include/aidge/backend/cpu.hpp

@@ -12,16 +12,58 @@
 #ifndef AIDGE_CPU_IMPORTS_H_
 #define AIDGE_CPU_IMPORTS_H_
 
-#include "aidge/backend/cpu/data/TensorImpl.hpp"
+#include "aidge/backend/cpu_version.h"
+
+#include "aidge/backend/cpu/operator/AbsImpl.hpp"
 #include "aidge/backend/cpu/operator/AddImpl.hpp"
+#include "aidge/backend/cpu/operator/AndImpl.hpp"
+#include "aidge/backend/cpu/operator/AtanImpl.hpp"
+
+#include "aidge/backend/cpu/operator/ArgMaxImpl.hpp"
 #include "aidge/backend/cpu/operator/AvgPoolingImpl.hpp"
+#include "aidge/backend/cpu/operator/MaxPoolingImpl.hpp"
 #include "aidge/backend/cpu/operator/BatchNormImpl.hpp"
+#include "aidge/backend/cpu/operator/BitShiftImpl.hpp"
+#include "aidge/backend/cpu/operator/ClipImpl.hpp"
 #include "aidge/backend/cpu/operator/ConvDepthWiseImpl.hpp"
 #include "aidge/backend/cpu/operator/ConvImpl.hpp"
+#include "aidge/backend/cpu/operator/ConvTransposeImpl.hpp"
+#include "aidge/backend/cpu/operator/ConstantOfShapeImpl.hpp"
+#include "aidge/backend/cpu/operator/CryptoHashImpl.hpp"
+#include "aidge/backend/cpu/operator/DivImpl.hpp"
+#include "aidge/backend/cpu/operator/DropoutImpl.hpp"
+#include "aidge/backend/cpu/operator/EqualImpl.hpp"
+#include "aidge/backend/cpu/operator/ErfImpl.hpp"
+#include "aidge/backend/cpu/operator/ExpandImpl.hpp"
 #include "aidge/backend/cpu/operator/FCImpl.hpp"
+#include "aidge/backend/cpu/operator/FoldImpl.hpp"
+#include "aidge/backend/cpu/operator/GlobalAveragePoolingImpl.hpp"
+#include "aidge/backend/cpu/operator/HeavisideImpl.hpp"
+#include "aidge/backend/cpu/operator/LRNImpl.hpp"
 #include "aidge/backend/cpu/operator/LeakyReLUImpl.hpp"
-#include "aidge/backend/cpu/operator/ProducerImpl.hpp"
+#include "aidge/backend/cpu/operator/LnImpl.hpp"
+#include "aidge/backend/cpu/operator/MatMulImpl.hpp"
+#include "aidge/backend/cpu/operator/MaxPoolingImpl.hpp"
+#include "aidge/backend/cpu/operator/ModImpl.hpp"
+#include "aidge/backend/cpu/operator/MulImpl.hpp"
+#include "aidge/backend/cpu/operator/PadImpl.hpp"
+#include "aidge/backend/cpu/operator/PaddedConvImpl.hpp"
+#include "aidge/backend/cpu/operator/PowImpl.hpp"
+#include "aidge/backend/cpu/operator/ReduceMeanImpl.hpp"
+#include "aidge/backend/cpu/operator/ReduceSumImpl.hpp"
+#include "aidge/backend/cpu/operator/ResizeImpl.hpp"
 #include "aidge/backend/cpu/operator/ReLUImpl.hpp"
+#include "aidge/backend/cpu/operator/RoundImpl.hpp"
+#include "aidge/backend/cpu/operator/ScalingImpl.hpp"
+#include "aidge/backend/cpu/operator/SigmoidImpl.hpp"
+#include "aidge/backend/cpu/operator/SqrtImpl.hpp"
+#include "aidge/backend/cpu/operator/SliceImpl.hpp"
 #include "aidge/backend/cpu/operator/SoftmaxImpl.hpp"
+#include "aidge/backend/cpu/operator/SubImpl.hpp"
+#include "aidge/backend/cpu/operator/TopKImpl.hpp"
+#include "aidge/backend/cpu/operator/TanhImpl.hpp"
+#include "aidge/backend/cpu/operator/WeightInterleavedImpl.hpp"
+
+#include "aidge/backend/cpu/data/TensorImpl.hpp"
 
 #endif /* AIDGE_CPU_IMPORTS_H_ */
\ No newline at end of file

include/aidge/backend/cpu/data/Broadcasting.hpp

+/********************************************************************************
+ * Copyright (c) 2024 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_CPU_DATA_BROADCASTING_H_
+#define AIDGE_CPU_DATA_BROADCASTING_H_
+
+#include <vector>
+
+namespace Aidge {
+
+// Function to broadCast an input dims vector into the same size as an outputDims vector
+
+    /**
+     * @brief  Broadcast an input dims vector into the same size as an outputDims vector
+     * @details The missing dimensions would be completed by 1
+     * @param outputDims The vector of dimensions to follow 
+     * @param dimsToBroadcast The vecotr of dimensions to braodcast
+     * @return std::vector<std::size_t> a broadcasted vector by addding 1 on the missing dimensions.
+     */
+    std::vector<std::size_t> getBroadcastedDims(const std::vector<std::size_t>& outputDims, const std::vector<std::size_t>& dimsToBroadcast);
+
+    /**
+     * @brief Get a vector of indexes along the dimensions vector from a flattened index
+     * @param dimensions The vector of dimensions we want the indexes on
+     * @param idx The flattened index
+     * @return std::vector<std::size_t> vector of indexes along dimensions.
+     */
+    std::vector<std::size_t> getMultiDimIndices(const std::vector<std::size_t>& dimensions, std::size_t idx);
+
+    // Function to get a flattened index from multi-dimensional indices
+    /**
+     * @brief Get a flattened index the dimensions vector from a given vector of indices on a broadcasted vector
+     * @param dimensions The vector of dimensions we want the flattened index on
+     * @param indices The vector of indices we want to flatten
+     * @return std::size_t The flattened index on the dimensions vector
+     */
+    std::size_t getFlattenedIndex(const std::vector<std::size_t>& dimensions, const std::vector<std::size_t>& indices);
+
+} // namespace Aidge
+
+#endif // AIDGE_CPU_DATA_BROADCASTING_H_
\ No newline at end of file

include/aidge/backend/cpu/data/Interpolation.hpp

+/********************************************************************************
+ * Copyright (c) 2024 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_CPU_DATA_INTERPOLATION_H_
+#define AIDGE_CPU_DATA_INTERPOLATION_H_
+
+#include <vector>
+
+#include <aidge/data/Interpolation.hpp>
+#include <aidge/utils/Types.h>
+
+namespace Aidge {
+class InterpolationCPU : public Interpolation {
+  public:
+    /*
+     * @brief Interpolates values given via input in given mode.
+     *
+     * Values are contiguously arranged in a "square" shape around the point to
+     * interpolate. Depending on interpolation mode.
+     * The point that will be interpolated is located right in the
+     * middle of all points.
+     * Immediate neighbours :
+     * 1D interp :     2D interp :
+     *                 . . . . . .
+     * . . 1 2 . .     . . . . . .
+     *                 . . 1 2 . .
+     *                 . . 3 4 . .
+     *                 . . . . . .
+     *                 . . . . . .
+     *
+     * 2 neighbours :
+     * 1D interp :         2D interp :
+     *                   .  .  .  .  .  .  . .
+     *                   .  .  .  .  .  .  . .
+     * . . 1 2 3 4 . .   .  .  1  2  3  4  . .
+     *                   .  .  5  6  7  8  . .
+     *                   .  .  9 10 11 12  . .
+     *                   .  . 13 14 15 16  . .
+     *                   .  .  .  .  .  .  . .
+     *                   .  .  .  .  .  .  . .
+     *
+     * @param[in] originalCoords: coord of the point to interpolate in the
+     * original picture. These coords are generated with
+     * Interpolation::untransformCoords(coordsInInterpolatedTensor)
+     * @param[in] points : points to interpolate, arranged in a vector of a
+     * pairs ((point_coord), value) :
+     * [[[X1, X2, ..., XN], Xval], ...., [[A1, A2, ..., AN],Aval]].
+     * With :
+     * - N: the number of dimensions.
+     * - A: the number of points of the grid to interpolate.
+     * - All coordinates expressed in originalTensor frame.
+     * @param[in] interpMode: interpolation mode
+     * @return interpolated value
+     */
+    template <typename T>
+    static T interpolate(const std::vector<float> &coordsToInterpolate,
+                         const std::set<Point<T>> &points,
+                         const Mode interpMode = Interpolation::Mode::Linear);
+
+    /**
+     * @brief performs linear interpolation on given points.
+     * @param[in] values: values to interpolate, since we only do an average of
+     * all values, their indexes isn't useful.
+     * @return interpolated value
+     */
+    template <typename T>
+    static T linear(const std::vector<float> &originalCoords,
+                    const std::set<Point<T>> &points);
+
+    /**
+     * @brief performs nearest interpolation on given points.
+     * @note it is a wrapper for linearRecurse() private method
+     * @param[in] coordsToInterpolate: coordinates to interpolate
+     * @param[in] points: points to interpolate
+     * @param[in] interpMode: interpolation method, must be a Nearest...
+     * otherwise function will throw an error.
+     * @return interpolated value
+     */
+    template <typename T>
+    static T nearest(const std::vector<float> &coordsToInterpolate,
+                     const std::set<Point<T>> &points,
+                     const Interpolation::Mode nearestMode);
+
+  private:
+    /**
+     * @brief actual linear interpolation function.
+     * will :
+     * - Split all points along each dimension depending of if their coords at
+     * idx alongDim are above or under coordsToInterpolate until they are
+     * 1-to-1.
+     * - Perform interpolation in 2 leftover points and return interpolated
+     * point to parent call with a set of size 1.
+     * - repeat until all dimensions have been interpolated.
+     * @param[in] coordsToInterpolate: coordinates to interpolate
+     * @param[in] points: points to interpolate
+     * @param[in] alongDim: discriminant on along which dimension are being
+     * segregated.
+     * @return
+     */
+    template <typename T>
+    static std::set<Interpolation::Point<T>>
+    linearRecurse(const std::vector<float> &coordsToInterpolate,
+                  const std::set<Point<T>> &points,
+                  const DimIdx_t alongDim = 0);
+};
+
+} // namespace Aidge
+
+#endif // AIDGE_CPU_DATA_INTERPOLATION_H_

