Analyze performance of OpenMP

[hmazhar]

Determine which thrust calls need to be parallelized. Often the work done is not enough to justify using a parallel version.

[hmazhar]

After running a parametric study the main conclusion was that parallelizing the transform reduce in the broadphase will always make the algorithm faster, and parallelizing the unique is never faster.

Need to test for different problem sizes if this holds true.

Currently only tested for up to 8 cores. testing on euler99 and lagrange will be next

[hmazhar]

It would be interesting to a simple genetic algorithm to try to find the optimal solution. Using a brute force approach works for the set of parameters currently available, but not for larger sets.
Tuning the optimal number of bins for a given problem size/distribution would be useful to know.