Merged with dev

.gitlab/ci/build.gitlab-ci.yml

@@ -136,6 +136,8 @@ build:ubuntu_python:
     - source venv/bin/activate
     - python3 -m pip install -r requirements.txt
     - python3 -m pip install .
+    - python3 -m pip install numpy unittest-xml-reporting
+    - python3 -m pip list
   artifacts:
     expire_in: 1 week
     paths:
@@ -160,9 +162,10 @@ build:ubuntu_python:
 #   script:
 #     # Download dependencies
 #     # aidge_core
-#     - 'curl "https://gitlab.eclipse.org/api/v4/projects/5139/jobs/artifacts/main/download?job=build:windows_cpp" -o build_artifacts.zip'
-#     - Expand-Archive -Path .\build_artifacts.zip -DestinationPath . -Force
-#     - Remove-Item .\build_cpp\ -Recurse
+#     - $DEPENDENCY_NAME="aidge_core"
+#     - $DEPENDENCY_JOB="build:windows_cpp"
+#     - !reference [.download_dependency_windows, script]
+#     - Remove-Item .\build_cpp\ -Recurse -Force -ErrorAction Ignore
 
 #     - $env:CMAKE_PREFIX_PATH = '../install_cpp'
 #     - mkdir -p build_cpp
@@ -196,8 +199,9 @@ build:ubuntu_python:
 #   script:
 #     # Download dependencies
 #     # aidge_core (Python)
-#     - 'curl "https://gitlab.eclipse.org/api/v4/projects/5139/jobs/artifacts/main/download?job=build:windows_python" -o build_artifacts.zip'
-#     - Expand-Archive -Path .\build_artifacts.zip -DestinationPath . -Force
+#     - $DEPENDENCY_NAME="aidge_core"
+#     - $DEPENDENCY_JOB="build:windows_python"
+#     - !reference [.download_dependency_windows, script]
 
 #     - python -m pip install virtualenv
 #     - virtualenv venv

.gitlab/ci/test.gitlab-ci.yml

@@ -18,9 +18,8 @@ test:ubuntu_python:
   script:
     - source venv/bin/activate
     - cd ${CI_PROJECT_NAME}
-    - python3 -m pip install numpy unittest-xml-reporting
-    - python3 -m pip list
-    # Run on discovery all tests located in core/unit_tests/python and discard the stdout 
+
+    # Run on discovery all tests located in core/unit_tests/python and discard the stdout
     # only to show the errors/warnings and the results of the tests
     - python3 -m xmlrunner discover -s unit_tests/ -v -b --output-file xmlrunner-results.xml
   artifacts:

aidge_backend_cpu/unit_tests/test_tensor.py

-import unittest
-import aidge_core
-import aidge_backend_cpu
-import numpy as np
-
-
-class test_tensor(unittest.TestCase):
-    """Test tensor binding
-    """
-    def setUp(self):
-        pass
-    def tearDown(self):
-        pass
-
-    def test_getavailable_backends(self):
-        self.assertTrue("cpu" in aidge_core.Tensor.get_available_backends())
-
-    def test_numpy_int_to_tensor(self):
-        np_array = np.arange(9).reshape(1,1,3,3).astype(np.int32)
-        # Numpy -> Tensor
-        t = aidge_core.Tensor(np_array)
-        self.assertEqual(t.dtype(), aidge_core.DataType.Int32)
-        for i_t, i_n in zip(t, np_array.flatten()):
-            self.assertTrue(i_t == i_n)
-        for i,j in zip(t.dims(), np_array.shape):
-            self.assertEqual(i,j)
-    def test_tensor_int_to_numpy(self):
-        np_array = np.arange(9).reshape(1,1,3,3)
-        # Numpy -> Tensor
-        t = aidge_core.Tensor(np_array)
-        # Tensor -> Numpy
-        nnarray = np.array(t)
-        for i_nn, i_n in zip(nnarray.flatten(), np_array.flatten()):
-            self.assertTrue(i_nn == i_n)
-        for i,j in zip(t.dims(), nnarray.shape):
-            self.assertEqual(i,j)
-
-    def test_numpy_int64_to_tensor(self):
-        np_array = np.arange(9).reshape(1,1,3,3).astype(np.int64)
-        # Numpy -> Tensor
-        t = aidge_core.Tensor(np_array)
-        self.assertEqual(t.dtype(), aidge_core.DataType.Int64)
-        for i_t, i_n in zip(t, np_array.flatten()):
-            self.assertTrue(i_t == i_n)
-        for i,j in zip(t.dims(), np_array.shape):
-            self.assertEqual(i,j)
-
-    def test_numpy_float_to_tensor(self):
-        t = aidge_core.Tensor()
-        np_array = np.random.rand(1, 1, 3, 3).astype(np.float32)
-        # Numpy -> Tensor
-        t = aidge_core.Tensor(np_array)
-        self.assertEqual(t.dtype(), aidge_core.DataType.Float32)
-        for i_t, i_n in zip(t, np_array.flatten()):
-            self.assertTrue(i_t == i_n) # TODO : May need to change this to a difference
-        for i,j in zip(t.dims(), np_array.shape):
-            self.assertEqual(i,j)
-
-    def test_get_set(self):
-        dims = [2,2,2]
-
-        np_array = np.arange(8).reshape(dims).astype(np.int32)
-        # Numpy -> Tensor
-        t = aidge_core.Tensor(np_array)
-        for i in range(8):
-            self.assertEqual(t[i], i)
-            t[i] = 5
-            self.assertEqual(t[i], 5)
-
-if __name__ == '__main__':
-    unittest.main()

