共享内存上的原子操作

Question

how can i do an atomic operation on a shared memory?
i have something similar to this:

__shared__ int a[10];
//set a
if(tid<5)
  a[2]++;

therefore 5 threads are incrementing a. how can i do this?
i know that in this way i am serializing the execution of 5 threads, but how does this effect the warp? would all the threads in the warp be serialized or just the first 5?

Answer 1

Replace a[2]++ with

atomicAdd( a + 2, 1 );

if you can make a unsigned, you might prefer to use atomicInc() instead, but either one is going to kill performance.