GPU code reorganization and tweaks

GPU code reorganization and tweaks

- Moved the warp size macro to the new arch utils header and converted
  it to static const;
- Changes some awkward macro names in CUDA and OpenCL;
- Change the inline modifier of a number of CUDA functions to
  __forceinline__ which means unconditional inlining.

Change-Id: I5b748a920927f9a51ada14e3d6e65b6b4d4dbd43