include/aidge/backend/cpu.hpp

@@ -12,7 +12,6 @@
 #ifndef AIDGE_CPU_IMPORTS_H_
 #define AIDGE_CPU_IMPORTS_H_
 
-#include "aidge/backend/cpu/data/TensorImpl.hpp"
 #include "aidge/backend/cpu/operator/AddImpl.hpp"
 #include "aidge/backend/cpu/operator/AvgPoolingImpl.hpp"
 #include "aidge/backend/cpu/operator/MaxPoolingImpl.hpp"
@@ -21,7 +20,9 @@
 #include "aidge/backend/cpu/operator/ConvDepthWiseImpl.hpp"
 #include "aidge/backend/cpu/operator/ConvImpl.hpp"
 #include "aidge/backend/cpu/operator/DivImpl.hpp"
+#include "aidge/backend/cpu/operator/ErfImpl.hpp"
 #include "aidge/backend/cpu/operator/FCImpl.hpp"
+#include "aidge/backend/cpu/operator/GatherImpl.hpp"
 #include "aidge/backend/cpu/operator/LeakyReLUImpl.hpp"
 #include "aidge/backend/cpu/operator/MatMulImpl.hpp"
 #include "aidge/backend/cpu/operator/MemorizeImpl.hpp"
@@ -29,7 +30,9 @@
 #include "aidge/backend/cpu/operator/PadImpl.hpp"
 #include "aidge/backend/cpu/operator/PopImpl.hpp"
 #include "aidge/backend/cpu/operator/PowImpl.hpp"
+#include "aidge/backend/cpu/operator/ReduceMeanImpl.hpp"
 #include "aidge/backend/cpu/operator/ReLUImpl.hpp"
+#include "aidge/backend/cpu/operator/ReshapeImpl.hpp"
 #include "aidge/backend/cpu/operator/ScalingImpl.hpp"
 #include "aidge/backend/cpu/operator/SigmoidImpl.hpp"
 #include "aidge/backend/cpu/operator/SliceImpl.hpp"
@@ -37,5 +40,9 @@
 #include "aidge/backend/cpu/operator/SoftmaxImpl.hpp"
 #include "aidge/backend/cpu/operator/SubImpl.hpp"
 #include "aidge/backend/cpu/operator/TanhImpl.hpp"
+#include "aidge/backend/cpu/operator/TransposeImpl.hpp"
+
+#include "aidge/backend/cpu/data/TensorImpl.hpp"
+
+#endif /* AIDGE_CPU_IMPORTS_H_ */
 
-#endif /* AIDGE_CPU_IMPORTS_H_ */
\ No newline at end of file

include/aidge/backend/cpu/data/GetCPUPtr.h

-/********************************************************************************
- * 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_DATA_GETCPUPTR_H_
-#define AIDGE_CPU_DATA_GETCPUPTR_H_
-
-#include "aidge/data/Tensor.hpp"
-
-namespace Aidge {
-inline void *getCPUPtr(std::shared_ptr<Aidge::Data> const &data) {
-  const auto tensor = std::static_pointer_cast<Tensor>(data);
-  return tensor->getImpl()->hostPtr(tensor->getImplOffset());
-}
-} // namespace Aidge
-
-#endif // AIDGE_CPU_DATA_GETCPUPTR_H_
\ No newline at end of file

