Merge remote-tracking branch 'EclipseRepo/dev' into feat/support_ASAN

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

+include:
+  - remote: 'https://gitlab.eclipse.org/eclipse/aidge/gitlab_shared_files/-/raw/main/.gitlab/ci/shared_script.gitlab-ci.yml'
+
 build:ubuntu_cpp:
   stage: build
   needs: []
@@ -6,9 +9,9 @@ build:ubuntu_cpp:
   script:
     # Download dependencies
     # aidge_core
-    - 'curl --location --output build_artifacts.zip "https://gitlab.eclipse.org/api/v4/projects/5139/jobs/artifacts/main/download?job=build:ubuntu_cpp"'
-    - unzip -o build_artifacts.zip -d .
-    - rm -rf build_cpp
+    - DEPENDENCY_NAME="aidge_core"
+    - DEPENDENCY_JOB="build:ubuntu_cpp"
+    - !reference [.download_dependency, script]
 
     # Build current module
     - export CMAKE_PREFIX_PATH=../install_cpp
@@ -32,9 +35,9 @@ build:ubuntu_cpp_g++10:
   script:
     # Download dependencies
     # aidge_core
-    - 'curl --location --output build_artifacts.zip "https://gitlab.eclipse.org/api/v4/projects/5139/jobs/artifacts/main/download?job=build:ubuntu_cpp"'
-    - unzip -o build_artifacts.zip -d .
-    - rm -rf build_cpp
+    - DEPENDENCY_NAME="aidge_core"
+    - DEPENDENCY_JOB="build:ubuntu_cpp"
+    - !reference [.download_dependency, script]
 
     # Build current module
     - export CMAKE_PREFIX_PATH=../install_cpp
@@ -55,9 +58,9 @@ build:ubuntu_cpp_g++12:
   script:
     # Download dependencies
     # aidge_core
-    - 'curl --location --output build_artifacts.zip "https://gitlab.eclipse.org/api/v4/projects/5139/jobs/artifacts/main/download?job=build:ubuntu_cpp"'
-    - unzip -o build_artifacts.zip -d .
-    - rm -rf build_cpp
+    - DEPENDENCY_NAME="aidge_core"
+    - DEPENDENCY_JOB="build:ubuntu_cpp"
+    - !reference [.download_dependency, script]
 
     # Build current module
     - export CMAKE_PREFIX_PATH=../install_cpp
@@ -78,9 +81,9 @@ build:ubuntu_cpp_clang12:
   script:
     # Download dependencies
     # aidge_core
-    - 'curl --location --output build_artifacts.zip "https://gitlab.eclipse.org/api/v4/projects/5139/jobs/artifacts/main/download?job=build:ubuntu_cpp"'
-    - unzip -o build_artifacts.zip -d .
-    - rm -rf build_cpp
+    - DEPENDENCY_NAME="aidge_core"
+    - DEPENDENCY_JOB="build:ubuntu_cpp"
+    - !reference [.download_dependency, script]
 
     # Build current module
     - export CMAKE_PREFIX_PATH=../install_cpp
@@ -101,9 +104,9 @@ build:ubuntu_cpp_clang15:
   script:
     # Download dependencies
     # aidge_core
-    - 'curl --location --output build_artifacts.zip "https://gitlab.eclipse.org/api/v4/projects/5139/jobs/artifacts/main/download?job=build:ubuntu_cpp"'
-    - unzip -o build_artifacts.zip -d .
-    - rm -rf build_cpp
+    - DEPENDENCY_NAME="aidge_core"
+    - DEPENDENCY_JOB="build:ubuntu_cpp"
+    - !reference [.download_dependency, script]
 
     # Build current module
     - export CMAKE_PREFIX_PATH=../install_cpp
@@ -120,17 +123,21 @@ build:ubuntu_python:
   needs: []
   tags:
     - docker
+
   script:
     # Download dependencies
     # aidge_core (Python)
-    - 'curl --location --output build_artifacts.zip "https://gitlab.eclipse.org/api/v4/projects/5139/jobs/artifacts/main/download?job=build:ubuntu_python"'
-    - unzip -o build_artifacts.zip -d .
+    - DEPENDENCY_NAME="aidge_core"
+    - DEPENDENCY_JOB="build:ubuntu_python"
+    - !reference [.download_dependency, script]
 
