Supposedly pretty complete, might receive minor final updates.

Improves performance on some NVIDIA GPUs by up to a factor of two.
Only kicks in if the matrix carries more than 12 nonzeros per row on average.

Strings are painfully hard to catch, whereas now it is a lot easier to
just catch std::exception (or a more specific inherited class as needed).

Makes it much easier to handle errors/exceptions.
Some more strings are thrown as exceptions at other locations,
will fix them in separate commit.

Checked with Clang 3.0.

Allows one to convert between {(u)int, (u)long, float, double} as needed.
Adds support for vectors and dense matrices (including proxies).
Support for viennacl::scalar<> already available via casts on host.
No support for sparse matrices for now, as no use case in sight.

Resolves #80.
Partially addresses #124: It is now easier to convert to the same types.

Affected OpenCL backend in generator:
"Unsupported reduction operator : no neutral element known"

Otherwise test fails on OpenCL devices without double precision support.

Based on pull request by cdeterman on GitHub. See discussion at #146.
A similar approach could also be applied for the non-symmetric case,
but that is not considered stable enough (complex?).

Problem was introduced with parent commit.

Includes updates to the checker-script in auxiliary-folder.

Addresses remaining issues in #145.

This code has not been used or tested in years. Time to clean up.

Also extended test suite such that this problem cannot show up again.
The cause was one declaration where 'float' was accidentally hard-coded.

Resolves #145.

Parameters not obtained from a full-fledged optimizer run,
but from careful manual tweaking. Obtained memory bandwidths
for BLAS levels 1 and 2 are about 70 GB/sec, which is okay given
that vector operations on the Xeon Phi are slow with OpenCL.
BLAS level 3 improves very mildly, peaks at about 40 GFLOP/sec.

Given that OpenCL for Xeon Phi (KNC) has limited use and that
everybody is eager for KNL, further tuning efforts are suspended.

Resolves #26.

Use viennacl/tools/timer.hpp instead.

New location: viennacl/tools/random.hpp

karlrupp/feature-improve-lanczos:
 Extends interface such that also eigenvectors are computed and returned.
 Removes all uBLAS dependencies (caused problems with some CUDA/Boost combinations).
 Improves performance for partial reorthogonalization.

Removed a couple of unnecessary host-device copies, removed unused counters.
Partial reorthogonalization now also computes eigenvectors if specified.

BLAS2 benchmark: Corrected calculation of bandwidth

Changed "BLAS3_M" and "BLAS3_N" to "BLAS2_M" and "BLAS2_N", respectively.

A library user can now call CG, BiCGStab, and GMRES directly
with STL types (vector<map<T, U> >, vector<U>).
Internally calculations are forwarded to the corresponding ViennaCL types.
Requires the user to include compressed_matrix.hpp and vector.hpp before
the respective solver implementation. This is not great, but the
best we can reasonably provide at this point.

Resolves #12.

size() now makes use of size1(), but was declared before size1().
Inverting the order fixes compilation problems with GCC 4.8 and Clang 3.3.

Simple random number generators now available in viennacl::tools:
 - uniform in closed interval [0, 1]
 - normally distributed
Both implementations based on rand(), hence not thread-safe.
However, they suffice for the time being.

Now demonstrating how to compute only eigenvalues and how to also
include eigenvectors.

Example works for all three backends.
Required a couple of internal reorganizations of prod(), handle(), and size().
Interface requirement for user-provided operators:
 - member function apply(x, y); to compute y = A * x;
 - member function size1(); to return the length of y.
The second interface requirement is technical, but will be needed
for matrix-free applications of rectangular matrices later.

Resolves #74.

The use of y -= prod(A, x) is troublesome for matrix-free applications,
hence the residual calculation
 residual = rhs;
 residual -= prod(A, current_guess);
was modified to
 residual = prod(A, current_guess);
 residual = rhs - residual;
to circumvent the problem.
As an extra benefit, this also improves performance since our current
implementation of y -= prod(A, x) for sparse matrices A relies
on a temporary for prod(A, x).

karlrupp/fix-and-improve-ilu:
 Resolves a dangling const-ref issue with preconditioner tags.
 Fixes a bug in ILU0 (values in U not computed correctly).
 Improves performance for ILUT by about three-fold.
 Removes unnecessary temporaries for Block-ILU.

Flaw was introduced recently in 51b61ba4.
L and U were setup as full matrices rather than just the triangular factors.

Fixed errors in bisection algorithm

Improves performance by about a factor of three to four.
Most of the optimization potential is now leveraged:
Each of the block matrices could also be removed by extending the
ILU0 and ILUT preconditioners accordingly. This, however, requires
larger refactoring and in view of the parallel Chow-Patel ILU
may not be worth the effort.

Resolves #40.

Now available via the solver tag, member function abs_tolerance().
Handy in cases where the first iterate is close to the solution already.
Resolves #123.

Now viennacl::cuda_arg() instead of viennacl::linalg::cuda::detail::cuda_arg.
Template argument for scalar, vector and matrix no longer required.
Reduces code volume and makes it easier (and more convenient) for
library users to inject their own custom kernels.

Resolves #132. Resolves #133.

Old interface returning only eigenvalue is still available.
A more MATLAB-like interface, e.g.
 viennacl::tie(eigenvalue, eigenvector) = eig(A, tag);
would be nice, but pollutes the code with a lot of template hackery.

Resolves request by Charles Determan on the viennacl-devel list.

Allows users to write e.g. A = prod(B+C, x + y);
Any expressions passed to prod() are converted to temporaries.
This is certainly desirable from a performance point of view for GEMM.
One could do better with expression templates for GEMV, but that would not work with OpenCL.
Fixes #126.

Changed std::size_t to unsigned int
Added an exception when there are more than 256 eigenvalues
Removed unnecessary line breaks

Changed some datatypes from const unsigned int to std::size_t
Fixed some comments
Deleted unnecessary #includes

Replaced VIENNACL_LINAL_BISECT_GPU by VIENNACL_LINALG_BISECT_GPU