Atomic operations Noob question.

Accelerated Computing CUDA CUDA Programming and Performance
1

All,

THe NVIDIA manual says “No other thread can update the memory” when an atomic operation is being done by a thread.
Goood…

But is this applicable only to threads in a block? OR all the threads in that kernel??

Thanks,

best Regards,
Sarnath

2

Atomic operations to global memory are atomic with respect to all other memory operations.

3

Vow! That sounds great!

So, if adjacent threads perform atomic operations on adjacent data, will coalescing n all still work??

I am just looking at the cost of an atomic operation…

Thanks,

best Regards,

Sarnath

4

All atomic operations are also uncoalesced. :)

5

And on >1.2 devices? Do they generate a single memory transaction for each thread, or do they get grouped within a half-warp?

6

Uncoalesced.