NVIDIA Fermi Tuning Guide 1.2 & CUDA C Programming Guide 3.1.1

[Stefan]

Fermi Tuning Guide Version 1.2

For CUDA Driver API applications, the use of 32-bit device code within a 64-bit host application is no longer recommended, as support for this mixed-bitness mode will be removed in CUDA 3.2.  (Note that the CUDA Runtime API already requires that the device code and host code have matching bitness.)

CUDA C Programming Guide Version 3.1.1

Removed from Sections 3.1.6 and 5.2.3 the paragraph about loading 32-bit
device code from 64-bit host code as this capability will no longer be supported
in the next toolkit release.
In Section 3.2.6.3, removed the reference to the canMapHostMemory property
and mentioned that all devices of compute capability greater than 1.0 now
support mapped page-locked host memory.
Mentioned in Section 3.2.7.1 that host   device memory copies of a memory
block of 64 KB or less are asynchronous.
Fixed the maximum size of a 3D texture reference for devices of compute
capability 2.0 (2048 instead of 4096) in Section G.1.
Updated the paragraph about __fdividef(x,y) in Section C.2.1 to clarify
behavior depending on compute capability and compilation flag.