sum over a matrix how to parallelize

Accelerated Computing CUDA CUDA Programming and Performance
1

I need to parallelize, a sum like this over a matrix: (threads index can be used for addressing elements of the matrix)

[codebox]for(unsigned int l=0;l<(blockSize*blockSize);l++)

{

    accsum+=*(sum + l);

    accsumsqr+=*(sumsqr + l);

    accsumqrt+=*(sumqrt + l);

}[/codebox]

in which way i can do this avoiding banck conflicts?

2

The “Reduction” program in Nvidia SDK is a good reference tutorial in solving your problem.

3

Also check these sources:
[url=“http://www.cse.chalmers.se/~billeter/papers.html”]http://www.cse.chalmers.se/~billeter/papers.html[/url] - compaction
[url=“http://www.cse.chalmers.se/~billeter/pub/pp/index.html”]http://www.cse.chalmers.se/~billeter/pub/pp/index.html[/url] - compaction, prefix sum, sorting

Those provide one of the fastest currently available algorithms.