Add avgPool2d implementation

include/backend.hpp

@@ -62,6 +62,16 @@ class Backend {
         CUDANet::Shape padding_shape,
         CUDANet::Shape output_shape
     ) = 0;
+
+    virtual CUDANet::Tensor& avgPool2d(
+        const CUDANet::Tensor& input,
+        CUDANet::Tensor& output,
+        CUDANet::Shape input_shape,
+        CUDANet::Shape pool_shape,
+        CUDANet::Shape stride_shape,
+        CUDANet::Shape padding_shape,
+        CUDANet::Shape output_shape
+    ) = 0;
 };
 
 }  // namespace CUDANet

include/backend/cuda.cuh

@@ -49,7 +49,7 @@ class CUDA : public Backend {
         const CUDANet::Shape out_shape
     ) override;
 
-    CUDANet::Tensor& CUDA::maxPool2d(
+    CUDANet::Tensor& maxPool2d(
         const CUDANet::Tensor& input,
         CUDANet::Tensor& output,
         CUDANet::Shape input_shape,
@@ -58,6 +58,16 @@ class CUDA : public Backend {
         CUDANet::Shape padding_shape,
         CUDANet::Shape output_shape
     ) override;
+
+    CUDANet::Tensor& avgPool2d(
+        const CUDANet::Tensor& input,
+        CUDANet::Tensor& output,
+        CUDANet::Shape input_shape,
+        CUDANet::Shape pool_shape,
+        CUDANet::Shape stride_shape,
+        CUDANet::Shape padding_shape,
+        CUDANet::Shape output_shape
+    ) = 0;
 };
 
 }  // namespace CUDANet::Backend

src/backends/cuda/kernels/pool.cu

@@ -1,6 +1,6 @@
 #include "cuda_helper.cuh"
 #include "layer.hpp"
-#include "pooling.cuh"
+#include "pool.cuh"
 
 using namespace CUDANet;
 

src/backends/cuda/layer_ops.cu

@@ -137,3 +137,29 @@ CUDANet::Tensor& CUDA::maxPool2d(
 
     return output;
 }
+
+CUDANet::Tensor& CUDA::avgPool2d(
+    const CUDANet::Tensor& input,
+    CUDANet::Tensor& output,
+    CUDANet::Shape input_shape,
+    CUDANet::Shape pool_shape,
+    CUDANet::Shape stride_shape,
+    CUDANet::Shape padding_shape,
+    CUDANet::Shape output_shape
+) {
+    dim3 block(8, 8, 8);
+    dim3 grid(
+        (output_shape[0] + block.x - 1) / block.x,
+        (output_shape[1] + block.y - 1) / block.y,
+        (output_shape[2] + block.z - 1) / block.z
+    );
+
+    Kernels::avg_pool<<<grid, block>>>(
+        input.data<float>(), output.data<float>(), input_shape, output_shape, pool_shape,
+        stride_shape, padding_shape
+    );
+    CUDA_CHECK(cudaGetLastError());
+    CUDA_CHECK(cudaDeviceSynchronize());
+
+    return output;
+}

src/backends/cuda/layers/avg_pooling.cu

@@ -1,45 +0,0 @@
-#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.cpp

@@ -49,23 +49,43 @@ AvgPool2d::AvgPool2d(
 
 AvgPool2d::~AvgPool2d() {}
 
-CUDANet::Tensor& AvgPool2d::forward(CUDANet::Tensor& input);
+CUDANet::Tensor& AvgPool2d::forward(CUDANet::Tensor& input) {
+    output.zero();
+    backend->avgPool2d(
+        input,
+        output,
+        in_shape,
+        pool_shape,
+        stride_shape,
+        padding_shape,
+        out_shape
+    );
+    return output;
+}
 
-CUDANet::Shape AvgPool2d::input_shape();
+CUDANet::Shape AvgPool2d::input_shape() {
+    return in_shape;
+}
 
-CUDANet::Shape AvgPool2d::output_shape();
+CUDANet::Shape AvgPool2d::output_shape() {
+    return out_shape;
+}
 
-size_t AvgPool2d::input_size();
+size_t AvgPool2d::input_size() {
+    return sizeof(float) * in_shape[0] * in_shape[1] * in_shape[2];
+}
 
-size_t AvgPool2d::output_size();
+size_t AvgPool2d::output_size() {
+    return sizeof(float) * out_shape[0] * out_shape[1] * out_shape[3];
+}
 
-void AvgPool2d::set_weights(void* input);
+void AvgPool2d::set_weights(void* input) {}
 
-CUDANet::Tensor& AvgPool2d::get_weights();
+CUDANet::Tensor& AvgPool2d::get_weights() {}
 
-void AvgPool2d::set_biases(void* input);
+void AvgPool2d::set_biases(void* input) {}
 
-CUDANet::Tensor& AvgPool2d::get_biases();
+CUDANet::Tensor& AvgPool2d::get_biases() {}
 
 
 AdaptiveAvgPool2d::AdaptiveAvgPool2d(
@@ -73,15 +93,29 @@ AdaptiveAvgPool2d::AdaptiveAvgPool2d(
     CUDANet::Shape        output_shape,
     CUDANet::Backend *backend
 )
-    : AvgPool2d(input_shape, {1, 1}, {1, 1}, {0, 0}, backend) {
+    : AvgPool2d(
-    stride_shape = {
+        input_shape,
+        // pool_shape
+        {
+            input_shape[0] - (output_shape[0] - 1) * (input_shape[0] / output_shape[0]),
+            input_shape[1] - (output_shape[1] - 1) * (input_shape[1] / output_shape[1])
+        },
+        // stride_shape
+        {
             input_shape[0] / output_shape[0],
             input_shape[1] / output_shape[1]
-    };
+        },
-    pool_shape = {
+        // padding_shape
-        input_shape[0] - (output_shape[0] - 1) * stride_shape[0],
+        {
-        input_shape[1] - (output_shape[1] - 1) * stride_shape[1]
+            (input_shape[0] - (output_shape[0] - 1) * (input_shape[0] / output_shape[0]) - 1) / 2,
-    };
+            (input_shape[1] - (output_shape[1] - 1) * (input_shape[1] / output_shape[1]) - 1) / 2
-    padding_shape    = {(pool_shape[0] - 1) / 2, (pool_shape[1] - 1) / 2};
+        },
+        backend
+    ) {
     out_shape = output_shape;
+
+    output = CUDANet::Tensor(
+        Shape{out_shape[0] * out_shape[1] * out_shape[2]},
+        CUDANet::DType::FLOAT32, backend
+    );
 }