Migrate avg pooling

include/cudanet.cuh

@@ -1,16 +1,17 @@
 #ifndef CUDANET_H
 #define CUDANET_H
 
-// Kernels
+#ifdef USE_CUDA
 #include "activation_functions.cuh"
 #include "convolution.cuh"
 #include "matmul.cuh"
 #include "pooling.cuh"
+#endif
 
 // Layers
 #include "activation.hpp"
-#include "add.cuh"
+#include "add.hpp"
-#include "avg_pooling.cuh"
+#include "avg_pooling.hpp"
 #include "batch_norm.cuh"
 #include "concat.cuh"
 #include "conv2d.cuh"
@@ -25,8 +26,10 @@
 #include "module.hpp"
 
 // Utils
-#include "cuda_helper.cuh"
 #include "imagenet.hpp"
+#ifdef USE_CUDA
+#include "cuda_helper.cuh"
 #include "vector.cuh"
+#endif
 
 #endif  // CUDANET_H

include/layers/avg_pooling.hpp

@@ -18,7 +18,7 @@ class AvgPooling2d : public SequentialLayer, public TwoDLayer {
     );
     ~AvgPooling2d();
 
-    float* forward(const float* d_input);
+    float* forward(const float* input);
 
     /**
      * @brief Get output size
@@ -45,14 +45,32 @@ class AvgPooling2d : public SequentialLayer, public TwoDLayer {
 
     shape2d outputSize;
 
-    float* d_output;
-
     Activation* activation;
+
+    float* forwardCPU(const float* input);
+
+#ifdef USE_CUDA
+    float* d_output;
+    float* forwardCUDA(const float* d_input);
+
+    void initCUDA();
+    void delCUDA();
+#endif
 };
 
 class AdaptiveAvgPooling2d : public AvgPooling2d {
   public:
-    AdaptiveAvgPooling2d(shape2d inputShape, int nChannels, shape2d outputShape, ActivationType activationType);
+    AdaptiveAvgPooling2d(
+        shape2d        inputShape,
+        int            nChannels,
+        shape2d        outputShape,
+        ActivationType activationType
+    );
+
+  private:
+#ifdef USE_CUDA
+    void initCUDA();
+#endif
 };
 
 }  // namespace CUDANet::Layers

src/backends/cuda/layers/avg_pooling.cu

@@ -0,0 +1,45 @@
+#include "avg_pooling.hpp"
+#include "cuda_helper.cuh"
+#include "pooling.cuh"
+
+using namespace CUDANet::Layers;
+
+void AvgPooling2d::initCUDA() {
+    d_output = nullptr;
+    CUDA_CHECK(cudaMalloc(
+        (void**)&d_output,
+        sizeof(float) * outputSize.first * outputSize.second * nChannels
+    ));
+}
+
+void AvgPooling2d::delCUDA() {
+    cudaFree(d_output);
+}
+
+float* AvgPooling2d::forwardCUDA(const float* d_input) {
+    dim3 block(8, 8, 8);
+    dim3 grid(
+        (outputSize.first + block.x - 1) / block.x,
+        (outputSize.second + block.y - 1) / block.y,
+        (nChannels + block.z - 1) / block.z
+    );
+
+    Kernels::avg_pooling<<<grid, block>>>(
+        d_input, d_output, inputSize, outputSize, nChannels, poolingSize,
+        stride, padding
+    );
+    CUDA_CHECK(cudaGetLastError());
+
+    activation->activate(d_output);
+    CUDA_CHECK(cudaDeviceSynchronize());
+
+    return d_output;
+}
+
+void AdaptiveAvgPooling2d::initCUDA() {
+    cudaFree(d_output);
+    cudaMalloc(
+        (void**)&d_output,
+        sizeof(float) * outputSize.first * outputSize.second * nChannels
+    );
+}

src/layers/avg_pooling.cu

@@ -1,6 +1,6 @@
-#include "avg_pooling.cuh"
+#include <stdexcept>
-#include "cuda_helper.cuh"
+
-#include "pooling.cuh"
+#include "avg_pooling.hpp"
 
 using namespace CUDANet::Layers;
 
