Implement max pooling layer

src/kernels/activation_functions.cu

@@ -3,9 +3,9 @@
 #include "activation_functions.cuh"
 #include "cuda_helper.cuh"
 
-using namespace CUDANet::Kernels;
+using namespace CUDANet;
 
-__global__ void sigmoid(
+__global__ void Kernels::sigmoid(
     const float* __restrict__ src,
     float* __restrict__ dst,
     const unsigned int len
@@ -18,7 +18,7 @@ __global__ void sigmoid(
     }
 }
 
-__global__ void relu(
+__global__ void Kernels::relu(
     const float* __restrict__ src,
     float* __restrict__ dst,
     const unsigned int len
@@ -31,7 +31,7 @@ __global__ void relu(
     }
 }
 
-__global__ void softmax_exp(
+__global__ void Kernels::softmax_exp(
     const float* __restrict__ src,
     float* __restrict__ dst,
     const unsigned int len
@@ -44,7 +44,7 @@ __global__ void softmax_exp(
     }
 }
 
-__global__ void softmax_sum(
+__global__ void Kernels::softmax_sum(
     const float* __restrict__ d_vector,
     float* __restrict__ d_output,
     const unsigned int w
@@ -66,7 +66,7 @@ __global__ void softmax_sum(
     }
 }
 
-__global__ void softmax_div(
+__global__ void Kernels::softmax_div(
     const float* __restrict__ src,
     float* __restrict__ dst,
     const float* __restrict__        sum,

src/kernels/convolution.cu

@@ -2,9 +2,9 @@
 
 #include "convolution.cuh"
 
-using namespace CUDANet::Kernels;
+using namespace CUDANet;
 
-__global__ void convolution(
+__global__ void Kernels::convolution(
     const float* __restrict__ d_input,
     const float* __restrict__ d_kernel,
     float* __restrict__ d_output,

src/kernels/matmul.cu

@@ -1,9 +1,9 @@
 #include "cuda_helper.cuh"
 #include "matmul.cuh"
 
-using namespace CUDANet::Kernels;
+using namespace CUDANet;
 
-__global__ void mat_vec_mul(
+__global__ void Kernels::mat_vec_mul(
     const float* __restrict__ d_matrix,
     const float* __restrict__ d_vector,
     float* __restrict__ d_output,
@@ -37,7 +37,7 @@ __global__ void mat_vec_mul(
     d_output[tid] = temp;
 }
 
-__global__ void vec_vec_add(
+__global__ void Kernels::vec_vec_add(
     const float* __restrict__ d_vector1,
     const float* __restrict__ d_vector2,
     float* __restrict__ d_output,

src/kernels/pooling.cu

@@ -0,0 +1,92 @@
+#include "pooling.cuh"
+
+#include "cuda_helper.cuh"
+
+using namespace CUDANet;
+
+__global__ void Kernels::max_pooling(
+    const float* __restrict__ d_input,
+    float* __restrict__ d_output,
+    const int inputSize,
+    const int nChannels,
+    const int poolingSize,
+    const int stride,
+    const int paddingSize
+) {
+    int tid = blockDim.x * blockIdx.x + threadIdx.x;
+    if (tid >= inputSize * inputSize * nChannels) {
+        return;
+    }
+
+    // Get output index
+    int c = tid / (inputSize * inputSize);
+    int i = tid % (inputSize * inputSize) / inputSize;
+    int j = tid % inputSize;
+
+    float max = 0.0f;
+
+    for (int k = 0; k < poolingSize; k++) {
+        for (int l = 0; l < poolingSize; l++) {
+
+            if (i * stride + k < paddingSize ||
+                i * stride + k >= (inputSize + paddingSize) ||
+                j * stride + l < paddingSize ||
+                j * stride + l >= (inputSize + paddingSize)) {
+                continue;
+            }
+
+
+            int inputIndex = c * inputSize * inputSize +
+                             (i * stride + k - paddingSize) * inputSize +
+                             (j * stride + l - paddingSize);
+
+            if (d_input[inputIndex] > max) {
+                max = d_input[inputIndex];
+            }
+        }
+    }
+
+    d_output[tid] = max;
+}
+
+__global__ void Kernels::avg_pooling(
+    const float* __restrict__ d_input,
+    float* __restrict__ d_output,
+    const int inputSize,
+    const int nChannels,
+    const int poolingSize,
+    const int stride,
+    const int paddingSize
+) {
+    int tid = blockDim.x * blockIdx.x + threadIdx.x;
+    if (tid >= inputSize * inputSize * nChannels) {
+        return;
+    }
+
+    // Get output index
+    int c = tid / (inputSize * inputSize);
+    int i = tid % (inputSize * inputSize) / inputSize;
+    int j = tid % inputSize;
+
+    float sum = 0.0f;
+
+    for (int k = 0; k < poolingSize; k++) {
+        for (int l = 0; l < poolingSize; l++) {
+
+            if (i * stride + k < paddingSize ||
+                i * stride + k >= (inputSize + paddingSize) ||
+                j * stride + l < paddingSize ||
+                j * stride + l >= (inputSize + paddingSize)) {
+                continue;
+            }
+
+            int inputIndex = c * inputSize * inputSize +
+                             (i * stride + k - paddingSize) * inputSize +
+                             (j * stride + l - paddingSize);
+
+            sum += d_input[inputIndex];
+        }
+    }
+
+    d_output[tid] = sum / (poolingSize * poolingSize);
+}

src/layers/conv2d.cu

@@ -39,7 +39,7 @@ Conv2d::Conv2d(
             break;
     }
 
-    activation = Layers::Activation(
+    activation = Activation(
         activationType, outputSize * outputSize * numFilters
     );
 

src/layers/max_pooling.cu

@@ -0,0 +1,59 @@
+#include "max_pooling.cuh"
+#include "cuda_helper.cuh"
+#include "pooling.cuh"
+
+using namespace CUDANet::Layers;
+
+
+MaxPooling2D::MaxPooling2D(
+        int            inputSize,
+        int            nChannels,
+        int            poolingSize,
+        int            stride,
+        Padding        padding,
+        ActivationType activationType
+    )
+    : inputSize(inputSize), nChannels(nChannels), poolingSize(poolingSize), stride(stride) {
+
+
+    switch (padding) {
+        case SAME:
+            outputSize  = inputSize;
+            paddingSize = ((stride - 1) * inputSize - stride + poolingSize) / 2;
+            break;
+
+        case VALID:
+            paddingSize = 0;
+            outputSize  = (inputSize - poolingSize) / stride + 1;
+            break;
+
+        default:
+            break;
+    }
+
+    activation = Activation(
+        activationType, outputSize * outputSize * nChannels
+    );
+
+    d_output = nullptr;
+    CUDA_CHECK(cudaMalloc(
+        (void**)&d_output, sizeof(float) * outputSize * outputSize * nChannels
+    ));
+
+    gridSize = (outputSize * outputSize * nChannels + BLOCK_SIZE - 1) / BLOCK_SIZE;
+
+}
+
+
+MaxPooling2D::~MaxPooling2D() {
+    cudaFree(d_output);
+}
+
+
+float* MaxPooling2D::forward(const float* d_input) {
+    Kernels::max_pooling<<<gridSize, BLOCK_SIZE>>>(
+        d_input, d_output, inputSize, nChannels, poolingSize, stride, paddingSize
+    );
+
+    return d_output;
+}