Mathis/CUDANet
1
0
Fork 0
mirror of https://github.com/lordmathis/CUDANet.git synced 2025-11-06 01:34:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add support for non square matrices

Browse Source
This commit is contained in:
Matúš Námešný
2024-05-20 15:20:43 +02:00
parent 6f8b5f4081
commit 74098b24e3
21 changed files with 314 additions and 299 deletions
Download Patch File Download Diff File Expand all files Collapse all files

52
src/kernels/pooling.cu
View File

@@ -1,4 +1,5 @@
#include "cuda_helper.cuh"
#include "layer.cuh"
#include "pooling.cuh"
using namespace CUDANet;
@@ -6,26 +7,27 @@ using namespace CUDANet;
__global__ void Kernels::max_pooling(
const float* __restrict__ d_input,
float* __restrict__ d_output,
const int inputSize,
const int outputSize,
const int nChannels,
const int poolingSize,
const int stride
const dim2d inputSize,
const dim2d outputSize,
const int nChannels,
const dim2d poolingSize,
const dim2d stride
) {
int j = blockDim.x * blockIdx.x + threadIdx.x;
int i = blockDim.y * blockIdx.y + threadIdx.y;
int c = blockDim.z * blockIdx.z + threadIdx.z;
if (i >= outputSize || j >= outputSize || c >= nChannels) {
if (i >= outputSize.first || j >= outputSize.second || c >= nChannels) {
return;
}
float max = 0.0f;
for (int k = 0; k < poolingSize; k++) {
for (int l = 0; l < poolingSize; l++) {
int inputIndex = c * inputSize * inputSize +
(i * stride + k) * inputSize + (j * stride + l);
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);
if (d_input[inputIndex] > max) {
max = d_input[inputIndex];
@@ -33,37 +35,41 @@ __global__ void Kernels::max_pooling(
}
}
d_output[c * outputSize * outputSize + i * outputSize + j] = max;
d_output
[c * outputSize.first * outputSize.second + i * outputSize.second + j] =
max;
}
__global__ void Kernels::avg_pooling(
const float* __restrict__ d_input,
float* __restrict__ d_output,
const int inputSize,
const int outputSize,
const int nChannels,
const int poolingSize,
const int stride
const dim2d inputSize,
const dim2d outputSize,
const int nChannels,
const dim2d poolingSize,
const dim2d stride
) {
int j = blockDim.x * blockIdx.x + threadIdx.x;
int i = blockDim.y * blockIdx.y + threadIdx.y;
int c = blockDim.z * blockIdx.z + threadIdx.z;
if (i >= outputSize || j >= outputSize || c >= outputSize) {
if (i >= outputSize.first || j >= outputSize.second || c >= nChannels) {
return;
}
float sum = 0.0f;
for (int k = 0; k < poolingSize; k++) {
for (int l = 0; l < poolingSize; l++) {
int inputIndex = c * inputSize * inputSize +
(i * stride + k) * inputSize + (j * stride + l);
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);
sum += d_input[inputIndex];
}
}
d_output[c * outputSize * outputSize + i * outputSize + j] =
sum / (poolingSize * poolingSize);
d_output
[c * outputSize.first * outputSize.second + i * outputSize.second + j] =
sum / (poolingSize.first * poolingSize.second);
}