ArrayFire nighty

API changes:
- approx split into approx1 (for 1D y=f(x)) and approx2 (for 2D z=f(x,y))

ArrayFire 1.9 -- 20 Nov 2012

CUDA 5.0 required (CUDA toolkit download now includes CUDA driver)
API changes:
- interp changed to approx
- ones/zeros changed to constant
- hostFree obsolete free() for all allocations: pinned(), host(), alloc()
- standardize convention for for af_morph_t, af_resize_t, af_rprops_t, af_interp_t, af_solve_t, etc..
New Image Processing Functions
- Image handling: loadimage() and saveimage() added
- Image filtering: bilateral filter added
- Image filtering: meanshift filter added
New pinned memory option for faster transfer
Better performance when using erode and dilate
resize, approx1, approx2 now support linear (L), nearest neighbor (N), and cubic (C) interpolation
Known issue: startup delay fixed in drivers 304.x and later
Bug fixes:
- approx
- ifft, ifft2, ifft3 now normalize outputs
- Parallelized loops: gfor bug for a few functions

CUDA 4.1 required
API changes
- switch operator*() to matrix multiply by default (instead of element-wise, see more)
- newdims() switched to reshape()
- convert() switched to as() and asfloat()
- alltrue() switched to all()
- allfalse() switched to any()
- mpow() switched to matpow()
- std() switched to stdev()
- afDevicePointer and afHostPointer have changed to afDevice and afHost
extended and organized documentation
simplified and verified all examples
- support for both Windows Visual 2008 and 2010
- use timeit() where possible
integer rand()
expanded licensing documentation
expanded Graphics documentation
various complex operations added
various subscripting bug fixes and performance enhancements
documented gfor usage and limitations
verified code samples in documentation
added common machine-independent constants: Pi, NaN, Inf
major work on Python and Fortran interfaces