Mathis/CUDANet
1
0
Fork 0
mirror of https://github.com/lordmathis/CUDANet.git synced 2025-11-05 17:34:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add padding to max pooling

Browse Source
This commit is contained in:
Matúš Námešný
2024-05-26 19:03:10 +02:00
parent 4a67b708f0
commit 94a16b4352
7 changed files with 78 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
src/kernels/pooling.cu
View File

@@ -11,7 +11,8 @@ __global__ void Kernels::max_pooling(
const dim2d outputSize,
const int nChannels,
const dim2d poolingSize,
const dim2d stride
const dim2d stride,
const dim2d padding
) {
int j = blockDim.x * blockIdx.x + threadIdx.x;
int i = blockDim.y * blockIdx.y + threadIdx.y;
@@ -25,12 +26,16 @@ __global__ void Kernels::max_pooling(
for (int k = 0; k < poolingSize.first; k++) {
for (int l = 0; l < poolingSize.second; l++) {
int inputIndex = c * inputSize.first * inputSize.second +
(i * stride.first + k) * inputSize.second +
(j * stride.second + l);
int inputRow = i * stride.first + k - padding.first;
int inputCol = j * stride.second + l - padding.second;
if (d_input[inputIndex] > max) {
max = d_input[inputIndex];
if (inputRow >= 0 && inputRow < inputSize.first && inputCol >= 0 &&
inputCol < inputSize.second) {
int inputIndex = c * inputSize.first * inputSize.second +
inputRow * inputSize.second + inputCol;
if (d_input[inputIndex] > max) {
max = d_input[inputIndex];
}
}
}
}
@@ -62,12 +67,11 @@ __global__ void Kernels::avg_pooling(
for (int k = 0; k < poolingSize.first; k++) {
for (int l = 0; l < poolingSize.second; l++) {
int inputRow = i * stride.first + k - padding.first;
int inputCol = j * stride.second + l - padding.second;
if (inputRow >= 0 && inputRow < inputSize.first &&
inputCol >= 0 && inputCol < inputSize.second) {
if (inputRow >= 0 && inputRow < inputSize.first && inputCol >= 0 &&
inputCol < inputSize.second) {
int inputIndex = c * inputSize.first * inputSize.second +
inputRow * inputSize.second + inputCol;
sum += d_input[inputIndex];

23
src/layers/max_pooling.cu
View File

@@ -9,23 +9,31 @@ MaxPooling2d::MaxPooling2d(
int nChannels,
dim2d poolingSize,
dim2d stride,
dim2d padding,
ActivationType activationType
)
: inputSize(inputSize),
nChannels(nChannels),
poolingSize(poolingSize),
stride(stride) {
stride(stride),
padding(padding) {
outputSize = {
(inputSize.first - poolingSize.first) / stride.first + 1,
(inputSize.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);
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
(void**)&d_output,
sizeof(float) * outputSize.first * outputSize.second * nChannels
));
}
@@ -43,7 +51,8 @@ float* MaxPooling2d::forward(const float* d_input) {
);
Kernels::max_pooling<<<grid, block>>>(
d_input, d_output, inputSize, outputSize, nChannels, poolingSize, stride
d_input, d_output, inputSize, outputSize, nChannels, poolingSize,
stride, padding
);
CUDA_CHECK(cudaGetLastError());