- windows support
- automatic caching and creation of shared library with all generated kernels
- restrict keyword and function prefixes are preprocessor macros now -> easier to generate one code for linux, cuda, windows