include/aidge/backend/cpu/data/TensorImpl.hpp

-#ifndef AIDGE_CPU_DATA_TENSORIMPL_H_
-#define AIDGE_CPU_DATA_TENSORIMPL_H_
-
-#include "aidge/backend/TensorImpl.hpp"
-#include "aidge/data/Tensor.hpp"
-#include "aidge/utils/Registrar.hpp"
-#include "aidge/utils/Types.h"
-
-namespace Aidge {
-template <class T>
-class TensorImpl_cpu : public TensorImpl {
-   private:
-    const Tensor &mTensor;  // Impl needs to access Tensor information, but is not
-                            // supposed to change it!
-    std::vector<T> mData;
-
-   public:
-    static constexpr const char *Backend = "cpu";
-
-    TensorImpl_cpu(const Tensor &tensor) : TensorImpl(Backend), mTensor(tensor) {}
-
-    bool operator==(const TensorImpl &otherImpl) const override final {
-        std::size_t i = 0;
-        for (; i < mTensor.size() &&
-               mData[i] == reinterpret_cast<const TensorImpl_cpu<T> &>(otherImpl).data()[i];
-             ++i) {
-        }
-        return i == mTensor.size();
-    }
-
-    static std::unique_ptr<TensorImpl_cpu> create(const Tensor &tensor) {
-        return std::make_unique<TensorImpl_cpu<T>>(tensor);
-    }
-
-    // native interface
-    const std::vector<T> &data() const { return mData; }
-
-    std::size_t scalarSize() const override { return sizeof(T); }
-
-    void copy(const void *src, NbElts_t length) override {
-        std::copy(static_cast<const T *>(src), static_cast<const T *>(src) + length,
-                  static_cast<T *>(rawPtr()));
-    }
-
-    void *rawPtr() override {
-        lazyInit(mData);
-        return mData.data();
-    };
-
-    virtual ~TensorImpl_cpu() = default;
-
-    void setRawPtr(void *ptr) override final {
-        T *newPtr = static_cast<T *>(ptr);
-        mData = std::vector<T>(newPtr, newPtr + mTensor.size());
-    };
-
-   private:
-    void lazyInit(std::vector<T> &data) {
-        assert(mTensor.dataType() == NativeType<T>::type);
-
-        if (data.size() != mTensor.size()) data.resize(mTensor.size());
-    }
-};
-
-namespace {
-static Registrar<Tensor> registrarTensorImpl_cpu_Float64(
-        {"cpu", DataType::Float64}, Aidge::TensorImpl_cpu<double>::create);
-static Registrar<Tensor> registrarTensorImpl_cpu_Float32(
-        {"cpu", DataType::Float32}, Aidge::TensorImpl_cpu<float>::create);
-static Registrar<Tensor> registrarTensorImpl_cpu_Int32(
-        {"cpu", DataType::Int32}, Aidge::TensorImpl_cpu<int>::create);
-}  // namespace
-}  // namespace Aidge
-
-#endif /* AIDGE_CPU_DATA_TENSORIMPL_H_ */

include/aidge/backend/cpu/operator/AbsImpl.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_CPU_OPERATOR_ABSIMPL_H_
+#define AIDGE_CPU_OPERATOR_ABSIMPL_H_
+
+#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
+#include "aidge/operator/Abs.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+#include <memory>
+#include <vector>
+
+namespace Aidge {
+// Operator implementation entry point for the backend
+using AbsImpl_cpu = OperatorImpl_cpu<Abs_Op,
+    void(const std::size_t, const void*, void*)>;
+
+// Implementation entry point registration to Operator
+REGISTRAR(Abs_Op, "cpu", Aidge::AbsImpl_cpu::create);
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ABSIMPL_H_ */

include/aidge/backend/cpu/operator/ReLUImpl_forward_kernels.hpp → include/aidge/backend/cpu/operator/AbsImpl_kernels.hpp

@@ -9,35 +9,39 @@
  *
  ********************************************************************************/
 
-#ifndef AIDGE_CPU_OPERATOR_RELUIMPL_FORWARD_KERNEL_H_
-#define AIDGE_CPU_OPERATOR_RELUIMPL_FORWARD_KERNEL_H_
+#ifndef AIDGE_CPU_OPERATOR_ABSIMPL_KERNELS_H_
+#define AIDGE_CPU_OPERATOR_ABSIMPL_KERNELS_H_
+
+#include <cmath>
 
 #include "aidge/utils/Registrar.hpp"
 
-#include "aidge/backend/cpu/operator/ReLUImpl.hpp"
+#include "aidge/backend/cpu/operator/AbsImpl.hpp"
 
 namespace Aidge {
 template <class I, class O>
-void ReLUImpl_cpu_forward_kernel(std::size_t inputLenght,
+void AbsImpl_cpu_forward_kernel(std::size_t inputLength,
                                      const void* input_,
                                      void* output_) {
 
     const I* input = static_cast<const I*>(input_);
     O* output = static_cast<O*>(output_);
 
-    for (std::size_t i = 0; i < inputLenght; ++i) {
-        output[i] = input[i] > 0 ? input[i] : 0;
+    for (std::size_t i = 0; i < inputLength; ++i) {
+        output[i] = std::abs(input[i]);
     }
 }
 
-namespace {
-static Registrar<ReLUImplForward_cpu> registrarReLUImplForward_cpu_Float32(
-        {DataType::Float32, DataType::Float32}, Aidge::ReLUImpl_cpu_forward_kernel<float, float>);
-static Registrar<ReLUImplForward_cpu> registrarReLUImplForward_cpu_Int32(
-        {DataType::Int32, DataType::Int32}, Aidge::ReLUImpl_cpu_forward_kernel<int, int>);
-static Registrar<ReLUImplForward_cpu> registrarReLUImplForward_cpu_Float64(
-        {DataType::Float64, DataType::Float64}, Aidge::ReLUImpl_cpu_forward_kernel<double, double>);
-}  // namespace
+// Kernels registration to implementation entry point
+REGISTRAR(AbsImpl_cpu,
+    {DataType::Float32},
+    {ProdConso::inPlaceModel, Aidge::AbsImpl_cpu_forward_kernel<float, float>, nullptr});
+REGISTRAR(AbsImpl_cpu,
+    {DataType::Float64},
+    {ProdConso::inPlaceModel, Aidge::AbsImpl_cpu_forward_kernel<double, double>, nullptr});
+REGISTRAR(AbsImpl_cpu,
+    {DataType::Int32},
+    {ProdConso::inPlaceModel, Aidge::AbsImpl_cpu_forward_kernel<std::int32_t, std::int32_t>, nullptr});
 }  // namespace Aidge
 
-#endif /* AIDGE_CPU_OPERATOR_RELUIMPL_FORWARD_KERNEL_H_ */
+#endif /* AIDGE_CPU_OPERATOR_ABSIMPL_KERNELS_H_ */

include/aidge/backend/cpu/operator/AddImpl.hpp

@@ -12,200 +12,33 @@
 #ifndef AIDGE_CPU_OPERATOR_ADDIMPL_H_
 #define AIDGE_CPU_OPERATOR_ADDIMPL_H_
 
-#include "aidge/backend/OperatorImpl.hpp"
+#include <cstddef>  // std::size_t
+#include <memory>   // std::unique_ptr, std::make_unique
+#include <string>
+#include <vector>
+
+#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
 #include "aidge/operator/Add.hpp"
 #include "aidge/utils/Registrar.hpp"
 #include "aidge/utils/Types.h"
-#include <memory>
-#include <vector>
 
