Thanks to Toby St Clere Smithe for reporting on viennacl-devel.

More satisfactory patch for NMF 0-matrix bug: compute norms of initial W and H, and if 0, set matrix/-ices to 1.0-valued-matrices

karlrupp/feature-pipelined-bicgstab:
Implements a pipelined version of BiCGStab with substantially improved
performance especially for smaller matrices. The algorithm is a slightly
rearranged version of BiCGStab as given in the book by Y. Saad
and only requires a single host<->device transfer in each step.
A fairly similar algorithm has been proposed in the paper
Jacques et al, "Electromagnetic scattering with the boundary integral
method on MIMD systems", LNCS 1593, Springer, 1999.

Some missing const qualifiers in the OpenCL kernels resulted in an almost
4x-performance drop. With equal const qualifiers, performance is now within
a roughly ten percent margin.

Implementation based on the BiCGStab listed in the book by Saad and
then optimized for minimum memory traffic and synchronization points
similar to the paper by Jacques et al.

karlrupp/feature-pipelined-cg:
Improvements of the conjugate gradient algorithm by using custom kernels
for full use of pipelining and avoiding kernel launch overheads to the extent
possible.
Follows algorithm 2.2 in the paper by Chronopoulos and Gear:
"s-step Iterative Methods for Symmetric Linear Systems",
J. Comp. and Appl. Math 25 (1989).

Support for other sparse matrices to be added later.

Provides a routine which generates a sparse matrix obtained from
the discretization of the Laplace equation using finite differences
and lexicographical ordering of the unknowns. One of the simplest,
symmetric and positive definite matrices.

Implementation for OpenCL and CUDA to come.
Reference for the implementation: Algorithm 2.2 in
Chronopoulos and Gear:
"s-step Iterative Methods for Symmetric Linear Systems",
Journal of Comp. and Appl. Math, 1989.

wip-sliced-ell-matrix: Support for Sliced ELL-matrix format.
Currently only matrix-vector products, but no
sparse-matrix-times-dense-matrix-products.

So far "-arch=sm_13" was hardcoded. Now the user can overwrite this
through the CUDA_ARCH_FLAG in CMake.

This format is an improved version of the ELL format,
aiming at maximized memory bandwidth for reading matrix values.
Currently only matrix-vector products are supported, no
sparse-matrix-by-dense-matrix products. To be added at some later time.
See paper
"A unified sparse matrix data format for efficient general sparse matrix-vector multiply on modern processors with wide SIMD units"
by Kreutzer et al. for details. We don't use any sorting of rows here.

Also works in the case that A and B are the same matrix.
Currently this operation is fairly slow, because the transposition is performed
in main memory. On the other hand, it's certainly better to have a slow
implementation rather than no implementation at all...

A subnamespace was ambiguous, at least according to MS' interpretation of
the standard.

Reported-by: Matthew Musto <matthew.musto@gmail.com> via viennacl-devel

Some nightly tests failed for element_pow() because numerical round-off errors
could become too severe. This commit reduces the exponent, which should
ultimately lead to no more spurious test failures.

For some reason the template argument is explicitly required here.
Doesn't hurt us, so let's use the more verbose variant here.

Probably requires more investigation, it may decrease the performance on some platforms, increase it on some others...

Device-specific / GEMM : Implemented the new FETCH_GLOBAL_CONTIGUOUS policy. Changed use_{A,B}_local from bool to an enum...

Device-specific : Added accessor to template::parameters(); rename template::parameters to template::parameters_type

Fixes #72. Under MinGW it is required to explicitly link with 'gomp'.

Could cause compilation errors on some platforms