Move some static functions into the FarnebackOpticalFlow class as well,
so they can access these new class variables.

oclMat objects cannot be declared statically, because their destructor
depends on the statically defined __module variable from cl_context.cpp.
Since statically defined variables in separate compilation units have
an undefined destruction order there is always the possibility the
__module will be destructed before an oclMat object, which results
in a segfault.

gauss_w_lut is a statically defined variable of type oclMat.  The oclMat
destructor calls openCLFree() which via getInitializationMutex() accesses
the __module variable which has been statically defined in cl_context.cpp

Since the destruction order of statically defined variables in different
compilation units is undefined, it is possible that __module will
be destructed before gauss_w_lut, which would result in a segfault when
getInitializationMutex() is called while destructing gauss_w_lut.

In order to avoid this issue, we need to make gauss_w_lut a private
member of the HOGDescriptors class so we know it will be destroyed
before __module.

ContextImpl::currentContext contains a reference to one of the
DeviceInfoImpl objects from:

static std::vector<DeviceInfoImpl> global_devices;

ContextImpl::currentContext is destroyed in the destructor
for the statically defined object __module, and relies on its
DeviceInfoImpl reference to query some hardware features while
being destroyed.

This means that we need to ensure that the global_devices vector is
destroyed affter __module, otherwise ContextImpl::currentContext's
DeviceInfoImpl reference will no longer be valid when __module is
destroyed.

Since these variables are all confined to a single compilation unit,
they will be destruct from bottom to top, so we need to make sure
that __module is the bottom definition so it can be destroyed first.

In C99 'inline' is not a hint to the compiler to inline the function,
it is an attribute that affects the linkage of the function.  'inline'
functions are required to have a definition in a different compiliation
unit, so compilers are free to delete 'inline' functions if they want to.

This issue can be seen in Clang when compiling at -O0.  Clang
will sometimes delete 'inline' functions which creates an invalid
program.

Issue 3746: http://code.opencv.org/issues/3746

Bugfix brute_force_match.cl (see http://code.opencv.org/issues/2837): wrong results for non-float descriptors in OpenCL BruteForceMatcher

Bugfix brute_force_match.cl (Bug #2837): wrong results for non-float descriptors in OpenCL BruteForceMatcher

This reverts commit 3dcddad8.

Conflicts:
	modules/ocl/src/opencl/haarobjectdetect.cl

This reverts commit 54ea5bba, reversing
changes made to 28e0d3d7.

Fixed an issue with weight assignment causing the resulting GMM weights to be unsorted in the CUDA and OCL versions of BackgroundSubtractorMOG2

The LLVM/Clang 3.3 has a bug when compile a cl kernel with
assignment of a scalar to a vector data type. This patch
could work around this bug.
Signed-off-by: Zhigang Gong <zhigang.gong@intel.com>

OpenCV Error: Unknown error code -6 (OpenCL function is not available: [clGetPlatformIDs]) in opencl_check_fn, file /home/ahb/software/opencv/modules/ocl/src/cl_runtime/cl_runtime.cpp, line 83

The issue results from modules/ocl/src/cl_runtime/cl_runtime.cpp checking for
"linux" instead of "__linux__" (cp.  http://sourceforge.net/p/predef/wiki/OperatingSystems/)

Adjust all other occurrences of "defined(linux)" as well.

the `map` buffer does not have the same size with CUDA and index starts at [1, 1] instead of [0, 0].

static function qualifier replaced on inline to enable kernel compilation with OpenCL 1.1 embedded profile.

Improve ocl cvt_color performance for the following conversions: RGB<->BGR, RGB->Gray, RGB<->XYZ, RGB<->YCrCb, RGB<->YUV, and mRGBA<->RGBA.
The improvement was done basically by processing more than 1 pixel by each work-item and using vector's operations.
new performance tests were added

Added "sep_filter_singlepass" kernel that performs separable filtering in one kernel call
Added appropriate host part - sepFilter2D_SinglePass function and SingleStepSeparableFilterEngine_GPU class
Changed function declarations to enable their usage

The oclMat::ptr() method was mistakenly used in ocl::BruteForceMatcher to pass a pointer to a oclMat object. The ptr() method returns a uchar pointer to the cl_mem data structure and this method will be removed.