* PTX kernel argument buffer now has naturally aligned arguments as required by PTX JIT compiler. * Change dynamic loading of CUDA driver API functions to load 32-bit or 64-bit versions of depending on the the host architecture. * Add ability to generate PTX kernels to be launched both on 32-bit and 64-bit hosts. * Use Unified Virtual Memory APIs to perform array argument marshalling. * PTX array storage test runs on the device and returns correct results. * More integer test failures on GPU fixed.