Compiler optimizations

[bkj]

This pull request includes two simple compiler optimizations:

• a) Use -O3 -ffast-math in CMakeLists.txt
• b) Inline the sqr(x) {return x * x} function from lib/utils.cpp

Testing

I've also added a build and benchmark script, so we can run w/:

./build_and_run.sh

Results

version	burnin	posterior sampling
original code	9.5s	1.8s
a only	3.7s	0.7s
a and b	1.2s	0.2s

So using both optimizations reduces runtime of this example by ~87%.

Notes

This is a PR to master, but the numbers and experiments above are based on comparison to commit c07eeae9394ab30ca8d984b2ec2e40ab4c2d2e08, per @manujinda 's recommendation. I have not tested on master, but it should be easy for you to copy/paste these changes to whichever branch is currently under development.

Possible Future Work

After these optimizations, profiling via callgrind reveals that the majority of runtime is spent computing calls to exp in KDE::pdf. To get additional speedups, we'd need to make exp run faster. This may be possible, depending on how much precision is required, by eg. using approximate exp functions like the ones described here.

[bkj]

Testing this some more -- it's possible that -ffast-math actually breaks something. If/before this gets merged -- the impact of -ffast-math should be tested. Speedup w/o -ffast-math is still substantial (2.9s).

[adarshp]

I'd like to request a couple of changes, just to keep things organized - can the 'test.py' and 'build_and_run.sh' scripts be moved to the 'scripts' directory?