Merge branch 'dev' of gitlab.eclipse.org:eclipse/aidge/aidge_backend_cpu into Fix/PowBackwardKernel

include/aidge/backend/cpu.hpp

@@ -19,6 +19,7 @@
 #include "aidge/backend/cpu/operator/AvgPoolingImpl.hpp"
 #include "aidge/backend/cpu/operator/MaxPoolingImpl.hpp"
 #include "aidge/backend/cpu/operator/BatchNormImpl.hpp"
+#include "aidge/backend/cpu/operator/BitShiftImpl.hpp"
 #include "aidge/backend/cpu/operator/ConvDepthWiseImpl.hpp"
 #include "aidge/backend/cpu/operator/ConvImpl.hpp"
 #include "aidge/backend/cpu/operator/ConstantOfShapeImpl.hpp"

include/aidge/backend/cpu/operator/BitShiftImpl.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_BITSHIFTIMPL_H_
+#define AIDGE_CPU_OPERATOR_BITSHIFTIMPL_H_
+
+#include "aidge/backend/cpu/operator/OperatorImpl.hpp"
+#include "aidge/operator/BitShift.hpp"
+#include "aidge/utils/Registrar.hpp"
+#include "aidge/utils/Types.h"
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
+#include <memory>
+#include <vector>
+
+namespace Aidge {
+// Operator implementation entry point for the backend
+using BitShiftImpl_cpu = OperatorImpl_cpu<BitShift_Op,
+    void(const BitShift_Op::BitShiftDirection,
+    const std::vector<std::size_t>&, 
+    const std::vector<std::size_t>&, 
+    const std::vector<std::size_t>&, 
+    const void*, 
+    const void*,
+    void*)>;
+    
+    // Implementation entry point registration to Operator
+    REGISTRAR(BitShift_Op,"cpu",Aidge::BitShiftImpl_cpu::create);
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_BITSHIFTIMPL_H_ */

include/aidge/backend/cpu/operator/BitShiftImpl_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_BITSHIFTIMPL_KERNELS_H_
+#define AIDGE_CPU_OPERATOR_BITSHIFTIMPL_KERNELS_H_
+
+#include "aidge/utils/Registrar.hpp"
+
+#include <cstdint>     // std::int32_t, std::int64_t
+#include "aidge/operator/BitShift.hpp"
+
+#include "aidge/backend/cpu/data/Broadcasting.hpp"
+#include "aidge/backend/cpu/operator/BitShiftImpl.hpp"
+
+
+
+namespace Aidge {
+template <class I1, class I2, class O>
+void BitShiftImpl_cpu_forward_kernel(
+                                const BitShift_Op::BitShiftDirection direction,
+                                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 size_t totalElements = std::accumulate(outputDims.begin(), outputDims.end(), std::size_t(1), std::multiplies<std::size_t>());
+    
+    for (std::size_t oIndex = 0; oIndex < totalElements; ++oIndex)
+    {
+        std::vector<size_t> indexes = getMultiDimIndices(outputDims, oIndex);
+        std::size_t idx1 = getFlattenedIndex(input1Dims, indexes);
+        std::size_t idx2 = getFlattenedIndex(input2Dims, indexes);
+        if(direction == BitShift_Op::BitShiftDirection::right)
+
+        {
+                output[oIndex]= input_1[idx1] >> input_2[idx2];
+        }
+        else
+        {
+                output[oIndex] = input_1[idx1] << input_2[idx2];
+        }
+    }
+}
+
+REGISTRAR(BitShiftImpl_cpu,
+{DataType::Int32},
+{ProdConso::inPlaceModel,Aidge::BitShiftImpl_cpu_forward_kernel<std::int32_t, std::int32_t, std::int32_t>,nullptr});
+REGISTRAR(BitShiftImpl_cpu,
+{DataType::Int64},
+{ProdConso::inPlaceModel,Aidge::BitShiftImpl_cpu_forward_kernel<std::int64_t, std::int64_t, std::int64_t>,nullptr});
+
+
+}  // namespace Aidge
+
+#endif /* AIDGE_CPU_OPERATOR_BitShiftIMPL_KERNELS_H_ */
\ No newline at end of file

