Dissecting GPU Memory Hierarchy through Microbenchmarking