# cuda_lib_test

**Repository Path**: oxy/cuda_lib_test

## Basic Information

- **Project Name**: cuda_lib_test
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2021-07-24
- **Last Updated**: 2021-07-24

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# 使用CMake构建DLL

项目使用Cmake工具构建和Clion IDE 编写代码 。运行测试环境win10 ,  cuda11.4

# 概要

1. 如何构建cuda项目

2. 如何构建dll库

3. 如何调用dll

   

## CMakeLists.txt简要说明

```cmake
cmake_minimum_required(VERSION 3.19)  //最低版本要求
project(cuda_demo5 CUDA)   //指定项目名称和所使用语言 ， 这里项目名：cuda_demo5  语言：CUDA

set(CMAKE_CUDA_STANDARD 14) //指定cuda版本

set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)  //构建dll关键所在

include_directories(.)   //包含头文件
include_directories(common/inc)   

add_subdirectory(test_lib) //添加子项目， 就是需要构建的dll库， 这里是test_lib.all  test_lib.lib


aux_source_directory(. SRC)  //获取当前文件夹下所有源文件名且赋值给SRC

add_executable(${PROJECT_NAME} ${SRC})  //构建.exe文件 ， 这里是cuda_demo5.exe

target_link_libraries(${PROJECT_NAME} PUBLIC test_lib)  //指定cuda_demo5.exe需要连接的dll（test_lib.all  test_lib.lib）

set_target_properties(
        ${PROJECT_NAME}
        PROPERTIES
        CUDA_SEPARABLE_COMPILATION ON)    //使用cuda编译


install(TARGETS ${PROJECT_NAME} test_lib
        RUNTIME DESTINATION ${CMAKE_BINARY_DIR}
        LIBRARY DESTINATION ${CMAKE_BINARY_DIR}
        ARCHIVE DESTINATION ${CMAKE_BINARY_DIR})   //指定安装目录
```



## 如何构建cuda项目

project(cuda_demo5 CUDA) 必须指定为cuda语言 

set(CMAKE_CUDA_STANDARD 14) //指定cuda版本

set_target_properties(${PROJECT_NAME} PROPERTIES CUDA_SEPARABLE_COMPILATION ON) 指定使用cuda编译



## 如何构建dll库

test_lib的CMakeLists.txt文件

```cmake
cmake_minimum_required(VERSION 3.19)
project(test_lib CUDA)

set(CMAKE_CUDA_STANDARD 14)

set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)

include_directories(.)
include_directories(../common/inc)

aux_source_directory(. SRC)

add_library(${PROJECT_NAME} SHARED ${SRC})
#add_executable(cuda_demo5 main.cu)

set_target_properties(
        ${PROJECT_NAME}
        PROPERTIES
        CUDA_SEPARABLE_COMPILATION ON)


```

set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON) 把宏CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS 设置为ON

CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS 来导出所有函数符号，通过.def文件来说明导出函数列表， .def是自动生成的。如果

CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS 设置为OFF则不会导出所有函数符号， 可以在代码中使用__declspec(dllexport)修饰某些需要被导出的函数。 

CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS 有利于代码跨平台移植



## 如何调用dll

test_lib包含有头文件test_lib.cuh和定义文件test_lib.cu

test_lib.cuh

```c++
//
// Created by ozcom on 2021/7/24.
//

#ifndef CUDA_DEMO5_TEST_LIB_CUH
#define CUDA_DEMO5_TEST_LIB_CUH

__global__ void matrix_mul_gpu(int *M, int* N, int* P, int width);

#endif //CUDA_DEMO5_TEST_LIB_CUH

```

test_lib.cu

```c++
//
// Created by ozcom on 2021/7/24.
//

#include "test_lib.cuh"


__global__ void matrix_mul_gpu(int *M, int* N, int* P, int width)
{
    int i = threadIdx.x + blockDim.x * blockIdx.x;
    int j = threadIdx.y + blockDim.y * blockIdx.y;

    int sum = 0;
    for(int k=0;k<width;k++)
    {
        int a = M[j*width+k];
        int b = N[k*width+i];
        sum += a*b;
    }
    P[j*width+i] = sum;
}
```

test_lib编译后会生成test_lib.dll和test_lib



### 调用dll

target_link_libraries(${PROJECT_NAME} PUBLIC test_lib)  指定cuda_demo5.exe需要连接的dll（test_lib.all  test_lib.lib）

main.cu

```c++
#include <cstdlib>
#include <iostream>
#include "cuda.h"
#include "helper_functions.h"
#include "helper_cuda.h"
#include <cmath>

#include "test_lib/test_lib.cuh"


#define Row  1024
#define Col 1024

int main() {

    int *A = (int *)malloc(sizeof(int) * Row * Col);
    int *B = (int *)malloc(sizeof(int) * Row * Col);
    int *C = (int *)malloc(sizeof(int) * Row * Col);
    //malloc device memory
    int *d_dataA, *d_dataB, *d_dataC;
    cudaMalloc((void**)&d_dataA, sizeof(int) *Row*Col);
    cudaMalloc((void**)&d_dataB, sizeof(int) *Row*Col);
    cudaMalloc((void**)&d_dataC, sizeof(int) *Row*Col);
    //set value
    for (int i = 0; i < Row*Col; i++) {
        A[i] = 90;
        B[i] = 10;
    }

    cudaMemcpy(d_dataA, A, sizeof(int) * Row * Col, cudaMemcpyHostToDevice);
    cudaMemcpy(d_dataB, B, sizeof(int) * Row * Col, cudaMemcpyHostToDevice);
    dim3 threadPerBlock(16, 16);
    dim3 blockNumber((Col+threadPerBlock.x-1)/ threadPerBlock.x, (Row+threadPerBlock.y-1)/ threadPerBlock.y );
    printf("Block(%d,%d)   Grid(%d,%d).\n", threadPerBlock.x, threadPerBlock.y, blockNumber.x, blockNumber.y);
    matrix_mul_gpu <<<blockNumber, threadPerBlock >>> (d_dataA, d_dataB, d_dataC, Col);
    //拷贝计算数据-一级数据指针
    cudaMemcpy(C, d_dataC, sizeof(int) * Row * Col, cudaMemcpyDeviceToHost);

    //释放内存
    free(A);
    free(B);
    free(C);
    cudaFree(d_dataA);
    cudaFree(d_dataB);
    cudaFree(d_dataC);

    return 0;
}

```

mian.cu需要包含test_lib.cuh文件，这要就可以愉快的使用了。