src/operator/BitShiftImpl.cpp

+/********************************************************************************
+ * 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
+ *
+ ********************************************************************************/
+
+#include <cassert>
+#include <chrono>  // std::chrono::milliseconds
+#include <numeric>
+#include <thread>  // std::this_thread::sleep_for
+#include <vector>
+
+
+#include "aidge/utils/Types.h"
+#include "aidge/backend/cpu/data/Broadcasting.hpp"
+#include "aidge/backend/cpu/data/GetCPUPtr.h"
+
+#include "aidge/backend/cpu/operator/BitShiftImpl.hpp"
+#include "aidge/backend/cpu/operator/BitShiftImpl_kernels.hpp"
+
+template<>
+void Aidge::BitShiftImpl_cpu::forward() {
+
+    const auto& op_ = dynamic_cast<const BitShift_Op&>(mOp);
+
+
+    const auto impl = Registrar<BitShiftImpl_cpu>::create(getBestMatch(getRequiredSpec()));
+
+
+    const std::vector<std::size_t> inputDims0 = getBroadcastedDims(std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dims(),
+                                                                   std::static_pointer_cast<Tensor>(mOp.getRawInput(0))->dims());
+    const std::vector<std::size_t> inputDims1 = getBroadcastedDims(std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dims(),
+                                                                   std::static_pointer_cast<Tensor>(mOp.getRawInput(1))->dims());
+
+    BitShift_Op::BitShiftDirection direction = op_.direction();
+
+    // Call kernel
+    impl.forward(
+        direction,
+        inputDims0,
+        inputDims1,
+        std::static_pointer_cast<Tensor>(mOp.getRawOutput(0))->dims(),
+        getCPUPtr(mOp.getRawInput(0)),
+        getCPUPtr(mOp.getRawInput(1)),
+        getCPUPtr(mOp.getRawOutput(0)));
+        
+}
+
+template <>
+void Aidge::BitShiftImpl_cpu::backward() {
+    AIDGE_THROW_OR_ABORT(std::runtime_error, "Backward not yet implemented for BitShift_Op on backend cpu");
+}
\ No newline at end of file

