Question

Consider the following code:

__global__ void kernel(int *something) {
    extern __shared__ int shared_array[];     

    // Some operations on shared_array here.
}

Is it possible to initialize the whole shared_array to some value – e.g. 0 – without explicitly addressing each cell in some thread?

Answer 1

No. Shared memory is uninitialised. You have to somehow initialise it yourself, one way or another…

From CUDA C Programming Guide 3.2, Section B.2.4.2, paragraph 2:

__shared__ variables cannot have an initialization as part of their declaration.

This also discards nontrivial default constructors for shared variables.