WIP Migrate Activation to Tensor

include/backend/backend.hpp

@@ -13,9 +13,10 @@ public:
     virtual void* allocate(size_t bytes) = 0;
     virtual void deallocate(void* ptr) = 0;
 
-    virtual void copyToDevice(void* devicePtr, const void* hostPtr, size_t bytes) = 0;
-    virtual void copyToHost(void* hostPtr, const void* devicePtr, size_t bytes) = 0;
-
+    // Layer operations
+    virtual void relu(CUDANet::Backend::Tensor &tensor) = 0;
+    virtual void sigmoid(CUDANet::Backend::Tensor &tensor) = 0;
+    virtual void softmax(CUDANet::Backend::Tensor &tensor, CUDANet::Backend::Tensor &temp_max, CUDANet::Backend::Tensor &temp_sum) = 0;
 };
 
 } // namespace CUDANet::Backend

include/backend/cuda_backend.cuh

@@ -0,0 +1,23 @@
+#pragma once
+
+#include "backend/backend.hpp"
+#include "backend/tensor.hpp"
+
+namespace CUDANet::Backend {
+
+class CUDABackend : public IBackend {
+public:
+    // Memory management
+    void* allocate(size_t bytes) override;
+    void deallocate(void* ptr) override;
+
+    // Layer operations
+    void relu(CUDANet::Backend::Tensor &tensor) override;
+    void sigmoid(CUDANet::Backend::Tensor &tensor) override;
+    void softmax(CUDANet::Backend::Tensor &tensor, CUDANet::Backend::Tensor &temp_max, CUDANet::Backend::Tensor &temp_sum) override;
+
+private:
+    static constexpr int BLOCK_SIZE = 256;
+};
+
+}  // namespace CUDANet::Backend

include/backend/tensor.hpp

@@ -18,6 +18,8 @@ typedef std::vector<size_t> Shape;
 class Tensor
 {
 public:
+
+    Tensor() = default;
     Tensor(Shape shape, DType dtype, IBackend* backend);
     ~Tensor();
 
@@ -27,6 +29,10 @@ public:
     void toDevice(const void* hostPtr);
     void toHost(void* hostPtr);
 
+    size_t size() const;
+    size_t numel() const;
+    void* data() const;
+
 private:
     Shape       shape;
     DType       dtype;

include/layers/activation.hpp

@@ -1,5 +1,7 @@
-#ifndef CUDANET_ACTIVATION_H
-#define CUDANET_ACTIVATION_H
+#pragma once
+
+#include "backend/tensor.hpp"
+#include "backend/backend.hpp"
 
 namespace CUDANet::Layers {
 
@@ -41,29 +43,16 @@ class Activation {
      * 
      * @param d_input Pointer to the input vector on the device
      */
-    void activate(float* d_input);
+    void activate(CUDANet::Backend::Tensor input);
 
 
   private:
+    CUDANet::Backend::IBackend* backend;
     ActivationType activationType;
     int length;
 
-    void activateCPU(float* input);
-
-#ifdef USE_CUDA
-    int gridSize;
-
-    float* d_softmax_sum;
-    float* d_max;
-
-    void activateCUDA(float* d_input);
-
-    void initCUDA();
-    void delCUDA();
-#endif
+    CUDANet::Backend::Tensor softmax_sum;
+    CUDANet::Backend::Tensor tensor_max;
 };
 
-
 }  // namespace CUDANet::Layers
-
-#endif  // CUDANET_ACTIVATION_H