 namespace Aidge {
-// class Add_Op<2>;
-
-// compute kernel registry for forward and backward
-template <DimIdx_t NUM>
-class AddImplForward_cpu;
-template <DimIdx_t NUM>
-class AddImplBackward_cpu;
-
-template <>
-class AddImplForward_cpu<1>
-    : public Registrable<AddImplForward_cpu<1>, std::tuple<DataType, DataType>, void(const std::size_t, const void*, void*)> {};
-template <>
-class AddImplBackward_cpu<1>
-    : public Registrable<AddImplBackward_cpu<1>, std::tuple<DataType, DataType>, void(const std::size_t, const void*, void*)> {};
-
-template <>
-class AddImplForward_cpu<2> : public Registrable<AddImplForward_cpu<2>, std::tuple<DataType, DataType, DataType>,
-                                                     void(const std::size_t, const void*, const void*, void*)> {};
-template <>
-class AddImplBackward_cpu<2> : public Registrable<AddImplBackward_cpu<2>, std::tuple<DataType, DataType, DataType>,
-                                                      void(const std::size_t, const void*, const void*, void*)> {};
-
-template <>
-class AddImplForward_cpu<3> : public Registrable<AddImplForward_cpu<3>, std::tuple<DataType, DataType, DataType, DataType>,
-                                                     void(const std::size_t, const void*, const void*, const void*, void*)> {
-};
-template <>
-class AddImplBackward_cpu<3>
-    : public Registrable<AddImplBackward_cpu<3>, std::tuple<DataType, DataType, DataType, DataType>,
-                         void(const std::size_t, const void*, const void*, const void*, void*)> {};
-
-template <DimIdx_t NUM>
-class AddImpl_cpu : public OperatorImpl {
-   private:
-    const Add_Op<NUM>& mOp;
-    std::array<NbElts_t, NUM> mNbConsumedData = {};
-    std::array<NbElts_t, 1> mNbProducedData = {};
-
-   public:
-    AddImpl_cpu(const Add_Op<NUM>& op) : mOp(op) {}
-
-    static std::unique_ptr<AddImpl_cpu<NUM>> create(const Add_Op<NUM>& op) {
-        return std::make_unique<AddImpl_cpu<NUM>>(op);
-    }
-
-   public:
-    NbElts_t getNbRequiredData(const IOIndex_t inputIdx) const override final {
-        assert(mOp.getInput(inputIdx) && "requires valid input");
-
-        // Requires the whole tensors
-        const auto& inputDims = std::static_pointer_cast<Tensor>(mOp.getInput(inputIdx))->dims();
-        return std::accumulate(inputDims.begin(), inputDims.end(), NbElts_t(1), std::multiplies<NbElts_t>());
-    }
-
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final {
-        // for the direct convolution algorithm, convolutions can be in-place, if there is no padding!
-        return 0;
-    }
-
-    NbElts_t getRequiredMemory(const IOIndex_t outputIdx, const std::vector<DimSize_t>& inputsSize) const override final {
-        // Requires the whole tensors, regardless of available data on inputs
-        assert(outputIdx == 0 && "operator has only one output");
-        (void) outputIdx;
-
-        const auto& outputDims = std::static_pointer_cast<Tensor>(mOp.getOutput(0))->dims();
-        return std::accumulate(outputDims.begin(), outputDims.end(), NbElts_t(1), std::multiplies<NbElts_t>());
-    }
-
-    NbElts_t getNbConsumedData(const IOIndex_t inputIdx) const override final {
-        assert(inputIdx < mNbConsumedData.size());
-        return mNbConsumedData[inputIdx];
-    }
-
-    NbElts_t getNbProducedData(const IOIndex_t outputIdx) const override final {
-        assert(outputIdx < mNbProducedData.size());
-        return mNbProducedData[outputIdx];
-    }
-    void updateConsummerProducer() override final;
-
-    void forward() {
-        // nothing
-    }
-
-    void backward() { printf("Not implemented yet.\n"); }
-};
-
-template <>
-class AddImpl_cpu<1> : public OperatorImpl {
-   private:
-    const Add_Op<1>& mOp;
-    std::array<NbElts_t, 1> mNbConsumedData;
-    std::array<NbElts_t, 1> mNbProducedData;
-
-   public:
-    AddImpl_cpu(const Add_Op<1>& op) : mOp(op), mNbConsumedData({0}), mNbProducedData({0}) {}
-
-    static std::unique_ptr<AddImpl_cpu<1>> create(const Add_Op<1>& op) {
-        return std::make_unique<AddImpl_cpu<1>>(op);
-    }
-
-   public:
-    NbElts_t getNbRequiredData(const IOIndex_t /*inputIdx*/) const override final;
-
-    NbElts_t getNbRequiredProtected(const IOIndex_t /*inputIdx*/) const override final;
-
-    NbElts_t getRequiredMemory(const IOIndex_t /*outputIdx*/,
-                               const std::vector<DimSize_t> &/*inputsSize*/) const override final;
-
-    NbElts_t getNbConsumedData(const IOIndex_t /*inputIdx*/) const override final;
-
-    NbElts_t getNbProducedData(const IOIndex_t /*outputIdx*/) const override final;
-    void updateConsummerProducer() override final;
-
-    void forward();
-
-    void backward();
-};
-
-template <>
-class AddImpl_cpu<2> : public OperatorImpl {
-   private:
-    const Add_Op<2>& mOp;
-    std::array<NbElts_t, 2> mNbConsumedData;
-    std::array<NbElts_t, 1> mNbProducedData;
-
-   public:
-    AddImpl_cpu(const Add_Op<2>& op) : mOp(op), mNbConsumedData({0, 0}), mNbProducedData({0}) {}
-
-    static std::unique_ptr<AddImpl_cpu<2>> create(const Add_Op<2>& op) {
-        return std::make_unique<AddImpl_cpu<2>>(op);
-    }
-
-   public:
-    NbElts_t getNbRequiredData(const IOIndex_t inputIdx) const override final;
-
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
-
-    NbElts_t getRequiredMemory(const IOIndex_t /*outputIdx*/,
-                               const std::vector<DimSize_t>& /*inputsSize*/) const override final;
-
-    NbElts_t getNbConsumedData(const IOIndex_t inputIdx) const override final;
-
-    NbElts_t getNbProducedData(const IOIndex_t /*outputIdx*/) const override final;
-    void updateConsummerProducer() override final;
-
-    void forward();
-
-    void backward();
-};
-
-template <>
-class AddImpl_cpu<3> : public OperatorImpl {
-   private:
-    const Add_Op<3>& mOp;
-    std::array<NbElts_t, 3> mNbConsumedData;
-    std::array<NbElts_t, 1> mNbProducedData;
-
-   public:
-    AddImpl_cpu(const Add_Op<3>& op) : mOp(op), mNbConsumedData({0, 0, 0}), mNbProducedData({0}) {}
-
-    static std::unique_ptr<AddImpl_cpu<3>> create(const Add_Op<3>& op) {
-        return std::make_unique<AddImpl_cpu<3>>(op);
-    }
-
-   public:
-    NbElts_t getNbRequiredData(const IOIndex_t inputIdx) const override final;
-
-    NbElts_t getNbRequiredProtected(const IOIndex_t /*inputIdx*/) const override final;
-
-    NbElts_t getRequiredMemory(const IOIndex_t /*outputIdx*/, const std::vector<DimSize_t>& /*inputsSize*/) const override final;
-
-    NbElts_t getNbConsumedData(const IOIndex_t inputIdx) const override final;
-
-    NbElts_t getNbProducedData(const IOIndex_t outputIdx) const override final;
-    void updateConsummerProducer() override final;
-
-    void forward();
-
-    void backward();
-};
-
-namespace {
-static Registrar<Add_Op<1>> registrarAddImpl1I_cpu("cpu", Aidge::AddImpl_cpu<1>::create);
-static Registrar<Add_Op<2>> registrarAddImpl2I_cpu("cpu", Aidge::AddImpl_cpu<2>::create);
-static Registrar<Add_Op<3>> registrarAddImpl3I_cpu("cpu", Aidge::AddImpl_cpu<3>::create);
-}  // namespace
+// Operator implementation entry point for the backend
+using AddImpl_cpu = OperatorImpl_cpu<Add_Op,
+    void(std::vector<std::size_t>, std::vector<std::size_t>, const std::vector<std::size_t>&, const void*, const void*, void*),
+    void(const std::size_t, 
+         const std::size_t, 
+         const std::size_t, 
+         const std::vector<std::size_t>&, 
+         const std::vector<std::size_t>&, 
+         const std::vector<std::size_t>&, 
+         const void*, 
+         void*, 
+         void*)
+>;
+
+// Implementation entry point registration to Operator
+REGISTRAR(Add_Op, "cpu", Aidge::AddImpl_cpu::create);
 }  // namespace Aidge
 
 #endif /* AIDGE_CPU_OPERATOR_ADDIMPL_H_ */

include/aidge/backend/cpu/operator/AddImpl_forward_kernels.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_CPU_OPERATOR_ADDIMPL_FORWARD_KERNEL_H_
-#define AIDGE_CPU_OPERATOR_ADDIMPL_FORWARD_KERNEL_H_
-
-#include "aidge/utils/Registrar.hpp"
-
-#include "aidge/backend/cpu/operator/AddImpl.hpp"
-
-namespace Aidge {
-
-template <class I1, class O>
-void AddImpl1I_cpu_forward_kernel(const std::size_t inputLength, const void* input1_, void* output_) {
-    // FIXME: missing Add attributes as arguments
-    const I1* input1 = static_cast<const I1*>(input1_);
-    O* output = static_cast<O*>(output_);
-
-    for (std::size_t oIndex = 0; oIndex < inputLength; ++oIndex) {
-        output[oIndex] = input1[oIndex];
-    }
-}
-
-template <class I1, class I2, class O>
-void AddImpl2I_cpu_forward_kernel(const std::size_t inputLength, const void* input1_, const void* input2_,
-                                      void* output_) {
-    // FIXME: missing Add attributes as arguments
-    const I1* input1 = static_cast<const I1*>(input1_);
-    const I2* input2 = static_cast<const I2*>(input2_);
-    O* output = static_cast<O*>(output_);
-
-    for (std::size_t oIndex = 0; oIndex < inputLength; ++oIndex) {
-        output[oIndex] = input1[oIndex] + input2[oIndex];
-    }
-}
-
-template <class I1, class I2, class I3, class O>
-void AddImpl3I_cpu_forward_kernel(const std::size_t inputLength, const void* input1_, const void* input2_,
-                                      const void* input3_, void* output_) {
-    // FIXME: missing Add attributes as arguments
-    const I1* input1 = static_cast<const I1*>(input1_);
-    const I2* input2 = static_cast<const I2*>(input2_);
-    const I3* input3 = static_cast<const I3*>(input3_);
-    O* output = static_cast<O*>(output_);
-
-    for (std::size_t oIndex = 0; oIndex < inputLength; ++oIndex) {
-        output[oIndex] = input1[oIndex] + input2[oIndex] + input3[oIndex];
-    }
-}
-
-namespace {
-static Registrar<AddImplForward_cpu<1>> registrarAddImpl1IForward_cpu_Float32(
-        {DataType::Float32, DataType::Float32}, Aidge::AddImpl1I_cpu_forward_kernel<float, float>);
-static Registrar<AddImplForward_cpu<1>> registrarAddImpl1IForward_cpu_Int32(
-        {DataType::Int32, DataType::Int32}, Aidge::AddImpl1I_cpu_forward_kernel<int, int>);
-static Registrar<AddImplForward_cpu<1>> registrarAddImpl1IForward_cpu_Float64(
-        {DataType::Float64, DataType::Float64}, Aidge::AddImpl1I_cpu_forward_kernel<double, double>);
-
-static Registrar<AddImplForward_cpu<2>> registrarAddImpl2IForward_cpu_Float32(
-        {DataType::Float32, DataType::Float32, DataType::Float32},
-        Aidge::AddImpl2I_cpu_forward_kernel<float, float, float>);
-static Registrar<AddImplForward_cpu<2>> registrarAddImpl2IForward_cpu_Int32(
-        {DataType::Int32, DataType::Int32, DataType::Int32}, Aidge::AddImpl2I_cpu_forward_kernel<int, int, int>);
-static Registrar<AddImplForward_cpu<2>> registrarAddImpl2IForward_cpu_Float64(
-        {DataType::Float64, DataType::Float64, DataType::Float64}, Aidge::AddImpl2I_cpu_forward_kernel<double, double, double>);
-
-static Registrar<AddImplForward_cpu<3>> registrarAddImpl3IForward_cpu_Float32(
-        {DataType::Float32, DataType::Float32, DataType::Float32, DataType::Float32},
-        Aidge::AddImpl3I_cpu_forward_kernel<float, float, float, float>);
-static Registrar<AddImplForward_cpu<3>> registrarAddImpl3IForward_cpu_Int32(
-        {DataType::Int32, DataType::Int32, DataType::Int32, DataType::Int32},
-        Aidge::AddImpl3I_cpu_forward_kernel<int, int, int, int>);
-static Registrar<AddImplForward_cpu<3>> registrarAddImpl3IForward_cpu_Float64(
-        {DataType::Float64, DataType::Float64, DataType::Float64, DataType::Float64},
-        Aidge::AddImpl3I_cpu_forward_kernel<double, double, double, double>);
-}  // namespace
-}  // namespace Aidge
-
-#endif /* AIDGE_CPU_OPERATOR_ADDIMPL_CPU_FORWARD_KERNEL_H_ */

