Using library code (specifically curand) with CUDALink

Question

Needs["CUDALink`"]
gausvSource = "
#include <curand_kernel.h>
__global__ void gausv( float *out  )
{
    curandState_t curand_state;
    curand_init ( 0, 0, 0, &curand_state );
    *out = curand_normal( &curand_state );
}";
CUDAFunctionLoad[gausvSource, "gausv", {{"Float", _, "Output"}}, 256]

yields a series of errors like:

curand_kernel.h(558): error: this declaration may not have extern "C" linkage

Wrapping the include, the function, or the whole thing in extern "C" { } changes the message to:

CUDALink encountered an invalid program.

So, how does one correctly use library code with CUDALink?

Answer 1

The compiler accepts the following code :

Needs["CUDALink`"]
gausvSource = "
}
#include <curand_kernel.h>
extern \"C\" {
__global__ void gausv( float *out  )
{
    curandState_t curand_state;
    curand_init ( 0, 0, 0, &curand_state );
    *out = curand_normal( &curand_state );
}";
cudaFun=CUDAFunctionLoad[gausvSource, "gausv", {{"Float", _, "Output"}}, 256]  

The function is callable :

x=1.
cudaFun[{x}]  

{{0.00459315}}

I have not tested further.

The idea is that Mathematica encloses the whole CUDA-C program in a extern { }, so one can exit it with a } (the third line of the code above).