How can I create and use a dynamic library ".so" from CUDA C++ and use it inside C# code under Linux environment (CentOS)?

Question

I am trying to create a dynamic library .so using CUDA C++ kernel to use it inside C# code under Linux environment (CentOS). I searched for a way yo do this, but unfortunately didn't find a complete clear solution for it. Some solutions only make partial part of it, like creating C++ shared library on Linux, or creating chain of libraries in CUDA using nvcc, but there was no method for creating a dynamic library from CUDA C++. The using of .so created from C++ seamed possible like in this solution.

Is there a way to create this dynamic library and use it successfully inside C# code?

score 0 · Accepted Answer · answered Nov 13 '20 at 01:59

After searching multiple different solutions, and trying to collect and test the available possibilities, I finally reached to this simple method.
The original C++ library can be made using gcc in one step like this answer.

gcc -shared -o dll.so -fPIC dllmain.cpp

but make sure to add extern "C" before the required function(s) inside .cpp file, like this:

#include <stdio.h>
extern "C" void func()
{
    // code
}

For CUDA C++, nvcc can be used in the same way similarly to this answer and this answer combined. Make sure to use .so instead of .dll and use the proper device architecture, I used 60 here as I am using "Tesla P100-PCIE-16GB".

nvcc -arch=sm_60 --compiler-options '-fPIC' -o dll.so --shared kernel.cu

The .cu file will be similar to this.

#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <stdio.h>

extern "C" void myfunc(int a, int b, ...);
__global__ void kernel(int a, int b, ...);

__global__ void kernel(int a, int b, ...)
{
    int i = threadIdx.x;
    // kernel code
}

void myfunc(int a, int b, ...)
{
    // code
}

Now the dynamic library .so is created and can be used inside C# code like this.

using System;
using System.Runtime.InteropServices;
class Program
{
    [DllImport("dll.so")]
    static extern myfunc(int a, int b, ...);

    private void Method()
    {
        int a, b;
        // code
        myfunc(a, b, ...);
    }
}

The C# code then is compiled using Mono.

mcs Program.cs
mono Program.exe

But it will probably be necessary to set the path of the used library like this.

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/path/to/library/

This worked for a simple CUDA C++ code, it will likely work for other ones, but some problems may arise depending on their complexity.

I recommend modifying your answer so that `myfunc` launches the CUDA kernel completing the full story. Right now `kernel()` is unused and could be removed from the source. — NoahR, Jul 21 '22 at 21:05

How can I create and use a dynamic library ".so" from CUDA C++ and use it inside C# code under Linux environment (CentOS)?

1 Answers1