Clean-up

include/aidge/backend/arrayfire/data/TensorImpl.hpp

@@ -159,6 +159,7 @@ public:
 private:
     void lazyInit() {
         if (static_cast<NbElts_t>(mData.elements()) < mNbElts) {
+            // Arrayfire convention for dims order is reversed compared to numpy
             const auto dims = std::vector<dim_t>(mDims.rbegin(), mDims.rend());
             mData = af::array(af::dim4(dims.size(), &dims[0]), static_cast<af_dtype>(af::dtype_traits<T>::af_type));
         }

include/aidge/backend/arrayfire/operator/ConvImpl.hpp

@@ -32,7 +32,6 @@ class ConvImpl2D_arrayfire : public OperatorImpl {
     }
 
    public:
-    Elts_t getNbRequiredProtected(const IOIndex_t inputIdx) const override final;
     void forward() override;
 };
 

src/backend/arrayfire/operator/ConvImpl.cpp

@@ -21,11 +21,6 @@
 #include "aidge/operator/Conv.hpp"
 #include "aidge/utils/Types.h"
 
-Aidge::Elts_t Aidge::ConvImpl2D_arrayfire::getNbRequiredProtected(IOIndex_t /*inputIdx*/) const {
-    // this implementation can be in-place
-    return Elts_t::DataElts(0);
-}
-
 void Aidge::ConvImpl2D_arrayfire::forward() {
     const auto& op_ = dynamic_cast<const Conv_Op<2>&>(mOp);
 
@@ -43,35 +38,21 @@ void Aidge::ConvImpl2D_arrayfire::forward() {
     const auto& input2 = (op_.getInput(2)) ? op_.getInput(2)->refCastFrom(input2Fallback, *op_.getOutput(0)) : Tensor();
 
     // Call kernel
+    // Arrayfire convention for dims order is reversed compared to numpy
     const auto strideDims = std::vector<dim_t>(op_.strideDims().rbegin(), op_.strideDims().rend());
     const auto paddingDims = std::vector<dim_t>(strideDims.size(), 0);
     const auto dilationDims = std::vector<dim_t>(op_.dilationDims().rbegin(), op_.dilationDims().rend());
 
     auto& output = std::dynamic_pointer_cast<TensorImpl_arrayfire_>(op_.getOutput(0)->getImpl())->data();
-    auto outputPtr = output.get();
-
-    // std::dynamic_pointer_cast<TensorImpl_arrayfire_>(op_.getOutput(0)->getImpl())->data() = af::convolve2NN(
-    //     std::dynamic_pointer_cast<TensorImpl_arrayfire_>(input0.getImpl())->data(),
-    //     std::dynamic_pointer_cast<TensorImpl_arrayfire_>(input1.getImpl())->data(),
-    //     af::dim4(strideDims.size(), &strideDims[0]),
-    //     af::dim4(paddingDims.size(), &paddingDims[0]),
-    //     af::dim4(dilationDims.size(), &dilationDims[0]));
-
-    af::print("i", std::dynamic_pointer_cast<TensorImpl_arrayfire_>(input0.getImpl())->data());
-
-    auto err = af_convolve2_nn(
-        static_cast<void**>(&outputPtr),
-        std::dynamic_pointer_cast<TensorImpl_arrayfire_>(input0.getImpl())->data().get(),
-        af::flip(af::flip(std::dynamic_pointer_cast<TensorImpl_arrayfire_>(input1.getImpl())->data(), 0), 1).get(),
-        strideDims.size(),
-        &strideDims[0],
-        paddingDims.size(),
-        &paddingDims[0],
-        dilationDims.size(),
-        &dilationDims[0]);
-    AIDGE_ASSERT(err == AF_SUCCESS, "af_convolve2_nn: {}", af_err_to_string(err));
 
-    output.set(outputPtr);
+    output = af::convolve2NN(
+        std::dynamic_pointer_cast<TensorImpl_arrayfire_>(input0.getImpl())->data(),
+        // convolve2NN() performs a true "convolution", not a correlation as everywhere else
+        // including Aidge. The kernels must therefore be flipped!
+        af::flip(af::flip(std::dynamic_pointer_cast<TensorImpl_arrayfire_>(input1.getImpl())->data(), 0), 1),
+        af::dim4(strideDims.size(), &strideDims[0]),
+        af::dim4(paddingDims.size(), &paddingDims[0]),
+        af::dim4(dilationDims.size(), &dilationDims[0]));
 
     if (op_.getInput(2) && input2.size() > 0) {
         // ArrayFire does not support broadcasting