Migrate concat layer

include/cudanet.cuh

@@ -13,7 +13,7 @@
 #include "add.hpp"
 #include "avg_pooling.hpp"
 #include "batch_norm.cuh"
-#include "concat.cuh"
+#include "concat.hpp"
 #include "conv2d.cuh"
 #include "dense.hpp"
 #include "input.cuh"

include/layers/concat.hpp

@@ -1,6 +1,8 @@
 #ifndef CUDANET_CONCAT_LAYER_H
 #define CUDANET_CONCAT_LAYER_H
 
+#include "layer.hpp"
+
 namespace CUDANet::Layers {
 
 /**
@@ -11,7 +13,7 @@ class Concat {
   public:
     /**
      * @brief Create a new Concat layer
-     * 
+     *
      * @param inputASize Size of the first input
      * @param inputBSize Size of the second input
      */
@@ -19,7 +21,7 @@ class Concat {
 
     /**
      * @brief Destroy the Concat layer
-     * 
+     *
      */
     ~Concat();
 
@@ -28,7 +30,7 @@ class Concat {
      *
      * @param d_input_A Device pointer to the first input
      * @param d_input_B Device pointer to the second input
-     * 
+     *
      * @return Device pointer to the output
      */
     float* forward(const float* d_input_A, const float* d_input_B);
@@ -39,7 +41,15 @@ class Concat {
     int inputASize;
     int inputBSize;
 
+    float* forwardCPU(const float* input_A, const float* input_B);
+
+#ifdef USE_CUDA
     float* d_output;
+    float* forwardCUDA(const float* d_input_A, const float* d_input_B);
+
+    void initCUDA();
+    void delCUDA();
+#endif
 };
 
 }  // namespace CUDANet::Layers