src/backends/tosaCommon/operatorMappings/QuantizeOperator.cpp

//
// Copyright © 2023-2024 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
// Copyright © 2020 The TensorFlow Authors. All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0
//

#include "QuantizeOperator.hpp"

#include "TosaRescaleOperatorUtils.hpp"

// This function is paraphrased from:
// tensorflow/compiler/mlir/tosa/transforms/legalize_common.cc from function convertQuantizeOp
TosaSerializationBasicBlock* ConvertQuantizeToTosaOperator(const Layer* layer,
                                                           const std::vector<const TensorInfo*>& inputs,
                                                           const std::vector<const TensorInfo*>& outputs)
{
    ARMNN_THROW_INVALIDARG_MSG_IF_FALSE( inputs.size() == 1,
                                         "ConvertQuantizeToTosaOperator: Quantize must have only one input" );
    ARMNN_THROW_INVALIDARG_MSG_IF_FALSE( outputs.size() == 1,
                                         "ConvertQuantizeToTosaOperator: Quantize must have only one output" );

    std::string inputName           = std::string("input0_");
    std::string outputName          = std::string("output0_");
    std::string blockName           = std::string("Op_QUANTIZE_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)
    {
        // Get the layers connected to the input slots and determine unique tensor names.
        Layer& connectedLayer = layer->GetInputSlot(0).GetConnectedOutputSlot()->GetOwningLayer();
        inputName = GenerateUniqueName(connectedLayer, 0);

        // Determine unique output tensor name.
        outputName = GenerateUniqueOutputName(*layer, 0);
    }

    const TensorInfo inputInfo = *inputs[0];
    const TensorInfo outputInfo = *outputs[0];

    // Extract quantization detail from Tensor
    float zeroPoint = static_cast<float>(outputInfo.GetQuantizationOffset());
    // No per axis support in Tensorflow TOSA code
    float scale = outputInfo.GetQuantizationScale();

    // As per the Tensorflow quantization specification
    // Tensorflow TOSA code calculates quantization using multiplication by scale
    // Armnn code calculates quantization using division by scale
    // Invert scale factor passed from Armnn for tf TOSA code
    scale = (scale != 0) ?  (1 / scale) : scale;

    std::vector<TosaSerializationTensor*> tensors;

    std::vector<int32_t> inputShape0 = GetTosaTensorShape(inputInfo.GetShape());
    DType inputDType0 = ArmNNToDType(inputInfo.GetDataType());
    float isFloatInput = inputDType0 == DType::DType_FP16 || inputDType0 == DType::DType_FP32;

    // 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.
    if(inputName.find("input0_") != std::string::npos)
    {
        tensors.push_back(new TosaSerializationTensor(inputName, inputShape0, inputDType0, {}));
    }

    std::vector<int32_t> outputShape0 = GetTosaTensorShape(outputInfo.GetShape());
    DType outputDType0 = ArmNNToDType(outputInfo.GetDataType());

    if (isFloatInput)
    {
        // quantize:
        // const_zeroPoint = constant(zeroPoint)
        // const_scale = constant(scale)
        // out_mul = mul(input, const_scale)
        // out_add = add(out_mul, const_zeroPoint)
        // output = cast<output_type>(out_add)

        std::string outputNameScale     = std::string("input1_") + GetUniqueTosaMappingID();
        std::string outputNameZeroPoint = std::string("input2_") + GetUniqueTosaMappingID();
        std::string outputNameMul       = std::string("intermediate0_") + GetUniqueTosaMappingID();
        std::string outputNameAdd       = std::string("intermediate1_") + GetUniqueTosaMappingID();

        // const_zeroPoint
        TosaSerializationOperator* zeroPointOp = nullptr;
        TosaSerializationTensor* zeroPointTensor = nullptr;
        CreateConstTosaOperator<float>(outputNameZeroPoint,
                                       zeroPoint,
                                       inputDType0,
                                       inputShape0,
                                       zeroPointOp,
                                       zeroPointTensor);
        tensors.push_back(zeroPointTensor);

        // const_scale
        TosaSerializationOperator *scaleOp = nullptr;
        TosaSerializationTensor* scaleTensor = nullptr;
        CreateConstTosaOperator<float>(outputNameScale,
                                       scale,
                                       inputDType0,
                                       inputShape0,
                                       scaleOp,
                                       scaleTensor);
        tensors.push_back(scaleTensor);

        // mul
        int32_t shift = 0;
        TosaMulAttribute mulAttribute(shift);
        TosaSerializationOperator* mulOp = new TosaSerializationOperator(Op_MUL,
                                                                         Attribute_MulAttribute,
                                                                         &mulAttribute,
                                                                         {inputName, outputNameScale},
                                                                         {outputNameMul});
        tensors.push_back(new TosaSerializationTensor(outputNameMul, inputShape0, inputDType0, {}));

        // add
        TosaSerializationOperator* addOp = new TosaSerializationOperator(Op_ADD,
                                                                         Attribute_NONE,
                                                                         nullptr,
                                                                         {outputNameMul, outputNameZeroPoint},
                                                                         {outputNameAdd});
        tensors.push_back(new TosaSerializationTensor(outputNameAdd, inputShape0, inputDType0, {}));

        // cast
        TosaSerializationOperator* castOp = new TosaSerializationOperator(Op_CAST,
                                                                          Attribute_NONE,
                                                                          nullptr,
                                                                          {outputNameAdd},
                                                                          {outputName});

        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
                                               {zeroPointOp, scaleOp, mulOp, addOp, castOp},    // operators
                                               tensors,                                         // tensors
                                               {inputName},                                     // inputs
                                               {outputName});                                   // outputs
    }
    else
    {
        double scale_alpha      = inputs[0]->GetQuantizationScale() / outputs[0]->GetQuantizationScale();
        int32_t input_zp        = inputs[0]->GetQuantizationOffset();
        int32_t output_zp       = outputs[0]->GetQuantizationOffset();

        TosaSerializationOperator* rescaleOp = nullptr;
        TosaSerializationTensor* rescaleTensor = nullptr;
        CreateRescaleTosaOperator(inputName,
                                  outputName,
                                  outputDType0,
                                  inputShape0,
                                  scale_alpha,
                                  input_zp,
                                  output_zp,
                                  true,
                                  true,
                                  &rescaleOp,
                                  &rescaleTensor);
        tensors.push_back(rescaleTensor);

        // 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
                                               {rescaleOp},    // operators
                                               tensors,        // tensors
                                               {inputName},    // inputs
                                               {outputName});  // outputs
    }
}