Isn't that a typical case where you don't have to share anything? Divide the image into N chunks, let N threads handle each one, no sharing, just need a single sync point at the end to wait on completion.