     - python3 -m pip install virtualenv
     - virtualenv venv
     - 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:
@@ -155,9 +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
@@ -191,8 +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_recipies.py → aidge_backend_cpu/unit_tests/test_recipes.py

@@ -15,7 +15,7 @@ import aidge_backend_cpu
 from functools import reduce
 import numpy as np
 
-class test_recipies(unittest.TestCase):
+class test_recipes(unittest.TestCase):
     def setUp(self):
         pass
 
@@ -33,12 +33,9 @@ class test_recipies(unittest.TestCase):
         conv = aidge_core.Conv2D(1, 1, [3, 3], name="Conv0")
         bn = aidge_core.BatchNorm2D(1, name="Add0")
 
-        graph_view = aidge_core.sequential([conv, bn])
+        graph_view = aidge_core.sequential([input_node, conv, bn])
 
         # Add random values to conv and BatchNorm parameters
-        input_node.add_child(graph_view)
-        input_node.get_operator().set_datatype(aidge_core.DataType.Float32)
-        input_node.get_operator().set_backend("cpu")
         graph_view.set_datatype(aidge_core.DataType.Float32)
         graph_view.set_backend("cpu")
 

aidge_backend_cpu/unit_tests/test_scheduler.py

@@ -40,18 +40,14 @@ class test_scheduler(unittest.TestCase):
         input_data =  np.array([0]).astype(np.float32)
         input_tensor = aidge_core.Tensor(input_data)
 
-        input_node = aidge_core.Producer(input_tensor, "X")
-
         graph_view = aidge_core.sequential([
+            aidge_core.Producer(input_tensor, "X"),
             aidge_core.FC(1, 50, name='0'),
             aidge_core.FC(50, 50, name='1'),
             aidge_core.FC(50, 10, name='2'),
         ])
         EXPECTED_SCHEDULE = ['0', '1', '2']
 
-        input_node.add_child(graph_view)
-        input_node.get_operator().set_datatype(aidge_core.DataType.Float32)
-        input_node.get_operator().set_backend("cpu")
         graph_view.set_datatype(aidge_core.DataType.Float32)
         graph_view.set_backend("cpu")
 
@@ -60,15 +56,17 @@ class test_scheduler(unittest.TestCase):
         scheduler = aidge_core.SequentialScheduler(graph_view)
         scheduler.generate_scheduling()
 
-        self.assertListEqual([i.name() for i in scheduler.get_static_scheduling()], EXPECTED_SCHEDULE)
+        self.assertEqual(len(scheduler.get_static_scheduling()), 10)
+        # Do not care about the order of execution of the producers
+        self.assertListEqual([i.name() for i in scheduler.get_static_scheduling()[-3:]], EXPECTED_SCHEDULE)
 
 
     def test_parallel_scheduling(self):
         input_data =  np.array([0]).astype(np.float32)
         input_tensor = aidge_core.Tensor(input_data)
 
-        input_node = aidge_core.Producer(input_tensor, "X")
         graph_view = aidge_core.sequential([
+            aidge_core.Producer(input_tensor, "X"),
             aidge_core.FC(1, 50, name='0'),
             aidge_core.parallel([aidge_core.FC(50, 50, name='1'), aidge_core.FC(50, 50, name='3')]),
             aidge_core.Add(2, name='2'),
@@ -76,9 +74,6 @@ class test_scheduler(unittest.TestCase):
 
         EXPECTED_SCHEDULE = [['0', '1', '3', '2'],  ['0', '3', '1', '2']] # Both scheduling are valid !
 
