Mathis/CUDANet
1
0
Fork 0
mirror of https://github.com/lordmathis/CUDANet.git synced 2025-12-22 14:24:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Migrate MaxPool2d layer to Tensors

Browse Source
This commit is contained in:
Matúš Námešný
2025-11-19 21:44:19 +01:00
parent 7896ff0e24
commit e4d05931d4
17 changed files with 215 additions and 454 deletions
Download Patch File Download Diff File Expand all files Collapse all files

68
src/backends/cuda/kernels/pooling.cu
View File

@@ -4,35 +4,34 @@
using namespace CUDANet;
__global__ void Kernels::max_pooling(
__global__ void Kernels::max_pool(
const float* __restrict__ d_input,
float* __restrict__ d_output,
const shape2d inputSize,
const shape2d outputSize,
const int nChannels,
const shape2d poolingSize,
const shape2d stride,
const shape2d padding
const Shape input_shape,
const Shape output_shape,
const Shape pool_shape,
const Shape stride_shape,
const Shape padding_shape
) {
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.first || j >= outputSize.second || c >= nChannels) {
if (i >= output_shape[0] || j >= output_shape[1] || c >= output_shape[2]) {
return;
}
float max = 0.0f;
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;
for (int k = 0; k < pool_shape[0]; k++) {
for (int l = 0; l < pool_shape[1]; l++) {
int inputRow = i * stride_shape[0] + k - padding_shape[0];
int inputCol = j * stride_shape[1] + l - padding_shape[1];
if (inputRow >= 0 && inputRow < inputSize.first && inputCol >= 0 &&
inputCol < inputSize.second) {
int inputIndex = c * inputSize.first * inputSize.second +
inputRow * inputSize.second + inputCol;
if (inputRow >= 0 && inputRow < input_shape[0] && inputCol >= 0 &&
inputCol < input_shape[1]) {
int inputIndex = c * input_shape[0] * input_shape[1] +
inputRow * input_shape[1] + inputCol;
if (d_input[inputIndex] > max) {
max = d_input[inputIndex];
}
@@ -41,45 +40,44 @@ __global__ void Kernels::max_pooling(
}
d_output
[c * outputSize.first * outputSize.second + i * outputSize.second + j] =
[c * output_shape[0] * output_shape[1] + i * output_shape[1] + j] =
max;
}
__global__ void Kernels::avg_pooling(
__global__ void Kernels::avg_pool(
const float* __restrict__ d_input,
float* __restrict__ d_output,
const shape2d inputSize,
const shape2d outputSize,
const int nChannels,
const shape2d poolingSize,
const shape2d stride,
const shape2d padding
const Shape input_shape,
const Shape output_shape,
const Shape pool_shape,
const Shape stride_shape,
const Shape padding_shape
) {
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.first || j >= outputSize.second || c >= nChannels) {
if (i >= output_shape[0] || j >= output_shape[1] || c >= output_shape[2]) {
return;
}
float sum = 0.0f;
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;
for (int k = 0; k < pool_shape[0]; k++) {
for (int l = 0; l < pool_shape[1]; l++) {
int inputRow = i * stride_shape[0] + k - padding_shape[0];
int inputCol = j * stride_shape[1] + l - padding_shape[1];
if (inputRow >= 0 && inputRow < inputSize.first && inputCol >= 0 &&
inputCol < inputSize.second) {
int inputIndex = c * inputSize.first * inputSize.second +
inputRow * inputSize.second + inputCol;
if (inputRow >= 0 && inputRow < input_shape[0] && inputCol >= 0 &&
inputCol < input_shape[1]) {
int inputIndex = c * input_shape[0] * input_shape[1] +
inputRow * input_shape[1] + inputCol;
sum += d_input[inputIndex];
}
}
}
d_output
[c * outputSize.first * outputSize.second + i * outputSize.second + j] =
sum / (poolingSize.first * poolingSize.second);
[c * output_shape[0] * output_shape[1] + i * output_shape[1] + j] =
sum / (pool_shape[0] * pool_shape[1]);
}

