CUDA在warp中对不同地址的原子操作性能

To my knowledge, if atomic operations are performed on same memory address location in a warp, the performance of the warp could be 32 times slower. To my knowledge, if atomic operations are perfo