diff --git a/include/aidge/backend/cpu/operator/AddImpl.hpp b/include/aidge/backend/cpu/operator/AddImpl.hpp
index e39c35b42fdb6065aa72aee092cd1cd23b2b1011..ca04dff9164ecc8492d9263b32b60272dbbad395 100644
--- a/include/aidge/backend/cpu/operator/AddImpl.hpp
+++ b/include/aidge/backend/cpu/operator/AddImpl.hpp
@@ -25,7 +25,19 @@
namespace Aidge {
// Operator implementation entry point for the backend
using AddImpl_cpu = OperatorImpl_cpu<Add_Op,
- void(std::vector<std::size_t>, std::vector<std::size_t>, const std::vector<std::size_t>&, const void*, const void*, void*)>;
+ void(std::vector<std::size_t>, std::vector<std::size_t>, const std::vector<std::size_t>&, const void*, const void*, void*),
+ void(const std::size_t,
+ const std::size_t,
+ const std::size_t,
+ const std::vector<std::size_t>&,
+ const std::vector<std::size_t>&,
+ const std::vector<std::size_t>&,
+ const void*,
+ const void*,
+ const void*,
+ void*,
+ void*)
+>;
// Implementation entry point registration to Operator
REGISTRAR(Add_Op, "cpu", Aidge::AddImpl_cpu::create);
diff --git a/include/aidge/backend/cpu/operator/AddImpl_kernels.hpp b/include/aidge/backend/cpu/operator/AddImpl_kernels.hpp
index e6d13fcf3699824a8410015d35ff766adf617c11..d6fff9b58d381895350998e11bae01684718ad3e 100644
--- a/include/aidge/backend/cpu/operator/AddImpl_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/AddImpl_kernels.hpp
@@ -147,25 +147,75 @@ void AddImpl_cpu_forward_kernel(std::vector<std::size_t> dims0,
}
}
+template <class I, class O>
+void AddImpl_cpu_backward_kernel(const std::size_t input0Length,
+ const std::size_t input1Length,
+ const std::size_t gradOutputLength,
+ const std::vector<std::size_t>& dims0,
+ const std::vector<std::size_t>& dims1,
+ const std::vector<std::size_t>& outputDims,
+ const void* input0_,
+ const void* input1_,
+ const void* grad_output_,
+ void* gradientInput0_,
+ void* gradientInput1_)
+{
+ // TODO: Remove input0/1 from the function
+ const I* input0 = static_cast<const I*>(input0_);
+ const I* input1 = static_cast<const I*>(input1_);
+ const O* gradOutput = static_cast<const O*>(grad_output_);
+ auto* gradInput0 = static_cast<I*>(gradientInput0_);
+ auto* gradInput1 = static_cast<I*>(gradientInput1_);
+
+ std::fill_n(gradInput0, input0Length, static_cast<I>(0));
+ std::fill_n(gradInput1, input1Length, static_cast<I>(0));
+
+ auto broadcastedDims0 = getBroadcastedDims(outputDims, dims0);
+ auto broadcastedDims1 = getBroadcastedDims(outputDims, dims1);
+
+ for (std::size_t i = 0; i < gradOutputLength; ++i) {
+ auto idxOutputGrad = getMultiDimIndices(outputDims, i);
+ std::vector<std::size_t> idxInput0(broadcastedDims0.size());
+ std::vector<std::size_t> idxInput1(broadcastedDims1.size());
+
+ for (std::size_t dimension = 0; dimension < broadcastedDims0.size(); ++dimension) {
+ idxInput0[dimension] = (broadcastedDims0[dimension] == 1) ? 0 : idxOutputGrad[dimension];
+ }
+
+ for (std::size_t dimension = 0; dimension < broadcastedDims1.size(); ++dimension) {
+ idxInput1[dimension] = (broadcastedDims1[dimension] == 1) ? 0 : idxOutputGrad[dimension];
+ }
+
+ auto idx0 = getFlattenedIndex(broadcastedDims0, idxInput0);
+ auto idx1 = getFlattenedIndex(broadcastedDims1, idxInput1);
+
+ // For addition: gradient of both inputs is just the output gradient
+ // (unlike multiplication where we need to multiply by the other input,
+ // or subtraction where we need to negate one of them)
+ gradInput0[idx0] += static_cast<I>(gradOutput[i]);
+ gradInput1[idx1] += static_cast<I>(gradOutput[i]);
+ }
+}
+
// Kernels registration to implementation entry point
REGISTRAR(AddImpl_cpu,
{ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Float32}},
- {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<float, float>, nullptr});
+ {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<float, float>, Aidge::AddImpl_cpu_backward_kernel<float, float>});
REGISTRAR(AddImpl_cpu,
{ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Float64}},
- {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<double, double>, nullptr});
+ {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<double, double>, Aidge::AddImpl_cpu_backward_kernel<double, double>});
REGISTRAR(AddImpl_cpu,
{ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Int8}},
- {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::int8_t, std::int8_t>, nullptr});
+ {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::int8_t, std::int8_t>, Aidge::AddImpl_cpu_backward_kernel<std::int8_t, std::int8_t>});
REGISTRAR(AddImpl_cpu,
{ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::UInt8}},
- {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::uint8_t, std::uint8_t>, nullptr});
+ {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::uint8_t, std::uint8_t>, Aidge::AddImpl_cpu_backward_kernel<std::uint8_t, std::uint8_t>});
REGISTRAR(AddImpl_cpu,
{ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Int32}},
- {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::int32_t, std::int32_t>, nullptr});
+ {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::int32_t, std::int32_t>, Aidge::AddImpl_cpu_backward_kernel<std::int32_t, std::int32_t>});
REGISTRAR(AddImpl_cpu,
{ImplSpec::IOSpec{DataType::Any}, ImplSpec::IOSpec{DataType::Int64}},
- {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::int64_t, std::int64_t>, nullptr});
+ {ProdConso::inPlaceModel, Aidge::AddImpl_cpu_forward_kernel<std::int64_t, std::int64_t>, Aidge::AddImpl_cpu_backward_kernel<std::int64_t, std::int64_t>});
} // namespace Aidge
-#endif /* AIDGE_CPU_OPERATOR_ADDIMPL_CPU_KERNELS_H_ */
\ No newline at end of file
+#endif /* AIDGE_CPU_OPERATOR_ADDIMPL_CPU_KERNELS_H_ */
diff --git a/src/operator/AddImpl.cpp b/src/operator/AddImpl.cpp
index 101743eccb606c998a38f49dd9b89f5ec279bcae..b027fb876c8e597c85d13fea0f1fb6dd1207a1d9 100644
--- a/src/operator/AddImpl.cpp
+++ b/src/operator/AddImpl.cpp
@@ -55,5 +55,28 @@ void Aidge::AddImpl_cpu::forward() {
template <>
void Aidge::AddImpl_cpu::backward() {
- AIDGE_THROW_OR_ABORT(std::runtime_error, "Backward not yet implemented for Add_Op on backend cpu");
+ const Add_Op& op_ = dynamic_cast<const Add_Op&>(mOp);
+
+ auto in0 = op_.getInput(0);
+ auto in1 = op_.getInput(1);
+ auto in0grad = op_.getInput(0)->grad();
+ auto in1grad = op_.getInput(1)->grad();
+ auto out0grad = op_.getOutput(0)->grad();
+
+ // Find the correct kernel type
+ const auto impl = Registrar<AddImpl_cpu>::create(getBestMatch(getRequiredSpec()));
+
+ // Call kernel
+ impl.backward(in0grad->size(),
+ in1grad->size(),
+ out0grad->size(),
+ in0->dims(),
+ in1->dims(),
+ out0grad->dims(),
+ getCPUPtr(in0),
+ getCPUPtr(in1),
+ getCPUPtr(out0grad),
+ getCPUPtr(in0grad),
+ getCPUPtr(in1grad));
+
}
diff --git a/unit_tests/operator/Test_AddImpl.cpp b/unit_tests/operator/Test_AddImpl.cpp
index bff9629be152163b2aa92bdc9d0c3029d7987b9b..4538b322a312457cf51e75a274724c5ab536f5ab 100644
--- a/unit_tests/operator/Test_AddImpl.cpp
+++ b/unit_tests/operator/Test_AddImpl.cpp
@@ -10,6 +10,7 @@
********************************************************************************/
#include <memory>
+#include <random>
#include <catch2/catch_test_macros.hpp>
@@ -19,6 +20,7 @@
#include "aidge/graph/Node.hpp"
#include "aidge/operator/Add.hpp"
#include "aidge/utils/ArrayHelpers.hpp"
+#include "aidge/utils/TensorUtils.hpp"
using namespace Aidge;
@@ -139,4 +141,275 @@ TEST_CASE("[cpu/operator] Add(forward)", "[Add][CPU]") {
Log::info("Expected Add_1 Tensor:\n{}", expectedOutput);
REQUIRE(*op_1->getOutput(0) == expectedOutput);
}
-}
\ No newline at end of file
+}
+
+TEST_CASE("[cpu/operator] Add(backward)", "[Add][CPU]") {
+ std::shared_ptr<Add_Op> op = std::make_shared<Add_Op>();
+ op->setDataType(DataType::Float32);
+ op->setBackend("cpu");
+
+ // NOTE: The first four tests use fixed values, the last one uses random values but static dimensions.