include/aidge/backend/cpu/operator/AddImpl_kernels.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_CPU_OPERATOR_ADDIMPL_KERNELS_H_
+#define AIDGE_CPU_OPERATOR_ADDIMPL_KERNELS_H_
+
+#include "aidge/utils/Registrar.hpp"
+
+#include <cstddef>  // std::size_t
+
+#include "aidge/backend/cpu/data/Broadcasting.hpp"
+#include "aidge/backend/cpu/operator/AddImpl.hpp"
+
+namespace Aidge {
+
+namespace {
+// suppose values are contiguous in memory
+template <class I, class O>
+void add_contiguous_arrays(const std::size_t input1size,
+                            const std::size_t input2size,
+                            const std::size_t output1size,
+                            const I* input1,
+                            const I* input2,
+                            O* output)
+{
+    for (std::size_t i = 0; i < output1size; ++i)
+    {
+        const std::size_t in1_id = (input1size != 1) ? i : 0;
+        const std::size_t in2_id = (input2size != 1) ? i : 0;
+        output[i] = static_cast<O>(input1[in1_id] + input2[in2_id]);
+    }
+}
+}
+
+template <class I, class O>
+void AddImpl_cpu_forward_kernel(std::vector<std::size_t> dims0,
+                                std::vector<std::size_t> dims1,
+                                const std::vector<std::size_t>& outputDims,
+                                const void* input0_,
+                                const void* input1_,
+                                void* output_) {
+
+    const I* input_0 = static_cast<const I*>(input0_);
+    const I* input_1 = static_cast<const I*>(input1_);
+    O* output = static_cast<O*>(output_);
+
+    // [5,2,1,7] & [2,6,7]
+    // 1. Same number of dimensions -> [5,2,1,7] & [1,2,6,7]
+    // 2. Find the highest equal dimension -> 3
+    //    Exception: if the first diverging dimension is the last one, then -> 4 (dims.size())
+    // 3. Compute the highest number of contiguous data -> 7
+    // 4. Compute stride and offset step for the broadcast mechanism
+    // 5. Call a simple kernel
+
+    // special case for equal dimensions, the kernel is called with the entire arrays at once
+    if (dims0 == dims1) {
+        const std::size_t input0_contiguous_size = std::accumulate(dims0.cbegin(), dims0.cend(), std::size_t(1), std::multiplies<std::size_t>());
+        for (std::size_t i = 0; i < input0_contiguous_size; ++i)
+        {
+            output[i] = static_cast<O>(input_0[i] + input_1[i]);
+        }
+        return;
+    }
+
+    // set dimensions to be of equal size by filling the smallest one with ones.
+    if (dims0.size() > dims1.size()) {
+        dims1.insert(dims1.cbegin(), dims0.size() - dims1.size(), std::size_t(1));
+    }
+    else if (dims1.size() > dims0.size()) {
+        dims0.insert(dims0.cbegin(), dims1.size() - dims0.size(), std::size_t(1));
+    }
+
+    const std::size_t nbDims = dims0.size();
+
+    // Find the highest equal dimension
+    // std::size_t contiguousIdx = nbDims - 1;
+    std::size_t contiguousIdx = nbDims;
+    while (contiguousIdx-- > 0) {
+    // for (; contiguousIdx+1 > 0; --contiguousIdx) {
+        if (dims0[contiguousIdx] != dims1[contiguousIdx]) {
+            if (contiguousIdx == (nbDims -1)) { // last dimensions of one of the input Tensor are of size 1
+                const std::vector<std::size_t>& dims = (dims0[contiguousIdx] == 1) ? dims0 : dims1;
+                while ((contiguousIdx+1 > 0) && (dims[contiguousIdx] == 1)) {
+                    --contiguousIdx;
+                }
+            }
+            break;
+        }
+    }
+    ++contiguousIdx;
+
+    // Compute the highest number of contiguous data for each Tensor
+    const std::size_t input0_contiguous_size = std::accumulate(dims0.cbegin()+contiguousIdx, dims0.cend(), std::size_t(1), std::multiplies<std::size_t>());
+    const std::size_t input1_contiguous_size = std::accumulate(dims1.cbegin()+contiguousIdx, dims1.cend(), std::size_t(1), std::multiplies<std::size_t>());
+    const std::size_t output_contiguous_size = std::accumulate(outputDims.cbegin()+contiguousIdx, outputDims.cend(), std::size_t(1), std::multiplies<std::size_t>());
+
+    // initialize strides to iterate through data because of broadcasting
+    std::unique_ptr<std::int32_t[]> stride_post0 = std::make_unique<std::int32_t[]>(contiguousIdx);
+    std::unique_ptr<std::int32_t[]> stride_post1 = std::make_unique<std::int32_t[]>(contiguousIdx);
+    std::unique_ptr<std::int32_t[]> stride_step0 = std::make_unique<std::int32_t[]>(contiguousIdx);
+    std::unique_ptr<std::int32_t[]> stride_step1 = std::make_unique<std::int32_t[]>(contiguousIdx);
+    if (contiguousIdx > 0) {
+        stride_post0[contiguousIdx - 1] = 1;
+        stride_post1[contiguousIdx - 1] = 1;
+        for (std::size_t i = contiguousIdx - 2; i != static_cast<std::size_t>(-1); --i) {
+            stride_post0[i] = stride_post0[i+1]*static_cast<std::int32_t>(dims0[i+1]);
+            stride_post1[i] = stride_post1[i+1]*static_cast<std::int32_t>(dims1[i+1]);
+        }
+        for (std::size_t i = 0; i != contiguousIdx; ++i) {
+            stride_step0[i] = (dims0[i] == 1) ? 1 - stride_post0[i] : 1;
+            stride_step1[i] = (dims1[i] == 1) ? 1 - stride_post1[i] : 1;
+        }
+    }
+
+    // variables for arrays offsets
+    std::size_t offsetIn0 = 0;
+    std::size_t offsetIn1 = 0;
+    std::size_t offsetOut = 0;
+
+
+    std::size_t dim = contiguousIdx - 1;
+    const std::size_t nbStacks = std::accumulate(outputDims.cbegin(), outputDims.cbegin() + contiguousIdx, std::size_t(1), std::multiplies<std::size_t>());
+    for (std::size_t stack = 0; stack < nbStacks;) {
+        add_contiguous_arrays<I,O>(input0_contiguous_size, input1_contiguous_size, output_contiguous_size,
+                    input_0 + offsetIn0*input0_contiguous_size,
+                    input_1 + offsetIn1*input1_contiguous_size,
+                    output + offsetOut*output_contiguous_size);
+        if (++stack < nbStacks) {
+            std::size_t tmp_stack = stack;
+            while(tmp_stack % outputDims[dim] == 0) {
+                tmp_stack /= outputDims[dim];
+                dim--;
+            }
+            offsetIn0 += stride_step0[dim];
+            offsetIn1 += stride_step1[dim];
+            ++offsetOut;
+            dim = contiguousIdx - 1;
+        }
+    }
+}
+
+template <class I, class O>
+void AddImpl_cpu_backward_kernel(const std::size_t input0Length,
+                               const std::size_t input1Length,
+                               const std::size_t gradOutputLength,
+                               const std::vector<std::size_t>& dims0,
+                               const std::vector<std::size_t>& dims1,
+                               const std::vector<std::size_t>& outputDims,
+                               const void* grad_output_,
+                               void* gradientInput0_,
+                               void* gradientInput1_)
+{
+    // TODO: Remove input0/1 from the function
+    const O* gradOutput = static_cast<const O*>(grad_output_);
+    auto* gradInput0 = static_cast<I*>(gradientInput0_);
+    auto* gradInput1 = static_cast<I*>(gradientInput1_);
+
+    std::fill_n(gradInput0, input0Length, static_cast<I>(0));
+    std::fill_n(gradInput1, input1Length, static_cast<I>(0));
+
+    auto broadcastedDims0 = getBroadcastedDims(outputDims, dims0);
+    auto broadcastedDims1 = getBroadcastedDims(outputDims, dims1);
+
+    for (std::size_t i = 0; i < gradOutputLength; ++i) {
+        auto idxOutputGrad = getMultiDimIndices(outputDims, i);
+        std::vector<std::size_t> idxInput0(broadcastedDims0.size());
+        std::vector<std::size_t> idxInput1(broadcastedDims1.size());
+
+        for (std::size_t dimension = 0; dimension < broadcastedDims0.size(); ++dimension) {
+            idxInput0[dimension] = (broadcastedDims0[dimension] == 1) ? 0 : idxOutputGrad[dimension];
+        }
+
+        for (std::size_t dimension = 0; dimension < broadcastedDims1.size(); ++dimension) {
+            idxInput1[dimension] = (broadcastedDims1[dimension] == 1) ? 0 : idxOutputGrad[dimension];
+        }
+
+        auto idx0 = getFlattenedIndex(broadcastedDims0, idxInput0);
+        auto idx1 = getFlattenedIndex(broadcastedDims1, idxInput1);
+
+        // For addition: gradient of both inputs is just the output gradient
+        // (unlike multiplication where we need to multiply by the other input,
+        // or subtraction where we need to negate one of them)
+        gradInput0[idx0] += static_cast<I>(gradOutput[i]);
+        gradInput1[idx1] += static_cast<I>(gradOutput[i]);
+    }
+}
+
+// Kernels registration to implementation entry point
+REGISTRAR(AddImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Float32}},
+    {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<float, float>, Aidge::AddImpl_cpu_backward_kernel<float, float>});
+REGISTRAR(AddImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Float64}},
+    {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<double, double>, Aidge::AddImpl_cpu_backward_kernel<double, double>});
+REGISTRAR(AddImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Int8}},
+    {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::int8_t, std::int8_t>, Aidge::AddImpl_cpu_backward_kernel<std::int8_t, std::int8_t>});
+REGISTRAR(AddImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::UInt8}},
+    {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::uint8_t, std::uint8_t>, Aidge::AddImpl_cpu_backward_kernel<std::uint8_t, std::uint8_t>});
+REGISTRAR(AddImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Int32}},
+    {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::int32_t, std::int32_t>, Aidge::AddImpl_cpu_backward_kernel<std::int32_t, std::int32_t>});
+REGISTRAR(AddImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Int64}},
+    {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::int64_t, std::int64_t>, Aidge::AddImpl_cpu_backward_kernel<std::int64_t, std::int64_t>});
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ADDIMPL_CPU_KERNELS_H_ */