include/aidge/backend/cpu/data/TensorImpl.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_DATA_TENSORIMPL_H_
-#define AIDGE_CPU_DATA_TENSORIMPL_H_
-
-#include "aidge/backend/TensorImpl.hpp"
-#include "aidge/data/Tensor.hpp"
-#include "aidge/data/half.hpp"
-#include "aidge/utils/Registrar.hpp"
-#include "aidge/utils/Types.h"
-#include "aidge/utils/ErrorHandling.hpp"
-#include "aidge/utils/future_std/span.hpp"
-
-namespace Aidge {
-
-template <class T>
-class TensorImpl_cpu : public TensorImpl {
-private:
-    /// Pointer to the data and its capacity
-    future_std::span<T> mData;
-    /// If this instance own the data, std::unique_ptr manages it
-    std::unique_ptr<T[]> mDataOwner;
-
-public:
-    static constexpr const char *Backend = "cpu";
-
-    TensorImpl_cpu(DeviceIdx_t device, NbElts_t length) : TensorImpl(Backend, device, length) {}
-
-    bool operator==(const TensorImpl &otherImpl) const override final {
-        const auto& typedOtherImpl = reinterpret_cast<const TensorImpl_cpu<T> &>(otherImpl);
-        AIDGE_INTERNAL_ASSERT(typedOtherImpl.size() >= mNbElts);
-
-        std::size_t i = 0;
-        for (; i < mNbElts &&
-               *static_cast<const T*>(rawPtr(i)) == *static_cast<const T*>(typedOtherImpl.rawPtr(i));
-               ++i) {
-        }
-        return i == mNbElts;
-    }
-
-    static std::shared_ptr<TensorImpl_cpu> create(DeviceIdx_t device, NbElts_t length) {
-        return std::make_shared<TensorImpl_cpu<T>>(device, length);
-    }
-
-    inline std::size_t scalarSize() const noexcept override final { return sizeof(T); }
-
-    void copy(const void *src, NbElts_t length, NbElts_t offset = 0) override final {
-        const T* srcT = static_cast<const T *>(src);
-        T* dstT = static_cast<T *>(rawPtr(offset));
-
-        AIDGE_ASSERT(length <= mData.size() || length <= mNbElts, "copy length is above capacity");
-        AIDGE_ASSERT(dstT < srcT || dstT >= srcT + length, "overlapping copy is not supported");
-        std::copy(srcT, srcT + length, dstT);
-    }
-
-    void copyCast(const void *src, const DataType srcDt, NbElts_t length, NbElts_t offset = 0) override final {
-        if (length == 0) {
-            return;
-        }
-
-        T* dstT = static_cast<T *>(rawPtr(offset));
-        AIDGE_ASSERT(length <= mData.size() || length <= mNbElts, "copy length is above capacity");
-        switch (srcDt)
-        {
-            case DataType::Float64:
-                std::copy(static_cast<const double*>(src), static_cast<const double*>(src) + length,
-                        dstT);
-                break;
-            case DataType::Float32:
-                std::copy(static_cast<const float*>(src), static_cast<const float*>(src) + length,
-                        dstT);
-                break;
-            case DataType::Float16:
-                std::copy(static_cast<const half_float::half*>(src), static_cast<const half_float::half*>(src) + length,
-                        dstT);
-                break;
-            case DataType::Int64:
-                std::copy(static_cast<const int64_t*>(src), static_cast<const int64_t*>(src) + length,
-                        dstT);
-                break;
-            case DataType::UInt64:
-                std::copy(static_cast<const uint64_t*>(src), static_cast<const uint64_t*>(src) + length,
-                        dstT);
-                break;
-            case DataType::Int32:
-                std::copy(static_cast<const int32_t*>(src), static_cast<const int32_t*>(src) + length,
-                        dstT);
-                break;
-            case DataType::UInt32:
-                std::copy(static_cast<const uint32_t*>(src), static_cast<const uint32_t*>(src) + length,
-                        dstT);
-                break;
-            case DataType::Int16:
-                std::copy(static_cast<const int16_t*>(src), static_cast<const int16_t*>(src) + length,
-                        dstT);
-                break;
-            case DataType::UInt16:
-                std::copy(static_cast<const uint16_t*>(src), static_cast<const uint16_t*>(src) + length,
-                        dstT);
-                break;
-            case DataType::Int8:
-                std::copy(static_cast<const int8_t*>(src), static_cast<const int8_t*>(src) + length,
-                        dstT);
-                break;
-            case DataType::UInt8:
-                std::copy(static_cast<const uint8_t*>(src), static_cast<const uint8_t*>(src) + length,
-                        dstT);
-                break;
-            default:
-                AIDGE_THROW_OR_ABORT(std::runtime_error, "Unsupported data type.");
-                break;
-        }
-    }
-
-    void copyFromDevice(const void *src, const std::pair<std::string, DeviceIdx_t>& device, NbElts_t length, NbElts_t offset = 0) override final {
-        AIDGE_ASSERT(device.first == Backend, "backend must match");
-        AIDGE_ASSERT(device.second == 0, "device cannot be != 0 for CPU backend");
-        copy(src, length, offset);
-    }
-
-    inline void copyFromHost(const void *src, NbElts_t length, NbElts_t offset = 0) override final {
-        copy(src, length, offset);
-    }
-
-    void copyToHost(void *dst, NbElts_t length, NbElts_t offset = 0) const override final {
-        const T* src = static_cast<const T*>(rawPtr(offset));
-        AIDGE_ASSERT(length <= mData.size() || length <= mNbElts, "copy length is above capacity");
-        std::copy(src, src + length, static_cast<T *>(dst));
-    }
-
-    void *rawPtr(NbElts_t offset = 0) override final {
-        lazyInit();
-        return (mData.data() + offset);
-    };
-
-    const void *rawPtr(NbElts_t offset = 0) const override final {
-        AIDGE_ASSERT(mData.size() >= mNbElts, "accessing uninitialized const rawPtr");
-        return (mData.data() + offset);
-    };
-
-    void *hostPtr(NbElts_t offset = 0) override final {
-        lazyInit();
-        return (mData.data() + offset);
-    };
-
-    const void *hostPtr(NbElts_t offset = 0) const override final {
-        AIDGE_ASSERT(mData.size() >= mNbElts, "accessing uninitialized const hostPtr");
-        return (mData.data() + offset);
-    };
-
-    void setRawPtr(void *ptr, NbElts_t length) override final {
-        AIDGE_ASSERT(length >= mNbElts, "trying to set raw pointer of insufficient capacity");
-        mData = future_std::span<T>(static_cast<T *>(ptr), length);
-        mDataOwner.reset();
-    };
-
-    virtual ~TensorImpl_cpu() = default;
-
-private:
-    void lazyInit() {
-        if (mData.size() < mNbElts) {
-            // Need more data, a re-allocation will occur
-            AIDGE_ASSERT(mData.empty() || mDataOwner != nullptr, "trying to enlarge non-owned data");
-            mDataOwner.reset(new T[mNbElts]);
-            mData = future_std::span<T>(mDataOwner.get(), mNbElts);
-        }
-    }
-};
-
-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_Float16(
-        {"cpu", DataType::Float16}, Aidge::TensorImpl_cpu<half_float::half>::create);
-static Registrar<Tensor> registrarTensorImpl_cpu_Int32(
-        {"cpu", DataType::Int32}, Aidge::TensorImpl_cpu<int>::create);
-static Registrar<Tensor> registrarTensorImpl_cpu_Int64(
-        {"cpu", DataType::Int64}, Aidge::TensorImpl_cpu<long>::create);
-}  // namespace
-}  // namespace Aidge
-
-#endif /* AIDGE_CPU_DATA_TENSORIMPL_H_ */

include/aidge/backend/cpu/operator/ErfImpl.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_ERFIMPL_H_
+#define AIDGE_CPU_OPERATOR_ERFIMPL_H_
+
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/operator/Erf.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+#include <memory>
+#include <vector>
+
+namespace Aidge {
+// class Erf_Op;
+
+// compute kernel registry for forward and backward
+class ErfImplForward_cpu
+    : public Registrable<ErfImplForward_cpu, std::tuple<DataType, DataType>, void(const std::size_t, const void*, void*)> {
+};
+class ErfImplBackward_cpu
+    : public Registrable<ErfImplBackward_cpu, std::tuple<DataType, DataType>, void(const std::size_t, const void*, void*)> {
+};
+
+class ErfImpl_cpu : public OperatorImpl {
+public:
+    ErfImpl_cpu(const Erf_Op& op) : OperatorImpl(op) {}
+
+    static std::unique_ptr<ErfImpl_cpu> create(const Erf_Op& op) {
+        return std::make_unique<ErfImpl_cpu>(op);
+    }
+
+    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    void forward() override;
+};
+
+namespace {
+static Registrar<Erf_Op> registrarErfImpl_cpu("cpu", Aidge::ErfImpl_cpu::create);
+}
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ERFIMPL_H_ */

