contention in atomics

Accelerated Computing CUDA CUDA Programming and Performance
1

Consider a warp where each thread is doing an atomic operation into one array based on an index in another array. E.g.

indexes: 0,0,1,1,2,2,3,3,4,4 ...
atomicSomething( &B[indexes[thread_id]], ...

In this example, thread 0 contends with 1, 2 contends with 3, etc.

Will some threads (e.g. 0,2,4…) operate in parallel, or will the entire warp serialize?

If it operates in parallel, would a more complicated index array (e.g. 0,1,2,3,3,3,3,4,4,5,6 …) cause the entire warp to serialize?

Thanks

2
  1. it depends on the memory type you are accessing - shared or global
  2. memory bank rules for shared memory, and coalescing rules for global memory apply to any sort of operation - load/store/atomic-update
3

This is all global memory.