include/aidge/backend/cpu/operator/AndImpl.hpp

+/********************************************************************************
+ * Copyright (c) 2024 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_CPU_OPERATOR_ANDIMPL_H_
+#define AIDGE_CPU_OPERATOR_ANDIMPL_H_
+
+#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
+#include "aidge/operator/And.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
+#include <memory>
+#include <vector>
+
+namespace Aidge {
+// Operator implementation entry point for the backend
+using AndImpl_cpu = OperatorImpl_cpu<And_Op,
+    void(std::vector<std::size_t>, std::vector<std::size_t>, const std::vector<std::size_t>&, const void*, const void*, void*)>;
+
+// Implementation entry point registration to Operator
+REGISTRAR(And_Op, "cpu", Aidge::AndImpl_cpu::create);
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ANDIMPL_H_ */

include/aidge/backend/cpu/operator/AndImpl_kernels.hpp

+/********************************************************************************
+ * Copyright (c) 2024 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_CPU_OPERATOR_ANDIMPL_KERNELS_H_
+#define AIDGE_CPU_OPERATOR_ANDIMPL_KERNELS_H_
+
+#include "aidge/backend/cpu/operator/AndImpl.hpp"
+#include "aidge/utils/Registrar.hpp"
+
+namespace Aidge {
+
+namespace {
+// suppose values are contiguous in memory
+template <class I, class O>
+void and_contiguous_arrays(const std::size_t input1size,
+                            const std::size_t input2size,
+                            const std::size_t output1size,
+                            const I* input1,
+                            const I* input2,
+                            O* output)
+{
+    for (std::size_t i = 0; i < output1size; ++i)
+    {
+        const std::size_t in1_id = (input1size != 1) ? i : 0;
+        const std::size_t in2_id = (input2size != 1) ? i : 0;
+        output[i] = static_cast<O>(input1[in1_id] && input2[in2_id]);
+    }
+}
+}
+
+
+template <class I, class O>
+void AndImpl_cpu_forward_kernel(std::vector<std::size_t> dims0,
+                                std::vector<std::size_t> dims1,
+                                const std::vector<std::size_t>& outputDims,
+                                const void* input0_,
+                                const void* input1_,
+                                void* output_) {
+
+    const I* input_0 = static_cast<const I*>(input0_);
+    const I* input_1 = static_cast<const I*>(input1_);
+    O* output = static_cast<O*>(output_);
+
+    // [5,2,1,7] & [2,6,7]
+    // 1. Same number of dimensions -> [5,2,1,7] & [1,2,6,7]
+    // 2. Find the highest equal dimension -> 3
+    //    Exception: if the first diverging dimension is the last one, then -> 4 (dims.size())
+    // 3. Compute the highest number of contiguous data -> 7
+    // 4. Compute stride and offset step for the broadcast mechanism
+    // 5. Call a simple kernel
+
+    // special case for equal dimensions, the kernel is called with the entire arrays at once
+    if (dims0 == dims1) {
+        const std::size_t input0_contiguous_size = std::accumulate(dims0.cbegin(), dims0.cend(), std::size_t(1), std::multiplies<std::size_t>());
+        for (std::size_t i = 0; i < input0_contiguous_size; ++i) {
+            output[i] = static_cast<O>(input_0[i] && input_1[i]);
+        }
+        return;
+    }
+
+    // set dimensions to be of equal size by filling the smallest one with ones.
+    if (dims0.size() > dims1.size()) {
+        dims1.insert(dims1.cbegin(), dims0.size() - dims1.size(), std::size_t(1));
+    }
+    else if (dims1.size() > dims0.size()) {
+        dims0.insert(dims0.cbegin(), dims1.size() - dims0.size(), std::size_t(1));
+    }
+
+    const std::size_t nbDims = dims0.size();
+
+    // Find the highest equal dimension
+    // std::size_t contiguousIdx = nbDims - 1;
+    std::size_t contiguousIdx = nbDims;
+    while (contiguousIdx-- > 0) {
+    // for (; contiguousIdx+1 > 0; --contiguousIdx) {
+        if (dims0[contiguousIdx] != dims1[contiguousIdx]) {
+            if (contiguousIdx == (nbDims -1)) { // last dimensions of one of the input Tensor are of size 1
+                const std::vector<std::size_t>& dims = (dims0[contiguousIdx] == 1) ? dims0 : dims1;
+                while ((contiguousIdx+1 > 0) && (dims[contiguousIdx] == 1)) {
+                    --contiguousIdx;
+                }
+            }
+            break;
+        }
+    }
+    ++contiguousIdx;
+
+    // Compute the highest number of contiguous data for each Tensor
+    const std::size_t input0_contiguous_size = std::accumulate(dims0.cbegin()+contiguousIdx, dims0.cend(), std::size_t(1), std::multiplies<std::size_t>());
+    const std::size_t input1_contiguous_size = std::accumulate(dims1.cbegin()+contiguousIdx, dims1.cend(), std::size_t(1), std::multiplies<std::size_t>());
+    const std::size_t output_contiguous_size = std::accumulate(outputDims.cbegin()+contiguousIdx, outputDims.cend(), std::size_t(1), std::multiplies<std::size_t>());
+
+    // initialize strides to iterate through data because of broadcasting
+    std::unique_ptr<std::int32_t[]> stride_post0 = std::make_unique<std::int32_t[]>(contiguousIdx);
+    std::unique_ptr<std::int32_t[]> stride_post1 = std::make_unique<std::int32_t[]>(contiguousIdx);
+    std::unique_ptr<std::int32_t[]> stride_step0 = std::make_unique<std::int32_t[]>(contiguousIdx);
+    std::unique_ptr<std::int32_t[]> stride_step1 = std::make_unique<std::int32_t[]>(contiguousIdx);
+    if (contiguousIdx > 0) {
+        stride_post0[contiguousIdx - 1] = 1;
+        stride_post1[contiguousIdx - 1] = 1;
+        for (std::size_t i = contiguousIdx - 2; i != static_cast<std::size_t>(-1); --i) {
+            stride_post0[i] = stride_post0[i+1]*static_cast<std::int32_t>(dims0[i+1]);
+            stride_post1[i] = stride_post1[i+1]*static_cast<std::int32_t>(dims1[i+1]);
+        }
+        for (std::size_t i = 0; i != contiguousIdx; ++i) {
+            stride_step0[i] = (dims0[i] == 1) ? 1 - stride_post0[i] : 1;
+            stride_step1[i] = (dims1[i] == 1) ? 1 - stride_post1[i] : 1;
+        }
+    }
+
+    // variables for arrays offsets
+    std::size_t offsetIn0 = 0;
+    std::size_t offsetIn1 = 0;
+    std::size_t offsetOut = 0;
+
+
+    std::size_t dim = contiguousIdx - 1;
+    const std::size_t nbStacks = std::accumulate(outputDims.cbegin(), outputDims.cbegin() + contiguousIdx, std::size_t(1), std::multiplies<std::size_t>());
+    for (std::size_t stack = 0; stack < nbStacks;) {
+        and_contiguous_arrays<I,O>(input0_contiguous_size, input1_contiguous_size, output_contiguous_size,
+                    input_0 + offsetIn0*input0_contiguous_size,
+                    input_1 + offsetIn1*input1_contiguous_size,
+                    output + offsetOut*output_contiguous_size);
+        if (++stack < nbStacks) {
+            std::size_t tmp_stack = stack;
+            while(tmp_stack % outputDims[dim] == 0) {
+                tmp_stack /= outputDims[dim];
+                dim--;
+            }
+            offsetIn0 += stride_step0[dim];
+            offsetIn1 += stride_step1[dim];
+            ++offsetOut;
+            dim = contiguousIdx - 1;
+        }
+    }
+}
+
+// Kernels registration to implementation entry point
+REGISTRAR(AndImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Float32}},
+    {ProdConso::inPlaceModel, Aidge::AndImpl_cpu_forward_kernel<float, float>, nullptr});
+REGISTRAR(AndImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Float64}},
+    {ProdConso::inPlaceModel, Aidge::AndImpl_cpu_forward_kernel<double, double>, nullptr});
+REGISTRAR(AndImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Int32}},
+    {ProdConso::inPlaceModel, Aidge::AndImpl_cpu_forward_kernel<std::int32_t, std::int32_t>, nullptr});
+REGISTRAR(AndImpl_cpu,
+    {ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Int64}},
+    {ProdConso::inPlaceModel, Aidge::AndImpl_cpu_forward_kernel<std::int64_t, std::int64_t>, nullptr});
+
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ANDIMPL_KERNELS_H_ */

include/aidge/backend/cpu/operator/ArgMaxImpl.hpp

+/********************************************************************************
+ * Copyright (c) 2024 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_CPU_OPERATOR_ARGMAXIMPL_H_
+#define AIDGE_CPU_OPERATOR_ARGMAXIMPL_H_
+
+#include <array>
+#include <memory>
+#include <tuple>
+#include <vector>
+
+#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
+#include "aidge/operator/ArgMax.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+
+namespace Aidge {
+// Operator implementation entry point for the backend
+using ArgMaxImpl_cpu = OperatorImpl_cpu<ArgMax_Op,
+    void(std::int32_t,
+        DimSize_t,
+        const std::vector<DimSize_t>&,
+        const void *,
+        void *)>;
+
+// Implementation entry point registration to Operator
+REGISTRAR(ArgMax_Op, "cpu", Aidge::ArgMaxImpl_cpu::create);
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ARGMAXIMPL_H_ */