include/aidge/backend/cpu/operator/ErfImpl_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_ERFIMPL_FORWARD_KERNEL_H_
+#define AIDGE_CPU_OPERATOR_ERFIMPL_FORWARD_KERNEL_H_
+
+#include <cmath>
+
+#include "aidge/utils/Registrar.hpp"
+
+#include "aidge/backend/cpu/operator/ErfImpl.hpp"
+
+namespace Aidge {
+template <class I, class O>
+void ErfImpl_cpu_forward_kernel(std::size_t inputLenght,
+                                     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] = std::erf(input[i]);
+    }
+}
+
+namespace {
+static Registrar<ErfImplForward_cpu> registrarErfImplForward_cpu_Float32(
+        {DataType::Float32, DataType::Float32}, Aidge::ErfImpl_cpu_forward_kernel<float, float>);
+static Registrar<ErfImplForward_cpu> registrarErfImplForward_cpu_Int32(
+        {DataType::Int32, DataType::Int32}, Aidge::ErfImpl_cpu_forward_kernel<int, int>);
+static Registrar<ErfImplForward_cpu> registrarErfImplForward_cpu_Float64(
+        {DataType::Float64, DataType::Float64}, Aidge::ErfImpl_cpu_forward_kernel<double, double>);
+}  // namespace
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_ERFIMPL_FORWARD_KERNEL_H_ */

include/aidge/backend/cpu/operator/GatherImpl.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_GATHERIMPL_H_
+#define AIDGE_CPU_OPERATOR_GATHERIMPL_H_
+
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/operator/Gather.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+#include <memory>
+#include <vector>
+
+namespace Aidge {
+// class Gather_Op;
+
+// compute kernel registry for forward and backward
+class GatherImplForward_cpu
+    : public Registrable<GatherImplForward_cpu, std::tuple<DataType, DataType>, void(const typename Gather_Op::Attrs&, const std::vector<DimSize_t>&, const void*, void*)> {
+};
+class GatherImplBackward_cpu
+    : public Registrable<GatherImplBackward_cpu, std::tuple<DataType, DataType>, void(const typename Gather_Op::Attrs&, const std::vector<DimSize_t>&, const void*, void*)> {
+};
+
+class GatherImpl_cpu : public OperatorImpl {
+public:
+    GatherImpl_cpu(const Gather_Op& op) : OperatorImpl(op) {}
+
+    static std::unique_ptr<GatherImpl_cpu> create(const Gather_Op& op) {
+        return std::make_unique<GatherImpl_cpu>(op);
+    }
+
+    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    void forward() override;
+};
+
+namespace {
+static Registrar<Gather_Op> registrarGatherImpl_cpu("cpu", Aidge::GatherImpl_cpu::create);
+}
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_GATHERIMPL_H_ */

include/aidge/backend/cpu/operator/GatherImpl_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_GATHERIMPL_FORWARD_KERNEL_H_
+#define AIDGE_CPU_OPERATOR_GATHERIMPL_FORWARD_KERNEL_H_
+
+#include "aidge/utils/Registrar.hpp"
+#include <cstddef>
+#include <cmath>
+#include "aidge/data/Data.hpp"
+#include "aidge/utils/Types.h"
+
+#include "aidge/backend/cpu/operator/GatherImpl.hpp"
+
+namespace Aidge {
+template <class I, class O>
+void GatherImpl_cpu_forward_kernel(const typename Gather_Op::Attrs& attrs, 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 axisIdx = std::get<2>(attrs)>=0 ?
+                                std::get<2>(attrs) :
+                                static_cast<std::size_t>(std::get<2>(attrs)) + inputDims.size();
+
+    std::size_t postAxisElems = 1;
+    for (std::size_t i = axisIdx + 1; i < inputDims.size(); ++i) {
+        postAxisElems *= inputDims[i];
+    }
+    std::size_t preAxisElems = 1;
+    for (std::size_t i = 0; i < axisIdx; ++i) {
+        preAxisElems *= inputDims[i];
+    }
+
+    const std::vector<std::int64_t> indices = std::get<0>(attrs);
+    for (std::size_t i=0; i<preAxisElems; ++i)
+    {
+        for(std::size_t j=0; j<indices.size(); ++j)
+        {
+            const std::size_t idx = indices[j] >= 0 ? indices[j] : static_cast<std::size_t>(indices[j]) + inputDims[axisIdx];
+            const I* startPtr = std::next(input, i * postAxisElems * inputDims[axisIdx] + idx * postAxisElems);
+            std::copy_n(startPtr, postAxisElems, output);
+            output += postAxisElems;
+        }
+    }
+}
+
+namespace {
+static Registrar<GatherImplForward_cpu> registrarGatherImplForward_cpu_Float32(
+        {DataType::Float32, DataType::Float32}, Aidge::GatherImpl_cpu_forward_kernel<float, float>);
+static Registrar<GatherImplForward_cpu> registrarGatherImplForward_cpu_Int32(
+        {DataType::Int32, DataType::Int32}, Aidge::GatherImpl_cpu_forward_kernel<int, int>);
+static Registrar<GatherImplForward_cpu> registrarGatherImplForward_cpu_Float64(
+        {DataType::Float64, DataType::Float64}, Aidge::GatherImpl_cpu_forward_kernel<double, double>);
+}  // namespace
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_GATHERIMPL_FORWARD_KERNEL_H_ */

