reusable software components for the CUDA programming model

 CUB provides state-of-the-art, reusable software components for every layer
 of the CUDA programming model:
 * Parallel primitives
   * Warp-wide "collective" primitives
   * Block-wide "collective" primitives
   * Device-wide primitives
 * Utilities
   * Fancy iterators
   * Thread and thread block I/O
   * PTX intrinsics
   * Device, kernel, and storage management