cuda c 混合编译

代码

文件1 : test1.cu

//文件：test1.cu
#include <stdio.h>
#include <stdlib.h>
#include <cuda_runtime.h>
 
#define ROWS 32
#define COLS 16
#define CHECK(res) if(res!=cudaSuccess){
      
      exit(-1);}
__global__ void Kerneltest(int **da, unsigned int rows, unsigned int cols)
{
    
    
  unsigned int row = blockDim.y*blockIdx.y + threadIdx.y;
  unsigned int col = blockDim.x*blockIdx.x + threadIdx.x;
  if (row < rows && col < cols)
  {
    
    
    da[row][col] = row*cols + col;
  }
}
 
extern "C" int func() // 注意这里定义形式
{
    
    
  int **da = NULL;
  int **ha = NULL;
  int *dc = NULL;
  int *hc = NULL;
  cudaError_t res;
  int r, c;
  bool is_right=true;
 
  res = cudaMalloc((void**)(&da), ROWS*sizeof(int*));CHECK(res)
  res = cudaMalloc((void**)(&dc), ROWS*COLS*sizeof(int));CHECK(res)
  ha = (int**)malloc(ROWS*sizeof(int*));
  hc = (int*)malloc(ROWS*COLS*sizeof(int));
 
  for (r = 0; r < ROWS; r++)
  {
    
    
    ha[r] = dc + r*COLS;
  }
  res = cudaMemcpy((void*)(da), (void*)(ha), ROWS*sizeof(int*), cudaMemcpyHostToDevice);CHECK(res)
  dim3 dimBlock(16,16);
  dim3 dimGrid((COLS+dimBlock.x-1)/(dimBlock.x), (ROWS+dimBlock.y-1)/(dimBlock.y));
  Kerneltest<<<dimGrid, dimBlock>>>(da, ROWS, COLS);
  res = cudaMemcpy((void*)(hc), (void*)(dc), ROWS*COLS*sizeof(int), cudaMemcpyDeviceToHost);CHECK(res)
 
  for (r = 0; r < ROWS; r++)
  {
    
    
    for (c = 0; c < COLS; c++)
    {
    
       
      printf("%4d ", hc[r*COLS+c]);
      if (hc[r*COLS+c] != (r*COLS+c))
      {
    
       
        is_right = false;
      }   
    }   
    printf("\n");
  }
  printf("the result is %s!\n", is_right? "right":"false");
 
  cudaFree((void*)da);
  cudaFree((void*)dc);
  free(ha);
  free(hc);
//  getchar();
  return 0;
}

文件2：test2.c

#include <stdio.h>
 
int func(); // 注意声明
int main()
{
    
    
	func();
	return 0;
}

文件3 ：test3.cpp

#include <iostream>
using namespace std;
 
extern "C" int func(); //注意这里的声明
int main()
{
    
    
	func();
	return 0;
}

---

编译

方案1：

将所有文件分别编译，最后统一合并！

对于C程序

[]$nvcc -c test1.cu
[]$gcc  -c test2.c
[]$gcc  -o testc test1.o test2.o -lcudart -L/usr/local/cuda/lib64

C++ 程序

[]$nvcc -c test1.cu
[]$g++  -c test3.cpp
[]$g++  -o testcpp test1.o test3.o -lcudart -L/usr/local/cuda/lib64

方案2：

将CUDA程序弄成静态库

对于C程序

[]$nvcc -lib test1.cu -o libtestcu.a
[]$gcc test2.c -ltestcu -L. -lcudart -L/usr/local/cuda/lib64 -o testc

特别注意：test2.c在链接库的前面
对于C++
完全域C类似，只要将gcc 换成g++， test2.c换成test3.cpp

方案3：

将CUDA程序弄成动态库

makefile

all : c cpp 
 
c : libtestcu.so
  gcc test2.c   -ltestcu -L. -lcudart -L/usr/local/cuda/lib64 -o testc
 
cpp : libtestcu.so
  g++ test3.cpp -ltestcu -L. -lcudart -L/usr/local/cuda/lib64 -o testcpp
 
libtestcu.so : test.cu
  nvcc -o libtestcu.so -shared -Xcompiler -fPIC test1.cu
 
 
clean :
  rm *.so testc testcpp  -f