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

@@ -25,7 +25,7 @@ void LeakyReLUImpl_cpu_forward_kernel(const LeakyReLU_Op::Attrs& attrs,
 
     const I* input = static_cast<const I*>(input_);
     O* output = static_cast<O*>(output_);
-    I negativeSlope = static_cast<I>(std::get<0>(attrs));
+    const I negativeSlope = static_cast<const I>(std::get<0>(attrs));
 
     for (std::size_t i = 0; i < inputLenght; ++i) {
         output[i] = input[i] >= 0 ? input[i] : input[i] * negativeSlope;

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

@@ -23,16 +23,14 @@
 #include "aidge/backend/cpu/data/GetCPUPtr.h"
 
 namespace Aidge {
-// class MatMul_Op;
 
-// compute kernel registry for forward and backward
 class MatMulImplForward_cpu
-    : public Registrable<MatMulImplForward_cpu, std::tuple<DataType, DataType, DataType>,
-                         void(const MatMul_Op::Attrs &, const DimSize_t, const DimSize_t,
+    : public Registrable<MatMulImplForward_cpu, std::tuple<DataType, DataType>,
+                         void(const std::size_t, const std::size_t, const std::size_t,
                               const void *, const void *, void *)> {};
 class MatMulImplBackward_cpu
-    : public Registrable<MatMulImplBackward_cpu, std::tuple<DataType, DataType, DataType>,
-                         void(const MatMul_Op::Attrs &, const DimSize_t, const DimSize_t,
+    : public Registrable<MatMulImplBackward_cpu, std::tuple<DataType, DataType>,
+                         void(const std::vector<DimSize_t>&, const std::vector<DimSize_t>&,
                               const void *, const void *, void *)> {};
 
 class MatMulImpl_cpu : public OperatorImpl {

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

@@ -12,45 +12,39 @@
 #ifndef AIDGE_CPU_OPERATOR_MATMULIMPL_FORWARD_KERNEL_H_
 #define AIDGE_CPU_OPERATOR_MATMULIMPL_FORWARD_KERNEL_H_
 
-#include "aidge/utils/Registrar.hpp"
-#include <algorithm>
-
 #include "aidge/backend/cpu/operator/MatMulImpl.hpp"
 
 namespace Aidge {
 
-template <class I, class W, class O>
-void MatMulImpl_cpu_forward_kernel(const MatMul_Op::Attrs& attrs, const DimSize_t batchSize, const DimSize_t oneInputSize,
-                                   const void* input_, const void* weights_, void* output_) {
+template <class I, class O>
+void MatMulImpl_cpu_forward_kernel(const std::size_t n, const std::size_t k, const std::size_t m,
+                                    const void* input1_, const void* input2_, void* output_) {
     // FIXME: missing MatMul parameters as arguments
-    const I* input = static_cast<const I*>(input_);
-    const W* weights = static_cast<const W*>(weights_);
+    const I* input1 = static_cast<const I*>(input1_);
+    const I* input2 = static_cast<const I*>(input2_);
     O* output = static_cast<O*>(output_);
 
-
-    std::fill(output, output+(batchSize*std::get<0>(attrs)), O(0));
-
-    for (std::size_t batch = 0; batch < batchSize; ++batch) {
-        for (std::size_t out = 0; out < std::get<0>(attrs); ++out) {
-            output[out + batch*std::get<0>(attrs)] = std::inner_product(input + batch*oneInputSize,
-                                                        input + (batch + 1)*oneInputSize,
-                                                        weights + out*oneInputSize,
-                                                        output[out + batch*std::get<0>(attrs)]);
+    for (std::size_t i = 0; i < n; ++i) {
+        for (std::size_t j = 0; j < m; ++j) {
+            O sum = O(0);
+            for (std::size_t l = 0; l < k; ++l) {
+                sum += static_cast<O>(input1[i*k + l] * input2[l*m + j]);
+            }
+            output[i*m + j] = sum;
         }
     }
 }
 
-
 namespace {
 static Registrar<MatMulImplForward_cpu> registrarMatMulImpl2DForward_cpu_Float32(
-        {DataType::Float32, DataType::Float32, DataType::Float32},
-        Aidge::MatMulImpl_cpu_forward_kernel<float, float, float>);
+        {DataType::Float32, DataType::Float32},
+        Aidge::MatMulImpl_cpu_forward_kernel<float, float>);
 static Registrar<MatMulImplForward_cpu> registrarMatMulImpl2DForward_cpu_Int32(
-        {DataType::Int32, DataType::Int32, DataType::Int32},
-        Aidge::MatMulImpl_cpu_forward_kernel<int, int, int>);
+        {DataType::Int32, DataType::Int32},
+        Aidge::MatMulImpl_cpu_forward_kernel<int, int>);
 static Registrar<MatMulImplForward_cpu> registrarMatMulImpl2DForward_cpu_Float64(
-        {DataType::Float64, DataType::Float64, DataType::Float64},
-        Aidge::MatMulImpl_cpu_forward_kernel<double, double, double>);
+        {DataType::Float64, DataType::Float64},
+        Aidge::MatMulImpl_cpu_forward_kernel<double, double>);
 }  // namespace
 