include/aidge/backend/cpu/operator/ArgMaxImpl_kernels.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_CPU_OPERATOR_ARGMAXIMPL_KERNELS_H_
+#define AIDGE_CPU_OPERATOR_ARGMAXIMPL_KERNELS_H_
+
+#include <algorithm>   // std::for_each
+#include <cstddef>     // std::size_t
+#include <cstdint>     // std::int32_t
+#include <functional>  //std::multiplies
+#include <numeric>     //std::accumulate
+#include <vector>
+#include <limits>
+
+#include "aidge/backend/cpu/operator/ArgMaxImpl.hpp"
+#include "aidge/data/Data.hpp"
+#include "aidge/operator/ArgMax.hpp"
+#include "aidge/utils/Registrar.hpp"
+
+namespace Aidge {
+template <class I, class O>
+void ArgMaxImpl_cpu_forward_kernel(std::int32_t axis_,
+                                    DimSize_t select_last_index,
+                                    const std::vector<DimSize_t>& inputDims,
+                                    const void* input_,
+                                    void* output_) {
+
+    const I* input = static_cast<const I*>(input_);
+    O* output = static_cast<O*>(output_);
+
+    const std::size_t axis = static_cast<std::size_t>(axis_);
+
+    std::size_t stride_post = 1;
+    for (std::size_t i = axis + 1; i < inputDims.size(); ++i) {
+        stride_post *= inputDims[i];
+    }
+    std::size_t stride_pre = 1;
+    for (std::size_t i = 0; i < axis; ++i) {
+        stride_pre *= inputDims[i];
+    }
+    const std::size_t dim_i = inputDims[axis];
+    for (std::size_t pre = 0; pre < stride_pre; ++pre) {
+        for (std::size_t post = 0; post < stride_post; ++post) {
+            const std::size_t idx_i = pre * dim_i * stride_post + post;
+            const std::size_t idx_o = pre * stride_post + post;
+            I max = std::numeric_limits<I>::min();
+            for (std::size_t i = 0; i < dim_i; ++i) {
+                I curr_value = input[idx_i + i*stride_post];
+                if (select_last_index) {
+                    if (curr_value>=max) {
+                        output[idx_o] = i;
+                        max = curr_value;
+                    }
+                }
+                else {
+                    if (curr_value > max) {
+                        output[idx_o] = i;
+                        max = curr_value;
+                    }
+                }
+            }
+        }
+    }
+
+}
+
+// Kernels registration to implementation entry point
+REGISTRAR(ArgMaxImpl_cpu,
+    {DataType::Float32},
+    {ProdConso::defaultModel, Aidge::ArgMaxImpl_cpu_forward_kernel<float, float>, nullptr});
+REGISTRAR(ArgMaxImpl_cpu,
+    {DataType::Float64},
+    {ProdConso::defaultModel, Aidge::ArgMaxImpl_cpu_forward_kernel<double, double>, nullptr});
+REGISTRAR(ArgMaxImpl_cpu,
+    {DataType::Int32},
+    {ProdConso::defaultModel, Aidge::ArgMaxImpl_cpu_forward_kernel<std::int32_t, std::int32_t>, nullptr});
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ARGMAXIMPL_KERNELS_H_ */

include/aidge/backend/cpu/operator/AtanImpl.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_CPU_OPERATOR_ATAN_H_
+#define AIDGE_CPU_OPERATOR_ATAN_H_
+
+#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
+#include "aidge/operator/Atan.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
+#include <memory>
+#include <vector>
+
+namespace Aidge {
+// Operator implementation entry point for the backend
+using AtanImpl_cpu = OperatorImpl_cpu<Atan_Op,
+    void(const std::size_t, const void*, void*),
+    void(const std::size_t, const void*, const void*, void*)>;
+
+// Implementation entry point registration to Operator
+REGISTRAR(Atan_Op, "cpu", Aidge::AtanImpl_cpu::create);
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ATAN_H_ */

include/aidge/backend/cpu/operator/AtanImpl_kernels.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_CPU_OPERATOR_ATANIMPL_KERNELS_H_
+#define AIDGE_CPU_OPERATOR_ATANIMPL_KERNELS_H_
+
+#include "aidge/utils/Registrar.hpp"
+
+#include "aidge/backend/cpu/operator/AtanImpl.hpp"
+#include <cmath>  // For atan()
+
+
+namespace Aidge {
+template <class I, class O>
+void AtanImpl_cpu_forward_kernel(std::size_t inputLength,
+                                    const void* input_,
+                                    void* output_) {
+    const I* input = static_cast<const I*>(input_);
+    O* output = static_cast<O*>(output_);
+
+    for (size_t i = 0; i < inputLength; ++i) {
+        output[i] = static_cast<O>(atan(input[i]));
+    }
+
+}
+
+template <class O, class GI, class GO>
+void AtanImpl_cpu_backward_kernel(const std::size_t inputLength,
+                                     const void* output_, const void* grad_output_,
+				     void* grad_input_) {
+    const O* output = static_cast<const O*>(output_);
+    const GO* grad_output = static_cast<const GO*>(grad_output_);
+    GI* grad_input = static_cast<GI*>(grad_input_);
+
+    // Apply the derivative of atan for each element in the input array
+    for (size_t i = 0; i < inputLength; ++i) {
+        // dx = dy * (1 / (1 + x^2))
+        grad_input[i] += grad_output[i] * static_cast<O>(1.0 / (1.0 + output[i] * output[i]));
+    }
+}
+
+
+// Kernels registration to implementation entry point
+REGISTRAR(AtanImpl_cpu,
+    {DataType::Float32},
+    {ProdConso::inPlaceModel, Aidge::AtanImpl_cpu_forward_kernel<float, float>, Aidge::AtanImpl_cpu_backward_kernel<float, float, float>});
+REGISTRAR(AtanImpl_cpu,
+    {DataType::Float64},
+    {ProdConso::inPlaceModel, Aidge::AtanImpl_cpu_forward_kernel<double, double>, Aidge::AtanImpl_cpu_backward_kernel<double, double, double>});
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ATANIMPL_KERNELS_H_ */

include/aidge/backend/cpu/operator/AvgPoolingImpl.hpp

@@ -17,54 +17,26 @@
 #include <tuple>
 #include <vector>
 
-#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
 #include "aidge/operator/AvgPooling.hpp"
 #include "aidge/utils/Registrar.hpp"
 #include "aidge/utils/Types.h"
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
 
 namespace Aidge {
-// class AvgPooling_Op;
-
-// compute kernel registry for forward and backward
-class AvgPoolingImpl2DForward_cpu
-    : public Registrable<AvgPoolingImpl2DForward_cpu,
-                         std::tuple<DataType, DataType>,
-                         void(const AvgPooling_Op<2>::Attrs &, const std::array<DimSize_t, 4> &, const void *, void *)> {};
-class AvgPoolingImpl2DBackward_cpu
-    : public Registrable<AvgPoolingImpl2DBackward_cpu,
-                         std::tuple<DataType, DataType>,
-                         void(const AvgPooling_Op<2>::Attrs &, const std::array<DimSize_t, 4> &, const void *, void *)> {};
-
-class AvgPoolingImpl2D_cpu : public OperatorImpl {
-   private:
-    const AvgPooling_Op<2> &mOp;
-    std::array<NbElts_t, 1> mNbConsumedData;
-    std::array<NbElts_t, 1> mNbProducedData;
-
-   public:
-    AvgPoolingImpl2D_cpu(const AvgPooling_Op<2> &op) : mOp(op), mNbConsumedData({0}), mNbProducedData({0}) {}
-
-    static std::unique_ptr<AvgPoolingImpl2D_cpu> create(const AvgPooling_Op<2> &op) {
-        return std::make_unique<AvgPoolingImpl2D_cpu>(op);
-    }
-
-   public:
-    NbElts_t getNbRequiredData(const IOIndex_t inputIdx) const override final;
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
-    NbElts_t getRequiredMemory(const IOIndex_t /*outputIdx*/, const std::vector<DimSize_t> &inputsSize) const override final;
-    NbElts_t getNbConsumedData(const IOIndex_t inputIdx) const override final;
-    NbElts_t getNbProducedData(const IOIndex_t outputIdx) const override final;
-    void updateConsummerProducer() override final;
-
-    void forward();
-
-    void backward();
-};
-
-namespace {
-// add cpu backend to AvgPooling_Op<2> implementation registry
-static Registrar<AvgPooling_Op<2>> registrarAvgPoolingImpl2D_cpu("cpu", Aidge::AvgPoolingImpl2D_cpu::create);
-}  // namespace
+// Operator implementation entry point for the backend
+using AvgPooling2D_Op = AvgPooling_Op<2>;
+using AvgPoolingImpl2D_cpu = OperatorImpl_cpu<AvgPooling_Op<2>,
+    void(const std::array<DimSize_t, 2>&,
+        const std::array<DimSize_t, 2>&,
+        const std::array<DimSize_t, 2>&,
+        const std::array<DimSize_t, 4>&,
+        bool,
+        const void *,
+        void *)>;
+
+// Implementation entry point registration to Operator
+REGISTRAR(AvgPooling2D_Op, "cpu", Aidge::AvgPoolingImpl2D_cpu::create);
 }  // namespace Aidge
 
