Migrate Dense layer

include/cudanet.cuh

@@ -8,15 +8,15 @@
 #include "pooling.cuh"
 
 // Layers
-#include "activation.cuh"
+#include "activation.hpp"
 #include "add.cuh"
 #include "avg_pooling.cuh"
 #include "batch_norm.cuh"
 #include "concat.cuh"
 #include "conv2d.cuh"
-#include "dense.cuh"
+#include "dense.hpp"
 #include "input.cuh"
-#include "layer.cuh"
+#include "layer.hpp"
 #include "max_pooling.cuh"
 #include "output.cuh"
 

include/kernels/convolution.cuh

@@ -2,7 +2,7 @@
 #define CUDANET_CONVOLUTION_H
 
 #include <cuda_runtime.h>
-#include "layer.cuh"
+#include "layer.hpp"
 
 namespace CUDANet::Kernels {
 

include/kernels/pooling.cuh

@@ -2,7 +2,7 @@
 #define CUDANET_POOLING_H
 
 #include <cuda_runtime.h>
-#include "layer.cuh"
+#include "layer.hpp"
 
 namespace CUDANet::Kernels {
 

include/layers/activation.hpp

@@ -48,6 +48,8 @@ class Activation {
     ActivationType activationType;
     int length;
 
+    void activateCPU(float* input);
+
 #ifdef USE_CUDA
     int gridSize;
 
@@ -58,10 +60,7 @@ class Activation {
 
     void initCUDA();
     void delCUDA();
-#else
-    void activateCPU(float* input);
 #endif
-
 };
 
 

include/layers/avg_pooling.cuh

@@ -1,8 +1,8 @@
 #ifndef CUDANET_AVG_POOLING_H
 #define CUDANET_AVG_POOLING_H
 
-#include "activation.cuh"
-#include "layer.cuh"
+#include "activation.hpp"
+#include "layer.hpp"
 
 namespace CUDANet::Layers {
 

include/layers/batch_norm.cuh

@@ -3,8 +3,8 @@
 
 #include <vector>
 
-#include "activation.cuh"
-#include "layer.cuh"
+#include "activation.hpp"
+#include "layer.hpp"
 
 namespace CUDANet::Layers {
 

include/layers/conv2d.cuh

@@ -3,9 +3,9 @@
 
 #include <vector>
 
-#include "activation.cuh"
+#include "activation.hpp"
 #include "convolution.cuh"
-#include "layer.cuh"
+#include "layer.hpp"
 
 namespace CUDANet::Layers {
 

include/layers/dense.hpp

@@ -3,8 +3,8 @@
 
 #include <vector>
 
-#include "activation.cuh"
-#include "layer.cuh"
+#include "activation.hpp"
+#include "layer.hpp"
 
 namespace CUDANet::Layers {
 
@@ -84,20 +84,11 @@ class Dense : public WeightedLayer {
     int inputSize;
     int outputSize;
 
-    float* d_output;
-
-    float* d_weights;
-    float* d_biases;
-
     std::vector<float> weights;
     std::vector<float> biases;
 
     Layers::Activation* activation;
 
-    // Precompute kernel launch parameters
-    int forwardGridSize;
-    int biasGridSize;
-
     /**
      * @brief Initialize the weights to zeros
      *
@@ -110,11 +101,30 @@ class Dense : public WeightedLayer {
      */
     void initializeBiases();
 
+    float* forwardCPU(const float* input);
+
+#ifdef USE_CUDA
+    float* d_output;
+
+    float* d_weights;
+    float* d_biases;
+
+    // Precompute kernel launch parameters
+    int forwardGridSize;
+    int biasGridSize;
+
     /**
      * @brief Copy the weights and biases to the device
      *
      */
     void toCuda();
+
+    void initCUDA();
+    void delCUDA();
+
+    float* forwardCUDA(const float* d_input);
+#endif
+
 };
 
 }  // namespace CUDANet::Layers

include/layers/input.cuh

@@ -1,7 +1,7 @@
 #ifndef CUDANET_INPUT_LAYER_H
 #define CUDANET_INPUT_LAYER_H
 
-#include "layer.cuh"
+#include "layer.hpp"
 
 namespace CUDANet::Layers {
 

include/layers/layer.hpp

@@ -111,10 +111,12 @@ class WeightedLayer : public SequentialLayer {
      */
     virtual void initializeBiases() = 0;
 
+#ifdef USE_CUDA
     /**
      * @brief Copy the weights and biases to the device
      */
     virtual void toCuda() = 0;
+#endif
 };
 