 }  // namespace Aidge

include/aidge/backend/cpu/operator/ReduceMeanImpl.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_REDUCEMEANIMPL_H_
+#define AIDGE_CPU_OPERATOR_REDUCEMEANIMPL_H_
+
+#include <array>
+#include <memory>
+#include <tuple>
+#include <vector>
+
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/operator/ReduceMean.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+
+namespace Aidge {
+// class ReduceMean_Op;
+
+// compute kernel registry for forward and backward
+// DIM 1
+class ReduceMeanImpl1DForward_cpu
+    : public Registrable<ReduceMeanImpl1DForward_cpu,
+                         std::tuple<DataType, DataType>,
+                         void(const ReduceMean_Op<1>::Attrs &, const std::vector<DimSize_t>&, const void *, void *)> {};
+class ReduceMeanImpl1DBackward_cpu
+    : public Registrable<ReduceMeanImpl1DBackward_cpu,
+                         std::tuple<DataType, DataType>,
+                         void(const ReduceMean_Op<1>::Attrs &, const std::vector<DimSize_t>&, const void *,  void *)> {};
+
+// DIM 2
+class ReduceMeanImpl2DForward_cpu
+    : public Registrable<ReduceMeanImpl2DForward_cpu,
+                         std::tuple<DataType, DataType>,
+                         void(const ReduceMean_Op<2>::Attrs &, const std::vector<DimSize_t>&, const void *, void *)> {};
+class ReduceMeanImpl2DBackward_cpu
+    : public Registrable<ReduceMeanImpl2DBackward_cpu,
+                         std::tuple<DataType, DataType>,
+                         void(const ReduceMean_Op<2>::Attrs &, const std::vector<DimSize_t>&, const void *,  void *)> {};
+// DIM 3
+class ReduceMeanImpl3DForward_cpu
+    : public Registrable<ReduceMeanImpl3DForward_cpu,
+                         std::tuple<DataType, DataType>,
+                         void(const ReduceMean_Op<3>::Attrs &, const std::vector<DimSize_t>&, const void *, void *)> {};
+class ReduceMeanImpl3DBackward_cpu
+    : public Registrable<ReduceMeanImpl3DBackward_cpu,
+                         std::tuple<DataType, DataType>,
+                         void(const ReduceMean_Op<3>::Attrs &, const std::vector<DimSize_t>&, const void *, void *)> {};
+
+class ReduceMeanImpl1D_cpu : public OperatorImpl {
+   public:
+    ReduceMeanImpl1D_cpu(const ReduceMean_Op<1>& op) : OperatorImpl(op) {}
+
+    static std::unique_ptr<ReduceMeanImpl1D_cpu> create(const ReduceMean_Op<1> &op) {
+        return std::make_unique<ReduceMeanImpl1D_cpu>(op);
+    }
+
+   public:
+    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    void forward() override;
+};
+
+class ReduceMeanImpl2D_cpu : public OperatorImpl {
+   public:
+    ReduceMeanImpl2D_cpu(const ReduceMean_Op<2>& op) : OperatorImpl(op) {}
+
+    static std::unique_ptr<ReduceMeanImpl2D_cpu> create(const ReduceMean_Op<2> &op) {
+        return std::make_unique<ReduceMeanImpl2D_cpu>(op);
+    }
+
+   public:
+    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    void forward() override;
+};
+
+class ReduceMeanImpl3D_cpu : public OperatorImpl {
+   public:
+    ReduceMeanImpl3D_cpu(const ReduceMean_Op<3>& op) : OperatorImpl(op) {}
+
+    static std::unique_ptr<ReduceMeanImpl3D_cpu> create(const ReduceMean_Op<3> &op) {
+        return std::make_unique<ReduceMeanImpl3D_cpu>(op);
+    }
+
+   public:
+    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    void forward() override;
+};
+namespace {
+// add cpu backend to ReduceMean_Op<2> implementation registry
+static Registrar<ReduceMean_Op<1>> registrarReduceMeanImpl1D_cpu("cpu", Aidge::ReduceMeanImpl1D_cpu::create);
+static Registrar<ReduceMean_Op<2>> registrarReduceMeanImpl2D_cpu("cpu", Aidge::ReduceMeanImpl2D_cpu::create);
+static Registrar<ReduceMean_Op<3>> registrarReduceMeanImpl3D_cpu("cpu", Aidge::ReduceMeanImpl3D_cpu::create);
+}  // namespace
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_REDUCEMEANIMPL_H_ */

