Creating a shared vector with block size?

Question

I need to create a shared vector, with the same size as the block.

__global__ func()
{   
    const int size = blockDim.x;
    __shared__ float* Vec[size];
..
}

I get this error

error : expression must have a constant value

I cannot understand where the problem is, since blockDim.x is "constant" for each block of threads?

Answer 1

Here's how you do it

__shared__ float Vec[size];

remove the star (*)

Answer 2

You have to have a compiler that supports C99 to use variable-length arrays. It would seem your compiler doesn't support VLAs, so you have to have an integer constant expression for your array size.

Answer 3

As far as I know, CUDA does not support variable length arrays (which is what you're trying to do here, regardless of the presence of the keyword const).

Answer 4

If you look at the section about __shared__ of the CUDA C++ Programming Guide, there's some text on how to specify a size for an extern declared shared array. Although it's a bit more complicated, this is the syntax on how to specify execution-time sized shared arrays. The way you're doing it won't work.