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

Refactor Backend and Layer interfaces

Browse Source
This commit is contained in:
Matúš Námešný
2025-11-18 18:27:57 +01:00
parent 25670f90c4
commit 6340b27055
23 changed files with 154 additions and 201 deletions
Download Patch File Download Diff File Expand all files Collapse all files

29
include/backend/backend.hpp
View File

@@ -1,29 +0,0 @@
#pragma once
#include <cstddef>
#include "backend/tensor.hpp"
namespace CUDANet::Backend
{
class IBackend
{
public:
// Memory management
virtual void* allocate(size_t bytes) = 0;
virtual void deallocate(void* ptr) = 0;
// Tensor ops
virtual void print(const CUDANet::Backend::Tensor &input) = 0;
virtual void clear(CUDANet::Backend::Tensor &input) = 0;
virtual void sum(const CUDANet::Backend::Tensor &input, CUDANet::Backend::Tensor &sum) = 0;
virtual void max(const CUDANet::Backend::Tensor &input, CUDANet::Backend::Tensor &max) = 0;
// Layer ops
virtual void relu(CUDANet::Backend::Tensor &tensor) = 0;
virtual void sigmoid(CUDANet::Backend::Tensor &tensor) = 0;
virtual void softmax(CUDANet::Backend::Tensor &tensor, CUDANet::Backend::Tensor &temp_max, CUDANet::Backend::Tensor &temp_sum) = 0;
};
} // namespace CUDANet::Backend

26
include/backend/cpu.hpp Normal file
View File

@@ -0,0 +1,26 @@
#pragma once
#include "backend.hpp"
#include "tensor.hpp"
namespace CUDANet::Backend {
class CPU : public Backend {
public:
// Memory management
void* allocate(size_t bytes) override;
void deallocate(void* ptr) override;
// Tensor ops
void print(const CUDANet::Tensor &input) override;
void clear(CUDANet::Tensor &input) override;
void sum(const CUDANet::Tensor &input, CUDANet::Tensor &sum) override;
void max(const CUDANet::Tensor &input, CUDANet::Tensor &max) override;
// Layer ops
void relu(CUDANet::Tensor &tensor) override;
void sigmoid(CUDANet::Tensor &tensor) override;
void softmax(CUDANet::Tensor &tensor, CUDANet::Tensor &temp_max, CUDANet::Tensor &temp_sum) override;
};
}

23
include/backend/cuda.cuh
View File

@@ -1,29 +1,26 @@
#pragma once
#include "backend/backend.hpp"
#include "backend/tensor.hpp"
#include "backend.hpp"
#include "tensor.hpp"
namespace CUDANet::Backend {
class CUDABackend : public IBackend {
class CUDA : public Backend {
public:
// Memory management
void* allocate(size_t bytes) override;
void deallocate(void* ptr) override;
// Tensor ops
void print(const CUDANet::Backend::Tensor &input) override;
void clear(CUDANet::Backend::Tensor &input) override;
void sum(const CUDANet::Backend::Tensor &input, CUDANet::Backend::Tensor &sum) override;
void max(const CUDANet::Backend::Tensor &input, CUDANet::Backend::Tensor &max) override;
void print(const CUDANet::Tensor &input) override;
void clear(CUDANet::Tensor &input) override;
void sum(const CUDANet::Tensor &input, CUDANet::Tensor &sum) override;
void max(const CUDANet::Tensor &input, CUDANet::Tensor &max) override;
// Layer ops
void relu(CUDANet::Backend::Tensor &tensor) override;
void sigmoid(CUDANet::Backend::Tensor &tensor) override;
void softmax(CUDANet::Backend::Tensor &tensor, CUDANet::Backend::Tensor &temp_max, CUDANet::Backend::Tensor &temp_sum) override;
private:
static constexpr int BLOCK_SIZE = 256;
void relu(CUDANet::Tensor &tensor) override;
void sigmoid(CUDANet::Tensor &tensor) override;
void softmax(CUDANet::Tensor &tensor, CUDANet::Tensor &temp_max, CUDANet::Tensor &temp_sum) override;
};
} // namespace CUDANet::Backend

46
include/backend/tensor.hpp
View File

@@ -1,46 +0,0 @@
#pragma once
#include <cstddef>
#include "backend/backend.hpp"
#include <vector>
namespace CUDANet::Backend
{
enum class DType
{
FLOAT32,
// FLOAT16, // Not implemented yet
// INT32, // Not implemented yet
};
typedef std::vector<size_t> Shape;
class Tensor
{
public:
Tensor() = default;
Tensor(Shape shape, DType dtype, IBackend* backend);
~Tensor();
size_t size() const;
size_t numel() const;
template <typename T>
const T* data() const;
template <typename T>
T* data();
private:
Shape shape;
DType dtype;
size_t total_elms;
size_t total_size;
IBackend* backend;
void* d_ptr;
};
} // namespace CUDANet::Backend