build_hist is (heavily) latency bounded

[Laurae2]

After the enhancements of #11, I am noticing the build_hist is heavily latency bounded when running in parallel.

Specifically, this line just eats all the CPU time in parallel: https://github.com/hcho3/xgboost-fast-hist-perf-lab/blob/master/src/build_hist.cc#L12

Multiple solutions from there:

• Find whether there is false sharing occurring (done), and fix it with padding, how?
• Get rid of the memory allocation by doing something else, how?
• Get rid of the data variable by writing directly to the hist, how?
• Rewrite the algorithm, what?

According to VTune:

In general:

The memory latency issue on the data allocation:

[Laurae2]

With a loop instead of fill.

gcc:

icc:

[Laurae2]

Found a workaround, new issue incoming to describe the workaround. But I had to ignore the private BuildHist variable private (public to main thread) though.