Optimize FactorizationMachinesNative.cpp

[sfilipi]

Optimize the FactorizationMachinesNative.cpp, by:
1- switching the control variables of the for loops from post-increment to pre-increment, to avoid the extra storage
2- Optimize the intermediate calculations on the current tight loops for better performance.

Example:
for (int fprime = 0; fprime < m; fprime++)
{
const int vBias = j * m * d + fprime * d;
const int qBias = f * m * d + fprime * d;

3- Potentially substitute AlignedArray with a regular array. See the discussion on PR #383

[shauheen]

Thanks Senja, could you please explain in more details what is the issue?

[sfilipi]

Updated the description. Thanks for pointing out.

[bojanmisic]

@sfilipi, I believe it should state from post-increment to pre-increment. Sorry for the annoyance.

[wschin]

Please do a comprehensive comparison if you need to do regular array. As computation cost is 200 times (for criteo) more than memory copy cost, it is a bit risky...

[codemzs]

Closing this until @wschin deems this is useful. Also please do a comparison before suggesting a perf change to be definitive.