Change forward function to return output pointer

include/layers/conv2d.cuh

@@ -26,10 +26,10 @@ class Conv2d : public ILayer {
     // Outputs
     int outputSize;
 
-    void forward(const float* d_input, float* d_output);
-    void setWeights(const float* weights_input);
-    void setBiases(const float* biases_input);
-    void host_conv(const float* input, float* output);
+    float* forward(const float* d_input);
+    void   setWeights(const float* weights_input);
+    void   setBiases(const float* biases_input);
+    void   host_conv(const float* input, float* output);
 
   private:
     // Inputs
@@ -47,6 +47,7 @@ class Conv2d : public ILayer {
     std::vector<float> biases;
 
     // Cuda
+    float* d_output;
     float* d_weights;
     float* d_biases;
     float* d_padded;

include/layers/dense.cuh

@@ -18,7 +18,7 @@ class Dense : public ILayer {
     );
     ~Dense();
 
-    void forward(const float* d_input, float* d_output);
+    float* forward(const float* d_input);
     void setWeights(const float* weights);
     void setBiases(const float* biases);
 
@@ -26,6 +26,8 @@ class Dense : public ILayer {
     int inputSize;
     int outputSize;
 
+    float* d_output;
+
     float* d_weights;
     float* d_biases;
 

include/layers/ilayer.cuh

@@ -6,24 +6,17 @@
 
 namespace Layers {
 
-enum Activation {
-    SIGMOID,
-    RELU,
-    NONE
-};
+enum Activation { SIGMOID, RELU, NONE };
 
-enum Padding {
-    SAME,
-    VALID
-};
+enum Padding { SAME, VALID };
 
 class ILayer {
   public:
     virtual ~ILayer() {}
 
-    virtual void forward(const float* input, float* output) = 0;
-    virtual void setWeights(const float* weights)           = 0;
-    virtual void setBiases(const float* biases)             = 0;
+    virtual float* forward(const float* input)      = 0;
+    virtual void   setWeights(const float* weights) = 0;
+    virtual void   setBiases(const float* biases)   = 0;
 
   private:
     virtual void initializeWeights() = 0;
@@ -34,6 +27,8 @@ class ILayer {
     int inputSize;
     int outputSize;
 
+    float* d_output;
+
     float* d_weights;
     float* d_biases;
 

include/utils/cuda_helper.cuh

@@ -2,6 +2,7 @@
 #define CUDA_HELPER_H
 
 #include <cuda_runtime.h>
+#include <cstdio>
 
 // CUDA error checking macro
 #define CUDA_CHECK(call) \