 }  // namespace CUDANet::Layers

include/layers/max_pooling.cuh

@@ -1,8 +1,8 @@
 #ifndef CUDANET_MAX_POOLING_H
 #define CUDANET_MAX_POOLING_H
 
-#include "activation.cuh"
-#include "layer.cuh"
+#include "activation.hpp"
+#include "layer.hpp"
 
 namespace CUDANet::Layers {
 

include/layers/output.cuh

@@ -1,7 +1,7 @@
 #ifndef CUDANET_OUTPUT_LAYER_H
 #define CUDANET_OUTPUT_LAYER_H
 
-#include "layer.cuh"
+#include "layer.hpp"
 
 namespace CUDANet::Layers {
 

include/model/model.hpp

@@ -6,7 +6,7 @@
 #include <vector>
 
 #include "input.cuh"
-#include "layer.cuh"
+#include "layer.hpp"
 #include "module.hpp"
 #include "output.cuh"
 

include/model/module.hpp

@@ -5,7 +5,7 @@
 #include <unordered_map>
 #include <vector>
 
-#include "layer.cuh"
+#include "layer.hpp"
 
 namespace CUDANet {
 

src/backends/cuda/kernels/pooling.cu

@@ -1,5 +1,5 @@
 #include "cuda_helper.cuh"
-#include "layer.cuh"
+#include "layer.hpp"
 #include "pooling.cuh"
 
 using namespace CUDANet;

src/backends/cuda/layers/dense.cu

@@ -6,26 +6,14 @@
 #include <iostream>
 
 #include "vector.cuh"
-#include "activation.cuh"
+#include "activation.hpp"
 #include "cuda_helper.cuh"
-#include "dense.cuh"
+#include "dense.hpp"
 #include "matmul.cuh"
 
 using namespace CUDANet::Layers;
 
-Dense::Dense(
-    int                inputSize,
-    int                outputSize,
-    ActivationType activationType
-)
-    : inputSize(inputSize), outputSize(outputSize) {
-    // Allocate memory for weights and biases
-    weights.resize(outputSize * inputSize);
-    biases.resize(outputSize);
-
-    initializeWeights();
-    initializeBiases();
-
+void Dense::initCUDA() {
     d_output = nullptr;
 
     CUDA_CHECK(cudaMalloc((void**)&d_output, sizeof(float) * outputSize));
@@ -44,27 +32,26 @@ Dense::Dense(
     forwardGridSize =
         (std::max(inputSize, outputSize) + BLOCK_SIZE - 1) / BLOCK_SIZE;
     biasGridSize = (outputSize + BLOCK_SIZE - 1) / BLOCK_SIZE;
-
-    activation = new Activation(activationType, outputSize);
 }
 
-Dense::~Dense() {
+void Dense::delCUDA() {
     cudaFree(d_output);
     cudaFree(d_weights);
     cudaFree(d_biases);
-    delete activation;
 }
 
-void Dense::initializeWeights() {
-    std::fill(weights.begin(), weights.end(), 0.0f);
+void Dense::toCuda() {
+    CUDA_CHECK(cudaMemcpy(
+        d_weights, weights.data(), sizeof(float) * inputSize * outputSize,
+        cudaMemcpyHostToDevice
+    ));
+    CUDA_CHECK(cudaMemcpy(
+        d_biases, biases.data(), sizeof(float) * outputSize,
+        cudaMemcpyHostToDevice
+    ));
 }
 
-void Dense::initializeBiases() {
-    std::fill(biases.begin(), biases.end(), 0.0f);
-}
-
-float* Dense::forward(const float* d_input) {
-
+float* Dense::forwardCUDA(const float* d_input) {
     Kernels::mat_vec_mul<<<forwardGridSize, BLOCK_SIZE>>>(
         d_weights, d_input, d_output, inputSize, outputSize
     );
@@ -80,40 +67,3 @@ float* Dense::forward(const float* d_input) {
 
     return d_output;
 }
-
-void Dense::toCuda() {
-    CUDA_CHECK(cudaMemcpy(
-        d_weights, weights.data(), sizeof(float) * inputSize * outputSize,
-        cudaMemcpyHostToDevice
-    ));
-    CUDA_CHECK(cudaMemcpy(
-        d_biases, biases.data(), sizeof(float) * outputSize,
-        cudaMemcpyHostToDevice
-    ));
-}
-
-void Dense::setWeights(const float* weights_input) {
-    std::copy(weights_input, weights_input + weights.size(), weights.begin());
-    toCuda();
-}
-
-std::vector<float> Dense::getWeights() {
-    return weights;
-}
-
-void Dense::setBiases(const float* biases_input) {
-    std::copy(biases_input, biases_input + biases.size(), biases.begin());
-    toCuda();
-}
-
-std::vector<float> Dense::getBiases() {
-    return biases;
-}
-
-int Dense::getOutputSize() {
-    return outputSize;
-}
-
-int Dense::getInputSize() {
-    return inputSize;
-}

