Mathis/CUDANet
1
0
Fork 0
mirror of https://github.com/lordmathis/CUDANet.git synced 2025-12-23 14:54:28 +00:00
Code Issues Packages Projects Releases Wiki Activity

WIP Migrate vector utils to Tesnor

Browse Source
This commit is contained in:
Matúš Námešný
2025-11-17 22:15:19 +01:00
parent 6133fb20af
commit 6744c8964f
8 changed files with 96 additions and 190 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/backends/cuda/cuda_backend.cu
View File

@@ -16,15 +16,6 @@ void CUDABackend::deallocate(void* ptr) {
CUDA_CHECK(cudaFree(ptr));
}
// void CUDABackend::copyToDevice(void* devicePtr, const void* hostPtr, size_t bytes) {
// CUDA_CHECK(cudaMemcpy(devicePtr, hostPtr, bytes, cudaMemcpyHostToDevice));
// CUDA_CHECK(cudaDeviceSynchronize());
// }
// void CUDABackend::copyToHost(void* hostPtr, const void* devicePtr, size_t bytes) {
// CUDA_CHECK(cudaMemcpy(hostPtr, devicePtr, bytes, cudaMemcpyDeviceToHost));
// CUDA_CHECK(cudaDeviceSynchronize());
// }
void CUDABackend::relu(Tensor &tensor) {
int gridSize = (tensor.numel() + BLOCK_SIZE - 1) / BLOCK_SIZE;

65
src/backends/cuda/tensor_ops.cu Normal file
View File

@@ -0,0 +1,65 @@
#include <iostream>
#include "backend/backend.hpp"
#include "backend/cuda.cuh"
#include "utils/cuda_helper.cuh"
#include "kernels/matmul.cuh"
using namespace CUDANet::Backend;
void CUDABackend::print(const CUDANet::Backend::Tensor &input) {
auto length = input.numel();
std::vector<float> h_vec(input.numel());
CUDA_CHECK(cudaMemcpy(
h_vec.data(), input.data<float>(), sizeof(float) * length, cudaMemcpyDeviceToHost
));
for (int i = 0; i < length; ++i) {
std::cout << h_vec[i] << ", ";
}
std::cout << std::endl;
}
void CUDABackend::clear(CUDANet::Backend::Tensor &input) {
CUDA_CHECK(cudaMemset(input.data<float>(), 0, sizeof(float) * input.numel()));
}
void CUDABackend::sum(const CUDANet::Backend::Tensor &input, CUDANet::Backend::Tensor &sum) {
auto length = input.numel();
const int gridSize = ( + BLOCK_SIZE - 1) / BLOCK_SIZE;
CUDANet::Kernels::sum_reduce<<<gridSize, BLOCK_SIZE>>>(
input.data<float>(), sum.data<float>(), length
);
CUDA_CHECK(cudaGetLastError());
int remaining = gridSize;
while (remaining > 1) {
int blocks_needed = (remaining + BLOCK_SIZE - 1) / BLOCK_SIZE;
CUDANet::Kernels::sum_reduce<<<blocks_needed, BLOCK_SIZE>>>(sum.data<float>(), sum.data<float>(), remaining);
CUDA_CHECK(cudaGetLastError());
remaining = blocks_needed;
}
}
void CUDABackend::max(const CUDANet::Backend::Tensor &input, CUDANet::Backend::Tensor &max) {
auto length = input.numel();
const int grid_size = (length + BLOCK_SIZE - 1) / BLOCK_SIZE;
Kernels::max_reduce<<<grid_size, BLOCK_SIZE>>>(input.data<float>(), max.data<float>(), length);
CUDA_CHECK(cudaGetLastError());
int remaining = grid_size;
while (remaining > 1) {
int blocks_needed = (remaining + BLOCK_SIZE - 1) / BLOCK_SIZE;
CUDANet::Kernels::max_reduce<<<blocks_needed, BLOCK_SIZE>>>(max.data<float>(), max.data<float>(), remaining);
CUDA_CHECK(cudaGetLastError());
remaining = blocks_needed;
}
}

107
src/backends/cuda/utils/vector.cu
View File

@@ -1,107 +0,0 @@
#include <iostream>
#include <vector>
#include "vector.cuh"
#include "matmul.cuh"
#include "cuda_helper.cuh"
using namespace CUDANet;
void Utils::print_vec(const float* d_vec, const unsigned int length) {
std::vector<float> h_vec(length);
CUDA_CHECK(cudaMemcpy(
h_vec.data(), d_vec, sizeof(float) * length, cudaMemcpyDeviceToHost
));
for (int i = 0; i < length; ++i) {
std::cout << h_vec[i] << ", ";
}
std::cout << std::endl;
}
void Utils::clear(float* d_vec, const unsigned int length) {
CUDA_CHECK(cudaMemset(d_vec, 0, sizeof(float) * length));
}
void Utils::max(const float* d_vec, float* d_max, const unsigned int length) {
const int grid_size = (length + BLOCK_SIZE - 1) / BLOCK_SIZE;
Kernels::max_reduce<<<grid_size, BLOCK_SIZE>>>(d_vec, d_max, length);
CUDA_CHECK(cudaGetLastError());
int remaining = grid_size;
while (remaining > 1) {
int blocks_needed = (remaining + BLOCK_SIZE - 1) / BLOCK_SIZE;
CUDANet::Kernels::max_reduce<<<blocks_needed, BLOCK_SIZE>>>(d_max, d_max, remaining);
CUDA_CHECK(cudaGetLastError());
remaining = blocks_needed;
}
}
void Utils::sum(const float* d_vec, float* d_sum, const unsigned int length) {
const int gridSize = (length + BLOCK_SIZE - 1) / BLOCK_SIZE;
CUDANet::Kernels::sum_reduce<<<gridSize, BLOCK_SIZE>>>(
d_vec, d_sum, length
);
CUDA_CHECK(cudaGetLastError());
int remaining = gridSize;
while (remaining > 1) {
int blocks_needed = (remaining + BLOCK_SIZE - 1) / BLOCK_SIZE;
CUDANet::Kernels::sum_reduce<<<blocks_needed, BLOCK_SIZE>>>(d_sum, d_sum, remaining);
CUDA_CHECK(cudaGetLastError());
remaining = blocks_needed;
}
}
void Utils::mean(const float* d_vec, float* d_mean, float *d_length, int length) {
Utils::sum(d_vec, d_mean, length);
const int gridSize = (length + BLOCK_SIZE - 1) / BLOCK_SIZE;
Kernels::vec_scalar_div<<<gridSize, BLOCK_SIZE>>>(
d_mean,
d_mean,
d_length,
length
);
CUDA_CHECK(cudaGetLastError());
}
void Utils::var(float* d_vec, float* d_var, float *d_length, const unsigned int length) {
const int gridSize = (length + BLOCK_SIZE - 1) / BLOCK_SIZE;
Kernels::vec_vec_mul<<<gridSize, BLOCK_SIZE>>>(
d_vec,
d_vec,
d_var,
length
);
CUDA_CHECK(cudaGetLastError());
// Sum over all differences
Utils::sum(
d_var,
d_var,
length
);
// Divide by difference sum / length -> variance
Kernels::vec_scalar_div<<<gridSize, BLOCK_SIZE>>>(
d_var,
d_var,
d_length,
length
);
CUDA_CHECK(cudaGetLastError());
}