include/aidge/backend/cpu/operator/ReduceMeanImpl_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_REDUCEMEANIMPL_FORWARD_KERNEL_H_
+#define AIDGE_CPU_OPERATOR_REDUCEMEANIMPL_FORWARD_KERNEL_H_
+
+#include <cstddef>
+#include <algorithm>   // std::copy, std::for_each
+#include <numeric>     //std::accumulate
+#include <functional>  //std::multiplies
+
+#include "aidge/backend/cpu/operator/ReduceMeanImpl.hpp"
+#include "aidge/data/Data.hpp"
+#include "aidge/operator/ReduceMean.hpp"
+#include "aidge/utils/Registrar.hpp"
+
+namespace Aidge {
+template <class I, class O, DimSize_t DIM>
+void ReduceMeanImpl_cpu_forward_kernel(const typename ReduceMean_Op<DIM>::Attrs& attrs,
+                                     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 nb_dims = inputDims.size();
+    const std::size_t totalElements = std::accumulate(inputDims.cbegin(), inputDims.cend(), 1, std::multiplies<std::size_t>());
+
+    if (DIM == 1) {
+        const std::size_t stride_pre = std::accumulate(inputDims.cbegin(), inputDims.cbegin() + std::get<0>(attrs)[0], 1, std::multiplies<std::size_t>());
+        const std::size_t stride_post = std::accumulate(inputDims.crbegin(), inputDims.crbegin() + nb_dims -1 - std::get<0>(attrs)[0], 1, std::multiplies<std::size_t>());
+
+        const std::size_t dim_i = inputDims[std::get<0>(attrs)[0]];
+        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;
+                output[idx_o] = input[idx_i];
+                for (std::size_t i = 1; i < dim_i; ++i) {
+                    output[idx_o] += input[idx_i + i*stride_post];
+                }
+                output[idx_o] /= dim_i;
+            }
+        }
+    } else {
+        std::size_t outputElements = totalElements;
+
+        std::size_t *stride_post = new std::size_t[nb_dims];
+        stride_post[nb_dims - 1] = 1;
+        for (std::size_t i = nb_dims-2; i != static_cast<std::size_t>(-1); --i) {
+            stride_post[i] = stride_post[i+1]*inputDims[i+1];
+        }
+        std::size_t *stride_pre = new std::size_t[nb_dims];
+        stride_pre[0] = 1;
+        for (std::size_t i = 1; i < nb_dims; ++i) {
+            stride_pre[i] = stride_pre[i-1]*inputDims[i-1];
+        }
+
+        const I* inputAccumulation = input;
+        I* outputAccumulation = nullptr;
+
+        for (const auto& axisInt : std::get<0>(attrs)) {
+            const std::size_t a = static_cast<std::size_t>(axisInt);
+            outputElements /= inputDims[a];
+            outputAccumulation = new I[outputElements];
+            const std::size_t dim_i = inputDims[a];
+            for (std::size_t pre = 0; pre < stride_pre[a]; ++pre) {
+                for (std::size_t post = 0; post < stride_post[a]; ++post) {
+                    const std::size_t idx_i = pre * dim_i * stride_post[a] + post;
+                    const std::size_t idx_o = pre * stride_post[a] + post;
+                    outputAccumulation[idx_o] = inputAccumulation[idx_i];
+                    for (std::size_t i = 1; i < dim_i; ++i) {
+                        outputAccumulation[idx_o] += inputAccumulation[idx_i + i*stride_post[a]];
+                    }
+                }
+            }
+            std::for_each(stride_pre+a+1, stride_pre+nb_dims, [dim_i] (std::size_t& val) { val /= dim_i; });
+            if (inputAccumulation != input) {
+                delete[] inputAccumulation;
+            }
+            inputAccumulation = outputAccumulation;
+        }
+
+        // Copy elements from inputAccumulation to output while dividing by divisor
+        I divisor = totalElements / outputElements;
+        std::transform(inputAccumulation, inputAccumulation + outputElements, output,
+                    [divisor](int element) { return element / divisor; });
+        if (outputAccumulation) {
+            delete[] outputAccumulation;
+        }
+        delete[] stride_post;
+        delete[] stride_pre;
+    }
+}
+
+namespace {
+// DIM = 1
+static Registrar<ReduceMeanImpl1DForward_cpu> registrarReduceMeanImplForward_1D_cpu_Float32(
+        {DataType::Float32, DataType::Float32}, Aidge::ReduceMeanImpl_cpu_forward_kernel<float, float,1>);
+static Registrar<ReduceMeanImpl1DForward_cpu> registrarReduceMeanImplForward_1D_cpu_Int32(
+        {DataType::Int32, DataType::Int32}, Aidge::ReduceMeanImpl_cpu_forward_kernel<int, int,1>);
+static Registrar<ReduceMeanImpl1DForward_cpu> registrarReduceMeanImplForward_1D_cpu_Float64(
+        {DataType::Float64, DataType::Float64}, Aidge::ReduceMeanImpl_cpu_forward_kernel<double, double,1>);
+
+// DIM = 2
+static Registrar<ReduceMeanImpl2DForward_cpu> registrarReduceMeanImplForward_2D_cpu_Float32(
+        {DataType::Float32, DataType::Float32}, Aidge::ReduceMeanImpl_cpu_forward_kernel<float, float,2>);
+static Registrar<ReduceMeanImpl2DForward_cpu> registrarReduceMeanImplForward_2D_cpu_Int32(
+        {DataType::Int32, DataType::Int32}, Aidge::ReduceMeanImpl_cpu_forward_kernel<int, int,2>);
+static Registrar<ReduceMeanImpl2DForward_cpu> registrarReduceMeanImplForward_2D_cpu_Float64(
+        {DataType::Float64, DataType::Float64}, Aidge::ReduceMeanImpl_cpu_forward_kernel<double, double,2>);
+
+// DIM = 3
+static Registrar<ReduceMeanImpl3DForward_cpu> registrarReduceMeanImplForward_3D_cpu_Float32(
+        {DataType::Float32, DataType::Float32}, Aidge::ReduceMeanImpl_cpu_forward_kernel<float, float,3>);
+static Registrar<ReduceMeanImpl3DForward_cpu> registrarReduceMeanImplForward_3D_cpu_Int32(
+        {DataType::Int32, DataType::Int32}, Aidge::ReduceMeanImpl_cpu_forward_kernel<int, int,3>);
+static Registrar<ReduceMeanImpl3DForward_cpu> registrarReduceMeanImplForward_3D_cpu_Float64(
+        {DataType::Float64, DataType::Float64}, Aidge::ReduceMeanImpl_cpu_forward_kernel<double, double,3>);
+}  // namespace
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_REDUCEMEANIMPL_FORWARD_KERNEL_H_ */

include/aidge/backend/cpu/operator/ReshapeImpl.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_RESHAPEIMPL_H_
+#define AIDGE_CPU_OPERATOR_RESHAPEIMPL_H_
+
+#include "aidge/backend/OperatorImpl.hpp"
+#include "aidge/operator/Reshape.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+#include <memory>
+#include <vector>
+
+namespace Aidge {
+// class Reshape_Op;
+
+// compute kernel registry for forward and backward
+class ReshapeImplForward_cpu
+    : public Registrable<ReshapeImplForward_cpu, std::tuple<DataType, DataType>, void(std::size_t, const void*, void*)> {
+};
+class ReshapeImplBackward_cpu
+    : public Registrable<ReshapeImplBackward_cpu, std::tuple<DataType, DataType>, void(std::size_t, const void*, void*)> {
+};
+
+class ReshapeImpl_cpu : public OperatorImpl {
+public:
+    ReshapeImpl_cpu(const Reshape_Op& op) : OperatorImpl(op) {}
+
+    static std::unique_ptr<ReshapeImpl_cpu> create(const Reshape_Op& op) {
+        return std::make_unique<ReshapeImpl_cpu>(op);
+    }
+
+    NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
+    void forward() override;
+};
+
+namespace {
+static Registrar<Reshape_Op> registrarReshapeImpl_cpu("cpu", Aidge::ReshapeImpl_cpu::create);
+}
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_RESHAPEIMPL_H_ */