-        input_node.add_child(graph_view)
-        input_node.get_operator().set_datatype(aidge_core.DataType.Float32)
-        input_node.get_operator().set_backend("cpu")
         graph_view.set_datatype(aidge_core.DataType.Float32)
         graph_view.set_backend("cpu")
 
@@ -87,7 +82,9 @@ class test_scheduler(unittest.TestCase):
         scheduler = aidge_core.SequentialScheduler(graph_view)
         scheduler.generate_scheduling()
 
-        self.assertTrue([i.name() for i in scheduler.get_static_scheduling()] in EXPECTED_SCHEDULE)
+        self.assertEqual(len(scheduler.get_static_scheduling()), 11)
+        # Do not care about the order of execution of the producers
+        self.assertTrue([i.name() for i in scheduler.get_static_scheduling()[-4:]] in EXPECTED_SCHEDULE)
 
 if __name__ == '__main__':
     unittest.main()

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"
@@ -21,18 +20,29 @@
 #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"
 #include "aidge/backend/cpu/operator/MulImpl.hpp"
 #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/ProducerImpl.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"
 #include "aidge/backend/cpu/operator/SqrtImpl.hpp"
 #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/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/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) {
-  return std::static_pointer_cast<Tensor>(data)->getImpl()->rawPtr();
-}
-} // 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:
-    const Tensor &mTensor;  // Impl needs to access Tensor information, but is not
-                            // supposed to change it!
-    /// 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(const Tensor &tensor) : TensorImpl(Backend), mTensor(tensor) {}
-
-    bool operator==(const TensorImpl &otherImpl) const override final {
-        const auto& typedOtherImpl = reinterpret_cast<const TensorImpl_cpu<T> &>(otherImpl);
-        AIDGE_INTERNAL_ASSERT(typedOtherImpl.size() >= mTensor.size());
-
-        std::size_t i = 0;
-        for (; i < mTensor.size() &&
-               *(mData.data()+i) == *static_cast<const T*>(typedOtherImpl.rawPtr(i));
-               ++i) {
-        }
-        return i == mTensor.size();
-    }
-
-    static std::unique_ptr<TensorImpl_cpu> create(const Tensor &tensor) {
-        return std::make_unique<TensorImpl_cpu<T>>(tensor);
-    }
-
-    inline std::size_t size() const noexcept override final { return mData.size(); }
-    inline std::size_t scalarSize() const noexcept override final { return sizeof(T); }
-
-    void setDevice(DeviceIdx_t device) override final {
-        AIDGE_ASSERT(device == 0, "device cannot be != 0 for CPU backend");
-    }
-
-    void copy(const void *src, NbElts_t length, NbElts_t offset = 0) override final {
-        AIDGE_ASSERT(length <= mData.size() || length <= mTensor.size(), "copy length is above capacity");
-        std::copy(static_cast<const T *>(src), static_cast<const T *>(src) + length,
-                  static_cast<T *>(rawPtr()) + offset);
-    }
-
-    void copyCast(const void *src, NbElts_t length, const DataType srcDt) override final {
-        if (length == 0) {
-            return;
-        }
-
-        AIDGE_ASSERT(length <= mData.size() || length <= mTensor.size(), "copy length is above capacity");
-        switch (srcDt)
-        {
-            case DataType::Float64:
-                std::copy(static_cast<const double*>(src), static_cast<const double*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::Float32:
-                std::copy(static_cast<const float*>(src), static_cast<const float*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::Float16:
-                std::copy(static_cast<const half_float::half*>(src), static_cast<const half_float::half*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::Int64:
-                std::copy(static_cast<const int64_t*>(src), static_cast<const int64_t*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::UInt64:
-                std::copy(static_cast<const uint64_t*>(src), static_cast<const uint64_t*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::Int32:
-                std::copy(static_cast<const int32_t*>(src), static_cast<const int32_t*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::UInt32:
-                std::copy(static_cast<const uint32_t*>(src), static_cast<const uint32_t*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::Int16:
-                std::copy(static_cast<const int16_t*>(src), static_cast<const int16_t*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::UInt16:
-                std::copy(static_cast<const uint16_t*>(src), static_cast<const uint16_t*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::Int8:
-                std::copy(static_cast<const int8_t*>(src), static_cast<const int8_t*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            case DataType::UInt8:
-                std::copy(static_cast<const uint8_t*>(src), static_cast<const uint8_t*>(src) + length,
-                        static_cast<T *>(rawPtr()));
-                break;
-            default:
-                AIDGE_THROW_OR_ABORT(std::runtime_error, "Unsupported data type.");
-                break;
-        }
-    }
-
-    void copyFromDevice(const void *src, NbElts_t length, const std::pair<std::string, DeviceIdx_t>& device) 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);
-    }
-
-    inline void copyFromHost(const void *src, NbElts_t length) override final {
-        copy(src, length);
-    }
-
-    void copyToHost(void *dst, NbElts_t length) const override final {
-        AIDGE_ASSERT(length <= mData.size() || length <= mTensor.size(), "copy length is above capacity");
-        const T* src = static_cast<const T*>(rawPtr());
-        std::copy(static_cast<const T *>(src), static_cast<const T *>(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() >= mTensor.size(), "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() >= mTensor.size(), "accessing uninitialized const hostPtr");
-        return (mData.data() + offset);
-    };
-
-    void setRawPtr(void *ptr, NbElts_t length) override final {
-        AIDGE_ASSERT(length >= mTensor.size(), "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() < mTensor.size()) {
-            // Need more data, a re-allocation will occur
-            AIDGE_ASSERT(mData.empty() || mDataOwner != nullptr, "trying to enlarge non-owned data");
-            mDataOwner.reset(new T[mTensor.size()]);
-            mData = future_std::span<T>(mDataOwner.get(), 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_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/AddImpl.hpp

