HardSwishOperator.cpp File Reference

#include "HardSwishOperator.hpp"
#include <gemmlowp/fixedpoint.h>

Include dependency graph for HardSwishOperator.cpp:

Go to the source code of this file.

Functions
std::vector< int16_t >	getTosaConstHardSwish8bitTable (float inputScale, int32_t inputZp, float outputScale, int32_t outputZp)
TosaSerializationBasicBlock *	ConvertHardSwishToTosaOperator (const Layer *layer, const std::vector< const TensorInfo * > &inputs, const std::vector< const TensorInfo * > &outputs, const ActivationDescriptor *desc)

Function Documentation

ConvertHardSwishToTosaOperator()

TosaSerializationBasicBlock * ConvertHardSwishToTosaOperator	(	const Layer *	layer,
		const std::vector< const TensorInfo * > &	inputs,
		const std::vector< const TensorInfo * > &	outputs,
		const ActivationDescriptor *	desc )

Definition at line 160 of file HardSwishOperator.cpp.

{
    if (inputs.size() != 1)
    {
        throw armnn::Exception("ConvertHardSwishToTosaOperator: 1 input tensors required.");
    }
 
    if (outputs.size() != 1)
    {
        throw armnn::Exception("ConvertHardSwishToTosaOperator: 1 output tensor required.");
    }
 
    if (desc->m_Function != ActivationFunction::HardSwish)
    {
        throw armnn::Exception("ConvertHardSwishToTosaOperator ActivationDescriptor only supports function HardSwish.");
    }
 
    std::string inputName  = std::string("input_");
    std::string outputName = std::string("output0_");
    std::string blockName  = std::string("Op_HARDSWISH_block_") + GetUniqueTosaMappingID();
 
    // If a layer is present then the block will be used for execution, so input and output names need to be determined
    // using the previous and following layers so the graph is connected correctly. For validation this doesn't matter.
    if (layer != nullptr)
    {
        inputName  = GenerateUniqueInputName(layer->GetInputSlot(0));
        outputName = GenerateUniqueOutputName(*layer);
    }
 
    std::vector<TosaSerializationTensor*> tensors;
    std::vector<TosaSerializationOperator*> operators;
 
    DataType inputDType = inputs[0]->GetDataType();
 
    bool isInt8 = (inputDType == DataType::QAsymmS8 || inputDType == DataType::QSymmS8);
    if (isInt8)
    {
        float inputScale = inputs[0]->GetQuantizationScale();
        float outputScale = outputs[0]->GetQuantizationScale();
        int32_t inputZp = inputs[0]->GetQuantizationOffset();
        int32_t outputZp = outputs[0]->GetQuantizationOffset();
 
        TosaTableAttribute attribute(
            getTosaConstHardSwish8bitTable(inputScale, inputZp, outputScale, outputZp));
        operators.push_back(new TosaSerializationOperator(tosa::Op_TABLE,
                                                          Attribute_TableAttribute,
                                                          &attribute,
                                                          {inputName},
                                                          {outputName}));
    }
    else
    {
        throw Exception("ConvertHardSwishToTosaOperator() type currently unimplemented.");
    }
 
    // Only add input tensors if connected layer is an input layer.
    // As intermediate or constant tensors will be created separately.
    // There also can't be duplicate tensor.
    std::vector<int32_t> inputShape0;
    DType inputDType0 = ArmNNToDType(inputs[0]->GetDataType());
    if(inputName.find("input_") != std::string::npos)
    {
        inputShape0 = GetTosaTensorShape(inputs[0]->GetShape());
        tensors.push_back(new TosaSerializationTensor(inputName, inputShape0, inputDType0, {}));
    }
 
    std::vector<int32_t> outputShape0 = GetTosaTensorShape(outputs[0]->GetShape());
    DType outputDType0 = ArmNNToDType(outputs[0]->GetDataType());
    tensors.push_back(new TosaSerializationTensor(outputName, outputShape0, outputDType0, {}));
 
    // operatorInputNames/operatorOutputNames ends up being the same as
    // blockInputNames/blockOutputNames for one-to-one ArmNN to Tosa mappings
    return new TosaSerializationBasicBlock(blockName,      // name
                                           mainName,       // region name
                                           operators,      // operators
                                           tensors,        // tensors
                                           {inputName},    // inputs
                                           {outputName});  // outputs
}

References ArmNNToDType(), GenerateUniqueInputName(), GenerateUniqueOutputName(), Layer::GetInputSlot(), getTosaConstHardSwish8bitTable(), GetTosaTensorShape(), GetUniqueTosaMappingID(), ActivationDescriptor::m_Function, and mainName.