include/aidge/backend/cpu/operator/ReshapeImpl_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_RESHAPEIMPL_FORWARD_KERNEL_H_
+#define AIDGE_CPU_OPERATOR_RESHAPEIMPL_FORWARD_KERNEL_H_
+
+#include "aidge/utils/Registrar.hpp"
+#include <cmath>
+
+#include "aidge/backend/cpu/operator/ReshapeImpl.hpp"
+
+namespace Aidge {
+template <class I, class O>
+void ReshapeImpl_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_);
+
+    std::copy_n(input, inputLength, output);
+}
+
+namespace {
+static Registrar<ReshapeImplForward_cpu> registrarReshapeImplForward_cpu_Float32(
+        {DataType::Float32, DataType::Float32},
+        Aidge::ReshapeImpl_cpu_forward_kernel<float, float>);
+static Registrar<ReshapeImplForward_cpu> registrarReshapeImplForward_cpu_Int32(
+        {DataType::Int32, DataType::Int32},
+        Aidge::ReshapeImpl_cpu_forward_kernel<int, int>);
+static Registrar<ReshapeImplForward_cpu> registrarReshapeImplForward_cpu_Float64(
+        {DataType::Float64, DataType::Float64},
+        Aidge::ReshapeImpl_cpu_forward_kernel<double, double>);
+}  // namespace
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_RESHAPEIMPL_FORWARD_KERNEL_H_ */

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

@@ -35,7 +35,7 @@ void SliceImpl_cpu_forward_kernel(const typename Slice_Op::Attrs& attrs,
         const std::int64_t axis_ = std::get<2>(attrs)[i];
         const std::int64_t start_ = std::get<0>(attrs)[i];
         const std::int64_t end_ = std::get<1>(attrs)[i];
-        const std::size_t axis = axis_ >= 0 ? axis_ : static_cast<std::size_t>(axis_ + static_cast<std::int32_t>(inputDims.size()));
+        const std::size_t axis = axis_ >= 0 ? axis_ : static_cast<std::size_t>(axis_) + inputDims.size();
         const std::size_t start = start_ >= 0 ? start_ : start_ + inputDims[axis];
         const std::size_t end = end_ >= 0 ? end_ : end_ + inputDims[axis];
         std::size_t stride = 1;

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

@@ -25,10 +25,10 @@ namespace Aidge {
 
 // compute kernel registry for forward and backward
 class SoftmaxImplForward_cpu
-    : public Registrable<SoftmaxImplForward_cpu, std::tuple<DataType, DataType>, void(const DimSize_t, const DimSize_t, const DimSize_t, const void*, void*)> {
+    : public Registrable<SoftmaxImplForward_cpu, std::tuple<DataType, DataType>, void(std::size_t, const std::vector<DimSize_t>&, const void*, void*)> {
 };
 class SoftmaxImplBackward_cpu
-    : public Registrable<SoftmaxImplBackward_cpu, std::tuple<DataType, DataType>, void(const std::size_t, const void*, void*)> {
+    : public Registrable<SoftmaxImplBackward_cpu, std::tuple<DataType, DataType>, void(std::size_t, const std::vector<DimSize_t>&, const void*, void*)> {
 };
 
 class SoftmaxImpl_cpu : public OperatorImpl {

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

@@ -23,30 +23,33 @@
 
 namespace Aidge {
 template <class I, class O>
-void SoftmaxImpl_cpu_forward_kernel(const DimSize_t batchSize,
-                                        const DimSize_t channelSize,
-                                        const DimSize_t featureSize,
-                                        const void* input_,
-                                        void* output_) {
-
+void SoftmaxImpl_cpu_forward_kernel(std::size_t axisIdx, 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_);
 
-    for (std::size_t batch = 0; batch < batchSize; ++batch) {
-        for (std::size_t feature = 0; feature < featureSize; ++feature) {
-            std::size_t ioIndex = batch*channelSize*featureSize + feature;
+    std::size_t postAxisElems = 1;
+    for (std::size_t i = axisIdx + 1; i < inputDims.size(); ++i) {
+        postAxisElems *= inputDims[i];
+    }
+    std::size_t preAxisElems = 1;
+    for (std::size_t i = 0; i < axisIdx; ++i) {
+        preAxisElems *= inputDims[i];
+    }
 
-            I sum(0.0);
-            for (std::size_t ch = 0; ch < channelSize; ++ch) {
-                output[ioIndex] = std::exp(input[ioIndex]);
-                sum += output[ioIndex];
-                ioIndex+=featureSize;
+    for (std::size_t i = 0; i < preAxisElems; ++i) {
+        for (std::size_t j = 0; j < postAxisElems; ++j) {
+            // Calculate sum of exponentials within the axis
+            I sumExp = 0;
+            for (std::size_t k = 0; k < inputDims[axisIdx]; ++k) {
+                std::size_t inIdx = i * inputDims[axisIdx] * postAxisElems + k * postAxisElems + j;
+                sumExp += std::exp(input[inIdx]);
             }
 
-            ioIndex = batch*channelSize*featureSize + feature;
-            for (std::size_t ch = 0; ch < channelSize; ++ch) {
-                output[ioIndex] /= sum;
-                ioIndex += featureSize;
+            // Calculate softmax for the current slice along the axis
+            for (std::size_t  k = 0; k < inputDims[axisIdx]; ++k) {
+                std::size_t inIdx = i * inputDims[axisIdx] * postAxisElems + k * postAxisElems + j;
+                output[inIdx] = std::exp(input[inIdx]) / sumExp;
             }
         }
     }