* Improved matrix-vector-test. Now checks all combinations of matrix/matrix-range/matrix-slice and vector/vector-range/vector-slice on rank-1-updates, matrix-vector products and triangular solves
* Removed redundant prod_impl(A, b)

* Reduced execution time of sparse-test by speeding up the reference uBLAS calculations (double-transpose-trick)

* Added inplace_qr_apply_trans_Q() to compute rhs of least-squares system R = Q^T b without setting up Q
* Fixed overloads for inplace_solve(A, b). More tests required, though.
* Unified use of viennacl::traits::clear(result) in CG and BiCGStab

* Added missing include directives for lu.hpp in one example and one test

* Reimplementation of LU factorization in viennacl/linalg/lu.hpp. Better performance, but still a lot of unused potential.
* Replaced slow generic CUDA matrix-matrix multiplication kernel by several semi-automatically generated kernels. Performance still only half of OpenCL, although code is virtually identical.
* Fixed a bug with C = prod(A, B) if C is a matrix_range or matrix_slice. An unnecessary temporary was introduced.
* CUDA-benchmarks now build correctly

* Generalized sparse_matrix_adapter. Now all types std::vector< std::map<T, U> > are supported (T was fixed to 'unsigned int' previously)

* Added level scheduling to ILUT, renamed routines from multifrontal_XYZ() to level_scheduling_XYZ()
* Fixed a couple of issues in block-ILU and improved performance. Now works well with CPU/OpenCL/CUDA, with the latter striving for higher block sizes than the default 8.

* CPU-fallback for ViennaCL-based block-ILU now working correctly.
* Removed old bicgstab-kernels (unused anyway)
* Eliminated 'potentially uninitialized variable' warnings in BiCGStab

* Improved OpenCL matrix-vector performance of coordinate_matrix (factor 2 on GTX 285)
* Added restart to BiCGStab if search direction vanishes or a certain number of iterations is reached.
* Added two missing operator-overloads for vector in order to handle b - prod(A,x)

* Fixed wrong estimated residual in BiCGStab as introduced with the previous commit
* Improved performance of block-ILU.

* Row- and Jacobi-preconditioner now work on CPU, OpenCL and CUDA
* Final summation in norm_1, norm_2, norm_inf is now carried out on GPU or CPU, depending on target (same as for inner_prod())
* Tweaked CG and BiCGStab to use norm_2 instead of inner_prod(v, v)

* Typesafe multi-backend transfer now working, making implementations based upon them nice and compact :-)
* added operator= to compressed_matrix<>
* moved viennacl::backend::memory_types to viennacl::memory_types

* Fixed some of the problems in the block preconditioners.
* Valgrind complains about uninitialized memory when using Cholesky with OpenCL. More investigations required. Maybe related to the AMD APP SDK bug on Trinity?

* Added workaround for AMD APP SDK 2.7 bug on Trinity APUs (Catalyst 12.8) to tests

Pimped ILU(0,T)-preconditioners, up to one order of magnitude faster with new low-level implementations.

Reimplementation of ILU0 for compressed_matrix. Using the new multi-backend, first tests indicate speedups of 10. That rocks. :-)

Sparse triangular solver for compressed_matrix now working on CPU. Moved in place for OpenCL and CUDA, testing required.

Some progress with triangular solver for compressed_matrix: unit-lower-solve and upper-solve working, but performance is rather poor. This is, however, expected, because the data structure is not well suited for that.

* Added initializer types for vectors: unit_vector, zero_vector, scalar_vector

* Unified preprocessor defines for external toolkit: VIENNACL_WITH_XYZ. Old uses of VIENNACL_HAVE_{UBLAS|EIGEN|MTL4} still work, but their use is deprecated.

* Matrix-matrix operations for CUDA now functional. Performance is lower than with OpenCL, though...

Fixed minor bugs in vector CUDA kernels (accidental uses of 'float' instead of template parameter 'T')

Vector operations and matrix operations now work for CUDA. No matrix-matrix products and no triangular solvers yet.

* Most parts of the ViennaCL core now compile without any OpenCL installation on the system.
* Improved build-system towards this new scenario, yet further love required.

Cleaned up norm-kernels. Now using a single kernel for norm_1, norm_2 and norm_inf. Also, final summations are generalized to a single kernel and now use shared memory.