Referenced by GetTosaMapping().

getTosaConstHardSwish8bitTable()

std::vector< int16_t > getTosaConstHardSwish8bitTable	(	float	inputScale,
		int32_t	inputZp,
		float	outputScale,
		int32_t	outputZp )

Definition at line 90 of file HardSwishOperator.cpp.

{
    const float hiresInputScale = (1.0f / 128.0f) * inputScale;
    const float outputMultiplier = hiresInputScale / outputScale;
    int outputMultiplierExponent;
    int16_t outputMultiplierFixedpointInt16;
    int32_t outputMultiplierFixedpointInt32;
 
    quantizeMultiplier(outputMultiplier, &outputMultiplierFixedpointInt32, &outputMultiplierExponent);
    downScaleInt32ToInt16Multiplier(outputMultiplierFixedpointInt32, &outputMultiplierFixedpointInt16);
 
    ARMNN_THROW_INVALIDARG_IF_FALSE(outputMultiplierExponent <= 0);
 
    const float reluishScale = 3.0f / 32768.0f;
    const float reluishMultiplier = hiresInputScale / reluishScale;
    int reluishMultiplierExponent;
    int16_t reluishMultiplierFixedpointInt16;
    int32_t reluishMultiplierFixedpointInt32;
 
    quantizeMultiplier(reluishMultiplier, &reluishMultiplierFixedpointInt32, &reluishMultiplierExponent);
    downScaleInt32ToInt16Multiplier(reluishMultiplierFixedpointInt32, &reluishMultiplierFixedpointInt16);
 
    std::vector<int16_t> table;
    table.reserve(256);
    for (int32_t i = -128; i < 128; i++)
    {
        const int16_t inputValue = static_cast<int16_t>(i - inputZp);
        const int16_t inputValueHiresInputScale = static_cast<int16_t>(inputValue * (1 << 7));
 
        int16_t reluishValue = inputValueHiresInputScale;
        if (reluishMultiplierExponent > 0)
        {
            reluishValue = gemmlowp::ShiftLeft(reluishValue, reluishMultiplierExponent - 1);
        }
 
        reluishValue = gemmlowp::SaturatingRoundingDoublingHighMul(reluishValue, reluishMultiplierFixedpointInt16);
 
        if (reluishMultiplierExponent > 0)
        {
            reluishValue = gemmlowp::ShiftLeft(reluishValue, 1);
        }
        else if (reluishMultiplierExponent < 0)
        {
            reluishValue = gemmlowp::RoundingDivideByPOT(reluishValue, -reluishMultiplierExponent);
        }
 
        reluishValue = static_cast<int16_t>((reluishValue + (1 << 15)) >> 1);
 
        const int16_t inputValPreshiftOutputScale =
            gemmlowp::SaturatingRoundingDoublingHighMul(inputValueHiresInputScale, outputMultiplierFixedpointInt16);
 
        const int16_t preshiftOutputValue = saturatingDoublingHighMul(reluishValue, inputValPreshiftOutputScale);
 
        int16_t outputValue = gemmlowp::RoundingDivideByPOT(preshiftOutputValue, -outputMultiplierExponent);
 
        outputValue = static_cast<int16_t>(outputValue + outputZp);
        outputValue = std::min<int16_t>(outputValue, std::numeric_limits<int8_t>::max());
        outputValue = std::max<int16_t>(outputValue, std::numeric_limits<int8_t>::min());
 
        table.push_back(outputValue);
    }
 
    return table;
}

References ARMNN_THROW_INVALIDARG_IF_FALSE.

Referenced by ConvertHardSwishToTosaOperator().