src/layers/batch_norm.cu

@@ -1,9 +1,9 @@
 #include <vector>
 
-#include "activation.cuh"
+#include "activation.hpp"
 #include "batch_norm.cuh"
 #include "cuda_helper.cuh"
-#include "layer.cuh"
+#include "layer.hpp"
 #include "matmul.cuh"
 #include "vector.cuh"
 

src/layers/conv2d.cu

@@ -1,11 +1,11 @@
 #include <iostream>
 #include <vector>
 
-#include "activation.cuh"
+#include "activation.hpp"
 #include "conv2d.cuh"
 #include "convolution.cuh"
 #include "cuda_helper.cuh"
-#include "layer.cuh"
+#include "layer.hpp"
 #include "matmul.cuh"
 #include "vector.cuh"
 

src/layers/dense.cpp

@@ -0,0 +1,80 @@
+#include "dense.hpp"
+
+#include <stdexcept>
+
+#include "activation.hpp"
+
+using namespace CUDANet::Layers;
+
+Dense::Dense(int inputSize, int outputSize, ActivationType activationType)
+    : inputSize(inputSize), outputSize(outputSize) {
+    // Allocate memory for weights and biases
+    weights.resize(outputSize * inputSize);
+    biases.resize(outputSize);
+
+    initializeWeights();
+    initializeBiases();
+
+    activation = new Activation(activationType, outputSize);
+
+#ifdef USE_CUDA
+    initCUDA();
+#endif
+}
+
+Dense::~Dense() {
+    delete activation;
+#ifdef USE_CUDA
+    delCUDA();
+#endif
+}
+
+void Dense::initializeWeights() {
+    std::fill(weights.begin(), weights.end(), 0.0f);
+}
+
+void Dense::initializeBiases() {
+    std::fill(biases.begin(), biases.end(), 0.0f);
+}
+
+float* Dense::forwardCPU(const float* input) {
+    throw std::logic_error("Not implemented");
+}
+
+float* Dense::forward(const float* input) {
+#ifdef USE_CUDA
+    return forwardCUDA(input);
+#else
+    return forwardCPU(input);
+#endif
+}
+
+void Dense::setWeights(const float* weights_input) {
+    std::copy(weights_input, weights_input + weights.size(), weights.begin());
+#ifdef USE_CUDA
+    toCuda();
+#endif
+}
+
+std::vector<float> Dense::getWeights() {
+    return weights;
+}
+
+void Dense::setBiases(const float* biases_input) {
+    std::copy(biases_input, biases_input + biases.size(), biases.begin());
+#ifdef USE_CUDA
+    toCuda();
+#endif
+}
+
+std::vector<float> Dense::getBiases() {
+    return biases;
+}
+
+int Dense::getOutputSize() {
+    return outputSize;
+}
+
+int Dense::getInputSize() {
+    return inputSize;
+}

src/model/model.cpp

@@ -8,7 +8,7 @@
 #include <vector>
 
 #include "input.cuh"
-#include "layer.cuh"
+#include "layer.hpp"
 #include "batch_norm.cuh"
 
 using namespace CUDANet;

test/cuda/layers/test_activation.cu

@@ -1,4 +1,4 @@
-#include "activation.cuh"
+#include "activation.hpp"
 #include <gtest/gtest.h>
 #include <cuda_runtime.h>
 #include <vector>

test/cuda/layers/test_batch_norm.cu

@@ -3,7 +3,7 @@
 
 #include <vector>
 
-#include "activation.cuh"
+#include "activation.hpp"
 #include "batch_norm.cuh"
 
 class BatchNormLayerTest : public ::testing::Test {

test/cuda/layers/test_dense.cu

@@ -3,8 +3,8 @@
 
 #include <iostream>
 
-#include "activation.cuh"
-#include "dense.cuh"
+#include "activation.hpp"
+#include "dense.hpp"
 
 class DenseLayerTest : public ::testing::Test {
   protected:

test/model/test_model.cpp

@@ -1,7 +1,7 @@
 #include <gtest/gtest.h>
 
 #include "conv2d.cuh"
-#include "dense.cuh"
+#include "dense.hpp"
 #include "max_pooling.cuh"
 #include "model.hpp"