Softmax.hpp

/********************************************************************************
 * Copyright (c) 2023 CEA-List
 *
 * This program and the accompanying materials are made available under the
 * terms of the Eclipse Public License 2.0 which is available at
 * http://www.eclipse.org/legal/epl-2.0.
 *
 * SPDX-License-Identifier: EPL-2.0
 *
 ********************************************************************************/

#ifndef AIDGE_CORE_OPERATOR_SOFTMAX_H_
#define AIDGE_CORE_OPERATOR_SOFTMAX_H_

#include <cassert>
#include <memory>
#include <vector>

#include "aidge/utils/Registrar.hpp"
#include "aidge/operator/Operator.hpp"
#include "aidge/backend/OperatorImpl.hpp"
#include "aidge/data/Tensor.hpp"
#include "aidge/data/Data.hpp"
#include "aidge/graph/Node.hpp"
#include "aidge/utils/Types.h"

namespace Aidge {

class Softmax_Op : public Operator,
    public Registrable<Softmax_Op, std::string, std::unique_ptr<OperatorImpl>(const Softmax_Op&)> {
public:
    // FIXME: change accessibility
    std::shared_ptr<Tensor> mInput = std::make_shared<Tensor>();
    const std::shared_ptr<Tensor> mOutput = std::make_shared<Tensor>();

public:
    static constexpr const char* Type = "Softmax";

    Softmax_Op()
            : Operator(Type)
    {
        setDatatype(DataType::Float32);
    }

    /**
     * @brief Copy-constructor. Copy the operator attributes and its output tensor(s), but not its input tensors (the new operator has no input associated).
     * @param op Operator to copy.
     */
    Softmax_Op(const Softmax_Op& op)
        : Operator(Type),
          mOutput(std::make_shared<Tensor>(*op.mOutput))
    {
        // cpy-ctor
        setDatatype(op.mOutput->dataType());
        mImpl = op.mImpl ? Registrar<Softmax_Op>::create(mOutput->getImpl()->backend())(*this) : nullptr;
    }

    /**
     * @brief Clone the operator using its copy-constructor.
     * @see Operator::Softmax_Op
     */
    std::shared_ptr<Operator> clone() const override {
        return std::make_shared<Softmax_Op>(*this);
    }

    void associateInput(const IOIndex_t inputIdx, std::shared_ptr<Data> data) override final {
        assert(inputIdx == 0 && "operator supports only 1 input");
        (void) inputIdx; // avoid unused warning
        assert(strcmp(data->type(), Tensor::Type)==0 && "input data must be of Tensor type");
        mInput = std::dynamic_pointer_cast<Tensor>(data);
    }

    void computeOutputDims() override final {
        if (!mInput->empty())
            mOutput->resize(mInput->dims());
    }

    bool outputDimsForwarded() const override final {
        return !(mOutput->empty());
    }


    inline Tensor& input(const IOIndex_t /*inputIdx*/) const override final { return *(mInput.get()); }
    inline Tensor& output(const IOIndex_t /*outputIdx*/) const override final { return *(mOutput.get()); }


    inline std::shared_ptr<Tensor> getInput(const IOIndex_t inputIdx) const override final {
        assert((inputIdx == 0) && "Softmax Operator has only 1 input");
        (void) inputIdx; // avoid unused warning
        return mInput;
    }
    inline std::shared_ptr<Tensor> getOutput(const IOIndex_t outputIdx) const override final {
        assert((outputIdx == 0) && "Softmax Operator has only 1 output");
        (void) outputIdx; // avoid unused warning
        return mOutput;
    }


    std::shared_ptr<Data> getRawInput(const IOIndex_t inputIdx) const override final {
        assert(inputIdx == 0 && "operator supports only 1 input");
        (void) inputIdx; // avoid unused warning
        return std::static_pointer_cast<Data>(mInput);
    }
    std::shared_ptr<Data> getRawOutput(const IOIndex_t outputIdx) const override final {
        assert(outputIdx == 0 && "operator supports only 1 output");
        (void) outputIdx; // avoid unused warning
        return std::static_pointer_cast<Data>(mOutput);
    }


    void setBackend(const std::string& name) override {
        mImpl = Registrar<Softmax_Op>::create(name)(*this);
        mOutput->setBackend(name);

        // FIXME: temporary workaround
        mInput->setBackend(name);
    }
    void setDatatype(const DataType& datatype) override {
        mOutput->setDatatype(datatype);

        // FIXME: temporary workaround
        mInput->setDatatype(datatype);
    }

    inline IOIndex_t nbInputs() const noexcept override final { return 1; }
    inline IOIndex_t nbDataInputs() const noexcept override final { return 1; }
    inline IOIndex_t nbOutputs() const noexcept override final { return 1; }
    static const std::vector<std::string> getInputsName(){
        return {"data_input"};
    }
    static const std::vector<std::string> getOutputsName(){
        return {"data_output"};
    }
};

inline std::shared_ptr<Node> Softmax(const std::string& name = "") {
    // FIXME: properly handle default w&b initialization in every cases
    return std::make_shared<Node>(std::make_shared<Softmax_Op>(), name);
}
}

#endif /* AIDGE_CORE_OPERATOR_SOFTMAX_H_ */