@@ -18,44 +18,40 @@ AvgPooling2d::AvgPooling2d(
       stride(stride),
       padding(padding) {
     outputSize = {
-        (inputSize.first + 2 * padding.first - poolingSize.first) / stride.first + 1,
+        (inputSize.first + 2 * padding.first - poolingSize.first) /
-        (inputSize.second + 2 * padding.second - poolingSize.second) / stride.second + 1
+                stride.first +
+            1,
+        (inputSize.second + 2 * padding.second - poolingSize.second) /
+                stride.second +
+            1
     };
 
     activation = new Activation(
         activationType, outputSize.first * outputSize.second * nChannels
     );
 
-    d_output = nullptr;
+#ifdef USE_CUDA
-    CUDA_CHECK(cudaMalloc(
+    initCUDA();
-        (void**)&d_output,
+#endif
-        sizeof(float) * outputSize.first * outputSize.second * nChannels
-    ));
 }
 
 AvgPooling2d::~AvgPooling2d() {
-    cudaFree(d_output);
+#ifdef USE_CUDA
+    delCUDA();
+#endif
     delete activation;
 }
 
-float* AvgPooling2d::forward(const float* d_input) {
+float* AvgPooling2d::forwardCPU(const float* input) {
-    dim3 block(8, 8, 8);
+    throw std::logic_error("Not implemented");
-    dim3 grid(
+}
-        (outputSize.first + block.x - 1) / block.x,
-        (outputSize.second + block.y - 1) / block.y,
-        (nChannels + block.z - 1) / block.z
-    );
 
-    Kernels::avg_pooling<<<grid, block>>>(
+float* AvgPooling2d::forward(const float* input) {
-        d_input, d_output, inputSize, outputSize, nChannels, poolingSize,
+#ifdef USE_CUDA
-        stride, padding
+    return forwardCUDA(input);
-    );
+#else
-    CUDA_CHECK(cudaGetLastError());
+    return forwardCPU(input);
-
+#endif
-    activation->activate(d_output);
-    CUDA_CHECK(cudaDeviceSynchronize());
-
-    return d_output;
 }
 
 int AvgPooling2d::getOutputSize() {
@@ -70,22 +66,36 @@ shape2d AvgPooling2d::getOutputDims() {
     return outputSize;
 }
 
-AdaptiveAvgPooling2d::AdaptiveAvgPooling2d(shape2d inputShape, int nChannels, shape2d outputShape, ActivationType activationType)
+AdaptiveAvgPooling2d::AdaptiveAvgPooling2d(
-    : AvgPooling2d(inputShape, nChannels, {1, 1}, {1, 1}, {0, 0}, activationType) {
+    shape2d        inputShape,
-    
+    int            nChannels,
-    stride = {inputShape.first / outputShape.first, inputShape.second / outputShape.second};
+    shape2d        outputShape,
+    ActivationType activationType
+)
+    : AvgPooling2d(
+          inputShape,
+          nChannels,
+          {1, 1},
+          {1, 1},
+          {0, 0},
+          activationType
+      ) {
+    stride = {
+        inputShape.first / outputShape.first,
+        inputShape.second / outputShape.second
+    };
     poolingSize = {
         inputShape.first - (outputShape.first - 1) * stride.first,
         inputShape.second - (outputShape.second - 1) * stride.second
     };
-    padding = {
+    padding    = {(poolingSize.first - 1) / 2, (poolingSize.second - 1) / 2};
-        (poolingSize.first - 1) / 2,
-        (poolingSize.second - 1) / 2
-    };
     outputSize = outputShape;
 
-    activation = new Activation(activationType, outputSize.first * outputSize.second * nChannels);
+    activation = new Activation(
+        activationType, outputSize.first * outputSize.second * nChannels
+    );
 
-    cudaFree(d_output);
+#ifdef USE_CUDA
-    cudaMalloc((void**)&d_output, sizeof(float) * outputSize.first * outputSize.second * nChannels);
+    initCUDA();
+#endif
 }

test/cuda/layers/test_avg_pooling.cu

@@ -3,7 +3,7 @@
 
 #include <vector>
 
-#include "avg_pooling.cuh"
+#include "avg_pooling.hpp"
 
 class AvgPoolingLayerTest : public ::testing::Test {
   protected: