This reverts commit 955ec252.
Other desktop systems resulted in reduced performance.

Otherwise we get complaints on Visual Studio, since OpenMP 2.0 only
supports signed integers.

AddCCompilerFlagIfSupported.cmake and AddCLinkerFlagIfSupported.cmake
were not included in the dist and dist-src targets.

Visual Studio 2012 ran into ambiguities with respect to conversions.
Adding tie-breaker overloads of operator= fixed the problems.

Most notably:
 - Added fine-grained ILU
 - Described custom compilation via command line
 - Better iterative solver description (including mixed-precision-CG)
 - Update the GPU support table.
 - Sparse matrix-matrix products.

Improves performance on my laptop by a factor 3.

Resulted in build failures on Visual Studio 2012.

Improves performance on NVIDIA GPUs by about 10 percent on average.
Also reduces memory footprint a little.

AMG: Fix a typo: coarseing -> coarsening.

OpenCL context handle was accidentally not set.

Same problem showed up with OpenCL earlier in
216a6ac4
I assume that we are hitting a bug in the CUDA stack here,
since the problem only shows up on some CUDA devices (e.g. K20m)
and only with certain build configurations. A debug build, for example,
does not show any issues.

See also the follow-up discussion in #147.

Since no standalone PDF manual is available anymore, this option became obsolete.

The respective tuning code for ViennaProfiler is no longer in ViennaCL,
so this optional dependency is obsolete.

This is to also support lines such as
 compressed_matrix<T> A = prod(B, C);
So far only operator= was supported.

Old API still supported. New API uses solver objects, where the initial
guess as well as the monitor callbacks are registered.
New tutorial for usage: iterative-custom

Resolves #97.

Resolves #147.

Flags used:
-Wall -pedantic -Wextra -Weverything
-Wno-exit-time-destructors -Wno-padded -Wno-global-constructors
-Wno-weak-vtables -Wno-unreachable-code

These routines compute the sum of a vector as well as the row- and
column-sums of a dense matrix, respectively. The implementation
reuses inner products for vectors and matrix-vector products for
matrices. Thus, there is some overhead when compared to super-optimized
routines involved, but this should be acceptable in almost all cases.

Replaces former attempts of a reduce<>() function.

Resolves #127.

Basd on request by Sumit Kumar on viennacl-devel.

The previous default of sm_13 is no longer supported with CUDA 7.
Not setting a default does not give a hint to the user about the value to set.

CMake: commented out default architecture options in ViennaCLCommon.cmake. Now relies on the default behavior of nvcc to set the default behavior of ViennaCL.

Couldn't find a portable way to detect the architecture portably at build time with CMake.

Leverages the new conversion routines, no custom-kernels required.

CUDA argument retrieval had to be updated.

karlrupp/feature-refurbished-amg:
 Full rewrite of AMG functionality.
 Fine-grained AMG as described by Bell et al. now available.
 Includes GPU-accelerated setup.
 New implementation available for all three backends.
 Setup on host and application on device also possible.

Resolves #16.

Conflicts:
	viennacl/linalg/detail/amg/amg_coarse.hpp
	viennacl/linalg/detail/amg/amg_debug.hpp
	viennacl/linalg/detail/amg/amg_interpol.hpp

When randomly feeding matrices from Matrix Market one barely gets good convergence
at the first attempt.

Rather than running into an assertion, the user now encounters an
exception explaining the reason why the AMG setup failed.
For additional diagnostics it is also possible to query the problem
sizes at each level.

karlrupp/feature-chow-patel-ilu:
 Implements a parallel incomplete LU preconditioner.
 Implements a parallel incomplete Cholesky factorization preconditioner.
 Both preconditioners use the same sparsity pattern as A.
 Proposed by Chow and Patel (Algorithms 2 and 3) in SIAM J. Sci. Comp.

Proposed by Chow and Patel in SIAM J. Sci. Comp. Vol. 37, No. 2, pp. C169–C193, 2015
in Algorithm 3. Rather than a column-major computation of U,
we compute the row-major L. This saves at least one (costly) transposition.

Only works with compressed_matrix.

Release 1.7.0 is getting closer.
Still a few more things to complete, though.

Supposedly pretty complete, might receive minor final updates.

These are left-overs from earlier coding.

Arguments now passed to getter/setter members rather than all
squeezed into the constructor. Interpolation weight is now the same
as the Jacobi weight, because the interpolation is just constructed
such that it complements the action of the smoother.

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.