GPU fluids in production: a compiler approach to parallelism