unit_tests/operator/Test_BitShift.cpp

+/********************************************************************************
+ * 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
+ *
+ ********************************************************************************/
+
+#include <catch2/catch_test_macros.hpp>
+#include <cstddef>   // std::size_t
+#include <cstdint>   // std::uint16_t
+#include <chrono>
+#include <iostream>
+#include <memory>
+#include <numeric>   
+#include <random>    // std::random_device, std::mt19937, std::uniform_real_distribution
+#include <iomanip>
+#include "aidge/data/Tensor.hpp"
+#include "aidge/operator/BitShift.hpp"
+#include "aidge/utils/TensorUtils.hpp"
+
+namespace Aidge {
+
+TEST_CASE("[cpu/operator] BitShift_TEST", "[BitShift][CPU]") {
+    constexpr std::uint16_t NBTRIALS = 15;
+    // Create a random number generator
+    std::random_device rd;
+    std::mt19937 gen(rd());
+    std::uniform_int_distribution<int> valueDist(-15, 15); 
+    std::uniform_int_distribution<std::size_t> dimSizeDist(std::size_t(2), std::size_t(5));
+    std::uniform_int_distribution<std::size_t> nbDimsDist(std::size_t(1), std::size_t(3));
+    std::uniform_int_distribution<int> boolDist(0,1);
+
+    BitShift_Op::BitShiftDirection direction = BitShift_Op::BitShiftDirection::left;
+
+    if(valueDist(gen) % 2 == 0)
+    {
+        direction = BitShift_Op::BitShiftDirection::right;
+    }
+
+    // Create BitShift Operator
+    std::shared_ptr<Node> myBitShift = BitShift(direction);
+    auto op = std::static_pointer_cast<OperatorTensor>(myBitShift-> getOperator());
+    op->setDataType(DataType::Int32);
+    op->setBackend("cpu");
+
+    // Create 2 input Tensors
+    std::shared_ptr<Tensor> T0 = std::make_shared<Tensor>();
+    op->associateInput(0,T0);
+    T0->setDataType(DataType::Int32);
+    T0->setBackend("cpu");
+    std::shared_ptr<Tensor> T1 = std::make_shared<Tensor>();
+    op -> associateInput(1,T1);
+    T1->setDataType(DataType::Int32);
+    T1->setBackend("cpu");
+
+    // Create results Tensor
+    std::shared_ptr<Tensor> Tres = std::make_shared<Tensor>();
+    Tres->setDataType(DataType::Int32);
+    Tres->setBackend("cpu");
+
+    // To measure execution time of 'BitShift_Op::forward()' member function call
+    std::chrono::time_point<std::chrono::system_clock> start;
+
+    std::chrono::time_point<std::chrono::system_clock> end;
+    std::chrono::duration<double, std::micro> duration{};
+
+    SECTION("BitShiftImpl_cpu::forward()") {
+        SECTION("Test Forward Kernel with same dimensions") {
+            std::size_t number_of_operation = 0;
+
+            for (std::uint16_t trial = 0; trial < NBTRIALS; ++trial) {
+                // generate 2 random Tensors
+                const std::size_t nbDims = nbDimsDist(gen);
+                std::vector<std::size_t> dims;
+                for (std::size_t i = 0; i < nbDims; ++i) {
+                    dims.push_back(dimSizeDist(gen));
+                }
+                const std::size_t nb_elements = std::accumulate(dims.cbegin(), dims.cend(), std::size_t(1), std::multiplies<std::size_t>());
+                number_of_operation += nb_elements;
+
+                // without broadcasting
+                int* array0 = new int[nb_elements];
+                int* array1 = new int[nb_elements];
+                int* result = new int[nb_elements];
+
+                for (std::size_t i = 0; i < nb_elements; ++i) {
+                    array0[i] = valueDist(gen);
+                    array1[i] = std::abs(valueDist(gen)); // bitshift is impossible with negative value
+                    if(direction == BitShift_Op::BitShiftDirection::left)
+                    {
+                        result[i] = array0[i] << array1[i];
+                    }
+                    else
+                    {
+                        result[i] = array0[i] >> array1[i];
+                    }
+                }
+
+                // input0
+                T0->resize(dims);
+                T0 -> getImpl() -> setRawPtr(array0, nb_elements);
+
+                // input1
+                T1->resize(dims);
+                T1 -> getImpl() -> setRawPtr(array1, nb_elements);
+
+                // results
+                Tres->resize(dims);
+                Tres -> getImpl() -> setRawPtr(result, nb_elements);
+
+                op->forwardDims();
+                start = std::chrono::system_clock::now();
+                myBitShift->forward();
+                end = std::chrono::system_clock::now();
+                duration += std::chrono::duration_cast<std::chrono::microseconds>(end - start);
+
+                bool is_eq = approxEq<int>(*(op->getOutput(0)), *Tres);
+
+                auto Output = *(op->getOutput(0));
+                auto prt = Output.getImpl()->rawPtr();
+
+                REQUIRE(is_eq);
+
+                delete[] array0;
+                delete[] array1;
+                delete[] result;
+
+
+            }
+            std::cout << "number of elements over time spent: " << (number_of_operation / duration.count())<< std::endl;
+            std::cout << "total time: " << duration.count() << "μs" << std::endl;
+        }
+        SECTION("Test BitShift kernels with Broadcasting") {
+            std::size_t number_of_operation = 0;
+
+            for (std::uint16_t trial = 0; trial < NBTRIALS; ++trial) {