 #endif /* AIDGE_CPU_OPERATOR_AVGPOOLINGIMPL_H_ */

include/aidge/backend/cpu/operator/AvgPoolingImpl_forward_kernels.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_CPU_OPERATOR_AVGPOOLINGIMPL_FORWARD_KERNEL_H_
-#define AIDGE_CPU_OPERATOR_AVGPOOLINGIMPL_FORWARD_KERNEL_H_
-
-#include "aidge/utils/Registrar.hpp"
-
-#include "aidge/backend/cpu/operator/AvgPoolingImpl.hpp"
-#include "aidge/utils/Types.h"
-#include "aidge/data/Data.hpp"
-#include <array>
-#include <tuple>
-#include <cmath>
-
-namespace Aidge {
-/**
- * @brief Forward kernel for 2D AvgPoolingolution on CPU backend.
- * @tparam I Input data type.
- * @tparam O Output data type.
- * @param params tuple of Attributes from the Operator
- * @param dims Array of input dimensions.
- * @param input_ const input Tensor.
- * @param output_ Output Tensor.
- */
-template <class I, class O>
-void AvgPoolingImpl2D_cpu_forward_kernel(const AvgPooling_Op<2>::Attrs &params,
-                                             const std::array<DimSize_t, 4> &dims,
-                                             const void *input_,
-                                             void *output_) {
-    // FIXME: missing convolution attributes as arguments
-    const I *input = static_cast<const I *>(input_);
-    O *output = static_cast<O *>(output_);
-
-
-    // output H size
-    const std::size_t oxSize =
-            static_cast<std::size_t>(std::floor(static_cast<float>(dims[2] + std::get<2>(params)[0] + std::get<2>(params)[2] - std::get<1>(params)[0] + std::get<0>(params)[0]) /
-                                static_cast<float>(std::get<0>(params)[0])));
-    // output W size
-    const std::size_t oySize =
-            static_cast<std::size_t>(std::floor(static_cast<float>(dims[3] + std::get<2>(params)[1] + std::get<2>(params)[3] - std::get<1>(params)[1] + std::get<0>(params)[1]) /
-                                static_cast<float>(std::get<0>(params)[1])));
-
-    // TODO: kernel computation
-    // output (batch, outCh, Xout, Yout)
-    // input  (batch, ch, Xin, Yin)
-    // weight (outCh, ch, kernelX, kernelY)
-    // does not take Dilation attribute into account
-    using signedsize = std::make_signed<std::size_t>::type;
-    for (std::size_t batch = 0; batch < dims[0]; ++batch) {
-        for (std::size_t ch = 0; ch < dims[1]; ++ch) {
-            const std::size_t oIndex = (ch + batch*dims[1]) * oxSize * oySize;
-            const std::size_t iIndex = (ch + batch*dims[1]) * dims[2] * dims[3];
-            for (std::size_t ox = 0; ox < oxSize; ++ox) {
-                const signedsize difx = static_cast<signedsize>(std::get<2>(params)[0] - ox * std::get<0>(params)[0]);
-                const std::size_t sxMin = static_cast<std::size_t>(std::max(difx, signedsize(0)));
-                const std::size_t sxMax = (static_cast<signedsize>(dims[2]) + difx) < 0 ? 0 : ((dims[2] + difx) > std::get<1>(params)[0] ? std::get<1>(params)[0] : dims[2] + difx);
-                for (std::size_t oy = 0; oy < oySize; ++oy) {
-                    const signedsize dify = static_cast<signedsize>(std::get<2>(params)[1] - oy * std::get<0>(params)[1]);
-                    const std::size_t syMin = static_cast<std::size_t>(std::max(dify, signedsize(0)));
-                    const std::size_t syMax = (static_cast<signedsize>(dims[3]) + dify) < 0 ? 0 : ((dims[3] + dify) > std::get<1>(params)[1] ? std::get<1>(params)[1] : dims[3] + dify);
-                    const std::size_t oIndexFull = oIndex + ox*oySize + oy;
-                    const std::size_t ix = ox * std::get<0>(params)[0];
-                    const std::size_t iy = oy * std::get<0>(params)[1];
-
-                    if (sxMin == 0 && syMin == 0 && sxMax == 3 && syMax == 3) {
-                        output[oIndexFull] += static_cast<O>(
-                                               input[iIndex + (ix+0)*dims[3] + (iy+0)] +
-                                               input[iIndex + (ix+0)*dims[3] + (iy+1)] +
-                                               input[iIndex + (ix+0)*dims[3] + (iy+2)] +
-                                               input[iIndex + (ix+1)*dims[3] + (iy+0)] +
-                                               input[iIndex + (ix+1)*dims[3] + (iy+1)] +
-                                               input[iIndex + (ix+1)*dims[3] + (iy+2)] +
-                                               input[iIndex + (ix+2)*dims[3] + (iy+0)] +
-                                               input[iIndex + (ix+2)*dims[3] + (iy+1)] +
-                                               input[iIndex + (ix+2)*dims[3] + (iy+2)]) / O(9);
-                    } else {
-                        for (std::size_t sx = sxMin; sx < sxMax; ++sx) {
-                            for (std::size_t sy = syMin; sy < syMax; ++sy) {
-                                output[oIndexFull] += input[iIndex + (ix+sx)*dims[3] + (iy+sy)];
-                            }
-                        }
-                        // padding not used
-                        output[oIndexFull] /= (sxMax - sxMin) * (syMax - syMin);
-                    }
-                }
-            }
-        }
-    }
-}
-
-namespace {
-static Registrar<AvgPoolingImpl2DForward_cpu> registrarAvgPoolingImpl2DForward_cpu_Float32(
-        std::tuple<DataType, DataType>({DataType::Float32, DataType::Float32}),
-        Aidge::AvgPoolingImpl2D_cpu_forward_kernel<float, float>);
-static Registrar<AvgPoolingImpl2DForward_cpu> registrarAvgPoolingImpl2DForward_cpu_Int32(
-        {DataType::Int32, DataType::Int32},
-        Aidge::AvgPoolingImpl2D_cpu_forward_kernel<int, int>);
-static Registrar<AvgPoolingImpl2DForward_cpu> registrarAvgPoolingImpl2DForward_cpu_Float64(
-        {DataType::Float64, DataType::Float64},
-        Aidge::AvgPoolingImpl2D_cpu_forward_kernel<double, double>);
-}  // namespace
-}  // namespace Aidge
-
-#endif /* AIDGE_CPU_OPERATOR_AVGPOOLINGIMPL_FORWARD_KERNEL_H_ */

include/aidge/backend/cpu/operator/AvgPoolingImpl_kernels.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_CPU_OPERATOR_AVGPOOLINGIMPL_KERNELS_H_
+#define AIDGE_CPU_OPERATOR_AVGPOOLINGIMPL_KERNELS_H_
+
+#include <array>
+#include <tuple>
+#include <cmath>
+
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
+#include "aidge/backend/cpu/operator/AvgPoolingImpl.hpp"
+#include "aidge/data/Data.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+
+namespace Aidge {
+
+template <typename T>
+using Acc_T = typename std::conditional<std::is_floating_point<T>::value, T, double>::type;
+
+template <typename T>
+typename std::enable_if<std::is_floating_point<T>::value, T>::type
+castFromFloat(T value) {
+  return value;
+}
+
+template <typename T>
+typename std::enable_if<!std::is_floating_point<T>::value, T>::type
+castFromFloat(double value) {
+  return static_cast<T>(std::nearbyint(value));
+}
+
+/**
+ * @brief Forward kernel for 2D AvgPoolingolution on CPU backend.
+ * @tparam I Input data type.
+ * @tparam O Output data type.
+ * @param params tuple of Attributes from the Operator
+ * @param dims Array of input dimensions.
+ * @param input_ const input Tensor.
+ * @param output_ Output Tensor.
+ */
+template <class I, class O>
+void AvgPoolingImpl2D_cpu_forward_kernel(const std::array<DimSize_t, 2>& strideDims,
+                                        const std::array<DimSize_t, 2>& kernelDims,
+                                        const std::array<DimSize_t, 2>& dilations,
+                                        const std::array<DimSize_t, 4> &dims,
+                                        bool ceilMode,
+                                        const void *input_,
+                                        void *output_) {
+    const I *input = static_cast<const I *>(input_);
+    O *output = static_cast<O *>(output_);
+
+    // output H size
+    const std::size_t oxSize = 
+        ceilMode 
+        ? static_cast<std::size_t>(std::ceil(static_cast<float>(dims[2] - (kernelDims[0] - 1) * dilations[0] - 1 + strideDims[0]) /
+                                            static_cast<float>(strideDims[0])))
+        : static_cast<std::size_t>(std::floor(static_cast<float>(dims[2] - (kernelDims[0] - 1) * dilations[0] - 1 + strideDims[0]) /
+                                            static_cast<float>(strideDims[0])));
+    // output W size
+    const std::size_t oySize = 
+        ceilMode 
+        ? static_cast<std::size_t>(std::ceil(static_cast<float>(dims[3] - (kernelDims[1] - 1) * dilations[1] - 1 + strideDims[1]) /
+                                            static_cast<float>(strideDims[1])))
+        : static_cast<std::size_t>(std::floor(static_cast<float>(dims[3] - (kernelDims[1] - 1) * dilations[1] - 1 + strideDims[1]) /
+                                            static_cast<float>(strideDims[1])));
+
+    using signedsize = std::make_signed<std::size_t>::type;
+
+#ifdef _OPENMP
+    #pragma omp parallel for collapse(2) if (dims[0] * dims[1] >= 16)
+#endif
+    for (int batch = 0; batch < static_cast<int>(dims[0]); ++batch) {
+        for (int ch = 0; ch < static_cast<int>(dims[1]); ++ch) {
+            const std::size_t oIndex = (ch + batch * dims[1]) * oxSize * oySize;
+            const std::size_t iIndex = (ch + batch * dims[1]) * dims[2] * dims[3];
+
+            for (std::size_t ox = 0; ox < oxSize; ++ox) {
+                const signedsize difx = static_cast<signedsize>(-ox * strideDims[0]);
+                const std::size_t sxMin = static_cast<std::size_t>(std::max(difx, signedsize(0)));
+                const std::size_t sxMax = (static_cast<signedsize>(dims[2]) + difx) < 0 ? 0 : ((dims[2] + difx) > kernelDims[0] ? kernelDims[0] : dims[2] + difx);
+
+                for (std::size_t oy = 0; oy < oySize; ++oy) {
+                    const signedsize dify = static_cast<signedsize>(-oy * strideDims[1]);
+                    const std::size_t syMin = static_cast<std::size_t>(std::max(dify, signedsize(0)));
+                    const std::size_t syMax = (static_cast<signedsize>(dims[3]) + dify) < 0 ? 0 : ((dims[3] + dify) > kernelDims[1] ? kernelDims[1] : dims[3] + dify);
+
+                    const std::size_t oIndexFull = oIndex + ox * oySize + oy;
+                    const std::size_t ix = ox * strideDims[0];
+                    const std::size_t iy = oy * strideDims[1];
+
+                    Acc_T<I> sum = static_cast<Acc_T<I>>(0);
+                    std::size_t count = 0;
+
+                    for (unsigned int sy = syMin; sy < syMax; ++sy) {
+                        for (unsigned int sx = sxMin; sx < sxMax; ++sx) {
+                            // Apply dilation factor
+                            const std::size_t dilated_sx = sx * dilations[0];
+                            const std::size_t dilated_sy = sy * dilations[1];
+
+                            // Ensure within bounds
+                            if ((ix + dilated_sx) < dims[2] && (iy + dilated_sy) < dims[3]) {
+                                sum += static_cast<Acc_T<I>>(input[iIndex + (ix + dilated_sx) * dims[3] + (iy + dilated_sy)]);
+                                ++count;
+                            }
+                        }
+                    }
+
+                    output[oIndexFull] = count > 0 ? castFromFloat<O>(sum / count) : 0;
+                }
+            }
+        }
+    }
+}
+
+// Kernels registration to implementation entry point
+REGISTRAR(AvgPoolingImpl2D_cpu,
+    {{DataType::Float32, DataFormat::NCHW}, {DataType::Float32, DataFormat::NCHW}},
+    {ProdConso::inPlaceModel, Aidge::AvgPoolingImpl2D_cpu_forward_kernel<float, float>, nullptr});
+REGISTRAR(AvgPoolingImpl2D_cpu,
+    {{DataType::Int32, DataFormat::NCHW}, {DataType::Int32, DataFormat::NCHW}},
+    {ProdConso::inPlaceModel, Aidge::AvgPoolingImpl2D_cpu_forward_kernel<std::int32_t, std::int32_t>, nullptr});
+REGISTRAR(AvgPoolingImpl2D_cpu,
+    {{DataType::Float64, DataFormat::NCHW}, {DataType::Float64, DataFormat::NCHW}},
+    {ProdConso::inPlaceModel, Aidge::AvgPoolingImpl2D_cpu_forward_kernel<double, double>, nullptr});
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_AVGPOOLINGIMPL_KERNELS_H_ */