27
src/backends/cuda/layer_ops.cu
View File

@@ -2,6 +2,7 @@
#include "kernels/activation_functions.cuh"
#include "kernels/convolution.cuh"
#include "kernels/matmul.cuh"
#include "kernels/pooling.cuh"
#include "utils/cuda_helper.cuh"
using namespace CUDANet::Backend;
@@ -108,5 +109,31 @@ CUDANet::Tensor& CUDA::conv2d(
CUDA_CHECK(cudaGetLastError());
CUDA_CHECK(cudaDeviceSynchronize());
return output;
}
CUDANet::Tensor& CUDA::maxPool2d(
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::max_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;
}

22
src/backends/cuda/layers/input.cu
View File

@@ -1,22 +0,0 @@
#include "cuda_helper.cuh"
#include "input.hpp"
using namespace CUDANet::Layers;
void Input::initCUDA() {
d_output = nullptr;
CUDA_CHECK(cudaMalloc((void**)&d_output, sizeof(float) * inputSize));
}
void Input::delCUDA() {
cudaFree(d_output);
}
float* Input::forwardCUDA(const float* input) {
CUDA_CHECK(cudaMemcpy(
d_output, input, sizeof(float) * inputSize, cudaMemcpyHostToDevice
));
CUDA_CHECK(cudaDeviceSynchronize());
return d_output;
}

38
src/backends/cuda/layers/max_pooling.cu
View File

@@ -1,38 +0,0 @@
#include "cuda_helper.cuh"
#include "max_pooling.hpp"
#include "pooling.cuh"
using namespace CUDANet::Layers;
void MaxPooling2d::initCUDA() {
d_output = nullptr;
CUDA_CHECK(cudaMalloc(
(void**)&d_output,
sizeof(float) * outputSize.first * outputSize.second * nChannels
));
}
void MaxPooling2d::delCUDA() {
cudaFree(d_output);
}
float* MaxPooling2d::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::max_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;
}

14
src/backends/cuda/layers/output.cu
View File

@@ -1,14 +0,0 @@
#include "output.hpp"
#include "cuda_helper.cuh"
using namespace CUDANet::Layers;
float* Output::forwardCUDA(const float* input) {
CUDA_CHECK(cudaMemcpy(
h_output, input, sizeof(float) * inputSize, cudaMemcpyDeviceToHost
));
CUDA_CHECK(cudaDeviceSynchronize());
return h_output;
}

2
src/layers/conv2d.cpp
View File

@@ -83,7 +83,7 @@ Conv2d::Conv2d(
Conv2d::~Conv2d() {}
CUDANet::Tensor& Conv2d::forward(CUDANet::Tensor& input) {
CUDANet::Tensor& Conv2d::forward( CUDANet::Tensor& input) {
output.zero();
backend->conv2d(
weights,

37
src/layers/input.cpp
View File

@@ -1,37 +0,0 @@
#include <stdexcept>
#include "input.hpp"
using namespace CUDANet::Layers;
Input::Input(int inputSize) : inputSize(inputSize) {
#ifdef USE_CUDA
initCUDA();
#endif
}
Input::~Input() {
#ifdef USE_CUDA
delCUDA();
#endif
}
float* Input::forwardCPU(const float* input) {
throw std::logic_error("Not implemented");
}
float* Input::forward(const float* input) {
#ifdef USE_CUDA
return forwardCUDA(input);
#else
return forwardCPU(input);
#endif
}
int Input::get_output_size() {
return inputSize;
}
int Input::getInputSize() {
return inputSize;
}

