I'm sure they can be broken down into smaller chunks, but is it more efficient if they aren't broken down and instead shared memory is used? If you want parallelism you're obviously already worried about performance.