Atomics or Bankconflicts

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi folks!

I have got a problem.

I have to decide if i use shared memory and risk bank conflicts, or make use of atomic functions.
The algorithem is not deterministic, so i have no chance to evade bank conflicts with help of padding or a well designed accses pattern. Its a kind of union find algorithem.

So whats the better decision use atomics, nearly in every thread, or risk some Bank conflicts ?

thanks for help and im sorry for my bad english.

2

I’m confused. Are you trying to decide between bank conflicts in shared memory and atomic operations on global memory? If that’s the case, pick the shared memory.

3

yes exactly,
thank you for help External Image .