include/aidge/backend/cpu/operator/BatchNormImpl.hpp

@@ -17,69 +17,29 @@
 #include <tuple>
 #include <vector>
 
-#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
 #include "aidge/operator/BatchNorm.hpp"
 #include "aidge/utils/Registrar.hpp"
 #include "aidge/utils/Types.h"
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
 
 namespace Aidge {
-// class BatchNorm_Op;
-
-// compute kernel registry for forward and backward
-class BatchNormImpl2DForward_cpu
-    : public Registrable<BatchNormImpl2DForward_cpu,
-                         std::tuple<DataType, DataType, DataType>,
-                         void(const BatchNorm_Op<2>::Attrs &,
-                              const std::array<DimSize_t, 4> &,
-                              const void *,
-                              const void *,
-                              const void *,
-                              void *,
-                              void *,
-                              void *,
-                              const bool)> {};
-class BatchNormImpl2DBackward_cpu
-    : public Registrable<BatchNormImpl2DBackward_cpu,
-                         std::tuple<DataType, DataType, DataType>,
-                         void(const BatchNorm_Op<2>::Attrs &,
-                              const std::array<DimSize_t, 4> &,
-                              const void *,
-                              const void *,
-                              const void *,
-                              void *,
-                              void *,
-                              void *)> {};
-
-class BatchNormImpl2D_cpu : public OperatorImpl {
-   private:
-    const BatchNorm_Op<2> &mOp;
-    std::array<NbElts_t, 5> mNbConsumedData;
-    std::array<NbElts_t, 1> mNbProducedData;
-
-   public:
-    BatchNormImpl2D_cpu(const BatchNorm_Op<2> &op) : mOp(op), mNbConsumedData({0, 0, 0, 0, 0}), mNbProducedData({0}) {}
-
-    static std::unique_ptr<BatchNormImpl2D_cpu> create(const BatchNorm_Op<2> &op) {
-        return std::make_unique<BatchNormImpl2D_cpu>(op);
-    }
-
-   public:
-    NbElts_t getNbRequiredData(const IOIndex_t inputIdx) const override final;
-    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
-    NbElts_t getRequiredMemory(const IOIndex_t /*outputIdx*/, const std::vector<DimSize_t> &inputsSize) const override final;
-    NbElts_t getNbConsumedData(const IOIndex_t inputIdx) const override final;
-    NbElts_t getNbProducedData(const IOIndex_t outputIdx) const override final;
-    void updateConsummerProducer() override final;
-
-    void forward();
-
-    void backward();
-};
-
-namespace {
-// add cpu backend to BatchNorm_Op<2> implementation registry
-static Registrar<BatchNorm_Op<2>> registrarBatchNormImpl2D_cpu("cpu", Aidge::BatchNormImpl2D_cpu::create);
-}  // namespace
+// Operator implementation entry point for the backend
+using BatchNorm2D_Op = BatchNorm_Op<2>;
+using BatchNormImpl2D_cpu = OperatorImpl_cpu<BatchNorm_Op<2>,
+    void(float,
+        float,
+        const std::vector<DimSize_t> &,
+        const void *,
+        const void *,
+        const void *,
+        void *,
+        void *,
+        void *,
+        const bool)>;
+
+// Implementation entry point registration to Operator
+REGISTRAR(BatchNorm2D_Op, "cpu", Aidge::BatchNormImpl2D_cpu::create);
 }  // namespace Aidge
 
 #endif /* AIDGE_CPU_OPERATOR_BATCHNORMIMPL_H_ */

include/aidge/backend/cpu/operator/BatchNormImpl_forward_kernels.hpp → include/aidge/backend/cpu/operator/BatchNormImpl_kernels.hpp

@@ -9,13 +9,14 @@
  *
  ********************************************************************************/
 
-#ifndef AIDGE_CPU_OPERATOR_BATCHNORMIMPL_FORWARD_KERNEL_H_
-#define AIDGE_CPU_OPERATOR_BATCHNORMIMPL_FORWARD_KERNEL_H_
+#ifndef AIDGE_CPU_OPERATOR_BATCHNORMIMPL_KERNELS_H_
+#define AIDGE_CPU_OPERATOR_BATCHNORMIMPL_KERNELS_H_
 
 #include "aidge/utils/Registrar.hpp"
 
 #include "aidge/backend/cpu/operator/BatchNormImpl.hpp"
 #include "aidge/utils/Types.h"
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
 #include <array>
 #include <cmath>
 #include <algorithm>
@@ -37,7 +38,7 @@ namespace Aidge {
  * @param output_ Output Tensor.
  */
 template <class I, class P, class O>
-void BatchNormImpl2D_cpu_forward_kernel(const BatchNorm_Op<2>::Attrs &params, const std::array<DimSize_t, 4> &dims,
+void BatchNormImpl2D_cpu_forward_kernel(float epsilon, float momentum, const std::vector<DimSize_t> &dims,
                                        const void *input_, const void *scale_, const void *shift_, void *batchMean_, void *batchVar_, void *output_, const bool freeze) {
     // FIXME: missing convolution attributes as arguments
     const I *input = static_cast<const I *>(input_);
@@ -48,16 +49,18 @@ void BatchNormImpl2D_cpu_forward_kernel(const BatchNorm_Op<2>::Attrs &params, co
     O *output = static_cast<O *>(output_);
 
     const DimSize_t nbBatch = dims[0];
-    const DimSize_t nbChannels = dims[1];
-    const DimSize_t featureMapSize = dims[2]*dims[3];
-
-
-    if ((freeze == true) || (std::get<1>(params) == 0.0f)) {
-        for (std::size_t batch = 0; batch < nbBatch; ++batch) {
-            for (std::size_t ch = 0; ch < nbChannels; ++ch) {
+    const DimSize_t nbChannels = (dims.size() > 1) ? dims[1] : 1;
+    const DimSize_t featureMapSize = (dims.size() > 2) ? std::accumulate(dims.begin() + 2, dims.end(), 1, std::multiplies<DimSize_t>()) : 1;
+
+    if ((freeze == true) || (momentum == 0.0f)) {
+#ifdef _OPENMP
+        #pragma omp parallel for collapse(2) if (nbBatch * nbChannels >= 16)
+#endif
+        for (int batch = 0; batch < static_cast<int>(nbBatch); ++batch) {
+            for (int ch = 0; ch < static_cast<int>(nbChannels); ++ch) {
                 const std::size_t ioIndex = (ch + batch*nbChannels) * featureMapSize;
                 std::fill(output + ioIndex, output + ioIndex + featureMapSize, shift[ch]);
-                const P var = std::sqrt(batchVar[ch] + static_cast<P>(std::get<0>(params)));
+                const P var = std::sqrt(batchVar[ch] + static_cast<P>(epsilon));
 
                 for (std::size_t feature = 0; feature<featureMapSize; ++feature) {
                     output[ioIndex + feature] += scale[ch] * (input[ioIndex + feature]-batchMean[ch]) / var;
@@ -81,10 +84,10 @@ void BatchNormImpl2D_cpu_forward_kernel(const BatchNorm_Op<2>::Attrs &params, co
             const I inputMean = sum / static_cast<I>(nbDataPerChannel);
             const I inputVar = sumSquare / static_cast<I>(nbDataPerChannel)  - inputMean*inputMean;
 
-            batchMean[ch] = batchMean[ch]*(1-std::get<1>(params)) + inputMean*std::get<1>(params);
-            batchVar[ch] = batchVar[ch]*(1-std::get<1>(params)) + inputVar*(static_cast<I>(nbDataPerChannel)/static_cast<I>(nbDataPerChannel-1))*std::get<1>(params);
+            batchMean[ch] = batchMean[ch]*(1-momentum) + inputMean*momentum;
+            batchVar[ch] = batchVar[ch]*(1-momentum) + inputVar*(static_cast<I>(nbDataPerChannel)/static_cast<I>(nbDataPerChannel-1))*momentum;
 
-            const P var = std::sqrt(inputVar + static_cast<P>(std::get<0>(params)));
+            const P var = std::sqrt(inputVar + static_cast<P>(epsilon));
             for (std::size_t batch = 0; batch < nbBatch; ++batch) {
                 const std::size_t ioIndex = (ch + batch*nbChannels) * featureMapSize;
                 for (std::size_t feature = 0; feature<featureMapSize; ++feature) {
@@ -95,15 +98,10 @@ void BatchNormImpl2D_cpu_forward_kernel(const BatchNorm_Op<2>::Attrs &params, co
     }
 }
 
-
-
-
-
-namespace {
-static Registrar<BatchNormImpl2DForward_cpu> registrarBatchNormImpl2DForward_cpu_Float32(
-        {DataType::Float32, DataType::Float32, DataType::Float32},
-        Aidge::BatchNormImpl2D_cpu_forward_kernel<float, float, float>);
-}  // namespace
+// Kernels registration to implementation entry point
+REGISTRAR(BatchNormImpl2D_cpu,
+    {{DataType::Float32, DataFormat::NCHW}, {DataType::Float32, DataFormat::NCHW}},
+    {ProdConso::inPlaceModel, Aidge::BatchNormImpl2D_cpu_forward_kernel<float, float, float>, nullptr});
 }  // namespace Aidge
 
-#endif /* AIDGE_CPU_OPERATOR_BATCHNORMIMPL_FORWARD_KERNEL_H_ */
+#endif /* AIDGE_CPU_OPERATOR_BATCHNORMIMPL_KERNELS_H_ */