Strip-Mining

Learned from the PMPP book while seeing Tiling for matrix multiplication.

Strip-mining takes a long-running loop and break it into phases.

Each phase consists of an inner loop that executes a number of consecutive iterations of the original loop. The original loop becomes an outer loop whose role is to iteratively invoke the inner loop so that all the iterations of the original loop are executed in their original order.

By adding barrier synchronizations before and after the inner loop, we force all threads in the same block to focus their work entirely on a section of their input data. Strip-mining can create the phases needed by tiling in data parallel programs.