This fixes compilation problems on Visual Studio, which (wrongly?) fails
to resolve an ambiguity correctly. The reasons seem to be that in-class
implementations get higher priority over operator overloads implemented as
free functions.

This is being migrated to python in an external repo

As some of these warnings tend to show up with some versions of Visual Studio,
it's better to be strict about it and fix all conversion warnings right away.
This also provides an indication about code which needs some improvements
because it exposes a poor handling of integer or floating points.

The OpenMP flag was only set for exectuables, leading to failures
when building libviennacl.so on MinGW.

Reported-by: skn123 on GitHub in #72

OCL : Now using basic program-caching mechanism if the environment variable VIENNACL_CACHE_PATH is defined

Now allocating the buffer dynamically, in case the log message is too big

Obtained with Clang 3.0 with -Wextra.

Clang: Fixed a bunch of sign conversion warnings, particularly for graph reordering implementations.

Obtained on Clang 3.0 with -Weverything

Reported-by: Pushkar Ratnalikar <pratnali@umail.iu.edu> via viennacl-support.

This addresses issue #64, (mildly) reduces the amount of code, and
simplifies certain automated testing methodologies.

Reported by

operator++ was incrementing the index incorrectly, while operator* was
returning a proxy which did not update any values if needed. Both
problems are now fixed and the respective tests are in place. Resolves #62

Reported-by: Karan Poddar via IRC.

No need to use more than the absolute minimum of Boost...

Shame on me for not taking enough care while merging my own branch

This fixes the compilation problems on MacOS X 10.6.8 (Snow Leopard)

This was a copy&paste bug introduced by the recent refactoring of matrix_base<T>

Reported-by: Paul Dufort via viennacl-support

* More compact and cleaner code (Column-Major Internally)
* New operators (reduce, row_reduce, col_reduce, element-wise, etc...)
* Better files architecture

The rationale for this change is to simplify the interfacing to external BLAS libraries,
which take the layout as a runtime argument. Another benefit is that we can
better unify kernels for row-major+non-transposed with column-major+transposed,
which is essentially the same operation. Compilation times
are mildly reduced for OpenMP and OpenCL, and mildly increased for CUDA.
OpenMP and OpenCL tests are known to pass, CUDA is not tested.
Code in the worker-backends can be further simplified in follow-up commits to
obtain a more compact code base.

No need to rely on OpenCL 1.1. MacOS X 10.6 for example only provides 1.0.

Since this is a fairly light-weight test, there's no need to have this split into two tests.
Supposed to reduce compilation times a little.

Apparently the tests were split to compile float and double separately to reduce compiler load,
but this was never fully completed. This reverts the split, compilation load is moderate
and therefore float and double can be tested in the same run.

Reduces amount of source code, compilation times (a bit) and simplifies calls to external libraries.

This replaces compile-time dispatches with respect to row_major and column_major tags
by runtime arguments (boolean flags). This reduces compilation times a little and
simplifies the incorporation of external BLAS backends.
Also, it reduces the amount of code we have to maintain :-)