+
+ SECTION("Case 1: 1D and 2D Tensors") {
+ const auto T0 = std::make_shared<Tensor>(
+ Array2D<cpptype_t<DataType::Float32>, 2, 3>({{{1, 2, 3}, {4, 5, 6}}}));
+
+ const auto T1 =
+ std::make_shared<Tensor>(Array1D<cpptype_t<DataType::Float32>, 3>({0.1, 0.2, 0.3}));
+
+ op->associateInput(0, T0);
+ op->associateInput(1, T1);
+ op->getOutput(0)->setGrad(std::make_shared<Tensor>(
+ Array2D<float, 2, 3>({{{1.0, 1.0, 1.0}, {1.0, 1.0, 1.0}}})));
+ op->forwardDims();
+
+ op->backward();
+
+ const Tensor expectedGrad0 =
+ Array2D<cpptype_t<DataType::Float32>, 2, 3>({{{1, 1, 1}, {1, 1, 1}}});
+
+ const Tensor expectedGrad1 = Array1D<cpptype_t<DataType::Float32>, 3>({2, 2, 2});
+
+
+ REQUIRE(approxEq<cpptype_t<DataType::Float32>>(*(op->getInput(0)->grad()), expectedGrad0));
+ REQUIRE(approxEq<cpptype_t<DataType::Float32>>(*(op->getInput(1)->grad()), expectedGrad1));
+ }
+
+ SECTION("Case 2: 3D and 1D tensors") {
+ const auto T0 = std::make_shared<Tensor>(Array3D<float, 2, 2, 3>(
+ {{{{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}},
+ {{7.0, 8.0, 9.0}, {10.0, 11.0, 12.0}}}}));
+
+ const auto T1 =
+ std::make_shared<Tensor>(Array1D<float, 3>({0.3, 0.2, 0.1}));
+
+ const auto newGrad = std::make_shared<Tensor>(Array3D<float, 2, 2, 3>(
+ {{{{1, 1, 1}, {1, 1, 1}}, {{1, 1, 1}, {1, 1, 1}}}}));
+
+ const Tensor expectedGrad0 =
+ Array3D<float, 2, 2, 3>({{{{1, 1, 1}, {1, 1, 1}},
+ {{1, 1, 1}, {1, 1, 1}}}});
+
+ const Tensor expectedGrad1 = Array1D<cpptype_t<DataType::Float32>, 3>({4, 4, 4});
+
+ op->associateInput(0, T0);
+ op->associateInput(1, T1);
+ op->getOutput(0)->setGrad(newGrad);
+ op->forwardDims();
+ op->backward();
+
+ REQUIRE(approxEq<cpptype_t<DataType::Float32>>(*(op->getInput(0)->grad()), expectedGrad0));
+ REQUIRE(approxEq<cpptype_t<DataType::Float32>>(*(op->getInput(1)->grad()), expectedGrad1));
+ }
+
+ SECTION("Case 3: 4D and 2D tensors") {
+ const auto T0 = std::make_shared<Tensor>(Array4D<cpptype_t<DataType::Float32>, 2, 2, 3, 3>(
+ {{{{{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}, {7.0, 8.0, 9.0}},
+ {{10.0, 11.0, 12.0}, {13.0, 14.0, 15.0}, {16.0, 17.0, 18.0}}},
+ {{{19.0, 20.0, 21.0}, {22.0, 23.0, 24.0}, {25.0, 26.0, 27.0}},
+ {{28.0, 29.0, 30.0},
+ {31.0, 32.0, 33.0},
+ {34.0, 35.0, 36.0}}}}}));
+
+ const auto T1 = std::make_shared<Tensor>(Array2D<cpptype_t<DataType::Float32>, 3, 3>(
+ {{{0.5, 0.3, 0.1}, {0.4, 0.2, 0.6}, {0.7, 0.8, 0.9}}}));
+
+ const auto newGrad =
+ std::make_shared<Tensor>(Array4D<cpptype_t<DataType::Float32>, 2, 2, 3, 3>(
+ {{{{{1.0, 1.0, 1.0}, {1.0, 1.0, 1.0}, {1.0, 1.0, 1.0}},
+ {{1.0, 1.0, 1.0}, {1.0, 1.0, 1.0}, {1.0, 1.0, 1.0}}},
+ {{{1.0, 1.0, 1.0}, {1.0, 1.0, 1.0}, {1.0, 1.0, 1.0}},
+ {{1.0, 1.0, 1.0}, {1.0, 1.0, 1.0}, {1.0, 1.0, 1.0}}}}}));
+
+ const Tensor expectedGrad0 =
+ Array4D<cpptype_t<DataType::Float32>, 2, 2, 3, 3>(
