adding array elements in shared memory

Accelerated Computing CUDA CUDA Programming and Performance
1

I am using a C870 which does not have atomic operations. To record particle interactions I am using a 0-1 array in shared memory and then adding the contents to record if an interaction took place between the particles just considered. I am currently using the thread with lowest threadid to perform this addition and to write the result to global memory. How else can this addition and writing be performed?

2

So you are essential doing a stream compaction? A scan is what you want then. See the SDK example for scan and search the forums for stream compaction to get the details: [url=“site:forums.nvidia.com "stream compaction" - Google Search”]Google

3

What is your block size (blockDim.x)? and how many elements do you want to add?

4

block size is currently 16, but I intend to increase this to 64 because the occupancy calculator suggests a block size of 64 gives the optimum occupancy, so I currently want to add 16 elements ( and possibly 64 in later codes).