Hi guys…i’m new with cuda and i found it awesome!!! I’m developping some functions that should be called from another big C program. This functions are composed by “floating point operations” so i would like to execute them with cuda but the main program is too big so i can’t rewrite in cuda so i would call the cuda program from the main C program…is that possible???
Thank u and sorry for my English!!! :D
bye bye
Hello, it seems to be possible, I’ve just completed writing simple example for self understanding.(It’s simple nut it works=) )
(It’s based on videos by Sarnath. Great thanks goes to this man External Media http://forums.nvidia.com/index.php?showtop…;hl=create+dll)
At first I’ve create standard DLL project in Visual Studio and added a *.cu file
(In properties for this file “Command line” property was specified according to the Sarnath’s advise as well: nvcc -I"$(CUDA_INC_PATH)" -c -o $(ConfigurationName)\cuda_kernel.obj cuda_kernel.cu)
#include <stdio.h>
__global__ void my_intcopy(int *src, int *dest, int n)
{
int i=threadIdx.x;
dest[i] = src[i]*2+1;
}
__declspec(dllexport)
void my_intcopy_caller(int *Source, int *Dest, int n)
{
printf("INITIALIZING \n");
dim3 grid(1), block(n);
int *d_S, *d_D;
cudaMalloc((void**) &d_S, n*sizeof(int));
cudaMalloc((void**) &d_D, n*sizeof(int));
cudaMemcpy(d_S, Source, n*sizeof(int), cudaMemcpyHostToDevice);
printf("RUNNING \n");
my_intcopy<<<grid,block>>>(d_S, d_D, n);
Dest[7]=777;
printf("DONE \n");
cudaMemcpy(Dest, d_D, n*sizeof(int), cudaMemcpyDeviceToHost);
cudaThreadSynchronize();
}
And then I’ve created a console application with following code
// cuda_app.cpp : Defines the entry point for the console application.
//
#include "stdafx.h"
#include <cuda_runtime_api.h>
#include <cuda.h>
#include <windows.h>
#include <atlconv.h>
int _tmain(int argc, _TCHAR* argv[])
{
int *h_src, *h_dest;//, *d_src, *d_dest;
char *c="0";
int k = 0;
HINSTANCE LoadMe;
typedef void (*funcPtr)(int*, int*, int);
funcPtr LibFunc;
h_src = (int*)malloc(10*sizeof(int));
h_dest = (int*)malloc(10*sizeof(int));
for(int i = 0; i<10; i++)
{
h_src[i]=i*i;
h_dest[i]=i;
}
LoadMe = LoadLibrary("CUDA_DLL.dll");//Here the DLL is loaded
if (LoadMe != 0)
printf("LoadMe library loaded!\n");
else
printf("LoadMe library failed to load!\n");
LibFunc = (funcPtr)GetProcAddress(LoadMe, "?my_intcopy_caller@@YAXPAH0H@Z");//Here the function is being loaded
//I was lazy for creating .def file for DLL, so I've just extracted function name from DLL =)
k=10;
LibFunc(h_src, h_dest, k);
for(int i=0; i<10; i++)
{
printf("%d",h_src[i]);
printf(" ");
printf("%d",h_dest[i]);
printf(" ");
}
while(scanf©)
{}
return 0;
}
I have to embed a few functions to a large project too, and in such an aproach it seems that I can use CUDA functions without any use of CUDA API in main project External Media . Is it true or some problems (like simultaneous acces to GPU) could appear?