Add missing operators for basic onnx model exporting

Context

While trying to export a hugging face onnx model to cpp, some operators were missing So in this merge request, I add a reshape operator as well as adding back the matmul operator (that was previously removed because of old api use) to the export_cpp module, to be able to correctly export a basic MLP onnx model to cpp. The onnx model can be found at: https://huggingface.co/EclipseAidge/mobilenet_v1

Modified files

• Added reshape.h, reshape.cpp and matching jinja kernel to add Operator reshape to export
• operator.py, to add reshape and matmul operator as well as refactor some duplicate code
• matmul.cpp, ,matmul.h, to adapt the operator to the new api interface

Detailed major modifications

Created a reshape operator for the export (Basic memcopy of the input buffer to the output buffer)

Adapted the matmul kernel and jinja for the new operator interface (inputs_dims -> in_dims for exemple) Then added back this matmul operator in the export

In operator.py refactor all the shared code for all the composed operator:

• setupConv2D for both Conv2D and PaddedConv2
• setupElemWiseOp for all the elemwise operators (Add, Sub, Mul, ...)
• setupPooling for all the pooling operators (Max, Avg, PaddedMax, ...)

TODO

• (ref discussion !31 (comment 3330814)) Refactor the reshape operator as a non op for the c++ side (as the reshape doesn't affect the buffer, only the forward dims of the buffer and those are already calculated with the scheduler in python)

aidge_export_cpp/kernels/convolution.hpp

@@ -65,8 +65,8 @@ void convolution_forward(
                 int oOffset = NB_OUTPUTS * oPos;
 
                 // <--
-
-                Bias_T weightedSum = biases[output];
+                // Check if the biases are defined
+                Bias_T weightedSum = biases ? biases[output] : 0;
 
                 for (int sy = 0; sy < KERNEL_HEIGHT; ++sy) {
                     if ((PADDING_Y != 0
@@ -116,4 +116,45 @@ void convolution_forward(
     }
 }
 
+// Template specialization when biases are not given to the convolution
+template<int NB_CHANNELS,
+         int CHANNELS_HEIGHT, int CHANNELS_WIDTH,
+         int NB_OUTPUTS,
+         int OUTPUTS_HEIGHT, int OUTPUTS_WIDTH,
+         int PADDING_Y, int PADDING_X,
+         int STRIDE_Y, int STRIDE_X,
+         int DILATION_Y, int DILATION_X,
+         int KERNEL_HEIGHT, int KERNEL_WIDTH,
+         ActivationFunction_T ACTIVATION,
+         typename Input_T, typename Output_T,
+         typename Weight_T,
+         typename Rescaling_T>
+__attribute__((always_inline)) inline
+void convolution_forward(
+    const Input_T* __restrict inputs,
+    Output_T* __restrict outputs,
+    const Weight_T* __restrict weights,
+    std::nullptr_t __restrict,
+    const Rescaling_T& __restrict rescaling)
+{
+    const float* b = nullptr;
+
+    convolution_forward<NB_CHANNELS,
+                        CHANNELS_HEIGHT,
+                        CHANNELS_WIDTH,
+                        NB_OUTPUTS,
+                        OUTPUTS_HEIGHT,
+                        OUTPUTS_WIDTH,
+                        PADDING_Y,
+                        PADDING_X,
+                        STRIDE_Y,
+                        STRIDE_X,
+                        DILATION_Y,
+                        DILATION_X,
+                        KERNEL_HEIGHT,
+                        KERNEL_WIDTH,
+                        ACTIVATION>
+                        (inputs, outputs, weights, b, rescaling);
+}
+
 #endif  // __AIDGE_EXPORT_CPP_KERNELS_CONVOLUTION__