFp16Impl::~ConvertFp32NetworkToFp16Impl().

    {
        if(layer.GetType() == LayerType::Input)
        {
            // if the outputs of this layer are DataType::Float32
            // add a ConvertFloat32ToFloat16 layer after each of the outputs
            if (layer.GetDataType() == DataType::Float32)
            {
                InsertConvertFp32ToFp16LayersAfter(graph, layer);
            }
        }
        else if (layer.GetType() == LayerType::Output)
        {
            // if the inputs of this layer are DataType::Float32
            // add a ConvertFloat16ToFloat32 layer before each of the inputs
            if (layer.GetDataType() == DataType::Float32)
            {
                // NOTE: We need to call InsertConvertFp16ToFp32LayersBefore with expectCorrectInputType = false
                // here, otherwise it will expect the inputs to be DataType::Float16
                InsertConvertFp16ToFp32LayersBefore(graph, layer, false);
            }
        }
        else if (layer.GetType() != LayerType::ConvertFp32ToFp16 && layer.GetType() != LayerType::ConvertFp16ToFp32)
        {
            // if the inputs/outputs of this layer are DataType::Float32
            // change the data type for all inputs and outputs to DataType::Float16
            for (auto&& input = layer.BeginInputSlots(); input != layer.EndInputSlots(); ++input)
            {
                // if it is connected to OutputSlot of the InputLayer do not change the DataType of connection
                // InputSlots of the current layer will be updated when conversion layer is inserted after InputLayer
                Layer& base = input->GetConnectedOutputSlot()->GetOwningLayer();
                if (base.GetType() != LayerType::Input)
                {
                    TensorInfo convertInfo = input->GetConnection()->GetTensorInfo();
                    if (convertInfo.GetDataType() == DataType::Float32)
                    {
                        convertInfo.SetDataType(DataType::Float16);
                        input->GetConnection()->SetTensorInfo(convertInfo);
                    }
                }
            }

            // change outputs to DataType::Float16
            for (auto&& output = layer.BeginOutputSlots(); output != layer.EndOutputSlots(); ++output)
            {
                TensorInfo convertInfo = output->GetTensorInfo();
                if (convertInfo.GetDataType() == DataType::Float32)
                {
                    convertInfo.SetDataType(DataType::Float16);
                    output->SetTensorInfo(convertInfo);
                }
            }
        }
    }

---

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

• src/armnn/optimizations/ConvertFp32NetworkToFp16.hpp