Can we use "AtomicAdd()" with GTX 8800? Any other option to do same thing...?

[Neeraj_Kulkarni]

sorry for not being clear, the reduction only works to the level of the block, meaning that if you have 1000 elements and your block size is 256 then buy the end of the first reduction you will still have 4 elements instead of one. if you have a much bigger number of elements you might need more kernel calls just of the reduction. What i did to minimize the overhead of launching multiple kernels (which can be significant)  is run the first reduction with at the end of the execution of the running kernel. and the last reduction at the begging of the next kernel (notice for this you are doing the same calculation in each block, so that some extra calculations. but i found that its faster then launching another kernel ).

and if my data set is big i run some more reduction kernels in between.

;)

[snapback]293849[/snapback]

I have been working on similar problems for the last few months.

I have programmed a few “hacks” that work for atomic computations at block level.

However the streaming architecture of the GPU is not designed for such constructs leading to possible dead locks!

I have already posted ways of doing this and is achieved by spin - loops + global writes!

http://forums.nvidia.com/index.php?showtopic=44144

Read from Memory

Work in parallel

Reduce in parallel (for threads within a single MP)

Reduce serially using the modified programming constructs.

Reduction + Block level synchronization + Memory optimization = very high performance gains

In short,Use the constructs as tools of getting around the problem not as a concrete reference!

I hope this helps

Cheers,

Neeraj