writing to the same global variable by different threads

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi, I am learning Cuda and trying to convert an existing program to GPU runable. The program is basically a Monte Carlo program but each thread needs to update the result after during computation.

My question is: is there a potential conflict if different threads trying to accumulate values to the same global variable? If yes, how to avoid the problem. I am thinking setting up the result variable for each individual thread but this method will consume a lot of memory space. Is there a better solution?

Thanks a lot!

2

Yes, writing to the same variable will introduce a race condition. Use the atomic update operations if your gpu supports sm_11.

3

thanks, that makes a lot of sense. One practical question, I just the atomic functions, it seems like there is no one deals with floating point numbers. How should I get around of that? Thanks a lot.

4

That is correct. To sum up floating point numbers, use a reduction (there is an example in the SDK)

5

I think I had the same problem. I found one solution using atomic operations (cmpxchg) to create a protected zone. Follow this [url=“The Official NVIDIA Forums | NVIDIA”]http://forums.nvidia.com/index.php?showtopic=152685[/url]