The HyCuda code-generator generates a template framework to easily compile different versions of a hybrid algorithm. When set up, you can switch between devices to execute part of the algorithm without having to worry about memory transfers. Based on a specification file that describes some properties of the algorithm, HyCuda generates C++11 header- and sourcefiles, only a few of which have to be modified by the programmer in order to implement the algorithm itself.
The generated code contains a lot of template metaprogramming (TMP), to make sure the decision for an algorithm to run on the CPU or GPU has already been made at compile-time. The same holds for the decisions whether or not to transfer data. This amounts to 0 runtime overhead.
Downloads:
0 This Week