+ {{{{{1, 1, 1}, {1, 1, 1}, {1, 1, 1}},
+ {{1, 1, 1}, {1, 1, 1}, {1, 1, 1}}},
+ {{{1, 1, 1}, {1, 1, 1}, {1, 1, 1}},
+ {{1, 1, 1}, {1, 1, 1}, {1, 1, 1}}}}});
+
+ const Tensor expectedGrad1 =
+ Array2D<cpptype_t<DataType::Float32>, 3, 3>({{
+ {4.0, 4.0, 4.0},
+ {4.0, 4.0, 4.0},
+ {4.0, 4.0, 4.0}}});
+
+ op->associateInput(0, T0);
+ op->associateInput(1, T1);
+ op->getOutput(0)->setGrad(newGrad);
+ op->forwardDims();
+
+ op->backward();
+
+ REQUIRE(approxEq<cpptype_t<DataType::Float32>>(*(op->getInput(0)->grad()), expectedGrad0));
+ REQUIRE(approxEq<cpptype_t<DataType::Float32>>(*(op->getInput(1)->grad()), expectedGrad1));
+ }
+
+ SECTION("Case 4: 3D and 2D tensors") {
+ const auto T0 = std::make_shared<Tensor>(
+ Array3D<float, 2, 3, 4>({{{
+ {1.0, 2.0, 3.0, 4.0},
+ {5.0, 6.0, 7.0, 8.0},
+ {9.0, 10.0, 11.0, 12.0},
+ },
+ {
+ {13.0, 14.0, 15.0, 16.0},
+ {17.0, 18.0, 19.0, 20.0},
+ {21.0, 22.0, 23.0, 24.0},
+ }}}));
+
+ const auto T1 = std::make_shared<Tensor>(
+ Array2D<cpptype_t<DataType::Float32>, 3, 4>({{{0.1, 0.2, 0.3, 0.4},
+ {0.5, 0.6, 0.7, 0.8},
+ {0.9, 1.0, 1.1, 1.2}}}));
+
+ const auto newGrad = std::make_shared<Tensor>(
+ Array3D<cpptype_t<DataType::Float32>, 2, 3, 4>({{{
+ {1.0, 1.0, 1.0, 1.0},
+ {1.0, 1.0, 1.0, 1.0},
+ {1.0, 1.0, 1.0, 1.0},
+ },
+ {
+ {1.0, 1.0, 1.0, 1.0},
+ {1.0, 1.0, 1.0, 1.0},
+ {1.0, 1.0, 1.0, 1.0},
+ }}}));
+
+ const Tensor expectedGrad0 =
+ Array3D<cpptype_t<DataType::Float32>, 2, 3, 4>({{{{1, 1, 1, 1},
+ {1, 1, 1, 1},
+ {1, 1, 1, 1}},
+ {{1, 1, 1, 1},
+ {1, 1, 1, 1},
+ {1, 1, 1, 1}}}});
+
+ const Tensor expectedGrad1 =
+ Array2D<cpptype_t<DataType::Float32>, 3, 4>({{{2.0, 2.0, 2.0, 2.0},
+ {2.0, 2.0, 2.0, 2.0},
+ {2.0, 2.0, 2.0, 2.0}}});
+
+ op->associateInput(0, T0);
+ op->associateInput(1, T1);
+ op->getOutput(0)->setGrad(newGrad);
+ op->forwardDims();
+
+ op->backward();
+
+ REQUIRE(approxEq<cpptype_t<DataType::Float32>>(*(op->getInput(0)->grad()), expectedGrad0));
+ REQUIRE(approxEq<cpptype_t<DataType::Float32>>(*(op->getInput(1)->grad()), expectedGrad1));
+ }
+
+ SECTION("Case 5: Tensors with random values") {
+
+ // Use random values
+ const std::vector<std::size_t> dims0 = {5, 2, 1, 7}; // First tensor
+ const std::vector<std::size_t> dims1 = {2, 6, 7}; // Second tensor
+ const std::vector<std::size_t> outputDims = {5, 2, 6, 7};
+
+ std::random_device rd;
+ std::mt19937 gen(rd());
+ std::uniform_real_distribution<float> dist(0.1f, 1.0f);
+
+ auto T0 = std::make_shared<Tensor>(dims0);
+ T0->setDataType(DataType::Float32);
+ T0->setBackend("cpu");
+ float* input0Data = static_cast<float*>(T0->getImpl()->rawPtr());
+ // Fill with random values
+ for (std::size_t i = 0; i < T0->size(); ++i) {
+ input0Data[i] = dist(gen);
+ }
+
+ auto T1 = std::make_shared<Tensor>(dims1);
+ T1->setDataType(DataType::Float32);
