how to generate DLL from cuda example code

Accelerated Computing CUDA CUDA Programming and Performance
1

I have successfully modified the simpleCUFFT.cu from the cuda example codes to fulfil my computation. I can use MS2017 to build exe file and run the exe file in CMD. But eventually I will call the functions in Labview. To do that I need a DLL file. now the question is how can I generate DLL file in MS2017? According to the discussion in stackoverflow thread : creating dll from cuda using nvcc, I played with the configuration type, but it still gives me exe file. And that exe file can not run in CMD. The reason I want to use MS 2017 instead of nvcc is that I basically have no experience in nvcc. And when I copy the command from MS2017 to nvcc command line to build the exe file, nothing is generated. So I personally prefer MS. Does anyone have experience on this? I highly appreciate.

3

Here is an extremely simple example, using three source files: cuda_dll.cu contains the DLL source code, cuda_dll.h exports a simple C-style API, and cuda_main.cu calls the DLL. This example uses the CUDA runtime.

cuda_dll.h:

#ifdef CUDADLL_EXPORTS
#define DLLEXPORT __declspec(dllexport)
#else
#define DLLEXPORT __declspec(dllimport)
#endif

extern "C" DLLEXPORT void wrapper (int n);

cuda_dll.cu:

#include <stdio.h>
#include <stdlib.h>
#include "cuda_dll.h"

__global__ void kernel (int n)
{
    printf ("kernel: n = %d\n", n);
}

void wrapper (int n)
{
    printf ("wrapper: calling kernel()\n");
    kernel<<<1,1>>>(n);
    cudaDeviceSynchronize();
}

cuda_main.cu:

#include <stdio.h>
#include <stdlib.h>
#include "cuda_dll.h"

int main (void)
{
    printf ("main: calling wrapper()\n");
    wrapper (5);
    fflush (stdout);
    return EXIT_SUCCESS;
}

Build the DLL, which creates two files (as is the custom on Windows): actual shared library cuda_dll.dll and interface stub library cuda_dll.lib:

nvcc --shared -DCUDADLL_EXPORTS -o cuda_dll.dll cuda_dll.cu

Build the main app and link with the stub library:

nvcc -o cuda_main.exe cuda_main.cu cuda_dll.lib

Run the app:

main: calling wrapper()
wrapper: calling kernel()
kernel: n = 5

Don’t ask me how to do this with the MSVS IDE. I have never used the IDE, so I don’t know. You do not have to build the calling app from CUDA source code compiled with nvcc. You can also call the DLL from regular C++ code compiled with MSVC. Simply copy cuda_main.cu to cpp_main.cpp, then build the main app like so:

cl /Fecpp_main.exe cpp_main.cpp cuda_dll.lib
4

Hello,
I see this to work needs a wrapper and i cannot just export the kernel.
Can you please point me to reason why tis is?
thank you.

5

To my knowledge (which may be dated!) C++ interfaces are insufficiently standardized to this day so C-style interfaces offer maximum robustness that let you interface with just about any software written in just about any programming language.

I provided a simple example to get the ball rolling. By no means did I mean to indicate that this is the only way to go about the issue. You should feel free to research and experiment on your own, and I expressly encourage you to do so. Based on my decades of experience (professional and otherwise), puzzling out solutions by oneself, while it may be painful, leads to accumulation of knowledge and experience much faster than asking random strangers on the internet.