Question

I am just wondering what is the difference between calling a CUDA kernel function with

<<<BLOCK_NUM, THREAD_NUM>>> 

and

<<<BLOCK_NUM, THREAD_NUM, THREAD_NUM * sizeof(float)>>> ?

Answer 1

The additional argument in the second case specifies the amount of dynamic shared memory to be allocated to each block in the kernel.

For example, if you have a kernel defined like this:

__global__ void kernel(...)
{

    extern __shared__ float buffer[];

}

The second kernel launch would attempt to allocate THREAD_NUM words for the shared memory array buffer for each block of the kernel. See Appendix B.17 of CUDA 4.1 programming guide for more information.