It was a bug in upstream function cpu::fast

initial gravity example

this macro takes care of asynchronous kernel launch and calls
sync on the queue when in debug mode.

Added perspective transform

This reverts commit 95d9346.

* Not necessary with the new queue class (see queue.hpp)
* Macros bad

Changes to make CPU backend functions asynchronous calls

This should speed things up when a lot of buffers are
present in the MemoryManager.

* Uses INTEL_MKL_ROOT as enviornment variable. Commonly /opt/intel/mkl
* If using RT, then add /opt/intel/mkl/lib and
  /opt/intel/compilers_and_libraries/mac/lib/ to DYLD_LIBRARY_PATH

* ArrayFire<Backend>.config and forge are now installed with the library component
* ArrayFireConfig.cmake files are installed by all repos
* Examples and Documentation and now independent components
* No changes to install paths
* When make osx_installer is called, it creates a new directory called
  osx_install_files which are then used to generate the installers
* osx_installer target depends on make install being called first

* libforge and ArrayFireConfig.cmake files are now in a common package
* This package is no visible at install time
* The package is enabled if any of the backends are enabled (like unified)
* This is done so that the common files are installed only once rather than by
  each backend package

- This behavior can be turned off optionally

- device id now occupies the last 8 bits.

…ated.

- Adding relevant tests and docs

Search for GLEWmx in default paths too

Change clBLAS/clFFT build tags to af3.3.0

Fixes to reduce memory usage for CPU backend in async mode.

Bug fix to set operations and minor document fixes

Changes to get device on which an array was created.

Fixes to and reorganization of OSX Installer

Memory manager fixes

Fixes for using MKL on OSX

Support to set visibility of windows programmatically

Release Notes for v3.3.0