Question

So, if I have a device (or global) function that creates/copies some data into shared memory and I later call another device function, like so:

__global__ void a(){ 
    __shared__ int blah=0;
    fun();
}
__device__ void fun(){
   blah = 1; //perform some operations
   //do whatever
}

I’m a bit rusty with my CUDA, I think you might have had to “redefine” shared variable (I assume the operation checked if a shared variable of that name exists, if so assigns it) – this had the effect of creating context – so basically the variable didn’t just come out of nowhere. Alternatively, if it’s similar to having a global variable in standard C/C++ and I can just reference it, like I did above, it’d be great.

I am familiar with memory hierarchy, I’m just rusty on the semantics of creating/referencing memory.

Please advise on whether the above sketch would work. Thanks.

Answer 1

No that won’t work in CUDA, any more that it would work in standard C99. Currently, the preferred method of __device__ function compilation is inline expansion (they are also compiled as standalone code objects for the Fermi architecture), but even so __device__ functions still must obey standard syntax and scope conventions of C99. So you need to pass arguments which don’t have compilation unit scope by reference to __device__ functions.