Mathis/CUDANet
1
0
Fork 0
mirror of https://github.com/lordmathis/CUDANet.git synced 2025-12-22 22:34:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

WIP Implement Tensor constructor and destructor

Browse Source
This commit is contained in:
Matúš Námešný
2025-11-17 22:36:29 +01:00
parent 6744c8964f
commit d231e515b1
6 changed files with 110 additions and 102 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
src/backends/cuda/layer_ops.cu Normal file
View File

@@ -0,0 +1,48 @@
#include "backend/cuda.cuh"
#include "utils/cuda_helper.cuh"
#include "kernels/activation_functions.cuh"
#include "kernels/matmul.cuh"
using namespace CUDANet::Backend;
void CUDABackend::relu(Tensor &tensor) {
int gridSize = (tensor.numel() + BLOCK_SIZE - 1) / BLOCK_SIZE;
Kernels::relu<<<gridSize, BLOCK_SIZE>>>(tensor.data<float>(), tensor.data<float>(), tensor.numel());
CUDA_CHECK(cudaGetLastError());
CUDA_CHECK(cudaDeviceSynchronize());
}
void CUDABackend::sigmoid(Tensor &tensor) {
int gridSize = (tensor.numel() + BLOCK_SIZE - 1) / BLOCK_SIZE;
Kernels::sigmoid<<<gridSize, BLOCK_SIZE>>>(tensor.data<float>(), tensor.data<float>(), tensor.numel());
CUDA_CHECK(cudaGetLastError());
CUDA_CHECK(cudaDeviceSynchronize());
}
void CUDABackend::softmax(Tensor &tensor, Tensor &temp_max, Tensor &temp_sum) {
int gridSize = (tensor.numel() + BLOCK_SIZE - 1) / BLOCK_SIZE;
// Find max value
max(tensor, temp_max);
// Subtract max value to improve numerical stability
Kernels::vec_scalar_sub<<<gridSize, BLOCK_SIZE>>>(
tensor.data<float>(), tensor.data<float>(), temp_max.data<float>(), tensor.numel()
);
CUDA_CHECK(cudaGetLastError());
// Compute exponentials
Kernels::vec_exp<<<gridSize, BLOCK_SIZE>>>(
tensor.data<float>(), tensor.data<float>(), tensor.numel()
);
CUDA_CHECK(cudaGetLastError());
// Find sum
sum(tensor, temp_sum);
Kernels::vec_scalar_div<<<gridSize, BLOCK_SIZE>>>(
tensor.data<float>(), tensor.data<float>(), temp_sum.data<float>(), tensor.numel()
);
CUDA_CHECK(cudaGetLastError());
CUDA_CHECK(cudaDeviceSynchronize());
}