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

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

@@ -14,12 +14,13 @@
 
 #include "aidge/utils/Registrar.hpp"
 
+#include "aidge/backend/cpu/data/Broadcasting.hpp"
 #include "aidge/backend/cpu/operator/AddImpl.hpp"
 
 namespace Aidge {
 
 template <class I, class O>
-void AddImpl_cpu_forward_kernel(const std::size_t inputLength, const std::vector<const void*> inputs_, void* output_) {
+void AddImpl_cpu_forward_kernel(const std::vector<const void*> inputs_, const std::vector<std::vector<std::size_t>>& inputDims, const std::size_t outputLength, const std::vector<std::size_t>& outDims, void* output_) {
     // FIXME: missing Add attributes as arguments
     std::vector<const I*> inputs;
     for (const auto& input_ : inputs_) {
@@ -27,12 +28,15 @@ void AddImpl_cpu_forward_kernel(const std::size_t inputLength, const std::vector
     }
     O* output = static_cast<O*>(output_);
 
-    for (std::size_t oIndex = 0; oIndex < inputLength; ++oIndex) {
+	for (std::size_t oIndex = 0; oIndex < outputLength; ++oIndex)
+	{
         output[oIndex] = 0;
-        for (std::size_t iIndex = 0; iIndex < inputs.size(); ++iIndex) {
-            output[oIndex] += inputs[iIndex][oIndex];
-        }
-    }
+		std::vector<size_t> indexes = getMultiDimIndices(outDims, oIndex);
+		for(std::size_t iIndex = 0; iIndex < inputs.size(); ++iIndex) {
+			std::size_t idx = getFlattenedIndex(inputDims[iIndex], indexes);
+            output[oIndex] += inputs[iIndex][idx];
+		}
+	}
 }
 
 namespace {

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

@@ -12,23 +12,24 @@
 #ifndef AIDGE_CPU_OPERATOR_DIVIMPL_H_
 #define AIDGE_CPU_OPERATOR_DIVIMPL_H_
 
+#include <memory>
+#include <tuple>
+#include <vector>
+
 #include "aidge/backend/OperatorImpl.hpp"
 #include "aidge/operator/Div.hpp"
 #include "aidge/utils/Registrar.hpp"
 #include "aidge/utils/Types.h"
-#include "aidge/backend/cpu/data/GetCPUPtr.h"
-#include <memory>
-#include <vector>
 
 namespace Aidge {
-// class Div_Op;
 
 // compute kernel registry for forward and backward
 class DivImplForward_cpu
-    : public Registrable<DivImplForward_cpu, std::tuple<DataType, DataType, DataType>, void(const std::size_t, const std::size_t, const void*, const void*,void*)> {
+    // : public Registrable<DivImplForward_cpu, std::tuple<DataType, DataType, DataType>, void(const std::vector<std::size_t>&, const std::vector<std::size_t>&, const std::vector<std::size_t>&, const void*, const void*,void*)> {
+    : public Registrable<DivImplForward_cpu, std::tuple<DataType, DataType, DataType>, void(const std::size_t, const std::size_t, const std::size_t, const void*, const void*,void*)> {
 };
 class DivImplBackward_cpu
-    : public Registrable<DivImplBackward_cpu, std::tuple<DataType, DataType, DataType>, void(const std::size_t, const std::size_t, const void*, const void*, void*)> {
+    : public Registrable<DivImplBackward_cpu, std::tuple<DataType, DataType, DataType>, void(const std::vector<std::size_t>&, const std::vector<std::size_t>&, const std::vector<std::size_t>&, const void*, const void*, void*)> {
 };
 
