Migrate avg pooling

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

@@ -1,15 +1,15 @@
-#include "avg_pooling.cuh"
-#include "cuda_helper.cuh"
-#include "pooling.cuh"
+#include <stdexcept>
+
+#include "avg_pooling.hpp"
 
 using namespace CUDANet::Layers;
 
 AvgPooling2d::AvgPooling2d(
-    shape2d          inputSize,
+    shape2d        inputSize,
     int            nChannels,
-    shape2d          poolingSize,
-    shape2d          stride,
-    shape2d          padding,
+    shape2d        poolingSize,
+    shape2d        stride,
+    shape2d        padding,
     ActivationType activationType
 )
     : inputSize(inputSize),
@@ -18,44 +18,40 @@ AvgPooling2d::AvgPooling2d(
       stride(stride),
       padding(padding) {
     outputSize = {
-        (inputSize.first + 2 * padding.first - poolingSize.first) / stride.first + 1,
-        (inputSize.second + 2 * padding.second - poolingSize.second) / stride.second + 1
+        (inputSize.first + 2 * padding.first - poolingSize.first) /
+                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;
-    CUDA_CHECK(cudaMalloc(
-        (void**)&d_output,
-        sizeof(float) * outputSize.first * outputSize.second * nChannels
-    ));
+#ifdef USE_CUDA
+    initCUDA();
+#endif
 }
 
 AvgPooling2d::~AvgPooling2d() {
-    cudaFree(d_output);
+#ifdef USE_CUDA
+    delCUDA();
+#endif
     delete activation;
 }
 
-float* AvgPooling2d::forward(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
-    );
+float* AvgPooling2d::forwardCPU(const float* input) {
+    throw std::logic_error("Not implemented");
+}
 
-    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;
+float* AvgPooling2d::forward(const float* input) {
+#ifdef USE_CUDA
+    return forwardCUDA(input);
+#else
+    return forwardCPU(input);
+#endif
 }
 
 int AvgPooling2d::getOutputSize() {
@@ -70,22 +66,36 @@ shape2d AvgPooling2d::getOutputDims() {
     return outputSize;
 }
 
-AdaptiveAvgPooling2d::AdaptiveAvgPooling2d(shape2d inputShape, int nChannels, shape2d outputShape, ActivationType activationType)
-    : AvgPooling2d(inputShape, nChannels, {1, 1}, {1, 1}, {0, 0}, activationType) {
-    
-    stride = {inputShape.first / outputShape.first, inputShape.second / outputShape.second};
+AdaptiveAvgPooling2d::AdaptiveAvgPooling2d(
+    shape2d        inputShape,
+    int            nChannels,
+    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 = {
-        (poolingSize.first - 1) / 2,
-        (poolingSize.second - 1) / 2
-    };
-    outputSize = outputShape;    
+    padding    = {(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);
-    cudaMalloc((void**)&d_output, sizeof(float) * outputSize.first * outputSize.second * nChannels);
+#ifdef USE_CUDA
+    initCUDA();
+#endif
 }