Clarify the working of rootbeer gpu compiler

I was reading the paper published on "Rootbeer: Seamlessly using GPUs from Java", couldn't understand the high level processing, could someone please clarify :

1. 
   

   In section III it is written, "After CUDA code is generated, Java bytecode is generated that can serialize all types accessible from current Kernel class", isn't the java bytecode converted to CUDA code (as mentioned in Section VI), which Java bytecode does the above line mention.

   
   


2. 
   

   When do we add the CompiledKernel interface to class supporting Kernel interface.

   
   


3. 
   

   Also didn't understood this statement, Once these high level steps are complete, the Soot Java Optimization Framework converts the in memory class representations back to Java Bytecode and Rootbeer packs these .class files and the required Rootbeer runtime .class files into a resulting .jar file.

   
   


4. 
   

   Most importantly, the performance results compares serial Java implementation with rootbeer right? Is rootbeer a wrapper on top of CUDA C, if not are their any test cases comparing CUDA C vs rootbeer like the once mentioned in the paper.