+                // generate 2 random Tensors
+                // handle dimensions, replace some dimensions with '1' to get broadcasting
+                constexpr std::size_t nbDims = 4;
+                std::vector<std::size_t> dims;
+                for (std::size_t i = 0; i < nbDims; ++i) {
+                    dims.push_back(dimSizeDist(gen));
+                }
+                std::vector<std::size_t> dims0 = dims;
+                std::vector<std::size_t> dims1 = dims;
+                std::vector<std::size_t> dimsOut = dims;
+                for (std::size_t i = 0; i < nbDims; ++i) {
+                    if (boolDist(gen)) {
+                        dims0[i] = 1;
+                    }
+                    if (boolDist(gen)) {
+                        dims1[i] = 1;
+                    }
+                    dimsOut[i] = (dims0[i] == 1) ? dims1[i] : dims0[i];
+                }
+
+                // create arrays and fill them with random values
+                int* array0 = new int[dims0[0]*dims0[1]*dims0[2]*dims0[3]];
+                int* array1 = new int[dims1[0]*dims1[1]*dims1[2]*dims1[3]];
+                int* result = new int[dimsOut[0]*dimsOut[1]*dimsOut[2]*dimsOut[3]];
+
+                for (std::size_t i = 0; i < dims0[0]*dims0[1]*dims0[2]*dims0[3]; ++i) {
+                    array0[i] = valueDist(gen);
+                }
+                for (std::size_t i = 0; i < dims1[0]*dims1[1]*dims1[2]*dims1[3]; ++i) {
+                    array1[i] = std::abs(valueDist(gen));
+                }
+
+                //True result with broadcast
+                const std::size_t strides0[nbDims] = {dims0[1]*dims0[2]*dims0[3], dims0[2]*dims0[3], dims0[3], 1};
+                const std::size_t strides1[nbDims] = {dims1[1]*dims1[2]*dims1[3], dims1[2]*dims1[3], dims1[3], 1};
+                for (std::size_t a = 0; a < dimsOut[0]; ++a) {
+                    for (std::size_t b = 0; b < dimsOut[1]; ++b) {
+                        const std::size_t idx0_0 = strides0[0] * ((dims0[0] > 1) ? a : 0)
+                                                    + strides0[1] * ((dims0[1] > 1) ? b : 0);
+                        const std::size_t idx1_0 = strides1[0] * ((dims1[0] > 1) ? a : 0)
+                                                    + strides1[1] * ((dims1[1] > 1) ? b : 0);
+                        for (std::size_t c = 0; c < dimsOut[2]; ++c) {
+                            const std::size_t idx_out = dimsOut[3] * (c + dimsOut[2] * (b + dimsOut[1] * a));
+                            for (std::size_t d = 0; d < dimsOut[3]; ++d) {
+                                std::size_t idx0 = idx0_0
+                                                    + strides0[2] * ((dims0[2] > 1) ? c : 0)
+                                                    + ((dims0[3] > 1) ? d : 0);
+                                std::size_t idx1 = idx1_0
+                                                    + strides1[2] * ((dims1[2] > 1) ? c : 0)
+                                                    + ((dims1[3] > 1) ? d : 0);
+                                if(direction == BitShift_Op::BitShiftDirection::left)
+                                {
+                                    result[idx_out + d] = array0[idx0] << array1[idx1];
+                                }
+                                else
+                                {
+                                    result[idx_out + d] = array0[idx0] >> array1[idx1];                               
+                                }
+                            }
+                        }
+                    }
+                }
+
+                // conversion to Aidge::Tensors
+                // input0
+                T0->resize(dims0);
+                T0 -> getImpl() -> setRawPtr(array0, dims0[0]*dims0[1]*dims0[2]*dims0[3]);
+
+                // input1
+                T1->resize(dims1);
+                T1 -> getImpl() -> setRawPtr(array1, dims1[0]*dims1[1]*dims1[2]*dims1[3]);
+
+                // results
+                Tres->resize(dimsOut);
+                Tres -> getImpl() -> setRawPtr(result, dimsOut[0]*dimsOut[1]*dimsOut[2]*dimsOut[3]);
+
+                // compute result
+                op->forwardDims();
+                start = std::chrono::system_clock::now();
+                myBitShift->forward();
+                end = std::chrono::system_clock::now();
+                duration += std::chrono::duration_cast<std::chrono::microseconds>(end - start);
+
+                // comparison between truth and computed result
+                bool equiv = (approxEq<int>(*(op->getOutput(0)), *Tres));
+                if(equiv == false)
+                {
+                    std::cout << "Problem\n";
+                }
+                REQUIRE(equiv);
+
+                delete[] array0;
+                delete[] array1;
+                delete[] result;
+
+                const std::size_t nb_elements = std::accumulate(dimsOut.cbegin(), dimsOut.cend(), std::size_t(1), std::multiplies<std::size_t>());
+                number_of_operation += nb_elements;
+            }
+            std::cout << "number of elements over time spent: " << (number_of_operation / duration.count())<< std::endl;
+            std::cout << "total time: " << duration.count() << "μs" << std::endl;
+        }
+
+}
+} // namespace Aidge
+}
\ No newline at end of file