Jerome Hue · Jerome Hue · Jerome Hue
--- a/include/aidge/backend/cpu/operator/AddImpl_kernels.hpp
+++ b/include/aidge/backend/cpu/operator/AddImpl_kernels.hpp
+        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)