 class DivImpl_cpu : public OperatorImpl {
@@ -40,7 +41,8 @@ public:
     }
 
     NbElts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
-    void forward() override;
+
+    void forward() override final;
 };
 
 namespace {

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

@@ -12,42 +12,64 @@
 #ifndef AIDGE_CPU_OPERATOR_DIVIMPL_FORWARD_KERNEL_H_
 #define AIDGE_CPU_OPERATOR_DIVIMPL_FORWARD_KERNEL_H_
 
+#include <numeric>     // std::accumulate
+#include <cstddef>     // std::size_t
+#include <functional>  // std::multiplies
+
 #include "aidge/utils/Registrar.hpp"
 
+#include "aidge/backend/cpu/data/Broadcasting.hpp"
 #include "aidge/backend/cpu/operator/DivImpl.hpp"
 
 namespace Aidge {
+// template <class I1, class I2, class O>
+// void DivImpl_cpu_forward_kernel(const std::vector<std::size_t>& input1Dims,
+//                                 const std::vector<std::size_t>& input2Dims,
+//                                 const std::vector<std::size_t>& outputDims,
+//                                 const void* input1_,
+//                                 const void* input2_,
+//                                 void* output_) {
+
+//     const I1* input_1 = static_cast<const I1*>(input1_);
+//     const I2* input_2 = static_cast<const I2*>(input2_);
+//     O* output = static_cast<O*>(output_);
+
+//     const std::size_t totalElements = std::accumulate(outputDims.cbegin(), outputDims.cend(), std::size_t(1), std::multiplies<std::size_t>());
+
+// 	for (std::size_t oIndex = 0; oIndex < totalElements; ++oIndex)
+// 	{
+// 		std::vector<std::size_t> indexes = getMultiDimIndices(outputDims, oIndex);
+
+// 		std::size_t idx1 = getFlattenedIndex(input1Dims, indexes);
+// 		std::size_t idx2 = getFlattenedIndex(input2Dims, indexes);
+
+//         // TODO assert if input_2 is bad?
+//         output[oIndex] = input_1[idx1] / input_2[idx2];
+//     }
+// }
+
 template <class I1, class I2, class O>
-void DivImpl_cpu_forward_kernel(std::size_t input1Length,
-                                     std::size_t input2Length,
-                                     const void* input1_,
-                                     const void* input2_,
-                                     void* output_) {
+constexpr void DivImpl_cpu_forward_kernel(const std::size_t input1size_,
+                                const std::size_t input2size_,
+                                const std::size_t output1size_,
+                                const void* input1_,
+                                const void* input2_,
+                                void* output_) {
 
     const I1* input_1 = static_cast<const I1*>(input1_);
     const I2* input_2 = static_cast<const I2*>(input2_);
     O* output = static_cast<O*>(output_);
-    if (input2Length == input1Length)
-    {
-        for (std::size_t i = 0; i < input1Length; ++i) {
-            output[i] = input_1[i] / input_2[i];
-        }
-    }
-    else if (input2Length == 1)
-    {
-        for (std::size_t i = 0; i < input1Length; ++i) {
-            output[i] = input_1[i] / input_2[0];
-        }
-    }
-    else // input_2 is 1d and of size the number of channels of input_1
-    {
-        for (std::size_t i = 0; i < input1Length; ++i) {
-            std::size_t channelIdx = i % input2Length;
-            output[i] = input_1[i] / input_2[channelIdx];
-        }
+
+// suppose values are contiguous in memory
+    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>(input_1[in1_id] / input_2[in2_id]);
     }
 }
 
+
+
 namespace {
 static Registrar<DivImplForward_cpu> registrarDivImplForward_cpu_Float32(
         {DataType::Float32, DataType::Float32, DataType::Float32},

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