* Added guard for not using AMD GEMM kernels if device has less than 20 kB of local memory.
* Fixed a warning for GEMM kernels (fast NVIDIA version).

Added fast GEMM kernels for AMD Tahiti based on input from Philippe's autotuner. Only works for square matrices with dimensions being a multiple of 256.

Fixed a warning (/* in comment) in CL/cl_gl_ext.h (staying with OpenCL 1.1. The fix is already in OpenCL 1.2 headers)

Replaced all uses of size_t by std::size_t (exception: viennacl/generator/*, which will be replaced soon anyway)

* Fixed some corner cases for BLAS-1-type operations on vectors.
* Removed bottleneck in sparse-test-XYZ (manually transposing ublas-matrix).

* Fixed an access violation in copy() from STL to device for compressed_matrix with empty rows.

Better OpenMP support for vector operations. Not using OpenMP below 5000 entries to avoid unnecessary startup-overhead.

Added support for Xeon Phi (using 128x128 work items). Adjusted number of work groups on CPU to power of two (six-core CPUs otherwise lead to problems with reductions).

Fixed assertions thrown for AMG in Debug mode (wrong matrix sizes passed over to compressed_matrix).

Fixed bug in AMG: Removed unnecessary .resize() on empty operator matrices on GPU. Thanks to Jakub Pola for reporting.

Added fix for segfaults on program exit when providing custom OpenCL queues (thanks to Denis Demidov for reporting)

* Fixed issues with copy back to host when OpenCL handles are passed to CTORs of vector, matrix, or compressed_matrix (thanks to Jakub Pola for reporting)
* vector.empty() is now const (as it is supposed to be)

* Added an overload for result_of::alignment for vector_expression (again thanks to Denis)
* Resolved a problem with the include order in the SPAI preconditioner when using g++ 4.7

* Added Doxygen comments to generated OpenCL kernel (source) files
* Ensured newline at end of each source file
* Specified AMD APP SDK 2.7 error: "on Linux"
* Makefile for manual now cleans up additional temporary files

* Improved preamble in tutorials and benchmarks
* Mixed CG tag now provides an inner_tolerance() for the low-precision loop

* Added Doxygen comments to all namespaces.
* Updated comments on host-based implementations to clearly state (optional) OpenMP usage.

* Removed MSVC-switch in tutorials and benchmarks for reading files (require users to run from build/ accross different OS)
* Updated old Eigen-code to version 3.x
* Fixed a few more warnings in Visual Studio, added /wd4996 flag to get rid of VC iterator advertisements
* Fixed an overly strict assert() on vector-reductions with OpenCL, including a clean initialization of reduction vector
* Changed STL overload of norm_X from enable-if to plain overloading, otherwise MSVC has problems.

* Added finish() before copy() in tests in order to resolve issues with AMD APP SDK