70
src/layers/max_pool.cpp Normal file
View File

@@ -0,0 +1,70 @@
#include "max_pool.hpp"
#include <stdexcept>
using namespace CUDANet::Layers;
MaxPool2d::MaxPool2d(
CUDANet::Shape input_shape,
CUDANet::Shape pooling_shape,
CUDANet::Shape stride_shape,
CUDANet::Shape padding_shape,
CUDANet::Backend* backend
)
: in_shape(input_shape),
pooling_shape(pooling_shape),
stride_shape(stride_shape),
padding_shape(padding_shape),
backend(backend) {
size_t out_h = (in_shape[0] + 2 * padding_shape[0] - pooling_shape[0]) /
stride_shape[0] +
1;
size_t out_w = (in_shape[1] + 2 * padding_shape[1] - pooling_shape[1]) /
stride_shape[1] +
1;
out_shape.resize(3);
out_shape[0] = out_h;
out_shape[1] = out_w;
out_shape[2] = in_shape[2];
output = CUDANet::Tensor(
Shape{out_shape[0] * out_shape[1] * out_shape[3]},
CUDANet::DType::FLOAT32, backend
);
}
MaxPool2d::~MaxPool2d() {}
CUDANet::Tensor& MaxPool2d::forward(CUDANet::Tensor& input) {
output.zero();
backend->maxPool2d(
input, output, in_shape, pooling_shape, stride_shape, padding_shape,
out_shape
);
return output;
}
CUDANet::Shape MaxPool2d::input_shape() {
return in_shape;
}
CUDANet::Shape MaxPool2d::output_shape() {
return out_shape;
}
size_t MaxPool2d::input_size() {
return sizeof(float) * in_shape[0] * in_shape[1] * in_shape[2];
}
size_t MaxPool2d::output_size() {
return sizeof(float) * out_shape[0] * out_shape[1] * out_shape[2];
}
void MaxPool2d::set_weights(void* input) {}
CUDANet::Tensor& MaxPool2d::get_weights() {}
void MaxPool2d::set_biases(void* input) {}
CUDANet::Tensor& MaxPool2d::get_biases() {}

67
src/layers/max_pooling.cpp
View File

@@ -1,67 +0,0 @@
#include "max_pooling.hpp"
#include <stdexcept>
using namespace CUDANet::Layers;
MaxPooling2d::MaxPooling2d(
shape2d inputSize,
int nChannels,
shape2d poolingSize,
shape2d stride,
shape2d padding,
ActivationType activationType
)
: inputSize(inputSize),
nChannels(nChannels),
poolingSize(poolingSize),
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
};
activation = new Activation(
activationType, outputSize.first * outputSize.second * nChannels
);
#ifdef USE_CUDA
initCUDA();
#endif
}
MaxPooling2d::~MaxPooling2d() {
#ifdef USE_CUDA
delCUDA();
#endif
delete activation;
}
float* MaxPooling2d::forwardCPU(const float* input) {
throw std::logic_error("Not implemented");
}
float* MaxPooling2d::forward(const float* input) {
#ifdef USE_CUDA
return forwardCUDA(input);
#else
return forwardCPU(input);
#endif
}
int MaxPooling2d::get_output_size() {
return outputSize.first * outputSize.second * nChannels;
}
int MaxPooling2d::getInputSize() {
return inputSize.first * inputSize.second * nChannels;
}
shape2d MaxPooling2d::getOutputDims() {
return outputSize;
}

34
src/layers/output.cu
View File

@@ -1,34 +0,0 @@
#include "output.hpp"
#include <stdexcept>
using namespace CUDANet::Layers;
Output::Output(int inputSize) : inputSize(inputSize) {
h_output = (float*) malloc(sizeof(float) * inputSize);
}
Output::~Output() {
free(h_output);
}
float* Output::forwardCPU(const float* input) {
throw std::logic_error("Not implemented");
}
float* Output::forward(const float* input) {
#ifdef USE_CUDA
return forwardCUDA(input);
#else
return forwardCPU(input);
#endif
}
int Output::get_output_size() {
return inputSize;
}
int Output::getInputSize() {
return inputSize;
}