+ T1->setBackend("cpu");
+ float* input1Data = static_cast<float*>(T1->getImpl()->rawPtr());
+ // Fill with random values
+ for (std::size_t i = 0; i < T1->size(); ++i) {
+ input1Data[i] = dist(gen);
+ }
+
+ op->associateInput(0, T0);
+ op->associateInput(1, T1);
+
+ op->forwardDims();
+ op->forward();
+
+ Tensor expectedOutput{outputDims};
+ expectedOutput.setBackend("cpu");
+ float* expectedOutputData = static_cast<float*>(expectedOutput.getImpl()->rawPtr());
+
+ for (std::size_t n = 0; n < 5; ++n) {
+ for (std::size_t c = 0; c < 2; ++c) {
+ for (std::size_t h = 0; h < 6; ++h) {
+ for (std::size_t w = 0; w < 7; ++w) {
+ std::size_t outIdx = w + 7 * (h + 6 * (c + 2 * n));
+ std::size_t in0Idx =
+ w + 7 * (0 + 1 * (c + 2 * n)); // middle dim is 1
+ std::size_t in1Idx =
+ w + 7 * (h + 6 * c); // no n dimension
+
+ expectedOutputData[outIdx] = input0Data[in0Idx] + input1Data[in1Idx];
+ }
+ }
+ }
+ }
+
+ auto outputTensor = op->getOutput(0);
+
+ REQUIRE(approxEq<float>(*outputTensor, expectedOutput));
+
+ // Backward pass
+ std::vector<float> gradOutputData(expectedOutput.size());
+ for (auto &val : gradOutputData) {
+ val = dist(gen);
+ }
+
+ op->getOutput(0)->setGrad(std::make_shared<Tensor>());
+ op->getOutput(0)->grad()->resize(outputDims);
+ op->getOutput(0)->grad()->getImpl()->setRawPtr(gradOutputData.data(),
+ expectedOutput.size());
+
+ // Compute reference gradients
+ std::vector<float> expectedGrad0(T0->size(), 0.0f);
+ std::vector<float> expectedGrad1(T1->size(), 0.0f);
+
+ for (std::size_t n = 0; n < 5; ++n) {
+ for (std::size_t c = 0; c < 2; ++c) {
+ for (std::size_t h = 0; h < 6; ++h) {
+ for (std::size_t w = 0; w < 7; ++w) {
+ std::size_t outIdx = w + 7 * (h + 6 * (c + 2 * n));
+ std::size_t in0Idx = w + 7 * (0 + 1 * (c + 2 * n));
+ std::size_t in1Idx = w + 7 * (h + 6 * c);
+
+ // Gradient for input0: just accumulate grad_output
+ expectedGrad0[in0Idx] += gradOutputData[outIdx];
+
+ // Gradient for input1: just accumulate grad_output
+ expectedGrad1[in1Idx] += gradOutputData[outIdx];
+ }
+ }
+ }
+ }
+
+ // Perform backward pass
+ op->backward();
+
+ auto expectedGrad0Tensor = std::make_shared<Tensor>();
+ expectedGrad0Tensor->resize(T0->dims());
+ expectedGrad0Tensor->setBackend("cpu");
+ expectedGrad0Tensor->setDataType(DataType::Float32);
+ expectedGrad0Tensor->getImpl()->setRawPtr(expectedGrad0.data(),
+ expectedGrad0.size());
+
+ auto expectedGrad1Tensor = std::make_shared<Tensor>(T1->dims());
+ expectedGrad1Tensor->setBackend("cpu");
+ expectedGrad1Tensor->setDataType(DataType::Float32);
+ expectedGrad1Tensor->getImpl()->setRawPtr(expectedGrad1.data(),
+ expectedGrad1.size());
+
+ // Verify backward pass
+ REQUIRE(approxEq<float>(*T0->grad(), *expectedGrad0Tensor));
+ REQUIRE(approxEq<float>(*T1->grad(), *expectedGrad1Tensor));
+ }
+}
+