In CUDA, I understand that the variable would be placed in shared memory if it was defined as __ shared __ and one would be placed in constant memory if it was defined as __ constant __.Also, those being allocated memory using cudamalloc() are put in GPU global memory. But where are those variable without prefixs like __ shared __ , __ constant __ and register placed? For example, the variable i as follow:

__device__ void func(){
   int i=0;
   return;
}

CodePudding user response：

Local variables are either placed in hardware registers or local memory (which is effectively global memory).

In your example, however, variable i will be removed by the compiler because it is unused.

CodePudding user response：

Automatic variables, i.e. variables without memory space specification within the scope of functions, are placed in one of the following locations:

1. When optimized away:

   1.1 Nowhere - if the variable isn't actually necessary. This actually happens a lot, since CUDA functions are often inlined, with some variables becoming copies of a variable in the calling function.

   For more information, consider reading: Local Memory and Register Spilling.