This patch adds ocl kernels to accelerate Dense Inverse Search
based optical flow algorithm, it acclerates 3 parts in the algorithm,
including 1) Structure tensor elements compute, 2) Patch inverse search,
3) Densification compute.

Perf and accuracy test are also added. The perf test shows it is 30%
faster than the current implementation.
Signed-off-by: Li Peng <peng.li@intel.com>

* Minor fixes

* Start adding correspondence finding

* Added finding of correspondences using GPC

* New evaluation tool for GPC

* Changed default parameters

* Display ground truth in the evaluation tool

* Added training tool for MPI Sintel dataset

* Added the training tool for Middlebury dataset

* Added some OpenCL optimization

* Added explanatory notes

* Minor improvements: time measurements + little ocl optimization

* Added demos

* Fixed warnings

* Make parameter struct assignable

* Fix warning

* Proper command line argument usage

* Prettified training tool, added parameters

* Fixed VS warning

* Fixed VS warning

* Using of compressed forest.yml.gz files by default to save space

* Added OpenCL flag to the evaluation tool

* Updated documentation

* Major speed and memory improvements:
1) Added new (optional) type of patch descriptors which are much faster. Retraining with option --descriptor-type=1 is required.
2) Got rid of hash table for descriptors, less memory usage.

* Fixed various floating point errors related to precision.
SIMD for dot product, forest traversing is a little bit faster now.

* Tolerant floating point comparison

* Triplets

* Added comment

* Choosing negative sample among nearest neighbors

* Fix warning

* Usage of parallel_for_() in critical places. Performance improvments.

* Simulated annealing heuristic

* Moved OpenCL kernel to separate file

* Moved implementation to source file

* Added basic accuracy tests for GPC and PCAFlow

* Fixing warnings

* Test accuracy constraints were too strict

* Test accuracy constraints were too strict

* Make tests more lightweight